Upload folder using huggingface_hub

icml26/19e76363-53a6-4f3c-8b50-844e1aea4e26/appendix_chunks.jsonl

@@ -0,0 +1,37 @@
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0100", "section": "A. Limitation", "page_start": 11, "page_end": 11, "type": "Text", "text": "While Med-TIV demonstrates substantial improvements over existing medical reasoning verification approaches, several limitations warrant discussion and suggest directions for future research.", "source": "marker_v2", "marker_block_id": "/page/10/Text/2"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0101", "section": "A. Limitation", "page_start": 11, "page_end": 11, "type": "Text", "text": "Process Supervision. Our current training paradigm relies solely on trace-level outcome rewards, providing no supervision on intermediate verification behaviors such as when to search, what queries to formulate, or how to integrate retrieved evidence. While this design eliminates the need for costly step-level annotations, it may lead to suboptimal search patterns or redundant retrieval operations. Future work could explore supervision for the verification task itself, or leverage techniques such as search behavior cloning from stronger models to provide denser optimization signals.", "source": "marker_v2", "marker_block_id": "/page/10/Text/3"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0102", "section": "A. Limitation", "page_start": 11, "page_end": 11, "type": "Text", "text": "Retrieval Corpus Coverage. Med-TIV's verification accuracy is inherently bounded by the coverage and quality of the underlying medical corpus. Our retrieval system indexes documents from PubMed abstracts and medical textbooks, which provides broad coverage of established medical knowledge but may lack recent findings, rare disease information, or region-specific clinical guidelines. Verification of reasoning traces involving cutting-edge treatments or highly specialized subspecialties may be limited by corpus gaps.", "source": "marker_v2", "marker_block_id": "/page/10/Text/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0103", "section": "A. Limitation", "page_start": 11, "page_end": 11, "type": "Text", "text": "Language and Domain Scope. All training and evaluation are conducted on English-language medical reasoning benchmarks. The generalization of Med-TIV to multilingual medical content or non-Western medical traditions remains unexplored. Additionally, while our benchmarks span multiple medical subdomains, certain specialized areas such as genomics, radiology interpretation, and surgical planning may require domain-adapted retrieval corpora for optimal verification performance.", "source": "marker_v2", "marker_block_id": "/page/10/Text/5"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0104", "section": "B.1. Hyperparameter Settings", "page_start": 11, "page_end": 11, "type": "Text", "text": "Table 4 provides comprehensive hyperparameter configurations for Med-TIV training across both iterations. We maintain mostly consistent settings between iterations to isolate the effect of iterative training from hyperparameter tuning.", "source": "marker_v2", "marker_block_id": "/page/10/Text/8"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0105", "section": "B.1. Hyperparameter Settings", "page_start": 11, "page_end": 11, "type": "TableGroup", "text": "Hyperparameters Iteration 1 Iteration 2 RL Algorithm Dr.GRPO Dr.GRPO Clip ratio (low / high) 0.2 / 0.3 0.2 / 0.3 Learning rate 1e-6 1e-6 Warmup steps 10 10 Training epochs 5 5 Global batch size 256 256 Mini-batch size 256 256 Group size (G) 5 8 Rollout sampling temperature 1.0 1.0 Rollout top-p 0.95 0.95 Curriculum filtering Enabled Enabled Table 4. Hyperparameter configurations for Med-TIV training across iterations.", "source": "marker_v2", "marker_block_id": "/page/10/TableGroup/309"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0106", "section": "B.2. Retrieval Setup", "page_start": 11, "page_end": 11, "type": "Text", "text": "We construct our retrieval infrastructure using a dense retrieval architecture optimized for medical domain queries. The corpus is derived from the MedRAG (Zhao et al., 2025) collection, specifically combining the PubMed and Textbooks subcorpora into a unified index. The PubMed subset contains approximately 23.9 million biomedical abstracts covering research publications, while the Textbooks subset includes content from standard medical textbooks spanning clinical medicine, pharmacology, pathology, and related disciplines. After deduplication and quality filtering, the combined corpus contains approximately 24 million snippets.", "source": "marker_v2", "marker_block_id": "/page/10/Text/12"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0107", "section": "B.2. Retrieval Setup", "page_start": 11, "page_end": 11, "type": "Text", "text": "We employ MedCPT (Jin et al., 2023) as our dense retrieval encoder, specifically the query encoder variant for encoding", "source": "marker_v2", "marker_block_id": "/page/10/Text/13"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0108", "section": "B.2. Retrieval Setup", "page_start": 12, "page_end": 12, "type": "Text", "text": "Table 5. Prompt template.", "source": "marker_v2", "marker_block_id": "/page/11/Text/1"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0109", "section": "User Prompt", "page_start": 12, "page_end": 12, "type": "Text", "text": "You are a reasoning validator for medical problems. Your task is to think step by step and evaluate whether the given reasoning trace of a medical problem contains errors.", "source": "marker_v2", "marker_block_id": "/page/11/Text/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0110", "section": "User Prompt", "page_start": 12, "page_end": 12, "type": "Text", "text": "First, you must always perform a step-by-step analysis to examine the entire reasoning process. Then, based on your analysis, you will make a definitive judgment.", "source": "marker_v2", "marker_block_id": "/page/11/Text/5"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0111", "section": "User Prompt", "page_start": 12, "page_end": 12, "type": "ListGroup", "text": "Use 1 if the reasoning trace is free of errors. Use 0 if the reasoning trace contains one or more errors.", "source": "marker_v2", "marker_block_id": "/page/11/ListGroup/387"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0112", "section": "Output Instruction", "page_start": 12, "page_end": 12, "type": "Text", "text": "You must conduct your step-by-step analysis inside <think> and </think> first every time you get new information. After reasoning, if you find you lack some knowledge, you can call a search engine by <search> query </search> and it will return the top searched results between <information> and </information>. You can search as many times as you want. If you find no further external knowledge needed, you can directly provide the answer inside <answer> and </answer>, without detailed illustrations.", "source": "marker_v2", "marker_block_id": "/page/11/Text/9"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0113", "section": "Output Instruction", "page_start": 12, "page_end": 12, "type": "Code", "text": "Medical Problem: {The full Medical Problem on one or more lines.} Reasoning Trace: {The full Reasoning Trace on one or more lines.}", "source": "marker_v2", "marker_block_id": "/page/11/Code/10"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0114", "section": "Output Instruction", "page_start": 12, "page_end": 12, "type": "Text", "text": "search queries and article encoder for encoding corpus snippets. Document embeddings are pre-computed and stored in a FAISS index using the Flat configuration for maximum retrieval accuracy, distributed across multiple GPUs using FAISS's GPU sharding capability to enable parallel similarity search. For each search query, we retrieve the top-3 most relevant documents for both training and inference.", "source": "marker_v2", "marker_block_id": "/page/11/Text/12"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0115", "section": "B.3. Baseline Setup", "page_start": 12, "page_end": 12, "type": "Text", "text": "We describe the configuration of reward model baselines used in our experiments. For Med-PRM, which employs static retrieval-augmented generation, we equip it with the same retrieval corpus, encoder, and top-k setting as our framework to ensure a controlled comparison. MedS 3 does not support external tool invocation and is therefore evaluated without retrieval augmentation. For confidence score extraction and inference hyperparameter settings, we follow the configurations specified in each baseline's original publication.", "source": "marker_v2", "marker_block_id": "/page/11/Text/16"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0116", "section": "B.4. Prompt Template", "page_start": 12, "page_end": 12, "type": "Text", "text": "We design a structured prompt template that guides the verifier through systematic reasoning with explicit tool invocation syntax. The complete prompt is shown in Table 5.", "source": "marker_v2", "marker_block_id": "/page/11/Text/20"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0117", "section": "C.1. Benchmarks", "page_start": 12, "page_end": 12, "type": "Text", "text": "We evaluate Med-TIV on four established medical reasoning benchmarks that collectively assess verification capability across varying difficulty levels and medical subdomains.", "source": "marker_v2", "marker_block_id": "/page/11/Text/25"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0118", "section": "C.1. Benchmarks", "page_start": 12, "page_end": 12, "type": "Text", "text": "• MedQA (Jin et al., 2020) : A dataset of multiple-choice questions derived from the United States Medical Licensing Examination (USMLE), designed to evaluate clinical reasoning and medical knowledge integration across diverse specialties.", "source": "marker_v2", "marker_block_id": "/page/11/Text/27"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0119", "section": "C.1. Benchmarks", "page_start": 12, "page_end": 12, "type": "Text", "text": "• MedMCQA (Pal et al., 2022) : A large-scale multi-subject benchmark sourced from Indian medical entrance examinations (AIIMS and NEET-PG), covering 21 medical subjects with emphasis on factual knowledge and clinical application.", "source": "marker_v2", "marker_block_id": "/page/11/Text/29"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0120", "section": "C.1. Benchmarks", "page_start": 12, "page_end": 12, "type": "Text", "text": "• MMLU-Med (Hendrycks et al., 2021) : An aggregation of medical-related subsets from the Massive Multitask Language Understanding benchmark, encompassing anatomy, clinical knowledge, college biology, college medicine, medical", "source": "marker_v2", "marker_block_id": "/page/11/Text/31"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0121", "section": "C.1. Benchmarks", "page_start": 13, "page_end": 13, "type": "Text", "text": "genetics, and professional medicine.", "source": "marker_v2", "marker_block_id": "/page/12/Text/1"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0122", "section": "C.1. Benchmarks", "page_start": 13, "page_end": 13, "type": "Text", "text": "• MedXpertQA (Zuo et al., 2025) : An expert-level benchmark featuring challenging questions that require multi-step clinical reasoning, differential diagnosis, and treatment planning at the level expected of practicing physicians.", "source": "marker_v2", "marker_block_id": "/page/12/Text/2"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0123", "section": "C.2. Baselines", "page_start": 13, "page_end": 13, "type": "Text", "text": "We compare Med-TIV against comprehensive baselines spanning proprietary systems, general-purpose models, and domainspecialized approaches.", "source": "marker_v2", "marker_block_id": "/page/12/Text/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0124", "section": "Proprietary Models.", "page_start": 13, "page_end": 13, "type": "ListGroup", "text": "GPT-4o-mini (OpenAI et al., 2024) : A compact variant of OpenAI's GPT-4o optimized for efficiency while maintaining strong reasoning capabilities across diverse tasks. Gemini-2.0-Flash: Google's efficient multimodal model designed for fast inference with competitive performance on knowledge-intensive benchmarks.", "source": "marker_v2", "marker_block_id": "/page/12/ListGroup/388"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0125", "section": "General Reasoning Models.", "page_start": 13, "page_end": 13, "type": "ListGroup", "text": "DeepSeek-R1 (Guo et al., 2025) : A 671B parameter reasoning model trained with RL, representing the current frontier of open-weight reasoning capabilities. R1-Distill-Qwen / R1-Distill-Llama: Distilled variants of DeepSeek-R1 at 7B and 8B scales respectively, designed to transfer reasoning capabilities to smaller architectures.", "source": "marker_v2", "marker_block_id": "/page/12/ListGroup/389"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0126", "section": "General Foundation Models.", "page_start": 13, "page_end": 13, "type": "ListGroup", "text": "Qwen2.5 (Yang et al., 2025) : A family of open-weight language models with strong multilingual and reasoning capabilities, evaluated at 7B and 32B parameter scales. Llama3.1 (Grattafiori et al., 2024) : Meta's open-source foundation model demonstrating competitive performance across diverse benchmarks, evaluated at the 8B scale.", "source": "marker_v2", "marker_block_id": "/page/12/ListGroup/390"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0127", "section": "Medical Domain Models.", "page_start": 13, "page_end": 13, "type": "ListGroup", "text": "AlphaMed (Liu et al., 2025a) : A medical reasoning model that employs RL with rule-based rewards to enhance clinical reasoning without reliance on distillation from larger models. UltraMedical (Zhang et al., 2024) : A specialized medical model combining high-quality instruction tuning on curated biomedical corpora with preference optimization for improved clinical accuracy. HuatuoGPT-o1 (Chen et al., 2024) : A medical reasoning model incorporating chain-of-thought reasoning with internal verification mechanisms to improve diagnostic accuracy.", "source": "marker_v2", "marker_block_id": "/page/12/ListGroup/391"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0128", "section": "Medical Reward Models.", "page_start": 13, "page_end": 13, "type": "ListGroup", "text": "MedS 3 (Jiang et al., 2025b) : A self-evolved soft dual-sided process supervision framework for medical reasoning that generates training signals through iterative self-improvement without external annotations. Med-PRM (Yun et al., 2025) : A process reward model for medical reasoning verification that provides step-level supervision using static retrieval-augmented generation with guideline-based verification.", "source": "marker_v2", "marker_block_id": "/page/12/ListGroup/392"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0129", "section": "D. Case Analysis", "page_start": 13, "page_end": 13, "type": "Text", "text": "Table 6 presents a complete verification example illustrating how a Med-TIV trained verifier identifies reasoning errors through dynamic evidence retrieval. The case involves a patient with bladder cancer who develops ototoxicity following chemotherapy. The generator's reasoning trace incorrectly attributes the symptoms to taxanes based on their known association with ototoxicity, concluding with answer (B). However, the model retrieves evidence establishing that cisplatin—the standard neoadjuvant therapy for transitional cell carcinoma—is the causative agent, and its mechanism involves DNA cross-linking rather than microtubule hyperstabilization. Through iterative search and reasoning, verifier correctly identifies the error, demonstrating the value of tool augmentation for catching subtle medical reasoning mistakes.", "source": "marker_v2", "marker_block_id": "/page/12/Text/22"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0130", "section": "D. Case Analysis", "page_start": 14, "page_end": 14, "type": "Text", "text": "Table 6. Complete verification demonstration. Given a medical problem and a reasoning trace, the verifier retrieves relevant evidence and correctly identifies the reasoning error within the trace.", "source": "marker_v2", "marker_block_id": "/page/13/Text/2"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0131", "section": "Medical Problem", "page_start": 14, "page_end": 14, "type": "Text", "text": "Question: A 67-year-old man with transitional cell carcinoma of the bladder comes to the physician because of a 2-day history of ringing sensation in his ear. He received this first course of neoadjuvant chemotherapy 1 week ago. Pure tone audiometry shows a sensorineural hearing loss of 45 dB. The expected beneficial effect of the drug that caused this patient's symptoms is most likely due to which of the following actions?", "source": "marker_v2", "marker_block_id": "/page/13/Text/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0132", "section": "Medical Problem", "page_start": 14, "page_end": 14, "type": "Text", "text": "Options: (A) Inhibition of proteasome (B) Hyperstabilization of microtubules (C) Generation of free radicals (D) Cross-linking of DNA", "source": "marker_v2", "marker_block_id": "/page/13/Text/5"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0133", "section": "Generator Reasoning Trace", "page_start": 14, "page_end": 14, "type": "Text", "text": "The patient's presentation of ringing in the ear, known as tinnitus, and a sensorineural hearing loss of 45 dB, suggest ototoxicity. Ototoxicity is a well-known side effect of certain chemotherapy drugs. Neoadjuvant chemotherapy refers to the use of chemotherapy before other treatments such as surgery. The patient has received a course of neoadjuvant chemotherapy 1 week ago, which aligns with the recent onset of ototoxic symptoms. To determine the cause of ototoxicity from the provided options, we need to identify which of the options is likely to cause sensorineural hearing loss. Option (A), Inhibition of proteasome, is associated with drugs like bortezomib (Velcade), which are less commonly associated with ototoxicity. Option (B), Hyperstabilization of microtubules, is a known mechanism of taxanes (e.g., paclitaxel, docetaxel). Taxanes are known to cause ototoxicity, making this a potential candidate. Option (C), Generation of free radicals, is a general mechanism that can cause cellular damage but is not specifically known to be associated with ototoxicity. Option (D), Cross-linking of DNA, is a known mechanism of alkylating agents like cyclophosphamide. While alkylating agents can cause ototoxicity, this mechanism is not as commonly linked to ototoxicity as other options. Considering the options, the one most commonly associated with ototoxicity is hyperstabilization of microtubules. Therefore, the drug that most likely caused this patient's symptoms is one that hyperstabilizes microtubules. The answer is (B).", "source": "marker_v2", "marker_block_id": "/page/13/Text/7"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0134", "section": "Verifier Judgment", "page_start": 14, "page_end": 14, "type": "Text", "text": "<think> Let's check the common chemotherapeutic agents used for bladder cancer and their known side effects. Transitional cell carcinoma of the bladder is often treated with cisplatin-based chemotherapy as neoadjuvant therapy. We need to determine if any of the options match the known side effects of these drugs. Let's search to confirm the side effects of cisplatin and other common bladder cancer chemotherapy drugs. </think>", "source": "marker_v2", "marker_block_id": "/page/13/Text/9"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0135", "section": "Verifier Judgment", "page_start": 14, "page_end": 14, "type": "ListGroup", "text": "<search> common side effects of cisplatin chemotherapy </search> <information> [Search Results regarding Cisplatin toxicity and mechanisms...] </information> <think> From the information, cisplatin is known to cause ototoxicity, which aligns with the patient's symptoms of sensorineural hearing loss. Now let's check the mechanisms of action for the options: Option (A): Inhibition of proteasome - This is associated with drugs like bortezomib and is not used for bladder cancer. Option (B): Hyperstabilization of microtubules - This is the mechanism of taxanes (e.g., paclitaxel, docetaxel). Option (C): Generation of free radicals - This is not specific to a single mechanism and is not commonly associated with ototoxicity. Option (D): Cross-linking of DNA - This is the mechanism of alkylating agents (e.g., cyclophosphamide). Cisplatin is an alkylating agent, and its mechanism involves cross-linking of DNA. Therefore, the correct answer should be (D). </think>", "source": "marker_v2", "marker_block_id": "/page/13/ListGroup/279"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0136", "section": "Verifier Judgment", "page_start": 14, "page_end": 14, "type": "Text", "text": "<answer>0</answer>", "source": "marker_v2", "marker_block_id": "/page/13/Text/13"}

icml26/19e76363-53a6-4f3c-8b50-844e1aea4e26/appendix_text_v3.txt

@@ -0,0 +1,110 @@
+[p. 11 | section: A. Limitation | type: Text]
+While Med-TIV demonstrates substantial improvements over existing medical reasoning verification approaches, several limitations warrant discussion and suggest directions for future research.
+
+[p. 11 | section: A. Limitation | type: Text]
+Process Supervision. Our current training paradigm relies solely on trace-level outcome rewards, providing no supervision on intermediate verification behaviors such as when to search, what queries to formulate, or how to integrate retrieved evidence. While this design eliminates the need for costly step-level annotations, it may lead to suboptimal search patterns or redundant retrieval operations. Future work could explore supervision for the verification task itself, or leverage techniques such as search behavior cloning from stronger models to provide denser optimization signals.
+
+[p. 11 | section: A. Limitation | type: Text]
+Retrieval Corpus Coverage. Med-TIV's verification accuracy is inherently bounded by the coverage and quality of the underlying medical corpus. Our retrieval system indexes documents from PubMed abstracts and medical textbooks, which provides broad coverage of established medical knowledge but may lack recent findings, rare disease information, or region-specific clinical guidelines. Verification of reasoning traces involving cutting-edge treatments or highly specialized subspecialties may be limited by corpus gaps.
+
+[p. 11 | section: A. Limitation | type: Text]
+Language and Domain Scope. All training and evaluation are conducted on English-language medical reasoning benchmarks. The generalization of Med-TIV to multilingual medical content or non-Western medical traditions remains unexplored. Additionally, while our benchmarks span multiple medical subdomains, certain specialized areas such as genomics, radiology interpretation, and surgical planning may require domain-adapted retrieval corpora for optimal verification performance.
+
+[p. 11 | section: B.1. Hyperparameter Settings | type: Text]
+Table 4 provides comprehensive hyperparameter configurations for Med-TIV training across both iterations. We maintain mostly consistent settings between iterations to isolate the effect of iterative training from hyperparameter tuning.
+
+[p. 11 | section: B.1. Hyperparameter Settings | type: TableGroup]
+Hyperparameters Iteration 1 Iteration 2 RL Algorithm Dr.GRPO Dr.GRPO Clip ratio (low / high) 0.2 / 0.3 0.2 / 0.3 Learning rate 1e-6 1e-6 Warmup steps 10 10 Training epochs 5 5 Global batch size 256 256 Mini-batch size 256 256 Group size (G) 5 8 Rollout sampling temperature 1.0 1.0 Rollout top-p 0.95 0.95 Curriculum filtering Enabled Enabled Table 4. Hyperparameter configurations for Med-TIV training across iterations.
+
+[p. 11 | section: B.2. Retrieval Setup | type: Text]
+We construct our retrieval infrastructure using a dense retrieval architecture optimized for medical domain queries. The corpus is derived from the MedRAG (Zhao et al., 2025) collection, specifically combining the PubMed and Textbooks subcorpora into a unified index. The PubMed subset contains approximately 23.9 million biomedical abstracts covering research publications, while the Textbooks subset includes content from standard medical textbooks spanning clinical medicine, pharmacology, pathology, and related disciplines. After deduplication and quality filtering, the combined corpus contains approximately 24 million snippets.
+
+[p. 11 | section: B.2. Retrieval Setup | type: Text]
+We employ MedCPT (Jin et al., 2023) as our dense retrieval encoder, specifically the query encoder variant for encoding
+
+[p. 12 | section: B.2. Retrieval Setup | type: Text]
+Table 5. Prompt template.
+
+[p. 12 | section: User Prompt | type: Text]
+You are a reasoning validator for medical problems. Your task is to think step by step and evaluate whether the given reasoning trace of a medical problem contains errors.
+
+[p. 12 | section: User Prompt | type: Text]
+First, you must always perform a step-by-step analysis to examine the entire reasoning process. Then, based on your analysis, you will make a definitive judgment.
+
+[p. 12 | section: User Prompt | type: ListGroup]
+Use 1 if the reasoning trace is free of errors. Use 0 if the reasoning trace contains one or more errors.
+
+[p. 12 | section: Output Instruction | type: Text]
+You must conduct your step-by-step analysis inside <think> and </think> first every time you get new information. After reasoning, if you find you lack some knowledge, you can call a search engine by <search> query </search> and it will return the top searched results between <information> and </information>. You can search as many times as you want. If you find no further external knowledge needed, you can directly provide the answer inside <answer> and </answer>, without detailed illustrations.
+
+[p. 12 | section: Output Instruction | type: Code]
+Medical Problem: {The full Medical Problem on one or more lines.} Reasoning Trace: {The full Reasoning Trace on one or more lines.}
+
+[p. 12 | section: Output Instruction | type: Text]
+search queries and article encoder for encoding corpus snippets. Document embeddings are pre-computed and stored in a FAISS index using the Flat configuration for maximum retrieval accuracy, distributed across multiple GPUs using FAISS's GPU sharding capability to enable parallel similarity search. For each search query, we retrieve the top-3 most relevant documents for both training and inference.
+
+[p. 12 | section: B.3. Baseline Setup | type: Text]
+We describe the configuration of reward model baselines used in our experiments. For Med-PRM, which employs static retrieval-augmented generation, we equip it with the same retrieval corpus, encoder, and top-k setting as our framework to ensure a controlled comparison. MedS 3 does not support external tool invocation and is therefore evaluated without retrieval augmentation. For confidence score extraction and inference hyperparameter settings, we follow the configurations specified in each baseline's original publication.
+
+[p. 12 | section: B.4. Prompt Template | type: Text]
+We design a structured prompt template that guides the verifier through systematic reasoning with explicit tool invocation syntax. The complete prompt is shown in Table 5.
+
+[p. 12 | section: C.1. Benchmarks | type: Text]
+We evaluate Med-TIV on four established medical reasoning benchmarks that collectively assess verification capability across varying difficulty levels and medical subdomains.
+
+[p. 12 | section: C.1. Benchmarks | type: Text]
+• MedQA (Jin et al., 2020) : A dataset of multiple-choice questions derived from the United States Medical Licensing Examination (USMLE), designed to evaluate clinical reasoning and medical knowledge integration across diverse specialties.
+
+[p. 12 | section: C.1. Benchmarks | type: Text]
+• MedMCQA (Pal et al., 2022) : A large-scale multi-subject benchmark sourced from Indian medical entrance examinations (AIIMS and NEET-PG), covering 21 medical subjects with emphasis on factual knowledge and clinical application.
+
+[p. 12 | section: C.1. Benchmarks | type: Text]
+• MMLU-Med (Hendrycks et al., 2021) : An aggregation of medical-related subsets from the Massive Multitask Language Understanding benchmark, encompassing anatomy, clinical knowledge, college biology, college medicine, medical
+
+[p. 13 | section: C.1. Benchmarks | type: Text]
+genetics, and professional medicine.
+
+[p. 13 | section: C.1. Benchmarks | type: Text]
+• MedXpertQA (Zuo et al., 2025) : An expert-level benchmark featuring challenging questions that require multi-step clinical reasoning, differential diagnosis, and treatment planning at the level expected of practicing physicians.
+
+[p. 13 | section: C.2. Baselines | type: Text]
+We compare Med-TIV against comprehensive baselines spanning proprietary systems, general-purpose models, and domainspecialized approaches.
+
+[p. 13 | section: Proprietary Models. | type: ListGroup]
+GPT-4o-mini (OpenAI et al., 2024) : A compact variant of OpenAI's GPT-4o optimized for efficiency while maintaining strong reasoning capabilities across diverse tasks. Gemini-2.0-Flash: Google's efficient multimodal model designed for fast inference with competitive performance on knowledge-intensive benchmarks.
+
+[p. 13 | section: General Reasoning Models. | type: ListGroup]
+DeepSeek-R1 (Guo et al., 2025) : A 671B parameter reasoning model trained with RL, representing the current frontier of open-weight reasoning capabilities. R1-Distill-Qwen / R1-Distill-Llama: Distilled variants of DeepSeek-R1 at 7B and 8B scales respectively, designed to transfer reasoning capabilities to smaller architectures.
+
+[p. 13 | section: General Foundation Models. | type: ListGroup]
+Qwen2.5 (Yang et al., 2025) : A family of open-weight language models with strong multilingual and reasoning capabilities, evaluated at 7B and 32B parameter scales. Llama3.1 (Grattafiori et al., 2024) : Meta's open-source foundation model demonstrating competitive performance across diverse benchmarks, evaluated at the 8B scale.
+
+[p. 13 | section: Medical Domain Models. | type: ListGroup]
+AlphaMed (Liu et al., 2025a) : A medical reasoning model that employs RL with rule-based rewards to enhance clinical reasoning without reliance on distillation from larger models. UltraMedical (Zhang et al., 2024) : A specialized medical model combining high-quality instruction tuning on curated biomedical corpora with preference optimization for improved clinical accuracy. HuatuoGPT-o1 (Chen et al., 2024) : A medical reasoning model incorporating chain-of-thought reasoning with internal verification mechanisms to improve diagnostic accuracy.
+
+[p. 13 | section: Medical Reward Models. | type: ListGroup]
+MedS 3 (Jiang et al., 2025b) : A self-evolved soft dual-sided process supervision framework for medical reasoning that generates training signals through iterative self-improvement without external annotations. Med-PRM (Yun et al., 2025) : A process reward model for medical reasoning verification that provides step-level supervision using static retrieval-augmented generation with guideline-based verification.
+
+[p. 13 | section: D. Case Analysis | type: Text]
+Table 6 presents a complete verification example illustrating how a Med-TIV trained verifier identifies reasoning errors through dynamic evidence retrieval. The case involves a patient with bladder cancer who develops ototoxicity following chemotherapy. The generator's reasoning trace incorrectly attributes the symptoms to taxanes based on their known association with ototoxicity, concluding with answer (B). However, the model retrieves evidence establishing that cisplatin—the standard neoadjuvant therapy for transitional cell carcinoma—is the causative agent, and its mechanism involves DNA cross-linking rather than microtubule hyperstabilization. Through iterative search and reasoning, verifier correctly identifies the error, demonstrating the value of tool augmentation for catching subtle medical reasoning mistakes.
+
+[p. 14 | section: D. Case Analysis | type: Text]
+Table 6. Complete verification demonstration. Given a medical problem and a reasoning trace, the verifier retrieves relevant evidence and correctly identifies the reasoning error within the trace.
+
+[p. 14 | section: Medical Problem | type: Text]
+Question: A 67-year-old man with transitional cell carcinoma of the bladder comes to the physician because of a 2-day history of ringing sensation in his ear. He received this first course of neoadjuvant chemotherapy 1 week ago. Pure tone audiometry shows a sensorineural hearing loss of 45 dB. The expected beneficial effect of the drug that caused this patient's symptoms is most likely due to which of the following actions?
+
+[p. 14 | section: Medical Problem | type: Text]
+Options: (A) Inhibition of proteasome (B) Hyperstabilization of microtubules (C) Generation of free radicals (D) Cross-linking of DNA
+
+[p. 14 | section: Generator Reasoning Trace | type: Text]
+The patient's presentation of ringing in the ear, known as tinnitus, and a sensorineural hearing loss of 45 dB, suggest ototoxicity. Ototoxicity is a well-known side effect of certain chemotherapy drugs. Neoadjuvant chemotherapy refers to the use of chemotherapy before other treatments such as surgery. The patient has received a course of neoadjuvant chemotherapy 1 week ago, which aligns with the recent onset of ototoxic symptoms. To determine the cause of ototoxicity from the provided options, we need to identify which of the options is likely to cause sensorineural hearing loss. Option (A), Inhibition of proteasome, is associated with drugs like bortezomib (Velcade), which are less commonly associated with ototoxicity. Option (B), Hyperstabilization of microtubules, is a known mechanism of taxanes (e.g., paclitaxel, docetaxel). Taxanes are known to cause ototoxicity, making this a potential candidate. Option (C), Generation of free radicals, is a general mechanism that can cause cellular damage but is not specifically known to be associated with ototoxicity. Option (D), Cross-linking of DNA, is a known mechanism of alkylating agents like cyclophosphamide. While alkylating agents can cause ototoxicity, this mechanism is not as commonly linked to ototoxicity as other options. Considering the options, the one most commonly associated with ototoxicity is hyperstabilization of microtubules. Therefore, the drug that most likely caused this patient's symptoms is one that hyperstabilizes microtubules. The answer is (B).
+
+[p. 14 | section: Verifier Judgment | type: Text]
+<think> Let's check the common chemotherapeutic agents used for bladder cancer and their known side effects. Transitional cell carcinoma of the bladder is often treated with cisplatin-based chemotherapy as neoadjuvant therapy. We need to determine if any of the options match the known side effects of these drugs. Let's search to confirm the side effects of cisplatin and other common bladder cancer chemotherapy drugs. </think>
+
+[p. 14 | section: Verifier Judgment | type: ListGroup]
+<search> common side effects of cisplatin chemotherapy </search> <information> [Search Results regarding Cisplatin toxicity and mechanisms...] </information> <think> From the information, cisplatin is known to cause ototoxicity, which aligns with the patient's symptoms of sensorineural hearing loss. Now let's check the mechanisms of action for the options: Option (A): Inhibition of proteasome - This is associated with drugs like bortezomib and is not used for bladder cancer. Option (B): Hyperstabilization of microtubules - This is the mechanism of taxanes (e.g., paclitaxel, docetaxel). Option (C): Generation of free radicals - This is not specific to a single mechanism and is not commonly associated with ototoxicity. Option (D): Cross-linking of DNA - This is the mechanism of alkylating agents (e.g., cyclophosphamide). Cisplatin is an alkylating agent, and its mechanism involves cross-linking of DNA. Therefore, the correct answer should be (D). </think>
+
+[p. 14 | section: Verifier Judgment | type: Text]
+<answer>0</answer>

icml26/19e76363-53a6-4f3c-8b50-844e1aea4e26/assets.json

@@ -0,0 +1,35 @@
+[
+  {
+    "filename": "_page_0_Figure_9.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/19e76363-53a6-4f3c-8b50-844e1aea4e26/marker_markdown/19e76363-53a6-4f3c-8b50-844e1aea4e26/_page_0_Figure_9.jpeg",
+    "bytes": 52819,
+    "width": 670,
+    "height": 462,
+    "aspect_ratio": 1.4502164502164503,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_0_Figure_9.jpeg"
+  },
+  {
+    "filename": "_page_6_Figure_2.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/19e76363-53a6-4f3c-8b50-844e1aea4e26/marker_markdown/19e76363-53a6-4f3c-8b50-844e1aea4e26/_page_6_Figure_2.jpeg",
+    "bytes": 57571,
+    "width": 1345,
+    "height": 381,
+    "aspect_ratio": 3.5301837270341205,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_6_Figure_2.jpeg"
+  },
+  {
+    "filename": "_page_7_Figure_1.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/19e76363-53a6-4f3c-8b50-844e1aea4e26/marker_markdown/19e76363-53a6-4f3c-8b50-844e1aea4e26/_page_7_Figure_1.jpeg",
+    "bytes": 19561,
+    "width": 638,
+    "height": 481,
+    "aspect_ratio": 1.3264033264033264,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_7_Figure_1.jpeg"
+  }
+]

icml26/19e76363-53a6-4f3c-8b50-844e1aea4e26/chunks_v3_anonymized.jsonl

@@ -0,0 +1,137 @@
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0000", "section": "Abstract", "page_start": 1, "page_end": 1, "type": "Text", "text": "Large language models have achieved strong performance on medical reasoning benchmarks, yet their deployment in clinical settings demands rigorous verification to ensure factual accuracy. While reward models offer a scalable approach for reasoning trace verification, existing methods face two limitations: they produce only scalar reward values without explicit justification, and they rely on single-pass retrieval that precludes adaptive knowledge access as verification unfolds. We introduce Med-TIV, an agentic framework that addresses these limitations by training medical reasoning verifiers to iteratively query external medical corpora during evaluation. Our approach combines tool-augmented verification with an iterative reinforcement learning paradigm that requires only trace-level supervision, alongside an adaptive curriculum mechanism that dynamically adjusts training data distribution. Across four medical reasoning benchmarks, Med-TIV achieves substantial gains over existing methods, improving MedQA accuracy by 23.5% and MedXpertQA by 32.0% relative to the base generator in particular. Crucially, Med-TIV demonstrates an 8× reduction in sampling budget requirement compared to prior reward model baselines. These findings establish that grounding verification in dynamically retrieved evidence offers a principled path toward more reliable medical reasoning systems.", "source": "marker_v2", "marker_block_id": "/page/0/Text/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0001", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "Large Language Models (LLMs) have demonstrated remarkable capabilities in medical reasoning, achieving competitive performance on clinical question answering, diagnostic", "source": "marker_v2", "marker_block_id": "/page/0/Text/6"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0002", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "FigureGroup", "text": "Figure 1. Comparison of medical reasoning verification paradigms. Text-based judges rely on parametric knowledge and may validate erroneous reasoning, while tool-integrated judges dynamically retrieve evidence to ground their judgments.", "source": "marker_v2", "marker_block_id": "/page/0/FigureGroup/498"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0003", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "inference, and medical knowledge benchmarks (Ji et al., 2025a; Xiao et al., 2026) . While these advances hold significant promise for augmenting clinical decision making and democratizing access to medical expertise, the deployment of LLMs in high-stakes clinical settings demands rigorous verification mechanisms to ensure that generated reasoning is both factually accurate and logically sound (Zhang et al., 2025; Wang et al., 2025) .", "source": "marker_v2", "marker_block_id": "/page/0/Text/11"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0004", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "Reward-based judges have therefore emerged as a scalable solution for evaluating model outputs, supporting both posttraining refinement via reinforcement learning from human feedback (RLHF) and inference-time scaling through tree search (Snell et al., 2024) . These judges can be broadly categorized by the granularity of their supervision. Outcome Reward Models (ORMs) provide sparse trace-level supervision that quantifies the quality of the entire output, while Process Reward Models (PRMs) offer dense step-level feedback that scores each intermediate reasoning step, enabling fine-grained credit assignment and precise error localization within multi-step reasoning. Recent work has adapted both paradigms to the medical domain to assess complex clinical reasoning traces. In parallel, advances in generative reward modeling have extended judge models beyond", "source": "marker_v2", "marker_block_id": "/page/0/Text/12"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0005", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "scalar scoring, enabling them to produce natural-language critiques that explicitly justify their decisions (Liu et al., 2025c; Xiong et al., 2025) .", "source": "marker_v2", "marker_block_id": "/page/1/Text/1"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0006", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "Despite their effectiveness, reward-based judges exhibit fundamental limitations when applied to clinical reasoning tasks (Yun et al., 2025) . A primary concern is the prevalence of hallucinations in critique traces, where judge models generate plausible yet factually incorrect assessments (Figure 1) . This issue is particularly noticeable in the medical domain, where reliable verification demands grounding in authoritative clinical evidence and established medical knowledge. Unverified judgments could lead to the propagation of incorrect diagnostic or treatment recommendations. Existing medical reasoning verifiers typically provide only scalar reward signals, offering little or no justification for their judgments and thus limiting interpretability (Jiang et al., 2025b) . Furthermore, these methods often rely on a static Retrieval-Augmented Generation (RAG) pipeline, in which a fixed set of retrieved documents is prefixed to the context and remains unchanged throughout evaluation (Yun et al., 2025) . Such static design precludes adaptive, multi-turn evidence gathering and forces the verifier to a fixed retrieval budget, thus limiting scalability.", "source": "marker_v2", "marker_block_id": "/page/1/Text/2"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0007", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "To address these issues, we propose Med-TIV (Medical Tool-Integrated reasoning Verifier), an agentic reinforcement learning (RL) framework that trains LLMs to leverage external knowledge bases for judging medical reasoning traces 1 . Med-TIV features three key design principles: (1) a tool-augmented verification paradigm that enables dynamic, iterative knowledge retrieval during the evaluation process; (2) an iterative RL approach that progressively improves verification capabilities without requiring step-level expert annotations; and (3) an adaptive curriculum formulation strategy that adjusts the data distribution in response to the evolving capability of the model. By equipping judge models with tool-use capabilities, Med-TIV grounds evaluation decisions in external evidence rather than relying solely on parametric knowledge, thereby mitigating hallucination, improving interpretability, and overcoming the limitations of static RAG (Ji et al., 2025b; Xia et al., 2025) .", "source": "marker_v2", "marker_block_id": "/page/1/Text/3"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0008", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "To verify the effectiveness of Med-TIV, we conduct extensive experiments on common medical reasoning benchmarks. Our results demonstrate that Med-TIV trains strong medical verifiers: when guiding inference-time search for a 7B generator model, our trained verifier achieves relative improvements of 23.5% on MedQA and 32.0% on MedXpertQA compared to the generator model alone. Moreover, Med-TIV consistently outperforms existing medical reward model baselines and surpasses the performance of models that are up to 4× larger in scale. Notably, Med-TIV", "source": "marker_v2", "marker_block_id": "/page/1/Text/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0009", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "also demonstrates an 8× gain in sampling efficiency compared to prior reward-based approaches, achieving equivalent accuracy with substantially fewer sampled reasoning traces during test-time search.", "source": "marker_v2", "marker_block_id": "/page/1/Text/6"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0010", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "Our main contributions are summarized as follows:", "source": "marker_v2", "marker_block_id": "/page/1/Text/7"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0011", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "ListGroup", "text": "We propose Med-TIV, a novel tool-integrated verification framework that enables dynamic, iterative knowledge retrieval during medical reasoning evaluation, providing both interpretable, fine-grained justification and improved factual grounding. We introduce an iterative RL paradigm with curriculumbased difficulty adaptation that progressively improves verification capabilities through self-bootstrapping, requiring only trace-level supervision rather than dense steplevel expert annotations. Med-TIV achieves state-of-the-art performance on four medical reasoning benchmarks, with comprehensive ablation studies that validate each component's contribution.", "source": "marker_v2", "marker_block_id": "/page/1/ListGroup/657"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0012", "section": "2.1. Problem Setup", "page_start": 2, "page_end": 2, "type": "Text", "text": "We define medical reasoning verification as the task of assessing the correctness of a multi-step reasoning trace generated in response to a medical question. Formally, given a medical question q ∈ Q and a multi-step reasoning trace τ = (s1, s2, . . . , sm) from a generator model, a verifier model determines whether τ contains any errors. We formulate this problem as binary classification, where the verifier Vθ(q, τ ) produces a judgment ℓ ∈ {0, 1}, where ℓ = 1 indicates a error-free reasoning trace, and ℓ = 0 indicates the presence of one or more errors. Unlike scalar reward models that output continuous scores, we adopt a generative judge paradigm in which the verifier produces a discrete judgment accompanied by a detailed critique trace that provides a structured justification for the decision.", "source": "marker_v2", "marker_block_id": "/page/1/Text/13"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0013", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 2, "page_end": 2, "type": "Text", "text": "Following prior works (Jin et al., 2025) , we extend the verifier with access to an external search engine E that retrieves top-k documents from a curated medical corpus (See Appendix B.2 for details). Retrieved documents are appended verbatim to the verifier context. Given a verification instance (q, τ ), the verifier constructs an iterative verfication trajectory. At step k, the trajectory is represented as:", "source": "marker_v2", "marker_block_id": "/page/1/Text/15"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0014", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 2, "page_end": 2, "type": "Equation", "text": "\\mathbf{t}_k = \\{r_1, a_1, o_1, \\dots, r_k, a_k, o_k\\},\\", "source": "marker_v2", "marker_block_id": "/page/1/Equation/16"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0015", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 2, "page_end": 2, "type": "Text", "text": "where r i denotes a natural language reasoning step analyzing the medical content, a i is a search query formulated to retrieve relevant medical knowledge, and o i = E(ai) represents the retrieved documents. The iterative verification", "source": "marker_v2", "marker_block_id": "/page/1/Text/17"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0016", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 2, "page_end": 2, "type": "Footnote", "text": "1 Code is available at PittNAIL/med-tiv", "source": "marker_v2", "marker_block_id": "/page/1/Footnote/5"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0017", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 3, "page_end": 3, "type": "Caption", "text": "Figure 2. Overview of Med-TIV. Left: Tool-integrated verification iteratively analyzes reasoning traces, formulates search queries, and retrieves medical evidence before producing correctness judgments. Middle: Curriculum formulation filters trivial and impossible instances, retaining boundary cases for RL training. Right: At inference time, the verifier evaluates candidate medical reasoning traces generated by a frozen model and final answers are selected via weighted self-consistency.", "source": "marker_v2", "marker_block_id": "/page/2/Caption/2"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0018", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 3, "page_end": 3, "type": "Text", "text": "process is defined as:", "source": "marker_v2", "marker_block_id": "/page/2/Text/3"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0019", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 3, "page_end": 3, "type": "Text", "text": "123124125", "source": "marker_v2", "marker_block_id": "/page/2/Text/29"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0020", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 3, "page_end": 3, "type": "Text", "text": "135136", "source": "marker_v2", "marker_block_id": "/page/2/Text/35"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0021", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 3, "page_end": 3, "type": "Text", "text": "137138", "source": "marker_v2", "marker_block_id": "/page/2/Text/36"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0022", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 3, "page_end": 3, "type": "Equation", "text": "(r_k, a_k) \\sim V_{\\theta}(q, \\tau, \\mathbf{t}_{k-1}), o_k = \\mathcal{E}(a_k), \\mathbf{t}_k = \\mathbf{t}_{k-1} \\oplus r_k \\oplus a_k \\oplus o_k,", "source": "marker_v2", "marker_block_id": "/page/2/Equation/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0023", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 3, "page_end": 3, "type": "Text", "text": "where \\oplus denotes sequence concatenation. This process continues until the verifier produces a final judgment \\ell \\sim V_{\\theta}(q,\\tau,\\mathbf{t}_T) at the terminal step T. By allowing multiple tool executions, the verifier dynamically retrievs medical knowledge as need to verify specific claims in the reasoning trace. Table 5 in the Appendix shows the explicit instruction used in our experiments.", "source": "marker_v2", "marker_block_id": "/page/2/Text/5"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0024", "section": "2.3. Test-Time Search", "page_start": 3, "page_end": 3, "type": "Text", "text": "Test-time search strategies improve reasoning performance by leveraging reward models to evaluate and select among multiple candidate solutions (Shi et al., 2024). Given a frozen generator model \\pi_{\\rm gen} and a question q, we first sample N independent reasoning traces:", "source": "marker_v2", "marker_block_id": "/page/2/Text/7"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0025", "section": "2.3. Test-Time Search", "page_start": 3, "page_end": 3, "type": "Equation", "text": "\\{\\tau^{(j)}\\}_{j=1}^N \\sim \\pi_{\\mathrm{gen}}(\\cdot \\mid q).", "source": "marker_v2", "marker_block_id": "/page/2/Equation/8"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0026", "section": "2.3. Test-Time Search", "page_start": 3, "page_end": 3, "type": "Text", "text": "A trained verifier V_{\\theta} then scores each candidate trace, and the final output is selected based on these scores. Common selection strategies include Best-of-N sampling, which", "source": "marker_v2", "marker_block_id": "/page/2/Text/9"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0027", "section": "2.3. Test-Time Search", "page_start": 3, "page_end": 3, "type": "Text", "text": "selects the trace with the highest score:", "source": "marker_v2", "marker_block_id": "/page/2/Text/10"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0028", "section": "2.3. Test-Time Search", "page_start": 3, "page_end": 3, "type": "Equation", "text": "\\hat{\\tau} = \\arg\\max_{\\tau^{(j)}} V_{\\theta}(q, \\tau^{(j)}),", "source": "marker_v2", "marker_block_id": "/page/2/Equation/11"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0029", "section": "2.3. Test-Time Search", "page_start": 3, "page_end": 3, "type": "Text", "text": "and verification-based majority voting, where candidate traces are first filtered by the verifier and the final answer is determined by consensus among verified traces. Med-TIV trains such a plug-in verifier that provides tool-grounded assessments that can be used to augment decision-making for any frozen generator model at inference time.", "source": "marker_v2", "marker_block_id": "/page/2/Text/12"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0030", "section": "3. Tool-Integrated Medical Reasoning Verifier", "page_start": 3, "page_end": 3, "type": "Text", "text": "Med-TIV is an agentic verification framework that trains models to leverage external knowledge bases for verifying whether a given medical reasoning trace contains errors. We adopt an iterative training approach based on dynamic curriculum learning, which requires no fine-grained step-level expert supervision and trains solely through multiple rounds of reinforcement learning (Figure 2). We next describe the training procedure of Med-TIV in details.", "source": "marker_v2", "marker_block_id": "/page/2/Text/14"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0031", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 3, "page_end": 3, "type": "Text", "text": "Data Construction. All training data across iterations is derived from the open-source Med-PRM dataset. Each original instance consists of a tuple (q, \\tau, \\ell_{\\text{step}}, \\ell_{\\text{trace}}) , where q is a medical question, \\tau is a multi-step reasoning trace, \\ell_{\\text{step}} de-", "source": "marker_v2", "marker_block_id": "/page/2/Text/16"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0032", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "notes step-level labels, and \\ell_{trace} is a trace-level correctness label<sup>2</sup>.", "source": "marker_v2", "marker_block_id": "/page/3/Text/1"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0033", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "At each training iteration, we only utilize the triplet (q,\\tau,\\ell_{\\text{trace}}) with human-annotated trace-level labels. Step-level labels \\ell_{\\text{step}} is intentionally excluded, as Med-TIV is designed to improve verification performance without replying on fine-grained supervision. For each training iteration, we fix the training data budget to 20K instances and enforce a balanced label distribution between correct ( \\ell_{\\text{trace}}=1 ) and incorrect ( \\ell_{\\text{trace}}=0 ) reasoning traces.", "source": "marker_v2", "marker_block_id": "/page/3/Text/2"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0034", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "Algorithm. We employ Dr. GRPO (Liu et al., 2025b) as the RL algorithm for training the verifier. Given a verification instance (q_i, \\tau_i) , we sample a group of G verification trajectories \\{\\mathbf{o}_i\\}_{i=1}^G from the current policy \\pi_\\theta . Each trajectory \\mathbf{o}_i = (o_i^1, \\dots, o_i^{|\\mathbf{o}_i|}) consists of reasoning tokens, search queries, retrieved documents, and a final judgment. The objective is:", "source": "marker_v2", "marker_block_id": "/page/3/Text/3"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0035", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Equation", "text": "\\frac{1}{G} \\sum_{i=1}^{G} \\sum_{t=1}^{|\\mathbf{o}_i|} \\left\\{ \\min \\left[ r_i^t \\hat{A}_i^t, \\operatorname{clip} \\left( r_i^t, 1 - \\epsilon_l, 1 + \\epsilon_h \\right) \\hat{A}_i^t \\right] \\right\\}, \\quad (1)", "source": "marker_v2", "marker_block_id": "/page/3/Equation/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0036", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "where r_i^t = \\frac{\\pi_\\theta(o_i^t|\\mathbf{q},\\mathbf{o}_i^{< t})}{\\pi_{\\theta_{\\mathrm{old}}}(o_i^t|\\mathbf{q},\\mathbf{o}_i^{< t})} , \\mathbf{q} = (q,\\tau) denotes the input prompt containing the question and reasoning trace, \\mathbf{o}_i^{< t} represents previously generated tokens, and \\epsilon_l and \\epsilon_h are the clipping parameters. The advantage term \\hat{A}_i^t is defined as:", "source": "marker_v2", "marker_block_id": "/page/3/Text/5"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0037", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Equation", "text": "\\hat{A}_i^t = R(\\mathbf{q}, \\mathbf{o}_i) - \\text{mean}\\left(\\left\\{R(\\mathbf{q}, \\mathbf{o}_1), \\dots, R(\\mathbf{q}, \\mathbf{o}_G)\\right\\}\\right).", "source": "marker_v2", "marker_block_id": "/page/3/Equation/6"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0038", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "Reward Designs. To facilitate multi-turn RL with tool execution, we design a structured reward covering two complementary objectives, following prior practices (Jin et al., 2025):", "source": "marker_v2", "marker_block_id": "/page/3/Text/7"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0039", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "(i) Correctness Reward (R_c) : This component measures whether the verifier's judgment aligns with the ground-truth label. Let \\mathbf{q} = (q, \\tau) denote the verification prompt and \\ell \\in \\{0, 1\\} the ground-truth label. We define:", "source": "marker_v2", "marker_block_id": "/page/3/Text/8"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0040", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Equation", "text": "R_c = \\mathbb{1}(\\text{extract}(\\mathbf{o}) = \\ell),", "source": "marker_v2", "marker_block_id": "/page/3/Equation/9"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0041", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "where \\mathbb{1}(\\cdot) is the indicator function and extract(o) parses the final judgment from the <answer> tags in the generated trajectory o. Intuitively, R_c=1 if the verifier's decision is correct, and R_c=0 otherwise.", "source": "marker_v2", "marker_block_id": "/page/3/Text/10"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0042", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "(ii) Format Reward (R_f) : To ensure reliable tool use and structured outputs, the verifier is required to adhere to a predefined format. Specifically, reasoning steps must be enclosed within <think> tags, search queries within <search> tags, and the final judgment within <answer>", "source": "marker_v2", "marker_block_id": "/page/3/Text/11"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0043", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "tags. To discourage degenerate outputs, we further penalize excessive tag usage. Specifically, R_f=1 if the output satisfies all formatting constraints and contains no more than 10 < answer> tag pairs; <math>R_f=0.25 if the output is correct but exhibits tag overflow; and R_f=0 otherwise.", "source": "marker_v2", "marker_block_id": "/page/3/Text/13"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0044", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "The final reward {\\cal R} is defined as the product of the two components:", "source": "marker_v2", "marker_block_id": "/page/3/Text/14"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0045", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Equation", "text": "R = R_c \\times R_f .", "source": "marker_v2", "marker_block_id": "/page/3/Equation/15"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0046", "section": "3.2. Training Strategies", "page_start": 4, "page_end": 4, "type": "Text", "text": "Adaptive Curriculum Formulation. A central challenge in RL for verification is ensuring that training data remains appropriately calibrated to the evolving capability of the model. Instances that are either trivially easy or impossibly difficult yields minimal learning signal, as the resulting policy gradients approach zero. To address this issue, we adopt a model-aware curriculum formulation mechanism that dynamically adapts the task distribution at each training iteration.", "source": "marker_v2", "marker_block_id": "/page/3/Text/17"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0047", "section": "3.2. Training Strategies", "page_start": 4, "page_end": 4, "type": "Text", "text": "Concretely, before each iteration t, we perform online filtering on the sampled batch \\mathcal{B}_t to construct an effective training set \\mathcal{D}_t :", "source": "marker_v2", "marker_block_id": "/page/3/Text/18"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0048", "section": "3.2. Training Strategies", "page_start": 4, "page_end": 4, "type": "Equation", "text": "\\mathcal{D}_t = \\{(q, \\tau, \\ell) \\in \\mathcal{B}_t : \\exists g, g' \\in \\{1, \\dots, G\\} \\text{ s.t. } r^{(g)} \\neq r^{(g')}\\}.", "source": "marker_v2", "marker_block_id": "/page/3/Equation/19"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0049", "section": "3.2. Training Strategies", "page_start": 4, "page_end": 4, "type": "Text", "text": "Here, for each candidate instance (q, \\tau, \\ell) \\in \\mathcal{B}_t , we sample G verification trajectories \\{o^{(g)}\\}_{g=1}^G from the current policy \\pi_{\\theta_t} . We then compute the corresponding rewards \\{r^{(g)}\\}_{g=1}^G . Finally, we retain only instances if any two rewards are different, i.e., reward variance is non-zero (Khatri et al., 2025).", "source": "marker_v2", "marker_block_id": "/page/3/Text/20"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0050", "section": "3.2. Training Strategies", "page_start": 4, "page_end": 4, "type": "Text", "text": "This criterion eliminates prompts where the model either consistently succeeds or consistently fails across all sampled trajectories. By filtering these zero-gradient instances, optimization is focused on decision-boundary cases where the verifier exhibits uncertainty.", "source": "marker_v2", "marker_block_id": "/page/3/Text/21"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0051", "section": "3.2. Training Strategies", "page_start": 4, "page_end": 4, "type": "Text", "text": "To maintain a fixed training budget per iteration, we iteratively resample additional instances from the labeled pool \\mathcal{B} and apply the same filtering criterion until |\\mathcal{D}_t|=20K . This dynamic curriculum evolves naturally across iterations as the verifier improves, eliminating the need for manually designed difficulty schedules.", "source": "marker_v2", "marker_block_id": "/page/3/Text/22"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0052", "section": "3.2. Training Strategies", "page_start": 4, "page_end": 4, "type": "Text", "text": "Iterative Training via Self-Bootstrapping. We adopt an iterative training approach that progressively improves verification capabilities through multiple rounds of RL. Unlike prior work that alternates between rejection sampling, supervised fine-tuning (SFT), and RL (Xu et al., 2025), our approach operates entirely through iterative RL, following the RL-Zero paradigm where the model reinforces its verification capabilities without requiring dense turn-level expert demonstrations for cold start.", "source": "marker_v2", "marker_block_id": "/page/3/Text/23"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0053", "section": "3.2. Training Strategies", "page_start": 4, "page_end": 4, "type": "Footnote", "text": "& lt;sup>2</sup>Dataset is available at", "source": "marker_v2", "marker_block_id": "/page/3/Footnote/12"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0054", "section": "3.2. Training Strategies", "page_start": 5, "page_end": 5, "type": "Text", "text": "Algorithm 1 Iterative Training of Tool-Integrated Medical Reasoning Verifier", "source": "marker_v2", "marker_block_id": "/page/4/Text/1"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0055", "section": "3.2. Training Strategies", "page_start": 5, "page_end": 5, "type": "Text", "text": "249250", "source": "marker_v2", "marker_block_id": "/page/4/Text/45"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0056", "section": "3.2. Training Strategies", "page_start": 5, "page_end": 5, "type": "Code", "text": "Require: Base verifier \\pi_{\\theta_0}, labeled dataset pool \\mathcal{D} = \\{(q_i, \\tau_i, \\ell_i)\\}_{i=1}^N, maximum iterations T_{\\text{max}}, batch size B, group size G, search engine \\mathcal{E} Ensure: Trained verifier \\pi_{\\theta^*} 1: for t = 1 to T_{\\text{max}} do Sample labeled batch \\mathcal{B}_t \\leftarrow \\text{SAMPLEBATCH}(\\mathcal{D}, B) 3: 4: \\mathcal{D}_t \\leftarrow \\emptyset ▷ Curriculum formulation 5: for each (q, \\tau, \\ell) \\in \\mathcal{B}_t do 6: Sample verification trajectories: 7: \\{\\hat{\\ell}^{(g)}\\}_{q=1}^G \\sim \\pi_{\\theta_t}(\\cdot \\mid q, \\tau, \\mathcal{E}) Compute rewards within group: 8: r^{(\\hat{g})} \\leftarrow \\mathbb{1}[\\hat{\\ell}^{(g)} = \\ell], \\text{ for } g \\in 1, \\dots, G if \\exists g \\neq g' such that r^{(g)} \\neq r^{(g')} then 9: Add (q, \\tau, \\ell) to curriculum set \\mathcal{D}_t 10: 11: end if 12: end for 13: ▶ RL optimization on curriculum data \\pi_{\\theta_{t+1}} \\leftarrow \\text{DR.GRPO}(\\pi_{\\theta_t}, \\mathcal{D}_t, \\mathcal{E}) 15: end for 16: Return \\pi_{\\theta_{T_{\\max}}}", "source": "marker_v2", "marker_block_id": "/page/4/Code/2"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0057", "section": "3.2. Training Strategies", "page_start": 5, "page_end": 5, "type": "Text", "text": "Starting from the base model \\pi_{\\theta_0} , we perform T_{\\text{max}} iterations. Each iteration consists of three stages:", "source": "marker_v2", "marker_block_id": "/page/4/Text/3"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0058", "section": "3.2. Training Strategies", "page_start": 5, "page_end": 5, "type": "Code", "text": "\\begin{split} \\mathcal{B}_t \\leftarrow \\text{SampleBatch}(\\mathcal{D}, B), \\\\ \\mathcal{D}_t \\leftarrow \\text{Filter}(\\mathcal{B}_t, \\pi_{\\theta_t}), \\\\ \\pi_{\\theta_{t+1}} \\leftarrow \\text{RL}(\\pi_{\\theta_t}, \\mathcal{D}_t). \\end{split}", "source": "marker_v2", "marker_block_id": "/page/4/Code/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0059", "section": "3.2. Training Strategies", "page_start": 5, "page_end": 5, "type": "Text", "text": "Each iteration draws a fresh batch \\mathcal{B}_t from the annotated pool \\mathcal{D} with trace-level labels, ensuring a balanced distribution of correct and incorrect reasoning traces. The curriculum filtering then constructs the training set \\mathcal{D}_t as described above, and RL optimization updates the policy based on the structured reward.", "source": "marker_v2", "marker_block_id": "/page/4/Text/5"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0060", "section": "3.2. Training Strategies", "page_start": 5, "page_end": 5, "type": "Text", "text": "The key insight underlying this iterative approach is the coevolution of model capability and training distribution. As the verifier improves, the filtering mechanism automatically removes instances that have become too easy, while the fresh sampling introduces new challenging cases. This creates a self-bootstrapping cycle: stronger models encounter harder verification tasks, which in turn drive further improvements. Since the trace-level correctness reward is deterministic and unambiguous, this self-bootstrapping process converges reliably without the instabilities that can arise from noisy synthetic step-level labels. We summarize the overall training procedure in Algorithm 1.", "source": "marker_v2", "marker_block_id": "/page/4/Text/6"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0061", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Text", "text": "Evaluation benchmarks. We evaluated Med-TIV on four open-source medical question-answering benchmarks: MedQA (Jin et al., 2020), MedMCQA (Pal et al., 2022), MMLU-Med (Hendrycks et al., 2021), and MedX-pertQA (Zuo et al., 2025), using accuracy as the evaluation metric. These benchmarks collectively assess the verifier's ability to distinguish correct from erroneous reasoning across varying difficulty levels and medical subdomains. Detailed descriptions of benchmarks are in Appendix C.1.", "source": "marker_v2", "marker_block_id": "/page/4/Text/9"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0062", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Text", "text": "Implementation details. We trained verifiers using two light-weight backbone models: Llama3.1-8B and Qwen2.5-7B, with Llama3.1-8B as the default for results reporting. All training was conducted using the VeRL-Tool framework (Jiang et al., 2025a). Detailed hyperparameters are shown in Appendix B.1. All experiments were conducted on 4 NVIDIA H100 GPUs with 80GB of memory. Due to computational constraints, we limit the maximum number of RL iterations to T_{\\rm max}=2 and we set the group size for curriculum formulation (Section 3.2) to G=8.", "source": "marker_v2", "marker_block_id": "/page/4/Text/10"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0063", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Text", "text": "For inference, we used the default sampling hyperparameters for all models. In reward-guided search experiments, unless otherwise specified, we used Qwen2.5-7B as the frozen generator and sampled up to 32 candidate reasoning traces per question. We applied Hard-Weighted Self-Consistency as the default test-time search strategy.", "source": "marker_v2", "marker_block_id": "/page/4/Text/11"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0064", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Text", "text": "Baselines. We compared Med-TIV against two groups of baselines. 1): Off-the-shelf LLMs : GPT-40-mini (OpenAI et al., 2024), Gemini-2.0-Flash, DeepSeek-R1 series (Guo et al., 2025), Qwen2.5 series (Yang et al., 2025), Llama3.1 (Grattafiori et al., 2024), AlphaMed (Liu et al., 2025a), UltraMedical (Zhang et al., 2024), and HuatuoGPT-01 (Chen et al., 2024). 2): Medical domain-specialized Reward Models : MedS<sup>3</sup> (Jiang et al., 2025b) and Med-PRM (Yun et al., 2025). Detailed descriptions of each reward model baseline are shown in Appendix B.3.", "source": "marker_v2", "marker_block_id": "/page/4/Text/12"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0065", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Text", "text": "Test-Time Search Strategies. We evaluated three test-time search strategies that leverage Med-TIV to improve the reasoning performance of frozen generators. Given a reasoning trace \\tau=(s_1,s_2,\\ldots,s_K) with K steps, our verifier assigns a confidence score r_{\\tau}\\in[0,1] for the entire trace, defined as the softmax probability of the 1 token over the logits of both 1 and 0 tokens.", "source": "marker_v2", "marker_block_id": "/page/4/Text/13"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0066", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Text", "text": "• Best-of-N. Given a question q, we sampled N candidate traces \\{\\tau^{(j)}\\}_{j=1}^N from the generator and selected the trace with the highest verifier confidence score:", "source": "marker_v2", "marker_block_id": "/page/4/Text/14"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0067", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Equation", "text": "\\hat{\\tau} = \\arg\\max_{\\tau^{(j)}} r_{\\tau^{(j)}}.", "source": "marker_v2", "marker_block_id": "/page/4/Equation/15"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0068", "section": "4.1. Experimental Setup", "page_start": 6, "page_end": 6, "type": "TableGroup", "text": "Table 1. Main evaluation results on medical reasoning benchmarks. We report accuracy (%) on MedQA, MedMCQA, MMLU-Med, and MedXpertQA. Bold numbers indicate the best results among the reward model group. ✓: Verifier supports external tools for judging; ✗: Verifier does not support external tools for judging. Baselines å |Train| Size MedQA MedMCQA MMLU-Med MedXpertQA Avg. Proprietary Models GPT-4o-mini - - - 79.03 68.20 87.79 17.84 63.22 Gemini-2.0-Flash - - - 87.51 72.60 92.01 20.57 68.17 General Reasoning Models DeepSeek-R1 - - 671B 90.34 78.80 94.40 37.76 75.33 R1-Distill-Qwen - - 7B 24.82 36.40 47.47 7.43 29.03 R1-Distill-Llama - - 8B 34.96 43.60 64.19 5.35 37.03 General Non-reasoning Models Qwen2.5 - - 32B 73.21 64.83 84.94 13.87 59.21 Qwen2.5 - - 7B 60.96 56.56 76.96 12.15 51.66 Llama3.1 - - 8B 70.93 61.60 78.97 13.02 56.13 Medical Reasoning Models AlphaMed - - 7B 71.01 61.46 81.16 19.16 58.20 UltraMedical - - 8B 72.66 62.60 79.61 15.25 57.53 HuatuoGPT-o1 - - 8B 72.19 63.60 75.30 16.84 56.98 Medical Reward Models MedS3 ✗ 225k 7B 64.89 58.91 80.53 12.90 54.31 Med-PRM ✓ 111k 7B 69.99 62.36 80.99 13.51 56.71 Med-TIV (Ours) ✓ 20k 7B 75.26 64.70 85.58 16.04 60.40", "source": "marker_v2", "marker_block_id": "/page/5/TableGroup/739"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0069", "section": "4.1. Experimental Setup", "page_start": 6, "page_end": 6, "type": "Text", "text": "• Hard-Weighted Self-Consistency. We first filtered traces by the verifier's binary judgment, keeping only those labeled correct (Vθ(q, τ ) = 1). Among the filtered traces, we applied majority voting to determine the final answer:", "source": "marker_v2", "marker_block_id": "/page/5/Text/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0070", "section": "4.1. Experimental Setup", "page_start": 6, "page_end": 6, "type": "Equation", "text": "\\hat{a} = \\arg\\max_{a} \\sum_{j=1}^{N} \\mathbb{1} \\left[ V_{\\theta}(q, \\tau^{(j)}) = 1 \\right] \\cdot \\mathbb{1} \\left[ \\operatorname{ans}(\\tau^{(j)}) = a \\right].", "source": "marker_v2", "marker_block_id": "/page/5/Equation/5"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0071", "section": "4.1. Experimental Setup", "page_start": 6, "page_end": 6, "type": "Text", "text": "• Soft-Weighted Self-Consistency. Instead of binary filtering, we weighted each trace's vote by the verifier's confidence score:", "source": "marker_v2", "marker_block_id": "/page/5/Text/6"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0072", "section": "4.1. Experimental Setup", "page_start": 6, "page_end": 6, "type": "Equation", "text": "\\hat{a} = \\arg\\max_{a} \\sum_{j=1}^{N} r_{\\tau^{(j)}} \\cdot \\mathbb{1} \\big[ \\mathrm{ans}(\\tau^{(j)}) = a \\big].", "source": "marker_v2", "marker_block_id": "/page/5/Equation/7"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0073", "section": "4.2. Main Results", "page_start": 6, "page_end": 6, "type": "Text", "text": "Table 1 presents the main results on four medical reasoning benchmarks. Models trained with Med-TIV consistently outperform existing baselines across all benchmarks. Specifically, under guided-search using a Med-TIVtrained verifier, Qwen2.5-7B attains accuracies of 75.26%", "source": "marker_v2", "marker_block_id": "/page/5/Text/9"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0074", "section": "4.2. Main Results", "page_start": 6, "page_end": 6, "type": "Text", "text": "on MedQA, 64.70% on MedMCQA, 85.58% on MMLU-Med, and 16.04% on MedXpertQA, yielding an average accuracy of 60.40%. Notably, Med-TIV enables this 7B generator to rival substantially larger models, even surpassing the base performance of Qwen2.5-32B despite using a generator that is approximately 4× smaller. Compared to domain-specialized medical reasoning models of similar scale, Med-TIV outperforms HuatuoGPT-o1-8B and UltraMedical-8B by 3.07% and 2.60% on MedQA, respectively, demonstrating the effectiveness of our tool-integrated verification. Case analysis in Appendix D further illustrates how Med-TIV identifies subtle reasoning errors.", "source": "marker_v2", "marker_block_id": "/page/5/Text/10"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0075", "section": "4.3. Analysis", "page_start": 6, "page_end": 6, "type": "Text", "text": "We conducted a series of ablation analyses to investigate six key research questions regarding our proposed framework.", "source": "marker_v2", "marker_block_id": "/page/5/Text/12"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0076", "section": "4.3. Analysis", "page_start": 6, "page_end": 6, "type": "Text", "text": "Q1: Does Med-TIV generalize across different generator models? To evaluate the generalizability of the trained verifier, we applied Med-TIV to guide test-time search across generator models of varying sizes and capa-", "source": "marker_v2", "marker_block_id": "/page/5/Text/13"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0077", "section": "4.3. Analysis", "page_start": 7, "page_end": 7, "type": "FigureGroup", "text": "Figure 3. Test-time scaling analysis across three medical reasoning benchmarks. Each plot shows accuracy versus sampling budget N \\in \\{1, 2, 4, 8, 16, 32\\} for four baselines. Med-TIV consistently outperforms baselines across all sampling budgets and benchmarks.", "source": "marker_v2", "marker_block_id": "/page/6/FigureGroup/226"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0078", "section": "4.3. Analysis", "page_start": 7, "page_end": 7, "type": "TableGroup", "text": "Table 2. Performance improvements from using Med-TIV as a verifier on MedQA. For each generator model, the first row indicates the accuracy over single sampled trace per question. Models MedQA Qwen2.5-7B 60.96 + Self-Consistency 66.38 (+5.42) + Best-of-N (Med-TIV) 72.35 (+11.39) + Soft Weighted SC (Med-TIV) 75.02 (+14.06) + Hard Weighted SC (Med-TIV) 75.26 (+14.30) AlphaMed-7B 71.01 + Self-Consistency 74.23 (+3.22) + Best-of-N (Med-TIV) 75.02 (+4.01) + Soft Weighted SC (Med-TIV) 75.33 (+4.32) + Hard Weighted SC (Med-TIV) 75.65 (+4.64) Qwen2.5-32B 73.21 + Self-Consistency 75.26 (+2.05) + Best-of-N (Med-TIV) 75.57 (+2.36) + Soft Weighted SC (Med-TIV) 75.96 (+2.75) + Hard Weighted SC (Med-TIV) 75.96 (+2.75)", "source": "marker_v2", "marker_block_id": "/page/6/TableGroup/227"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0079", "section": "4.3. Analysis", "page_start": 7, "page_end": 7, "type": "Text", "text": "bilities. As shown in Table 2, when using Qwen2.5-7B as the generator, Hard-Weighted Self-Consistency yields a relative improvement of 23.5% over the base model's single-sample accuracy, substantially outperforming the 12.2% gain achieved by standard Self-Consistency. Notably, the domain-specialized AlphaMed-7B model also benefits from verifier guidance with a 6.5% relative improvement, indicating that our verifier provides complementary verification capabilities beyond domain-specific fine-tuning. The improvements extend to larger models as well: Qwen2.5-32B achieves a 3.8% relative gain during test-time search, demonstrating that a light-weight 8B verifier can effectively guide models that are significantly larger than itself. This", "source": "marker_v2", "marker_block_id": "/page/6/Text/6"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0080", "section": "4.3. Analysis", "page_start": 7, "page_end": 7, "type": "Text", "text": "cross-scale generalization suggests that Med-TIV learns transferable verification patterns rather than overfitting to specific generator characteristics.", "source": "marker_v2", "marker_block_id": "/page/6/Text/7"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0081", "section": "4.3. Analysis", "page_start": 7, "page_end": 7, "type": "Text", "text": "Q2: How do different test-time search strategies compare under Med-TIV? We then systematically compare different test-time search strategies under verifier guidance to identify the most effective approach for leveraging verification signals. As shown in Table 2, Hard-Weighted Self-Consistency consistently achieves the highest accuracy across all generators, followed by Soft-Weighted Self-Consistency and Best-of-N selection. On Qwen2.5-7B, Hard-Weighted Self-Consistency outperforms Best-of-N by 3% absolute accuracy, suggesting that majority voting among verified traces provides more robust answer selection than simply choosing the highest-confidence individual trace.", "source": "marker_v2", "marker_block_id": "/page/6/Text/8"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0082", "section": "4.3. Analysis", "page_start": 7, "page_end": 7, "type": "Text", "text": "Q3: Can Med-TIV reduce the sampling budget required to achieve state-of-the-art performance compared to existing baselines? Next, we investigated how verification performance scales with sampling budget, a critical consideration for deployment under varying computational constraints. As shown in Figure 3, Med-TIV achieves substantial efficiency advantage over existing medical reward models across all three benchmarks. In particular, Med-TIV matches the performance of baselines using only 4 samples, whereas the baselines require 32 samples, representing an 8× reduction in sampling budget. On MedOA, Med-TIV achieves 72.1% accuracy at N=4, while Med-PRM requires the full N=32 budget to reach 70.0% accuracy. Since inference cost scales approximately linearly with the number of sampled traces, this translates to equivalent performance at one-eighth the generator inference cost in practical deployment settings.", "source": "marker_v2", "marker_block_id": "/page/6/Text/9"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0083", "section": "4.3. Analysis", "page_start": 7, "page_end": 7, "type": "Text", "text": "Q4: Does Med-TIV generalize across different base models? To assess the generality of our proposed framework, we compared verification performance using two", "source": "marker_v2", "marker_block_id": "/page/6/Text/10"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0084", "section": "4.3. Analysis", "page_start": 8, "page_end": 8, "type": "FigureGroup", "text": "Figure 4. Ablation on base model selection and training iterations.", "source": "marker_v2", "marker_block_id": "/page/7/FigureGroup/401"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0085", "section": "4.3. Analysis", "page_start": 8, "page_end": 8, "type": "Text", "text": "distinct verifier backbones: Llama3.1-8B and Qwen2.5-7B. As shown in Figure 4, both backbones achieve strong performance after two training iterations. Llama3.1-8B consistently outperforms Qwen2.5-7B by approximately 3.5% absolute accuracy on MedQA, achieving 75.86% versus 72.35% after 2 iterations of training. The parallel performance gains observed across both models indicate that Med-TIV is agnostic to backbone architectures.", "source": "marker_v2", "marker_block_id": "/page/7/Text/3"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0086", "section": "4.3. Analysis", "page_start": 8, "page_end": 8, "type": "Text", "text": "Q5: What is the impact of iterative training? Figure 4 presents ablation results examining the impact of iterative training with adaptive curriculum formulation. Llama3.1-8B improves from 60.96% to 75.26% after iteration 1, with marginal gains to 75.86% at iteration 2. Qwen2.5-7B follows a similar pattern, reaching 72.35% after two iterations. The rapid convergence suggest that the majority of verification capability is acquired in the first round, with subsequent iterations refining boundary cases.", "source": "marker_v2", "marker_block_id": "/page/7/Text/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0087", "section": "4.3. Analysis", "page_start": 8, "page_end": 8, "type": "Text", "text": "O6 : How does RL and tool integration impact verification performance? Table 3 highlights the dual benefits of our framework across two generators. RL", "source": "marker_v2", "marker_block_id": "/page/7/Text/5"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0088", "section": "4.3. Analysis", "page_start": 8, "page_end": 8, "type": "TableGroup", "text": "Table 3. Ablation on RL and tool integration. Models MedQA Qwen2.5-7B 60.96 + Med-TIV (RL) 69.60 + Med-TIV (RL + Tool) 70.54 AlphaMed-7B 71.01 + Med-TIV (RL) 76.12 + Med-TIV (RL + Tool) 77.14", "source": "marker_v2", "marker_block_id": "/page/7/TableGroup/402"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0089", "section": "4.3. Analysis", "page_start": 8, "page_end": 8, "type": "Text", "text": "training drives the primary gain, boosting MedQA accuracy of Qwen2.5-7B by 8.64%, confirming that the verifier effectively internalizes reasoning patterns. Tool integration provides a critical secondary boost, further elevating accuracy to 70.54%. A similar cumulative trend is observed with AlphaMed-7B. This demonstrates that while RL", "source": "marker_v2", "marker_block_id": "/page/7/Text/8"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0090", "section": "4.3. Analysis", "page_start": 8, "page_end": 8, "type": "Text", "text": "anchors logical verification, dynamic retrieval is essential for resolving knowledge-intensive boundary cases beyond the model's parametric memory.", "source": "marker_v2", "marker_block_id": "/page/7/Text/9"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0091", "section": "5. Related Work", "page_start": 8, "page_end": 8, "type": "Text", "text": "Medical Reasoning Models. The application of large language models to medical reasoning has attracted considerable attention. Early efforts focused on domain-adaptive pretraining and instruction tuning on medical corpora (Wu et al., 2023; Singhal et al., 2025; Chen et al., 2023). More recent work has explored reasoning-enhanced medical models. HuatuoGPT-o1 (Chen et al., 2024) incorporates chainof-thought reasoning with verification mechanisms, and UltraMedical (Zhang et al., 2024) combines high-quality instruction data with preference optimization. AlphaMed (Liu et al., 2025a) employs RL to improve medical reasoning capabilities. Despite these advances, most existing approaches focus on improving the generator model itself, whereas our work addresses the complementary problem of training a plug-and-play verifier that can improve any frozen generator through test-time search.", "source": "marker_v2", "marker_block_id": "/page/7/Text/11"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0092", "section": "5. Related Work", "page_start": 8, "page_end": 8, "type": "Text", "text": "Tool-Assisted Reward and Judge Models. Standard LLM-based judges typically function as passive scorers limited by parametric knowledge. Recent work addresses this through agentic reward modeling, equipping verifiers with executable tools. Themis (Li et al., 2024) established the foundational framework by enabling access to calculators, search engines, and knowledge bases through structured tool-calling traces. TIR-Judge (Xu et al., 2025) advanced this paradigm in the general domain by integrating code execution to judge paired responses. TIM-PRM (Kuang et al., 2025) introduced independent tool queries for multi-modal verification to eliminate confirmation bias. The concept has further expanded to the Agent-as-a-Judge paradigm (You et al., 2026), which employs dynamic planning, tool augmentation and multi-agent coordination to decompose complex evaluation tasks. Our work instantiates this agentic paradigm within the medical domain, moving beyond static retrieval to iterative, evidence-grounded clinical verification.", "source": "marker_v2", "marker_block_id": "/page/7/Text/12"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0093", "section": "6. Conclusion", "page_start": 8, "page_end": 8, "type": "Text", "text": "We presented Med-TIV, an agentic RL framework for medical reasoning verification. Our approach addresses key limitations of existing medical reward models by offering explicit critique traces and enabling dynamic knowledge retrieval during verification. Empirical evaluations across four medical reasoning benchmarks demonstrate that Med-TIV substantially outperforms prior approaches. More broadly, Med-TIV introduces a general paradigm for training toolaugmented verifiers that can be extended to other highstakes domains requiring evidence-grounded evaluation.", "source": "marker_v2", "marker_block_id": "/page/7/Text/14"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0094", "section": "Impact Statement", "page_start": 9, "page_end": 9, "type": "Text", "text": "This paper introduces research aimed at improving the reliability of large language models for medical reasoning tasks. We believe our work contributes positively to the development of trustworthy medical AI systems by providing mechanisms to verify reasoning correctness before clinical deployment. Med-TIV holds potential to enhance the safety of LLM-assisted clinical decision support by reducing erroneous reasoning outputs through systematic verification. By grounding judgments in retrieved medical evidence, our approach offers improved transparency compared to opaque scalar reward models, enabling practitioners to better understand and audit verification decisions. The efficiency gains demonstrated by Med-TIV could democratize access to reliable medical reasoning verification, making robust verification feasible even in resource-constrained settings.", "source": "marker_v2", "marker_block_id": "/page/8/Text/2"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0095", "section": "References", "page_start": 9, "page_end": 9, "type": "ListGroup", "text": "Chen, J., Cai, Z., Ji, K., Wang, X., Liu, W., Wang, R., Hou, J., and Wang, B. Huatuogpt-o1, towards medical complex reasoning with llms, 2024. URL https:// arxiv.org/abs/2412.18925 . Chen, Z., Cano, A. H., Romanou, A., Bonnet, A., Matoba, K., Salvi, F., Pagliardini, M., Fan, S., Kopf, A., Mo- ¨ htashami, A., Sallinen, A., Sakhaeirad, A., Swamy, V., Krawczuk, I., Bayazit, D., Marmet, A., Montariol, S., Hartley, M.-A., Jaggi, M., and Bosselut, A. Meditron-70b: Scaling medical pretraining for large language models, 2023. URL 16079 . Grattafiori, A. et al. The llama 3 herd of models, 2024. URL . Guo, D. et al. Deepseek-r1 incentivizes reasoning in llms through reinforcement learning. Nature, 645(8081), 2025. doi: 10.1038/s41586-025-09422-z. Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., and Steinhardt, J. Measuring massive multitask language understanding, 2021. URL https: //arxiv.org/abs/2009.03300 . Ji, Y., Ma, W., Sivarajkumar, S., et al. Mitigating the risk of health inequity exacerbated by large language models. npj Digital Medicine, 8:246, 2025a. doi: 10.1038/s41746-025-01576-4. Ji, Y., Zhang, H., and Wang, Y. Bias evaluation and mitigation in retrieval-augmented medical question-answering systems, 2025b. URL 2503.15454 .", "source": "marker_v2", "marker_block_id": "/page/8/ListGroup/529"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0096", "section": "References", "page_start": 9, "page_end": 9, "type": "ListGroup", "text": "Jiang, D., Lu, Y., Li, Z., Lyu, Z., Nie, P., Wang, H., Su, A., Chen, H., Zou, K., Du, C., Pang, T., and Chen, W. Verltool: Towards holistic agentic reinforcement learning with tool use, 2025a. URL abs/2509.01055 . Jiang, S., Liao, Y., Chen, Z., Zhang, Y., Wang, Y., and Wang, Y. Meds 3 : Towards medical slow thinking with self-evolved soft dual-sided process supervision, 2025b. URL . Jin, B., Zeng, H., Yue, Z., Yoon, J., Arik, S., Wang, D., Zamani, H., and Han, J. Search-r1: Training llms to reason and leverage search engines with reinforcement learning, 2025. URL 2503.09516 . Jin, D., Pan, E., Oufattole, N., Weng, W.-H., Fang, H., and Szolovits, P. What disease does this patient have? a large-scale open domain question answering dataset from medical exams, 2020. URL abs/2009.13081 . Jin, Q., Kim, W., Chen, Q., Comeau, D. C., Yeganova, L., Wilbur, W. J., and Lu, Z. Medcpt: Contrastive pre-trained transformers with large-scale pubmed search logs for zero-shot biomedical information retrieval. Bioinformatics, 39(11), November 2023. ISSN 1367-4811. doi: 10.1093/bioinformatics/ btad651. URL bioinformatics/btad651 . Khatri, D., Madaan, L., Tiwari, R., Bansal, R., Duvvuri, S. S., Zaheer, M., Dhillon, I. S., Brandfonbrener, D., and Agarwal, R. The art of scaling reinforcement learning compute for llms, 2025. URL abs/2510.13786 . Kuang, P., Wang, X., Liu, W., Dong, J., and Xu, K. Timprm: Verifying multimodal reasoning with tool-integrated prm, 2025. URL 22998 . Li, L., Chai, Y., Wang, S., Sun, Y., Tian, H., Zhang, N., and Wu, H. Tool-augmented reward modeling, 2024. URL . Liu, C., Wang, H., Pan, J., Wan, Z., Dai, Y., Lin, F., Bai, W., Rueckert, D., and Arcucci, R. Beyond distillation: Pushing the limits of medical llm reasoning with minimalist rule-based rl, 2025a. URL https: //arxiv.org/abs/2505.17952 . Liu, Z., Chen, C., Li, W., Qi, P., Pang, T., Du, C., Lee, W. S., and Lin, M. Understanding r1-zero-like training: A critical perspective, 2025b. URL org/abs/2503.20783 .", "source": "marker_v2", "marker_block_id": "/page/8/ListGroup/530"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0097", "section": "References", "page_start": 10, "page_end": 10, "type": "Text", "text": "495 496 497 498 Liu, Z., Wang, P., Xu, R., Ma, S., Ruan, C., Li, P., Liu, Y., and Wu, Y. Inference-time scaling for generalist reward modeling, 2025c. URL 2504.02495 .", "source": "marker_v2", "marker_block_id": "/page/9/Text/441"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0098", "section": "References", "page_start": 10, "page_end": 10, "type": "ListGroup", "text": "OpenAI et al. Gpt-4o system card, 2024. URL https: //arxiv.org/abs/2410.21276 . Pal, A., Umapathi, L. K., and Sankarasubbu, M. Medmcqa : A large-scale multi-subject multi-choice dataset for medical domain question answering, 2022. URL https: //arxiv.org/abs/2203.14371 . Shi, W., Xu, R., Zhuang, Y., Yu, Y., Sun, H., Wu, H., Yang, C., and Wang, M. D. Medadapter: Efficient test-time adaptation of large language models towards medical reasoning, 2024. URL 2405.03000 . Singhal, K., Tu, T., Gottweis, J., et al. Toward expertlevel medical question answering with large language models. Nature Medicine, 31:943–950, 2025. doi: 10. 1038/s41591-024-03423-7. Snell, C., Lee, J., Xu, K., and Kumar, A. Scaling llm testtime compute optimally can be more effective than scaling model parameters, 2024. URL org/abs/2408.03314 . Wang, K., Fu, Z., Xin, W., Zhou, L., and Chandrappa, S. K. Digital voices of survival: From social media disclosures to support provisions for domestic violence victims. arXiv preprint arXiv:2509.12288, 2025. Wu, C., Lin, W., Zhang, X., Zhang, Y., Wang, Y., and Xie, W. Pmc-llama: Towards building open-source language models for medicine, 2023. URL org/abs/2304.14454 . Xia, C., Wu, Q., Tian, S., and Hao, Y. Parallelism meets adaptiveness: Scalable documents understanding in multi-agent llm systems, 2025. URL https: //arxiv.org/abs/2507.17061 . Xiao, W., Lian, J. J., Ouyang, K., Gu, S., Ke, Z., Wei, D., Sha, X., Wang, J., Fu, S., Qiu, M., and Xu, C. Newton downhill optimizer with application to engineering optimization and breast cancer feature selection. Biomedical Signal Processing and Control, 117:109184, 2026. ISSN 1746-8094. doi: URL science/article/pii/S1746809425016957 . Xiong, W., Zhao, W., Yuan, W., Golovneva, O., Zhang, T., Weston, J., and Sukhbaatar, S. Stepwiser: Stepwise generative judges for wiser reasoning, 2025. URL .", "source": "marker_v2", "marker_block_id": "/page/9/ListGroup/439"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0099", "section": "References", "page_start": 10, "page_end": 10, "type": "ListGroup", "text": "Xu, R., Chen, J., Ye, J., Wu, Y., Yan, J., Yang, C., and Yu, H. Incentivizing agentic reasoning in llm judges via tool-integrated reinforcement learning, 2025. URL . Yang, A. et al. Qwen2.5 technical report, 2025. You, R., Cai, H., Zhang, C., Xu, Q., Liu, M., Yu, T., Li, Y., and Li, W. Agent-as-a-judge, 2026. URL https: //arxiv.org/abs/2601.05111 . Yun, J., Sohn, J., Park, J., Kim, H., Tang, X., Shao, Y., Koo, Y., Ko, M., Chen, Q., Gerstein, M., Moor, M., and Kang, J. Med-prm: Medical reasoning models with stepwise, guideline-verified process rewards, 2025. URL https: //arxiv.org/abs/2506.11474 . Zhang, H., Lou, Q., and Wang, Y. Towards safe ai clinicians: A comprehensive study on large language model jailbreaking in healthcare, 2025. URL https: //arxiv.org/abs/2501.18632 . Zhang, K., Zeng, S., Hua, E., Ding, N., Chen, Z.-R., Ma, Z., Li, H., Cui, G., Qi, B., Zhu, X., Lv, X., Jinfang, H., Liu, Z., and Zhou, B. Ultramedical: Building specialized generalists in biomedicine, 2024. URL . Zhao, X., Liu, S., Yang, S.-Y., and Miao, C. Medrag: Enhancing retrieval-augmented generation with knowledge graph-elicited reasoning for healthcare copilot, 2025. URL . Zuo, Y., Qu, S., Li, Y., Chen, Z., Zhu, X., Hua, E., Zhang, K., Ding, N., and Zhou, B. Medxpertqa: Benchmarking expert-level medical reasoning and understanding, 2025. URL .", "source": "marker_v2", "marker_block_id": "/page/9/ListGroup/440"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0100", "section": "A. Limitation", "page_start": 11, "page_end": 11, "type": "Text", "text": "While Med-TIV demonstrates substantial improvements over existing medical reasoning verification approaches, several limitations warrant discussion and suggest directions for future research.", "source": "marker_v2", "marker_block_id": "/page/10/Text/2"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0101", "section": "A. Limitation", "page_start": 11, "page_end": 11, "type": "Text", "text": "Process Supervision. Our current training paradigm relies solely on trace-level outcome rewards, providing no supervision on intermediate verification behaviors such as when to search, what queries to formulate, or how to integrate retrieved evidence. While this design eliminates the need for costly step-level annotations, it may lead to suboptimal search patterns or redundant retrieval operations. Future work could explore supervision for the verification task itself, or leverage techniques such as search behavior cloning from stronger models to provide denser optimization signals.", "source": "marker_v2", "marker_block_id": "/page/10/Text/3"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0102", "section": "A. Limitation", "page_start": 11, "page_end": 11, "type": "Text", "text": "Retrieval Corpus Coverage. Med-TIV's verification accuracy is inherently bounded by the coverage and quality of the underlying medical corpus. Our retrieval system indexes documents from PubMed abstracts and medical textbooks, which provides broad coverage of established medical knowledge but may lack recent findings, rare disease information, or region-specific clinical guidelines. Verification of reasoning traces involving cutting-edge treatments or highly specialized subspecialties may be limited by corpus gaps.", "source": "marker_v2", "marker_block_id": "/page/10/Text/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0103", "section": "A. Limitation", "page_start": 11, "page_end": 11, "type": "Text", "text": "Language and Domain Scope. All training and evaluation are conducted on English-language medical reasoning benchmarks. The generalization of Med-TIV to multilingual medical content or non-Western medical traditions remains unexplored. Additionally, while our benchmarks span multiple medical subdomains, certain specialized areas such as genomics, radiology interpretation, and surgical planning may require domain-adapted retrieval corpora for optimal verification performance.", "source": "marker_v2", "marker_block_id": "/page/10/Text/5"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0104", "section": "B.1. Hyperparameter Settings", "page_start": 11, "page_end": 11, "type": "Text", "text": "Table 4 provides comprehensive hyperparameter configurations for Med-TIV training across both iterations. We maintain mostly consistent settings between iterations to isolate the effect of iterative training from hyperparameter tuning.", "source": "marker_v2", "marker_block_id": "/page/10/Text/8"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0105", "section": "B.1. Hyperparameter Settings", "page_start": 11, "page_end": 11, "type": "TableGroup", "text": "Hyperparameters Iteration 1 Iteration 2 RL Algorithm Dr.GRPO Dr.GRPO Clip ratio (low / high) 0.2 / 0.3 0.2 / 0.3 Learning rate 1e-6 1e-6 Warmup steps 10 10 Training epochs 5 5 Global batch size 256 256 Mini-batch size 256 256 Group size (G) 5 8 Rollout sampling temperature 1.0 1.0 Rollout top-p 0.95 0.95 Curriculum filtering Enabled Enabled Table 4. Hyperparameter configurations for Med-TIV training across iterations.", "source": "marker_v2", "marker_block_id": "/page/10/TableGroup/309"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0106", "section": "B.2. Retrieval Setup", "page_start": 11, "page_end": 11, "type": "Text", "text": "We construct our retrieval infrastructure using a dense retrieval architecture optimized for medical domain queries. The corpus is derived from the MedRAG (Zhao et al., 2025) collection, specifically combining the PubMed and Textbooks subcorpora into a unified index. The PubMed subset contains approximately 23.9 million biomedical abstracts covering research publications, while the Textbooks subset includes content from standard medical textbooks spanning clinical medicine, pharmacology, pathology, and related disciplines. After deduplication and quality filtering, the combined corpus contains approximately 24 million snippets.", "source": "marker_v2", "marker_block_id": "/page/10/Text/12"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0107", "section": "B.2. Retrieval Setup", "page_start": 11, "page_end": 11, "type": "Text", "text": "We employ MedCPT (Jin et al., 2023) as our dense retrieval encoder, specifically the query encoder variant for encoding", "source": "marker_v2", "marker_block_id": "/page/10/Text/13"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0108", "section": "B.2. Retrieval Setup", "page_start": 12, "page_end": 12, "type": "Text", "text": "Table 5. Prompt template.", "source": "marker_v2", "marker_block_id": "/page/11/Text/1"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0109", "section": "User Prompt", "page_start": 12, "page_end": 12, "type": "Text", "text": "You are a reasoning validator for medical problems. Your task is to think step by step and evaluate whether the given reasoning trace of a medical problem contains errors.", "source": "marker_v2", "marker_block_id": "/page/11/Text/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0110", "section": "User Prompt", "page_start": 12, "page_end": 12, "type": "Text", "text": "First, you must always perform a step-by-step analysis to examine the entire reasoning process. Then, based on your analysis, you will make a definitive judgment.", "source": "marker_v2", "marker_block_id": "/page/11/Text/5"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0111", "section": "User Prompt", "page_start": 12, "page_end": 12, "type": "ListGroup", "text": "Use 1 if the reasoning trace is free of errors. Use 0 if the reasoning trace contains one or more errors.", "source": "marker_v2", "marker_block_id": "/page/11/ListGroup/387"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0112", "section": "Output Instruction", "page_start": 12, "page_end": 12, "type": "Text", "text": "You must conduct your step-by-step analysis inside <think> and </think> first every time you get new information. After reasoning, if you find you lack some knowledge, you can call a search engine by <search> query </search> and it will return the top searched results between <information> and </information>. You can search as many times as you want. If you find no further external knowledge needed, you can directly provide the answer inside <answer> and </answer>, without detailed illustrations.", "source": "marker_v2", "marker_block_id": "/page/11/Text/9"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0113", "section": "Output Instruction", "page_start": 12, "page_end": 12, "type": "Code", "text": "Medical Problem: {The full Medical Problem on one or more lines.} Reasoning Trace: {The full Reasoning Trace on one or more lines.}", "source": "marker_v2", "marker_block_id": "/page/11/Code/10"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0114", "section": "Output Instruction", "page_start": 12, "page_end": 12, "type": "Text", "text": "search queries and article encoder for encoding corpus snippets. Document embeddings are pre-computed and stored in a FAISS index using the Flat configuration for maximum retrieval accuracy, distributed across multiple GPUs using FAISS's GPU sharding capability to enable parallel similarity search. For each search query, we retrieve the top-3 most relevant documents for both training and inference.", "source": "marker_v2", "marker_block_id": "/page/11/Text/12"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0115", "section": "B.3. Baseline Setup", "page_start": 12, "page_end": 12, "type": "Text", "text": "We describe the configuration of reward model baselines used in our experiments. For Med-PRM, which employs static retrieval-augmented generation, we equip it with the same retrieval corpus, encoder, and top-k setting as our framework to ensure a controlled comparison. MedS 3 does not support external tool invocation and is therefore evaluated without retrieval augmentation. For confidence score extraction and inference hyperparameter settings, we follow the configurations specified in each baseline's original publication.", "source": "marker_v2", "marker_block_id": "/page/11/Text/16"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0116", "section": "B.4. Prompt Template", "page_start": 12, "page_end": 12, "type": "Text", "text": "We design a structured prompt template that guides the verifier through systematic reasoning with explicit tool invocation syntax. The complete prompt is shown in Table 5.", "source": "marker_v2", "marker_block_id": "/page/11/Text/20"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0117", "section": "C.1. Benchmarks", "page_start": 12, "page_end": 12, "type": "Text", "text": "We evaluate Med-TIV on four established medical reasoning benchmarks that collectively assess verification capability across varying difficulty levels and medical subdomains.", "source": "marker_v2", "marker_block_id": "/page/11/Text/25"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0118", "section": "C.1. Benchmarks", "page_start": 12, "page_end": 12, "type": "Text", "text": "• MedQA (Jin et al., 2020) : A dataset of multiple-choice questions derived from the United States Medical Licensing Examination (USMLE), designed to evaluate clinical reasoning and medical knowledge integration across diverse specialties.", "source": "marker_v2", "marker_block_id": "/page/11/Text/27"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0119", "section": "C.1. Benchmarks", "page_start": 12, "page_end": 12, "type": "Text", "text": "• MedMCQA (Pal et al., 2022) : A large-scale multi-subject benchmark sourced from Indian medical entrance examinations (AIIMS and NEET-PG), covering 21 medical subjects with emphasis on factual knowledge and clinical application.", "source": "marker_v2", "marker_block_id": "/page/11/Text/29"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0120", "section": "C.1. Benchmarks", "page_start": 12, "page_end": 12, "type": "Text", "text": "• MMLU-Med (Hendrycks et al., 2021) : An aggregation of medical-related subsets from the Massive Multitask Language Understanding benchmark, encompassing anatomy, clinical knowledge, college biology, college medicine, medical", "source": "marker_v2", "marker_block_id": "/page/11/Text/31"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0121", "section": "C.1. Benchmarks", "page_start": 13, "page_end": 13, "type": "Text", "text": "genetics, and professional medicine.", "source": "marker_v2", "marker_block_id": "/page/12/Text/1"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0122", "section": "C.1. Benchmarks", "page_start": 13, "page_end": 13, "type": "Text", "text": "• MedXpertQA (Zuo et al., 2025) : An expert-level benchmark featuring challenging questions that require multi-step clinical reasoning, differential diagnosis, and treatment planning at the level expected of practicing physicians.", "source": "marker_v2", "marker_block_id": "/page/12/Text/2"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0123", "section": "C.2. Baselines", "page_start": 13, "page_end": 13, "type": "Text", "text": "We compare Med-TIV against comprehensive baselines spanning proprietary systems, general-purpose models, and domainspecialized approaches.", "source": "marker_v2", "marker_block_id": "/page/12/Text/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0124", "section": "Proprietary Models.", "page_start": 13, "page_end": 13, "type": "ListGroup", "text": "GPT-4o-mini (OpenAI et al., 2024) : A compact variant of OpenAI's GPT-4o optimized for efficiency while maintaining strong reasoning capabilities across diverse tasks. Gemini-2.0-Flash: Google's efficient multimodal model designed for fast inference with competitive performance on knowledge-intensive benchmarks.", "source": "marker_v2", "marker_block_id": "/page/12/ListGroup/388"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0125", "section": "General Reasoning Models.", "page_start": 13, "page_end": 13, "type": "ListGroup", "text": "DeepSeek-R1 (Guo et al., 2025) : A 671B parameter reasoning model trained with RL, representing the current frontier of open-weight reasoning capabilities. R1-Distill-Qwen / R1-Distill-Llama: Distilled variants of DeepSeek-R1 at 7B and 8B scales respectively, designed to transfer reasoning capabilities to smaller architectures.", "source": "marker_v2", "marker_block_id": "/page/12/ListGroup/389"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0126", "section": "General Foundation Models.", "page_start": 13, "page_end": 13, "type": "ListGroup", "text": "Qwen2.5 (Yang et al., 2025) : A family of open-weight language models with strong multilingual and reasoning capabilities, evaluated at 7B and 32B parameter scales. Llama3.1 (Grattafiori et al., 2024) : Meta's open-source foundation model demonstrating competitive performance across diverse benchmarks, evaluated at the 8B scale.", "source": "marker_v2", "marker_block_id": "/page/12/ListGroup/390"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0127", "section": "Medical Domain Models.", "page_start": 13, "page_end": 13, "type": "ListGroup", "text": "AlphaMed (Liu et al., 2025a) : A medical reasoning model that employs RL with rule-based rewards to enhance clinical reasoning without reliance on distillation from larger models. UltraMedical (Zhang et al., 2024) : A specialized medical model combining high-quality instruction tuning on curated biomedical corpora with preference optimization for improved clinical accuracy. HuatuoGPT-o1 (Chen et al., 2024) : A medical reasoning model incorporating chain-of-thought reasoning with internal verification mechanisms to improve diagnostic accuracy.", "source": "marker_v2", "marker_block_id": "/page/12/ListGroup/391"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0128", "section": "Medical Reward Models.", "page_start": 13, "page_end": 13, "type": "ListGroup", "text": "MedS 3 (Jiang et al., 2025b) : A self-evolved soft dual-sided process supervision framework for medical reasoning that generates training signals through iterative self-improvement without external annotations. Med-PRM (Yun et al., 2025) : A process reward model for medical reasoning verification that provides step-level supervision using static retrieval-augmented generation with guideline-based verification.", "source": "marker_v2", "marker_block_id": "/page/12/ListGroup/392"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0129", "section": "D. Case Analysis", "page_start": 13, "page_end": 13, "type": "Text", "text": "Table 6 presents a complete verification example illustrating how a Med-TIV trained verifier identifies reasoning errors through dynamic evidence retrieval. The case involves a patient with bladder cancer who develops ototoxicity following chemotherapy. The generator's reasoning trace incorrectly attributes the symptoms to taxanes based on their known association with ototoxicity, concluding with answer (B). However, the model retrieves evidence establishing that cisplatin—the standard neoadjuvant therapy for transitional cell carcinoma—is the causative agent, and its mechanism involves DNA cross-linking rather than microtubule hyperstabilization. Through iterative search and reasoning, verifier correctly identifies the error, demonstrating the value of tool augmentation for catching subtle medical reasoning mistakes.", "source": "marker_v2", "marker_block_id": "/page/12/Text/22"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0130", "section": "D. Case Analysis", "page_start": 14, "page_end": 14, "type": "Text", "text": "Table 6. Complete verification demonstration. Given a medical problem and a reasoning trace, the verifier retrieves relevant evidence and correctly identifies the reasoning error within the trace.", "source": "marker_v2", "marker_block_id": "/page/13/Text/2"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0131", "section": "Medical Problem", "page_start": 14, "page_end": 14, "type": "Text", "text": "Question: A 67-year-old man with transitional cell carcinoma of the bladder comes to the physician because of a 2-day history of ringing sensation in his ear. He received this first course of neoadjuvant chemotherapy 1 week ago. Pure tone audiometry shows a sensorineural hearing loss of 45 dB. The expected beneficial effect of the drug that caused this patient's symptoms is most likely due to which of the following actions?", "source": "marker_v2", "marker_block_id": "/page/13/Text/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0132", "section": "Medical Problem", "page_start": 14, "page_end": 14, "type": "Text", "text": "Options: (A) Inhibition of proteasome (B) Hyperstabilization of microtubules (C) Generation of free radicals (D) Cross-linking of DNA", "source": "marker_v2", "marker_block_id": "/page/13/Text/5"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0133", "section": "Generator Reasoning Trace", "page_start": 14, "page_end": 14, "type": "Text", "text": "The patient's presentation of ringing in the ear, known as tinnitus, and a sensorineural hearing loss of 45 dB, suggest ototoxicity. Ototoxicity is a well-known side effect of certain chemotherapy drugs. Neoadjuvant chemotherapy refers to the use of chemotherapy before other treatments such as surgery. The patient has received a course of neoadjuvant chemotherapy 1 week ago, which aligns with the recent onset of ototoxic symptoms. To determine the cause of ototoxicity from the provided options, we need to identify which of the options is likely to cause sensorineural hearing loss. Option (A), Inhibition of proteasome, is associated with drugs like bortezomib (Velcade), which are less commonly associated with ototoxicity. Option (B), Hyperstabilization of microtubules, is a known mechanism of taxanes (e.g., paclitaxel, docetaxel). Taxanes are known to cause ototoxicity, making this a potential candidate. Option (C), Generation of free radicals, is a general mechanism that can cause cellular damage but is not specifically known to be associated with ototoxicity. Option (D), Cross-linking of DNA, is a known mechanism of alkylating agents like cyclophosphamide. While alkylating agents can cause ototoxicity, this mechanism is not as commonly linked to ototoxicity as other options. Considering the options, the one most commonly associated with ototoxicity is hyperstabilization of microtubules. Therefore, the drug that most likely caused this patient's symptoms is one that hyperstabilizes microtubules. The answer is (B).", "source": "marker_v2", "marker_block_id": "/page/13/Text/7"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0134", "section": "Verifier Judgment", "page_start": 14, "page_end": 14, "type": "Text", "text": "<think> Let's check the common chemotherapeutic agents used for bladder cancer and their known side effects. Transitional cell carcinoma of the bladder is often treated with cisplatin-based chemotherapy as neoadjuvant therapy. We need to determine if any of the options match the known side effects of these drugs. Let's search to confirm the side effects of cisplatin and other common bladder cancer chemotherapy drugs. </think>", "source": "marker_v2", "marker_block_id": "/page/13/Text/9"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0135", "section": "Verifier Judgment", "page_start": 14, "page_end": 14, "type": "ListGroup", "text": "<search> common side effects of cisplatin chemotherapy </search> <information> [Search Results regarding Cisplatin toxicity and mechanisms...] </information> <think> From the information, cisplatin is known to cause ototoxicity, which aligns with the patient's symptoms of sensorineural hearing loss. Now let's check the mechanisms of action for the options: Option (A): Inhibition of proteasome - This is associated with drugs like bortezomib and is not used for bladder cancer. Option (B): Hyperstabilization of microtubules - This is the mechanism of taxanes (e.g., paclitaxel, docetaxel). Option (C): Generation of free radicals - This is not specific to a single mechanism and is not commonly associated with ototoxicity. Option (D): Cross-linking of DNA - This is the mechanism of alkylating agents (e.g., cyclophosphamide). Cisplatin is an alkylating agent, and its mechanism involves cross-linking of DNA. Therefore, the correct answer should be (D). </think>", "source": "marker_v2", "marker_block_id": "/page/13/ListGroup/279"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0136", "section": "Verifier Judgment", "page_start": 14, "page_end": 14, "type": "Text", "text": "<answer>0</answer>", "source": "marker_v2", "marker_block_id": "/page/13/Text/13"}

icml26/19e76363-53a6-4f3c-8b50-844e1aea4e26/dataset_meta.json

@@ -0,0 +1,60 @@
+{
+  "paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26",
+  "pipeline": "Paper2Markdown-V3",
+  "ok": true,
+  "page_count": 14,
+  "chunk_count": 137,
+  "main_body_chunk_count": 95,
+  "appendix_chunk_count": 37,
+  "reference_chunk_count": 5,
+  "model_text_chars": 39229,
+  "raw_markdown_chars": 64704,
+  "sanitized_chars": 58960,
+  "page_provenance": {
+    "min_page": 1,
+    "max_page": 14,
+    "invalid_count": 0
+  },
+  "marker_block_type_counts": {
+    "Caption": 1,
+    "Code": 3,
+    "Equation": 12,
+    "FigureGroup": 3,
+    "Footnote": 3,
+    "ListGroup": 12,
+    "PageFooter": 14,
+    "PageHeader": 17,
+    "SectionHeader": 39,
+    "TableGroup": 4,
+    "Text": 352
+  },
+  "asset_count_raw": 3,
+  "asset_count_model_kept": 3,
+  "asset_count_rejected": 0,
+  "asset_reject_reasons": {
+    "kept": 3
+  },
+  "artifact_leak_audit": {
+    "ok": true,
+    "hits": {
+      "Anonymous Authors": [],
+      "ACKNOWLEDGMENT": [],
+      "OpenReview": [],
+      "\"accept_label\"": [],
+      "\"decision\"": [],
+      "\"decision_tier\"": [],
+      "\"source_status\"": [],
+      "Meta-review": [],
+      "Official Review": [],
+      "official_reviews": [],
+      "meta_reviews": [],
+      "suggested_verdict_score": []
+    },
+    "artifact_count": 2
+  },
+  "default_model_input": "model_text_v3.txt",
+  "appendix_input": "appendix_text_v3.txt",
+  "reference_input": "reference_text_v3.txt",
+  "source": "koala_icml26_due_queue",
+  "run_name": "icml26_20260429_1952_duequeue"
+}

icml26/19e76363-53a6-4f3c-8b50-844e1aea4e26/main_body_chunks.jsonl

@@ -0,0 +1,95 @@
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0000", "section": "Abstract", "page_start": 1, "page_end": 1, "type": "Text", "text": "Large language models have achieved strong performance on medical reasoning benchmarks, yet their deployment in clinical settings demands rigorous verification to ensure factual accuracy. While reward models offer a scalable approach for reasoning trace verification, existing methods face two limitations: they produce only scalar reward values without explicit justification, and they rely on single-pass retrieval that precludes adaptive knowledge access as verification unfolds. We introduce Med-TIV, an agentic framework that addresses these limitations by training medical reasoning verifiers to iteratively query external medical corpora during evaluation. Our approach combines tool-augmented verification with an iterative reinforcement learning paradigm that requires only trace-level supervision, alongside an adaptive curriculum mechanism that dynamically adjusts training data distribution. Across four medical reasoning benchmarks, Med-TIV achieves substantial gains over existing methods, improving MedQA accuracy by 23.5% and MedXpertQA by 32.0% relative to the base generator in particular. Crucially, Med-TIV demonstrates an 8× reduction in sampling budget requirement compared to prior reward model baselines. These findings establish that grounding verification in dynamically retrieved evidence offers a principled path toward more reliable medical reasoning systems.", "source": "marker_v2", "marker_block_id": "/page/0/Text/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0001", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "Large Language Models (LLMs) have demonstrated remarkable capabilities in medical reasoning, achieving competitive performance on clinical question answering, diagnostic", "source": "marker_v2", "marker_block_id": "/page/0/Text/6"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0002", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "FigureGroup", "text": "Figure 1. Comparison of medical reasoning verification paradigms. Text-based judges rely on parametric knowledge and may validate erroneous reasoning, while tool-integrated judges dynamically retrieve evidence to ground their judgments.", "source": "marker_v2", "marker_block_id": "/page/0/FigureGroup/498"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0003", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "inference, and medical knowledge benchmarks (Ji et al., 2025a; Xiao et al., 2026) . While these advances hold significant promise for augmenting clinical decision making and democratizing access to medical expertise, the deployment of LLMs in high-stakes clinical settings demands rigorous verification mechanisms to ensure that generated reasoning is both factually accurate and logically sound (Zhang et al., 2025; Wang et al., 2025) .", "source": "marker_v2", "marker_block_id": "/page/0/Text/11"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0004", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "Reward-based judges have therefore emerged as a scalable solution for evaluating model outputs, supporting both posttraining refinement via reinforcement learning from human feedback (RLHF) and inference-time scaling through tree search (Snell et al., 2024) . These judges can be broadly categorized by the granularity of their supervision. Outcome Reward Models (ORMs) provide sparse trace-level supervision that quantifies the quality of the entire output, while Process Reward Models (PRMs) offer dense step-level feedback that scores each intermediate reasoning step, enabling fine-grained credit assignment and precise error localization within multi-step reasoning. Recent work has adapted both paradigms to the medical domain to assess complex clinical reasoning traces. In parallel, advances in generative reward modeling have extended judge models beyond", "source": "marker_v2", "marker_block_id": "/page/0/Text/12"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0005", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "scalar scoring, enabling them to produce natural-language critiques that explicitly justify their decisions (Liu et al., 2025c; Xiong et al., 2025) .", "source": "marker_v2", "marker_block_id": "/page/1/Text/1"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0006", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "Despite their effectiveness, reward-based judges exhibit fundamental limitations when applied to clinical reasoning tasks (Yun et al., 2025) . A primary concern is the prevalence of hallucinations in critique traces, where judge models generate plausible yet factually incorrect assessments (Figure 1) . This issue is particularly noticeable in the medical domain, where reliable verification demands grounding in authoritative clinical evidence and established medical knowledge. Unverified judgments could lead to the propagation of incorrect diagnostic or treatment recommendations. Existing medical reasoning verifiers typically provide only scalar reward signals, offering little or no justification for their judgments and thus limiting interpretability (Jiang et al., 2025b) . Furthermore, these methods often rely on a static Retrieval-Augmented Generation (RAG) pipeline, in which a fixed set of retrieved documents is prefixed to the context and remains unchanged throughout evaluation (Yun et al., 2025) . Such static design precludes adaptive, multi-turn evidence gathering and forces the verifier to a fixed retrieval budget, thus limiting scalability.", "source": "marker_v2", "marker_block_id": "/page/1/Text/2"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0007", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "To address these issues, we propose Med-TIV (Medical Tool-Integrated reasoning Verifier), an agentic reinforcement learning (RL) framework that trains LLMs to leverage external knowledge bases for judging medical reasoning traces 1 . Med-TIV features three key design principles: (1) a tool-augmented verification paradigm that enables dynamic, iterative knowledge retrieval during the evaluation process; (2) an iterative RL approach that progressively improves verification capabilities without requiring step-level expert annotations; and (3) an adaptive curriculum formulation strategy that adjusts the data distribution in response to the evolving capability of the model. By equipping judge models with tool-use capabilities, Med-TIV grounds evaluation decisions in external evidence rather than relying solely on parametric knowledge, thereby mitigating hallucination, improving interpretability, and overcoming the limitations of static RAG (Ji et al., 2025b; Xia et al., 2025) .", "source": "marker_v2", "marker_block_id": "/page/1/Text/3"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0008", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "To verify the effectiveness of Med-TIV, we conduct extensive experiments on common medical reasoning benchmarks. Our results demonstrate that Med-TIV trains strong medical verifiers: when guiding inference-time search for a 7B generator model, our trained verifier achieves relative improvements of 23.5% on MedQA and 32.0% on MedXpertQA compared to the generator model alone. Moreover, Med-TIV consistently outperforms existing medical reward model baselines and surpasses the performance of models that are up to 4× larger in scale. Notably, Med-TIV", "source": "marker_v2", "marker_block_id": "/page/1/Text/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0009", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "also demonstrates an 8× gain in sampling efficiency compared to prior reward-based approaches, achieving equivalent accuracy with substantially fewer sampled reasoning traces during test-time search.", "source": "marker_v2", "marker_block_id": "/page/1/Text/6"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0010", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "Our main contributions are summarized as follows:", "source": "marker_v2", "marker_block_id": "/page/1/Text/7"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0011", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "ListGroup", "text": "We propose Med-TIV, a novel tool-integrated verification framework that enables dynamic, iterative knowledge retrieval during medical reasoning evaluation, providing both interpretable, fine-grained justification and improved factual grounding. We introduce an iterative RL paradigm with curriculumbased difficulty adaptation that progressively improves verification capabilities through self-bootstrapping, requiring only trace-level supervision rather than dense steplevel expert annotations. Med-TIV achieves state-of-the-art performance on four medical reasoning benchmarks, with comprehensive ablation studies that validate each component's contribution.", "source": "marker_v2", "marker_block_id": "/page/1/ListGroup/657"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0012", "section": "2.1. Problem Setup", "page_start": 2, "page_end": 2, "type": "Text", "text": "We define medical reasoning verification as the task of assessing the correctness of a multi-step reasoning trace generated in response to a medical question. Formally, given a medical question q ∈ Q and a multi-step reasoning trace τ = (s1, s2, . . . , sm) from a generator model, a verifier model determines whether τ contains any errors. We formulate this problem as binary classification, where the verifier Vθ(q, τ ) produces a judgment ℓ ∈ {0, 1}, where ℓ = 1 indicates a error-free reasoning trace, and ℓ = 0 indicates the presence of one or more errors. Unlike scalar reward models that output continuous scores, we adopt a generative judge paradigm in which the verifier produces a discrete judgment accompanied by a detailed critique trace that provides a structured justification for the decision.", "source": "marker_v2", "marker_block_id": "/page/1/Text/13"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0013", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 2, "page_end": 2, "type": "Text", "text": "Following prior works (Jin et al., 2025) , we extend the verifier with access to an external search engine E that retrieves top-k documents from a curated medical corpus (See Appendix B.2 for details). Retrieved documents are appended verbatim to the verifier context. Given a verification instance (q, τ ), the verifier constructs an iterative verfication trajectory. At step k, the trajectory is represented as:", "source": "marker_v2", "marker_block_id": "/page/1/Text/15"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0014", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 2, "page_end": 2, "type": "Equation", "text": "\\mathbf{t}_k = \\{r_1, a_1, o_1, \\dots, r_k, a_k, o_k\\},\\", "source": "marker_v2", "marker_block_id": "/page/1/Equation/16"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0015", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 2, "page_end": 2, "type": "Text", "text": "where r i denotes a natural language reasoning step analyzing the medical content, a i is a search query formulated to retrieve relevant medical knowledge, and o i = E(ai) represents the retrieved documents. The iterative verification", "source": "marker_v2", "marker_block_id": "/page/1/Text/17"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0016", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 2, "page_end": 2, "type": "Footnote", "text": "1 Code is available at PittNAIL/med-tiv", "source": "marker_v2", "marker_block_id": "/page/1/Footnote/5"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0017", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 3, "page_end": 3, "type": "Caption", "text": "Figure 2. Overview of Med-TIV. Left: Tool-integrated verification iteratively analyzes reasoning traces, formulates search queries, and retrieves medical evidence before producing correctness judgments. Middle: Curriculum formulation filters trivial and impossible instances, retaining boundary cases for RL training. Right: At inference time, the verifier evaluates candidate medical reasoning traces generated by a frozen model and final answers are selected via weighted self-consistency.", "source": "marker_v2", "marker_block_id": "/page/2/Caption/2"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0018", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 3, "page_end": 3, "type": "Text", "text": "process is defined as:", "source": "marker_v2", "marker_block_id": "/page/2/Text/3"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0019", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 3, "page_end": 3, "type": "Text", "text": "123124125", "source": "marker_v2", "marker_block_id": "/page/2/Text/29"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0020", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 3, "page_end": 3, "type": "Text", "text": "135136", "source": "marker_v2", "marker_block_id": "/page/2/Text/35"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0021", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 3, "page_end": 3, "type": "Text", "text": "137138", "source": "marker_v2", "marker_block_id": "/page/2/Text/36"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0022", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 3, "page_end": 3, "type": "Equation", "text": "(r_k, a_k) \\sim V_{\\theta}(q, \\tau, \\mathbf{t}_{k-1}), o_k = \\mathcal{E}(a_k), \\mathbf{t}_k = \\mathbf{t}_{k-1} \\oplus r_k \\oplus a_k \\oplus o_k,", "source": "marker_v2", "marker_block_id": "/page/2/Equation/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0023", "section": "2.2. Tool-Augmented Reasoning Verifier", "page_start": 3, "page_end": 3, "type": "Text", "text": "where \\oplus denotes sequence concatenation. This process continues until the verifier produces a final judgment \\ell \\sim V_{\\theta}(q,\\tau,\\mathbf{t}_T) at the terminal step T. By allowing multiple tool executions, the verifier dynamically retrievs medical knowledge as need to verify specific claims in the reasoning trace. Table 5 in the Appendix shows the explicit instruction used in our experiments.", "source": "marker_v2", "marker_block_id": "/page/2/Text/5"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0024", "section": "2.3. Test-Time Search", "page_start": 3, "page_end": 3, "type": "Text", "text": "Test-time search strategies improve reasoning performance by leveraging reward models to evaluate and select among multiple candidate solutions (Shi et al., 2024). Given a frozen generator model \\pi_{\\rm gen} and a question q, we first sample N independent reasoning traces:", "source": "marker_v2", "marker_block_id": "/page/2/Text/7"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0025", "section": "2.3. Test-Time Search", "page_start": 3, "page_end": 3, "type": "Equation", "text": "\\{\\tau^{(j)}\\}_{j=1}^N \\sim \\pi_{\\mathrm{gen}}(\\cdot \\mid q).", "source": "marker_v2", "marker_block_id": "/page/2/Equation/8"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0026", "section": "2.3. Test-Time Search", "page_start": 3, "page_end": 3, "type": "Text", "text": "A trained verifier V_{\\theta} then scores each candidate trace, and the final output is selected based on these scores. Common selection strategies include Best-of-N sampling, which", "source": "marker_v2", "marker_block_id": "/page/2/Text/9"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0027", "section": "2.3. Test-Time Search", "page_start": 3, "page_end": 3, "type": "Text", "text": "selects the trace with the highest score:", "source": "marker_v2", "marker_block_id": "/page/2/Text/10"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0028", "section": "2.3. Test-Time Search", "page_start": 3, "page_end": 3, "type": "Equation", "text": "\\hat{\\tau} = \\arg\\max_{\\tau^{(j)}} V_{\\theta}(q, \\tau^{(j)}),", "source": "marker_v2", "marker_block_id": "/page/2/Equation/11"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0029", "section": "2.3. Test-Time Search", "page_start": 3, "page_end": 3, "type": "Text", "text": "and verification-based majority voting, where candidate traces are first filtered by the verifier and the final answer is determined by consensus among verified traces. Med-TIV trains such a plug-in verifier that provides tool-grounded assessments that can be used to augment decision-making for any frozen generator model at inference time.", "source": "marker_v2", "marker_block_id": "/page/2/Text/12"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0030", "section": "3. Tool-Integrated Medical Reasoning Verifier", "page_start": 3, "page_end": 3, "type": "Text", "text": "Med-TIV is an agentic verification framework that trains models to leverage external knowledge bases for verifying whether a given medical reasoning trace contains errors. We adopt an iterative training approach based on dynamic curriculum learning, which requires no fine-grained step-level expert supervision and trains solely through multiple rounds of reinforcement learning (Figure 2). We next describe the training procedure of Med-TIV in details.", "source": "marker_v2", "marker_block_id": "/page/2/Text/14"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0031", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 3, "page_end": 3, "type": "Text", "text": "Data Construction. All training data across iterations is derived from the open-source Med-PRM dataset. Each original instance consists of a tuple (q, \\tau, \\ell_{\\text{step}}, \\ell_{\\text{trace}}) , where q is a medical question, \\tau is a multi-step reasoning trace, \\ell_{\\text{step}} de-", "source": "marker_v2", "marker_block_id": "/page/2/Text/16"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0032", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "notes step-level labels, and \\ell_{trace} is a trace-level correctness label<sup>2</sup>.", "source": "marker_v2", "marker_block_id": "/page/3/Text/1"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0033", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "At each training iteration, we only utilize the triplet (q,\\tau,\\ell_{\\text{trace}}) with human-annotated trace-level labels. Step-level labels \\ell_{\\text{step}} is intentionally excluded, as Med-TIV is designed to improve verification performance without replying on fine-grained supervision. For each training iteration, we fix the training data budget to 20K instances and enforce a balanced label distribution between correct ( \\ell_{\\text{trace}}=1 ) and incorrect ( \\ell_{\\text{trace}}=0 ) reasoning traces.", "source": "marker_v2", "marker_block_id": "/page/3/Text/2"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0034", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "Algorithm. We employ Dr. GRPO (Liu et al., 2025b) as the RL algorithm for training the verifier. Given a verification instance (q_i, \\tau_i) , we sample a group of G verification trajectories \\{\\mathbf{o}_i\\}_{i=1}^G from the current policy \\pi_\\theta . Each trajectory \\mathbf{o}_i = (o_i^1, \\dots, o_i^{|\\mathbf{o}_i|}) consists of reasoning tokens, search queries, retrieved documents, and a final judgment. The objective is:", "source": "marker_v2", "marker_block_id": "/page/3/Text/3"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0035", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Equation", "text": "\\frac{1}{G} \\sum_{i=1}^{G} \\sum_{t=1}^{|\\mathbf{o}_i|} \\left\\{ \\min \\left[ r_i^t \\hat{A}_i^t, \\operatorname{clip} \\left( r_i^t, 1 - \\epsilon_l, 1 + \\epsilon_h \\right) \\hat{A}_i^t \\right] \\right\\}, \\quad (1)", "source": "marker_v2", "marker_block_id": "/page/3/Equation/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0036", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "where r_i^t = \\frac{\\pi_\\theta(o_i^t|\\mathbf{q},\\mathbf{o}_i^{< t})}{\\pi_{\\theta_{\\mathrm{old}}}(o_i^t|\\mathbf{q},\\mathbf{o}_i^{< t})} , \\mathbf{q} = (q,\\tau) denotes the input prompt containing the question and reasoning trace, \\mathbf{o}_i^{< t} represents previously generated tokens, and \\epsilon_l and \\epsilon_h are the clipping parameters. The advantage term \\hat{A}_i^t is defined as:", "source": "marker_v2", "marker_block_id": "/page/3/Text/5"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0037", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Equation", "text": "\\hat{A}_i^t = R(\\mathbf{q}, \\mathbf{o}_i) - \\text{mean}\\left(\\left\\{R(\\mathbf{q}, \\mathbf{o}_1), \\dots, R(\\mathbf{q}, \\mathbf{o}_G)\\right\\}\\right).", "source": "marker_v2", "marker_block_id": "/page/3/Equation/6"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0038", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "Reward Designs. To facilitate multi-turn RL with tool execution, we design a structured reward covering two complementary objectives, following prior practices (Jin et al., 2025):", "source": "marker_v2", "marker_block_id": "/page/3/Text/7"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0039", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "(i) Correctness Reward (R_c) : This component measures whether the verifier's judgment aligns with the ground-truth label. Let \\mathbf{q} = (q, \\tau) denote the verification prompt and \\ell \\in \\{0, 1\\} the ground-truth label. We define:", "source": "marker_v2", "marker_block_id": "/page/3/Text/8"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0040", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Equation", "text": "R_c = \\mathbb{1}(\\text{extract}(\\mathbf{o}) = \\ell),", "source": "marker_v2", "marker_block_id": "/page/3/Equation/9"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0041", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "where \\mathbb{1}(\\cdot) is the indicator function and extract(o) parses the final judgment from the <answer> tags in the generated trajectory o. Intuitively, R_c=1 if the verifier's decision is correct, and R_c=0 otherwise.", "source": "marker_v2", "marker_block_id": "/page/3/Text/10"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0042", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "(ii) Format Reward (R_f) : To ensure reliable tool use and structured outputs, the verifier is required to adhere to a predefined format. Specifically, reasoning steps must be enclosed within <think> tags, search queries within <search> tags, and the final judgment within <answer>", "source": "marker_v2", "marker_block_id": "/page/3/Text/11"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0043", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "tags. To discourage degenerate outputs, we further penalize excessive tag usage. Specifically, R_f=1 if the output satisfies all formatting constraints and contains no more than 10 < answer> tag pairs; <math>R_f=0.25 if the output is correct but exhibits tag overflow; and R_f=0 otherwise.", "source": "marker_v2", "marker_block_id": "/page/3/Text/13"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0044", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Text", "text": "The final reward {\\cal R} is defined as the product of the two components:", "source": "marker_v2", "marker_block_id": "/page/3/Text/14"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0045", "section": "3.1. Tool-Integrated RL with Verifiable Rewards", "page_start": 4, "page_end": 4, "type": "Equation", "text": "R = R_c \\times R_f .", "source": "marker_v2", "marker_block_id": "/page/3/Equation/15"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0046", "section": "3.2. Training Strategies", "page_start": 4, "page_end": 4, "type": "Text", "text": "Adaptive Curriculum Formulation. A central challenge in RL for verification is ensuring that training data remains appropriately calibrated to the evolving capability of the model. Instances that are either trivially easy or impossibly difficult yields minimal learning signal, as the resulting policy gradients approach zero. To address this issue, we adopt a model-aware curriculum formulation mechanism that dynamically adapts the task distribution at each training iteration.", "source": "marker_v2", "marker_block_id": "/page/3/Text/17"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0047", "section": "3.2. Training Strategies", "page_start": 4, "page_end": 4, "type": "Text", "text": "Concretely, before each iteration t, we perform online filtering on the sampled batch \\mathcal{B}_t to construct an effective training set \\mathcal{D}_t :", "source": "marker_v2", "marker_block_id": "/page/3/Text/18"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0048", "section": "3.2. Training Strategies", "page_start": 4, "page_end": 4, "type": "Equation", "text": "\\mathcal{D}_t = \\{(q, \\tau, \\ell) \\in \\mathcal{B}_t : \\exists g, g' \\in \\{1, \\dots, G\\} \\text{ s.t. } r^{(g)} \\neq r^{(g')}\\}.", "source": "marker_v2", "marker_block_id": "/page/3/Equation/19"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0049", "section": "3.2. Training Strategies", "page_start": 4, "page_end": 4, "type": "Text", "text": "Here, for each candidate instance (q, \\tau, \\ell) \\in \\mathcal{B}_t , we sample G verification trajectories \\{o^{(g)}\\}_{g=1}^G from the current policy \\pi_{\\theta_t} . We then compute the corresponding rewards \\{r^{(g)}\\}_{g=1}^G . Finally, we retain only instances if any two rewards are different, i.e., reward variance is non-zero (Khatri et al., 2025).", "source": "marker_v2", "marker_block_id": "/page/3/Text/20"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0050", "section": "3.2. Training Strategies", "page_start": 4, "page_end": 4, "type": "Text", "text": "This criterion eliminates prompts where the model either consistently succeeds or consistently fails across all sampled trajectories. By filtering these zero-gradient instances, optimization is focused on decision-boundary cases where the verifier exhibits uncertainty.", "source": "marker_v2", "marker_block_id": "/page/3/Text/21"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0051", "section": "3.2. Training Strategies", "page_start": 4, "page_end": 4, "type": "Text", "text": "To maintain a fixed training budget per iteration, we iteratively resample additional instances from the labeled pool \\mathcal{B} and apply the same filtering criterion until |\\mathcal{D}_t|=20K . This dynamic curriculum evolves naturally across iterations as the verifier improves, eliminating the need for manually designed difficulty schedules.", "source": "marker_v2", "marker_block_id": "/page/3/Text/22"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0052", "section": "3.2. Training Strategies", "page_start": 4, "page_end": 4, "type": "Text", "text": "Iterative Training via Self-Bootstrapping. We adopt an iterative training approach that progressively improves verification capabilities through multiple rounds of RL. Unlike prior work that alternates between rejection sampling, supervised fine-tuning (SFT), and RL (Xu et al., 2025), our approach operates entirely through iterative RL, following the RL-Zero paradigm where the model reinforces its verification capabilities without requiring dense turn-level expert demonstrations for cold start.", "source": "marker_v2", "marker_block_id": "/page/3/Text/23"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0053", "section": "3.2. Training Strategies", "page_start": 4, "page_end": 4, "type": "Footnote", "text": "& lt;sup>2</sup>Dataset is available at", "source": "marker_v2", "marker_block_id": "/page/3/Footnote/12"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0054", "section": "3.2. Training Strategies", "page_start": 5, "page_end": 5, "type": "Text", "text": "Algorithm 1 Iterative Training of Tool-Integrated Medical Reasoning Verifier", "source": "marker_v2", "marker_block_id": "/page/4/Text/1"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0055", "section": "3.2. Training Strategies", "page_start": 5, "page_end": 5, "type": "Text", "text": "249250", "source": "marker_v2", "marker_block_id": "/page/4/Text/45"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0056", "section": "3.2. Training Strategies", "page_start": 5, "page_end": 5, "type": "Code", "text": "Require: Base verifier \\pi_{\\theta_0}, labeled dataset pool \\mathcal{D} = \\{(q_i, \\tau_i, \\ell_i)\\}_{i=1}^N, maximum iterations T_{\\text{max}}, batch size B, group size G, search engine \\mathcal{E} Ensure: Trained verifier \\pi_{\\theta^*} 1: for t = 1 to T_{\\text{max}} do Sample labeled batch \\mathcal{B}_t \\leftarrow \\text{SAMPLEBATCH}(\\mathcal{D}, B) 3: 4: \\mathcal{D}_t \\leftarrow \\emptyset ▷ Curriculum formulation 5: for each (q, \\tau, \\ell) \\in \\mathcal{B}_t do 6: Sample verification trajectories: 7: \\{\\hat{\\ell}^{(g)}\\}_{q=1}^G \\sim \\pi_{\\theta_t}(\\cdot \\mid q, \\tau, \\mathcal{E}) Compute rewards within group: 8: r^{(\\hat{g})} \\leftarrow \\mathbb{1}[\\hat{\\ell}^{(g)} = \\ell], \\text{ for } g \\in 1, \\dots, G if \\exists g \\neq g' such that r^{(g)} \\neq r^{(g')} then 9: Add (q, \\tau, \\ell) to curriculum set \\mathcal{D}_t 10: 11: end if 12: end for 13: ▶ RL optimization on curriculum data \\pi_{\\theta_{t+1}} \\leftarrow \\text{DR.GRPO}(\\pi_{\\theta_t}, \\mathcal{D}_t, \\mathcal{E}) 15: end for 16: Return \\pi_{\\theta_{T_{\\max}}}", "source": "marker_v2", "marker_block_id": "/page/4/Code/2"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0057", "section": "3.2. Training Strategies", "page_start": 5, "page_end": 5, "type": "Text", "text": "Starting from the base model \\pi_{\\theta_0} , we perform T_{\\text{max}} iterations. Each iteration consists of three stages:", "source": "marker_v2", "marker_block_id": "/page/4/Text/3"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0058", "section": "3.2. Training Strategies", "page_start": 5, "page_end": 5, "type": "Code", "text": "\\begin{split} \\mathcal{B}_t \\leftarrow \\text{SampleBatch}(\\mathcal{D}, B), \\\\ \\mathcal{D}_t \\leftarrow \\text{Filter}(\\mathcal{B}_t, \\pi_{\\theta_t}), \\\\ \\pi_{\\theta_{t+1}} \\leftarrow \\text{RL}(\\pi_{\\theta_t}, \\mathcal{D}_t). \\end{split}", "source": "marker_v2", "marker_block_id": "/page/4/Code/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0059", "section": "3.2. Training Strategies", "page_start": 5, "page_end": 5, "type": "Text", "text": "Each iteration draws a fresh batch \\mathcal{B}_t from the annotated pool \\mathcal{D} with trace-level labels, ensuring a balanced distribution of correct and incorrect reasoning traces. The curriculum filtering then constructs the training set \\mathcal{D}_t as described above, and RL optimization updates the policy based on the structured reward.", "source": "marker_v2", "marker_block_id": "/page/4/Text/5"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0060", "section": "3.2. Training Strategies", "page_start": 5, "page_end": 5, "type": "Text", "text": "The key insight underlying this iterative approach is the coevolution of model capability and training distribution. As the verifier improves, the filtering mechanism automatically removes instances that have become too easy, while the fresh sampling introduces new challenging cases. This creates a self-bootstrapping cycle: stronger models encounter harder verification tasks, which in turn drive further improvements. Since the trace-level correctness reward is deterministic and unambiguous, this self-bootstrapping process converges reliably without the instabilities that can arise from noisy synthetic step-level labels. We summarize the overall training procedure in Algorithm 1.", "source": "marker_v2", "marker_block_id": "/page/4/Text/6"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0061", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Text", "text": "Evaluation benchmarks. We evaluated Med-TIV on four open-source medical question-answering benchmarks: MedQA (Jin et al., 2020), MedMCQA (Pal et al., 2022), MMLU-Med (Hendrycks et al., 2021), and MedX-pertQA (Zuo et al., 2025), using accuracy as the evaluation metric. These benchmarks collectively assess the verifier's ability to distinguish correct from erroneous reasoning across varying difficulty levels and medical subdomains. Detailed descriptions of benchmarks are in Appendix C.1.", "source": "marker_v2", "marker_block_id": "/page/4/Text/9"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0062", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Text", "text": "Implementation details. We trained verifiers using two light-weight backbone models: Llama3.1-8B and Qwen2.5-7B, with Llama3.1-8B as the default for results reporting. All training was conducted using the VeRL-Tool framework (Jiang et al., 2025a). Detailed hyperparameters are shown in Appendix B.1. All experiments were conducted on 4 NVIDIA H100 GPUs with 80GB of memory. Due to computational constraints, we limit the maximum number of RL iterations to T_{\\rm max}=2 and we set the group size for curriculum formulation (Section 3.2) to G=8.", "source": "marker_v2", "marker_block_id": "/page/4/Text/10"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0063", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Text", "text": "For inference, we used the default sampling hyperparameters for all models. In reward-guided search experiments, unless otherwise specified, we used Qwen2.5-7B as the frozen generator and sampled up to 32 candidate reasoning traces per question. We applied Hard-Weighted Self-Consistency as the default test-time search strategy.", "source": "marker_v2", "marker_block_id": "/page/4/Text/11"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0064", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Text", "text": "Baselines. We compared Med-TIV against two groups of baselines. 1): Off-the-shelf LLMs : GPT-40-mini (OpenAI et al., 2024), Gemini-2.0-Flash, DeepSeek-R1 series (Guo et al., 2025), Qwen2.5 series (Yang et al., 2025), Llama3.1 (Grattafiori et al., 2024), AlphaMed (Liu et al., 2025a), UltraMedical (Zhang et al., 2024), and HuatuoGPT-01 (Chen et al., 2024). 2): Medical domain-specialized Reward Models : MedS<sup>3</sup> (Jiang et al., 2025b) and Med-PRM (Yun et al., 2025). Detailed descriptions of each reward model baseline are shown in Appendix B.3.", "source": "marker_v2", "marker_block_id": "/page/4/Text/12"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0065", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Text", "text": "Test-Time Search Strategies. We evaluated three test-time search strategies that leverage Med-TIV to improve the reasoning performance of frozen generators. Given a reasoning trace \\tau=(s_1,s_2,\\ldots,s_K) with K steps, our verifier assigns a confidence score r_{\\tau}\\in[0,1] for the entire trace, defined as the softmax probability of the 1 token over the logits of both 1 and 0 tokens.", "source": "marker_v2", "marker_block_id": "/page/4/Text/13"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0066", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Text", "text": "• Best-of-N. Given a question q, we sampled N candidate traces \\{\\tau^{(j)}\\}_{j=1}^N from the generator and selected the trace with the highest verifier confidence score:", "source": "marker_v2", "marker_block_id": "/page/4/Text/14"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0067", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Equation", "text": "\\hat{\\tau} = \\arg\\max_{\\tau^{(j)}} r_{\\tau^{(j)}}.", "source": "marker_v2", "marker_block_id": "/page/4/Equation/15"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0068", "section": "4.1. Experimental Setup", "page_start": 6, "page_end": 6, "type": "TableGroup", "text": "Table 1. Main evaluation results on medical reasoning benchmarks. We report accuracy (%) on MedQA, MedMCQA, MMLU-Med, and MedXpertQA. Bold numbers indicate the best results among the reward model group. ✓: Verifier supports external tools for judging; ✗: Verifier does not support external tools for judging. Baselines å |Train| Size MedQA MedMCQA MMLU-Med MedXpertQA Avg. Proprietary Models GPT-4o-mini - - - 79.03 68.20 87.79 17.84 63.22 Gemini-2.0-Flash - - - 87.51 72.60 92.01 20.57 68.17 General Reasoning Models DeepSeek-R1 - - 671B 90.34 78.80 94.40 37.76 75.33 R1-Distill-Qwen - - 7B 24.82 36.40 47.47 7.43 29.03 R1-Distill-Llama - - 8B 34.96 43.60 64.19 5.35 37.03 General Non-reasoning Models Qwen2.5 - - 32B 73.21 64.83 84.94 13.87 59.21 Qwen2.5 - - 7B 60.96 56.56 76.96 12.15 51.66 Llama3.1 - - 8B 70.93 61.60 78.97 13.02 56.13 Medical Reasoning Models AlphaMed - - 7B 71.01 61.46 81.16 19.16 58.20 UltraMedical - - 8B 72.66 62.60 79.61 15.25 57.53 HuatuoGPT-o1 - - 8B 72.19 63.60 75.30 16.84 56.98 Medical Reward Models MedS3 ✗ 225k 7B 64.89 58.91 80.53 12.90 54.31 Med-PRM ✓ 111k 7B 69.99 62.36 80.99 13.51 56.71 Med-TIV (Ours) ✓ 20k 7B 75.26 64.70 85.58 16.04 60.40", "source": "marker_v2", "marker_block_id": "/page/5/TableGroup/739"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0069", "section": "4.1. Experimental Setup", "page_start": 6, "page_end": 6, "type": "Text", "text": "• Hard-Weighted Self-Consistency. We first filtered traces by the verifier's binary judgment, keeping only those labeled correct (Vθ(q, τ ) = 1). Among the filtered traces, we applied majority voting to determine the final answer:", "source": "marker_v2", "marker_block_id": "/page/5/Text/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0070", "section": "4.1. Experimental Setup", "page_start": 6, "page_end": 6, "type": "Equation", "text": "\\hat{a} = \\arg\\max_{a} \\sum_{j=1}^{N} \\mathbb{1} \\left[ V_{\\theta}(q, \\tau^{(j)}) = 1 \\right] \\cdot \\mathbb{1} \\left[ \\operatorname{ans}(\\tau^{(j)}) = a \\right].", "source": "marker_v2", "marker_block_id": "/page/5/Equation/5"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0071", "section": "4.1. Experimental Setup", "page_start": 6, "page_end": 6, "type": "Text", "text": "• Soft-Weighted Self-Consistency. Instead of binary filtering, we weighted each trace's vote by the verifier's confidence score:", "source": "marker_v2", "marker_block_id": "/page/5/Text/6"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0072", "section": "4.1. Experimental Setup", "page_start": 6, "page_end": 6, "type": "Equation", "text": "\\hat{a} = \\arg\\max_{a} \\sum_{j=1}^{N} r_{\\tau^{(j)}} \\cdot \\mathbb{1} \\big[ \\mathrm{ans}(\\tau^{(j)}) = a \\big].", "source": "marker_v2", "marker_block_id": "/page/5/Equation/7"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0073", "section": "4.2. Main Results", "page_start": 6, "page_end": 6, "type": "Text", "text": "Table 1 presents the main results on four medical reasoning benchmarks. Models trained with Med-TIV consistently outperform existing baselines across all benchmarks. Specifically, under guided-search using a Med-TIVtrained verifier, Qwen2.5-7B attains accuracies of 75.26%", "source": "marker_v2", "marker_block_id": "/page/5/Text/9"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0074", "section": "4.2. Main Results", "page_start": 6, "page_end": 6, "type": "Text", "text": "on MedQA, 64.70% on MedMCQA, 85.58% on MMLU-Med, and 16.04% on MedXpertQA, yielding an average accuracy of 60.40%. Notably, Med-TIV enables this 7B generator to rival substantially larger models, even surpassing the base performance of Qwen2.5-32B despite using a generator that is approximately 4× smaller. Compared to domain-specialized medical reasoning models of similar scale, Med-TIV outperforms HuatuoGPT-o1-8B and UltraMedical-8B by 3.07% and 2.60% on MedQA, respectively, demonstrating the effectiveness of our tool-integrated verification. Case analysis in Appendix D further illustrates how Med-TIV identifies subtle reasoning errors.", "source": "marker_v2", "marker_block_id": "/page/5/Text/10"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0075", "section": "4.3. Analysis", "page_start": 6, "page_end": 6, "type": "Text", "text": "We conducted a series of ablation analyses to investigate six key research questions regarding our proposed framework.", "source": "marker_v2", "marker_block_id": "/page/5/Text/12"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0076", "section": "4.3. Analysis", "page_start": 6, "page_end": 6, "type": "Text", "text": "Q1: Does Med-TIV generalize across different generator models? To evaluate the generalizability of the trained verifier, we applied Med-TIV to guide test-time search across generator models of varying sizes and capa-", "source": "marker_v2", "marker_block_id": "/page/5/Text/13"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0077", "section": "4.3. Analysis", "page_start": 7, "page_end": 7, "type": "FigureGroup", "text": "Figure 3. Test-time scaling analysis across three medical reasoning benchmarks. Each plot shows accuracy versus sampling budget N \\in \\{1, 2, 4, 8, 16, 32\\} for four baselines. Med-TIV consistently outperforms baselines across all sampling budgets and benchmarks.", "source": "marker_v2", "marker_block_id": "/page/6/FigureGroup/226"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0078", "section": "4.3. Analysis", "page_start": 7, "page_end": 7, "type": "TableGroup", "text": "Table 2. Performance improvements from using Med-TIV as a verifier on MedQA. For each generator model, the first row indicates the accuracy over single sampled trace per question. Models MedQA Qwen2.5-7B 60.96 + Self-Consistency 66.38 (+5.42) + Best-of-N (Med-TIV) 72.35 (+11.39) + Soft Weighted SC (Med-TIV) 75.02 (+14.06) + Hard Weighted SC (Med-TIV) 75.26 (+14.30) AlphaMed-7B 71.01 + Self-Consistency 74.23 (+3.22) + Best-of-N (Med-TIV) 75.02 (+4.01) + Soft Weighted SC (Med-TIV) 75.33 (+4.32) + Hard Weighted SC (Med-TIV) 75.65 (+4.64) Qwen2.5-32B 73.21 + Self-Consistency 75.26 (+2.05) + Best-of-N (Med-TIV) 75.57 (+2.36) + Soft Weighted SC (Med-TIV) 75.96 (+2.75) + Hard Weighted SC (Med-TIV) 75.96 (+2.75)", "source": "marker_v2", "marker_block_id": "/page/6/TableGroup/227"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0079", "section": "4.3. Analysis", "page_start": 7, "page_end": 7, "type": "Text", "text": "bilities. As shown in Table 2, when using Qwen2.5-7B as the generator, Hard-Weighted Self-Consistency yields a relative improvement of 23.5% over the base model's single-sample accuracy, substantially outperforming the 12.2% gain achieved by standard Self-Consistency. Notably, the domain-specialized AlphaMed-7B model also benefits from verifier guidance with a 6.5% relative improvement, indicating that our verifier provides complementary verification capabilities beyond domain-specific fine-tuning. The improvements extend to larger models as well: Qwen2.5-32B achieves a 3.8% relative gain during test-time search, demonstrating that a light-weight 8B verifier can effectively guide models that are significantly larger than itself. This", "source": "marker_v2", "marker_block_id": "/page/6/Text/6"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0080", "section": "4.3. Analysis", "page_start": 7, "page_end": 7, "type": "Text", "text": "cross-scale generalization suggests that Med-TIV learns transferable verification patterns rather than overfitting to specific generator characteristics.", "source": "marker_v2", "marker_block_id": "/page/6/Text/7"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0081", "section": "4.3. Analysis", "page_start": 7, "page_end": 7, "type": "Text", "text": "Q2: How do different test-time search strategies compare under Med-TIV? We then systematically compare different test-time search strategies under verifier guidance to identify the most effective approach for leveraging verification signals. As shown in Table 2, Hard-Weighted Self-Consistency consistently achieves the highest accuracy across all generators, followed by Soft-Weighted Self-Consistency and Best-of-N selection. On Qwen2.5-7B, Hard-Weighted Self-Consistency outperforms Best-of-N by 3% absolute accuracy, suggesting that majority voting among verified traces provides more robust answer selection than simply choosing the highest-confidence individual trace.", "source": "marker_v2", "marker_block_id": "/page/6/Text/8"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0082", "section": "4.3. Analysis", "page_start": 7, "page_end": 7, "type": "Text", "text": "Q3: Can Med-TIV reduce the sampling budget required to achieve state-of-the-art performance compared to existing baselines? Next, we investigated how verification performance scales with sampling budget, a critical consideration for deployment under varying computational constraints. As shown in Figure 3, Med-TIV achieves substantial efficiency advantage over existing medical reward models across all three benchmarks. In particular, Med-TIV matches the performance of baselines using only 4 samples, whereas the baselines require 32 samples, representing an 8× reduction in sampling budget. On MedOA, Med-TIV achieves 72.1% accuracy at N=4, while Med-PRM requires the full N=32 budget to reach 70.0% accuracy. Since inference cost scales approximately linearly with the number of sampled traces, this translates to equivalent performance at one-eighth the generator inference cost in practical deployment settings.", "source": "marker_v2", "marker_block_id": "/page/6/Text/9"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0083", "section": "4.3. Analysis", "page_start": 7, "page_end": 7, "type": "Text", "text": "Q4: Does Med-TIV generalize across different base models? To assess the generality of our proposed framework, we compared verification performance using two", "source": "marker_v2", "marker_block_id": "/page/6/Text/10"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0084", "section": "4.3. Analysis", "page_start": 8, "page_end": 8, "type": "FigureGroup", "text": "Figure 4. Ablation on base model selection and training iterations.", "source": "marker_v2", "marker_block_id": "/page/7/FigureGroup/401"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0085", "section": "4.3. Analysis", "page_start": 8, "page_end": 8, "type": "Text", "text": "distinct verifier backbones: Llama3.1-8B and Qwen2.5-7B. As shown in Figure 4, both backbones achieve strong performance after two training iterations. Llama3.1-8B consistently outperforms Qwen2.5-7B by approximately 3.5% absolute accuracy on MedQA, achieving 75.86% versus 72.35% after 2 iterations of training. The parallel performance gains observed across both models indicate that Med-TIV is agnostic to backbone architectures.", "source": "marker_v2", "marker_block_id": "/page/7/Text/3"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0086", "section": "4.3. Analysis", "page_start": 8, "page_end": 8, "type": "Text", "text": "Q5: What is the impact of iterative training? Figure 4 presents ablation results examining the impact of iterative training with adaptive curriculum formulation. Llama3.1-8B improves from 60.96% to 75.26% after iteration 1, with marginal gains to 75.86% at iteration 2. Qwen2.5-7B follows a similar pattern, reaching 72.35% after two iterations. The rapid convergence suggest that the majority of verification capability is acquired in the first round, with subsequent iterations refining boundary cases.", "source": "marker_v2", "marker_block_id": "/page/7/Text/4"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0087", "section": "4.3. Analysis", "page_start": 8, "page_end": 8, "type": "Text", "text": "O6 : How does RL and tool integration impact verification performance? Table 3 highlights the dual benefits of our framework across two generators. RL", "source": "marker_v2", "marker_block_id": "/page/7/Text/5"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0088", "section": "4.3. Analysis", "page_start": 8, "page_end": 8, "type": "TableGroup", "text": "Table 3. Ablation on RL and tool integration. Models MedQA Qwen2.5-7B 60.96 + Med-TIV (RL) 69.60 + Med-TIV (RL + Tool) 70.54 AlphaMed-7B 71.01 + Med-TIV (RL) 76.12 + Med-TIV (RL + Tool) 77.14", "source": "marker_v2", "marker_block_id": "/page/7/TableGroup/402"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0089", "section": "4.3. Analysis", "page_start": 8, "page_end": 8, "type": "Text", "text": "training drives the primary gain, boosting MedQA accuracy of Qwen2.5-7B by 8.64%, confirming that the verifier effectively internalizes reasoning patterns. Tool integration provides a critical secondary boost, further elevating accuracy to 70.54%. A similar cumulative trend is observed with AlphaMed-7B. This demonstrates that while RL", "source": "marker_v2", "marker_block_id": "/page/7/Text/8"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0090", "section": "4.3. Analysis", "page_start": 8, "page_end": 8, "type": "Text", "text": "anchors logical verification, dynamic retrieval is essential for resolving knowledge-intensive boundary cases beyond the model's parametric memory.", "source": "marker_v2", "marker_block_id": "/page/7/Text/9"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0091", "section": "5. Related Work", "page_start": 8, "page_end": 8, "type": "Text", "text": "Medical Reasoning Models. The application of large language models to medical reasoning has attracted considerable attention. Early efforts focused on domain-adaptive pretraining and instruction tuning on medical corpora (Wu et al., 2023; Singhal et al., 2025; Chen et al., 2023). More recent work has explored reasoning-enhanced medical models. HuatuoGPT-o1 (Chen et al., 2024) incorporates chainof-thought reasoning with verification mechanisms, and UltraMedical (Zhang et al., 2024) combines high-quality instruction data with preference optimization. AlphaMed (Liu et al., 2025a) employs RL to improve medical reasoning capabilities. Despite these advances, most existing approaches focus on improving the generator model itself, whereas our work addresses the complementary problem of training a plug-and-play verifier that can improve any frozen generator through test-time search.", "source": "marker_v2", "marker_block_id": "/page/7/Text/11"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0092", "section": "5. Related Work", "page_start": 8, "page_end": 8, "type": "Text", "text": "Tool-Assisted Reward and Judge Models. Standard LLM-based judges typically function as passive scorers limited by parametric knowledge. Recent work addresses this through agentic reward modeling, equipping verifiers with executable tools. Themis (Li et al., 2024) established the foundational framework by enabling access to calculators, search engines, and knowledge bases through structured tool-calling traces. TIR-Judge (Xu et al., 2025) advanced this paradigm in the general domain by integrating code execution to judge paired responses. TIM-PRM (Kuang et al., 2025) introduced independent tool queries for multi-modal verification to eliminate confirmation bias. The concept has further expanded to the Agent-as-a-Judge paradigm (You et al., 2026), which employs dynamic planning, tool augmentation and multi-agent coordination to decompose complex evaluation tasks. Our work instantiates this agentic paradigm within the medical domain, moving beyond static retrieval to iterative, evidence-grounded clinical verification.", "source": "marker_v2", "marker_block_id": "/page/7/Text/12"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0093", "section": "6. Conclusion", "page_start": 8, "page_end": 8, "type": "Text", "text": "We presented Med-TIV, an agentic RL framework for medical reasoning verification. Our approach addresses key limitations of existing medical reward models by offering explicit critique traces and enabling dynamic knowledge retrieval during verification. Empirical evaluations across four medical reasoning benchmarks demonstrate that Med-TIV substantially outperforms prior approaches. More broadly, Med-TIV introduces a general paradigm for training toolaugmented verifiers that can be extended to other highstakes domains requiring evidence-grounded evaluation.", "source": "marker_v2", "marker_block_id": "/page/7/Text/14"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0094", "section": "Impact Statement", "page_start": 9, "page_end": 9, "type": "Text", "text": "This paper introduces research aimed at improving the reliability of large language models for medical reasoning tasks. We believe our work contributes positively to the development of trustworthy medical AI systems by providing mechanisms to verify reasoning correctness before clinical deployment. Med-TIV holds potential to enhance the safety of LLM-assisted clinical decision support by reducing erroneous reasoning outputs through systematic verification. By grounding judgments in retrieved medical evidence, our approach offers improved transparency compared to opaque scalar reward models, enabling practitioners to better understand and audit verification decisions. The efficiency gains demonstrated by Med-TIV could democratize access to reliable medical reasoning verification, making robust verification feasible even in resource-constrained settings.", "source": "marker_v2", "marker_block_id": "/page/8/Text/2"}

icml26/19e76363-53a6-4f3c-8b50-844e1aea4e26/marker_meta.json

@@ -0,0 +1,1668 @@
+{
+  "table_of_contents": [
+    {
+      "title": "Scaling Medical Reasoning Verification via Tool-Integrated Reinforcement\nLearning",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          70.5234375,
+          88.9453125
+        ],
+        [
+          525.8800659179688,
+          88.9453125
+        ],
+        [
+          525.8800659179688,
+          122.191162109375
+        ],
+        [
+          70.5234375,
+          122.191162109375
+        ]
+      ]
+    },
+    {
+      "title": "Anonymous Authors1",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          239.0625,
+          159.67803955078125
+        ],
+        [
+          330.7948913574219,
+          159.67803955078125
+        ],
+        [
+          330.7948913574219,
+          170.8990478515625
+        ],
+        [
+          239.0625,
+          170.8990478515625
+        ]
+      ]
+    },
+    {
+      "title": "Abstract",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          148.21875,
+          192.97265625
+        ],
+        [
+          194.6832733154297,
+          192.97265625
+        ],
+        [
+          194.6832733154297,
+          205.94830322265625
+        ],
+        [
+          148.21875,
+          205.94830322265625
+        ]
+      ]
+    },
+    {
+      "title": "1. Introduction",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          54.685546875,
+          592.83984375
+        ],
+        [
+          132.27606201171875,
+          592.83984375
+        ],
+        [
+          132.27606201171875,
+          605.7343597412109
+        ],
+        [
+          54.685546875,
+          605.7343597412109
+        ]
+      ]
+    },
+    {
+      "title": "2. Preliminaries",
+      "heading_level": null,
+      "page_id": 1,
+      "polygon": [
+        [
+          306.59765625,
+          311.30859375
+        ],
+        [
+          388.2452087402344,
+          311.30859375
+        ],
+        [
+          388.2452087402344,
+          323.56036376953125
+        ],
+        [
+          306.59765625,
+          323.56036376953125
+        ]
+      ]
+    },
+    {
+      "title": "2.1. Problem Setup",
+      "heading_level": null,
+      "page_id": 1,
+      "polygon": [
+        [
+          306.0,
+          332.50750732421875
+        ],
+        [
+          388.06732177734375,
+          332.50750732421875
+        ],
+        [
+          388.06732177734375,
+          342.4700927734375
+        ],
+        [
+          306.0,
+          342.4700927734375
+        ]
+      ]
+    },
+    {
+      "title": "2.2. Tool-Augmented Reasoning Verifier",
+      "heading_level": null,
+      "page_id": 1,
+      "polygon": [
+        [
+          307.1953125,
+          530.96484375
+        ],
+        [
+          476.9296875,
+          530.96484375
+        ],
+        [
+          476.9296875,
+          541.6730804443359
+        ],
+        [
+          307.1953125,
+          541.6730804443359
+        ]
+      ]
+    },
+    {
+      "title": "2.3. Test-Time Search",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          52.59375,
+          570.75
+        ],
+        [
+          147.0,
+          570.75
+        ],
+        [
+          147.0,
+          579.69140625
+        ],
+        [
+          52.59375,
+          579.69140625
+        ]
+      ]
+    },
+    {
+      "title": "3. Tool-Integrated Medical Reasoning Verifier",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          306.0,
+          523.5
+        ],
+        [
+          542.25,
+          523.5
+        ],
+        [
+          542.25,
+          534.05859375
+        ],
+        [
+          306.0,
+          534.05859375
+        ]
+      ]
+    },
+    {
+      "title": "3.1. Tool-Integrated RL with Verifiable Rewards",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          305.25,
+          651.62109375
+        ],
+        [
+          513.0,
+          651.62109375
+        ],
+        [
+          513.0,
+          661.5
+        ],
+        [
+          305.25,
+          661.5
+        ]
+      ]
+    },
+    {
+      "title": "3.2. Training Strategies",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          306.0,
+          184.078125
+        ],
+        [
+          407.25,
+          184.078125
+        ],
+        [
+          407.25,
+          193.359375
+        ],
+        [
+          306.0,
+          193.359375
+        ]
+      ]
+    },
+    {
+      "title": "4. Experiments",
+      "heading_level": null,
+      "page_id": 4,
+      "polygon": [
+        [
+          304.505859375,
+          68.44921875
+        ],
+        [
+          385.5,
+          68.44921875
+        ],
+        [
+          385.5,
+          78.50390625
+        ],
+        [
+          304.505859375,
+          78.50390625
+        ]
+      ]
+    },
+    {
+      "title": "4.1. Experimental Setup",
+      "heading_level": null,
+      "page_id": 4,
+      "polygon": [
+        [
+          303.908203125,
+          88.5
+        ],
+        [
+          410.25,
+          88.5
+        ],
+        [
+          410.25,
+          99.0
+        ],
+        [
+          303.908203125,
+          99.0
+        ]
+      ]
+    },
+    {
+      "title": "4.2. Main Results",
+      "heading_level": null,
+      "page_id": 5,
+      "polygon": [
+        [
+          54.087890625,
+          639.24609375
+        ],
+        [
+          129.0536651611328,
+          639.24609375
+        ],
+        [
+          129.0536651611328,
+          650.9041137695312
+        ],
+        [
+          54.087890625,
+          650.9041137695312
+        ]
+      ]
+    },
+    {
+      "title": "4.3. Analysis",
+      "heading_level": null,
+      "page_id": 5,
+      "polygon": [
+        [
+          305.701171875,
+          610.2935028076172
+        ],
+        [
+          360.8594665527344,
+          610.2935028076172
+        ],
+        [
+          360.8594665527344,
+          620.256103515625
+        ],
+        [
+          305.701171875,
+          620.256103515625
+        ]
+      ]
+    },
+    {
+      "title": "5. Related Work",
+      "heading_level": null,
+      "page_id": 7,
+      "polygon": [
+        [
+          306.0,
+          120.75
+        ],
+        [
+          391.5,
+          120.75
+        ],
+        [
+          391.5,
+          130.5
+        ],
+        [
+          306.0,
+          130.5
+        ]
+      ]
+    },
+    {
+      "title": "6. Conclusion",
+      "heading_level": null,
+      "page_id": 7,
+      "polygon": [
+        [
+          306.0,
+          578.25
+        ],
+        [
+          377.25,
+          578.25
+        ],
+        [
+          377.25,
+          588.97265625
+        ],
+        [
+          306.0,
+          588.97265625
+        ]
+      ]
+    },
+    {
+      "title": "Impact Statement",
+      "heading_level": null,
+      "page_id": 8,
+      "polygon": [
+        [
+          52.59375,
+          67.84710693359375
+        ],
+        [
+          146.74185180664062,
+          67.84710693359375
+        ],
+        [
+          146.74185180664062,
+          79.80230712890625
+        ],
+        [
+          52.59375,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "References",
+      "heading_level": null,
+      "page_id": 8,
+      "polygon": [
+        [
+          53.490234375,
+          283.15814208984375
+        ],
+        [
+          110.98385620117188,
+          283.15814208984375
+        ],
+        [
+          110.98385620117188,
+          295.11334228515625
+        ],
+        [
+          53.490234375,
+          295.11334228515625
+        ]
+      ]
+    },
+    {
+      "title": "A. Limitation",
+      "heading_level": null,
+      "page_id": 10,
+      "polygon": [
+        [
+          54.087890625,
+          67.67578125
+        ],
+        [
+          124.51715850830078,
+          67.67578125
+        ],
+        [
+          124.51715850830078,
+          79.80230712890625
+        ],
+        [
+          54.087890625,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "B. Additional Implementation Details",
+      "heading_level": null,
+      "page_id": 10,
+      "polygon": [
+        [
+          54.38671875,
+          344.1796875
+        ],
+        [
+          246.73516845703125,
+          344.1796875
+        ],
+        [
+          246.73516845703125,
+          356.96038818359375
+        ],
+        [
+          54.38671875,
+          356.96038818359375
+        ]
+      ]
+    },
+    {
+      "title": "B.1. Hyperparameter Settings",
+      "heading_level": null,
+      "page_id": 10,
+      "polygon": [
+        [
+          54.38671875,
+          365.90753173828125
+        ],
+        [
+          182.62254333496094,
+          365.90753173828125
+        ],
+        [
+          182.62254333496094,
+          376.27734375
+        ],
+        [
+          54.38671875,
+          376.27734375
+        ]
+      ]
+    },
+    {
+      "title": "B.2. Retrieval Setup",
+      "heading_level": null,
+      "page_id": 10,
+      "polygon": [
+        [
+          55.44000244140625,
+          611.0535125732422
+        ],
+        [
+          140.44921875,
+          611.0535125732422
+        ],
+        [
+          140.44921875,
+          621.01611328125
+        ],
+        [
+          55.44000244140625,
+          621.01611328125
+        ]
+      ]
+    },
+    {
+      "title": "User Prompt:",
+      "heading_level": null,
+      "page_id": 11,
+      "polygon": [
+        [
+          64.248046875,
+          97.83984375
+        ],
+        [
+          125.2335433959961,
+          97.83984375
+        ],
+        [
+          125.2335433959961,
+          108.82611083984375
+        ],
+        [
+          64.248046875,
+          108.82611083984375
+        ]
+      ]
+    },
+    {
+      "title": "Output Instruction:",
+      "heading_level": null,
+      "page_id": 11,
+      "polygon": [
+        [
+          65.443359375,
+          182.6414794921875
+        ],
+        [
+          143.1463165283203,
+          182.6414794921875
+        ],
+        [
+          143.1463165283203,
+          192.60406494140625
+        ],
+        [
+          65.443359375,
+          192.60406494140625
+        ]
+      ]
+    },
+    {
+      "title": "B.3. Baseline Setup",
+      "heading_level": null,
+      "page_id": 11,
+      "polygon": [
+        [
+          53.7890625,
+          416.109375
+        ],
+        [
+          136.8144989013672,
+          416.109375
+        ],
+        [
+          136.8144989013672,
+          426.9541015625
+        ],
+        [
+          53.7890625,
+          426.9541015625
+        ]
+      ]
+    },
+    {
+      "title": "B.4. Prompt Template",
+      "heading_level": null,
+      "page_id": 11,
+      "polygon": [
+        [
+          53.19140625,
+          508.5975341796875
+        ],
+        [
+          148.97885131835938,
+          508.5975341796875
+        ],
+        [
+          148.97885131835938,
+          518.5601196289062
+        ],
+        [
+          53.19140625,
+          518.5601196289062
+        ]
+      ]
+    },
+    {
+      "title": "C. Benchmarks and Baselines",
+      "heading_level": null,
+      "page_id": 11,
+      "polygon": [
+        [
+          54.38671875,
+          564.99609375
+        ],
+        [
+          206.8885040283203,
+          564.99609375
+        ],
+        [
+          206.8885040283203,
+          578.1953582763672
+        ],
+        [
+          54.38671875,
+          578.1953582763672
+        ]
+      ]
+    },
+    {
+      "title": "C.1. Benchmarks",
+      "heading_level": null,
+      "page_id": 11,
+      "polygon": [
+        [
+          53.19140625,
+          587.1425170898438
+        ],
+        [
+          128.7747039794922,
+          587.1425170898438
+        ],
+        [
+          128.7747039794922,
+          598.25390625
+        ],
+        [
+          53.19140625,
+          598.25390625
+        ]
+      ]
+    },
+    {
+      "title": "C.2. Baselines",
+      "heading_level": null,
+      "page_id": 12,
+      "polygon": [
+        [
+          52.59375,
+          118.3359375
+        ],
+        [
+          114.38870239257812,
+          118.3359375
+        ],
+        [
+          114.38870239257812,
+          128.38311767578125
+        ],
+        [
+          52.59375,
+          128.38311767578125
+        ]
+      ]
+    },
+    {
+      "title": "Proprietary Models.",
+      "heading_level": null,
+      "page_id": 12,
+      "polygon": [
+        [
+          53.19140625,
+          172.4765625
+        ],
+        [
+          141.5865936279297,
+          172.4765625
+        ],
+        [
+          141.5865936279297,
+          183.64813232421875
+        ],
+        [
+          53.19140625,
+          183.64813232421875
+        ]
+      ]
+    },
+    {
+      "title": "General Reasoning Models.",
+      "heading_level": null,
+      "page_id": 12,
+      "polygon": [
+        [
+          54.685546875,
+          252.140625
+        ],
+        [
+          172.5005340576172,
+          252.140625
+        ],
+        [
+          172.5005340576172,
+          262.1041259765625
+        ],
+        [
+          54.685546875,
+          262.1041259765625
+        ]
+      ]
+    },
+    {
+      "title": "General Foundation Models.",
+      "heading_level": null,
+      "page_id": 12,
+      "polygon": [
+        [
+          54.087890625,
+          330.2578125
+        ],
+        [
+          177.24273681640625,
+          330.2578125
+        ],
+        [
+          177.24273681640625,
+          340.5591125488281
+        ],
+        [
+          54.087890625,
+          340.5591125488281
+        ]
+      ]
+    },
+    {
+      "title": "Medical Domain Models.",
+      "heading_level": null,
+      "page_id": 12,
+      "polygon": [
+        [
+          53.7890625,
+          407.21484375
+        ],
+        [
+          161.9800262451172,
+          407.21484375
+        ],
+        [
+          161.9800262451172,
+          419.01513671875
+        ],
+        [
+          53.7890625,
+          419.01513671875
+        ]
+      ]
+    },
+    {
+      "title": "Medical Reward Models.",
+      "heading_level": null,
+      "page_id": 12,
+      "polygon": [
+        [
+          53.7890625,
+          511.41851806640625
+        ],
+        [
+          161.97007751464844,
+          511.41851806640625
+        ],
+        [
+          161.97007751464844,
+          521.381103515625
+        ],
+        [
+          53.7890625,
+          521.381103515625
+        ]
+      ]
+    },
+    {
+      "title": "D. Case Analysis",
+      "heading_level": null,
+      "page_id": 12,
+      "polygon": [
+        [
+          53.19140625,
+          592.2521514892578
+        ],
+        [
+          140.549072265625,
+          592.2521514892578
+        ],
+        [
+          140.549072265625,
+          604.2073516845703
+        ],
+        [
+          53.19140625,
+          604.2073516845703
+        ]
+      ]
+    },
+    {
+      "title": "Medical Problem:",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          65.443359375,
+          170.15625
+        ],
+        [
+          143.5049591064453,
+          170.15625
+        ],
+        [
+          143.5049591064453,
+          180.83203125
+        ],
+        [
+          65.443359375,
+          180.83203125
+        ]
+      ]
+    },
+    {
+      "title": "Generator Reasoning Trace:",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          65.14453125,
+          261.421875
+        ],
+        [
+          188.04774475097656,
+          261.421875
+        ],
+        [
+          188.04774475097656,
+          272.4891357421875
+        ],
+        [
+          65.14453125,
+          272.4891357421875
+        ]
+      ]
+    },
+    {
+      "title": "Verifier Judgment:",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          65.7421875,
+          447.43359375
+        ],
+        [
+          147.16087341308594,
+          447.43359375
+        ],
+        [
+          147.16087341308594,
+          459.7861022949219
+        ],
+        [
+          65.7421875,
+          459.7861022949219
+        ]
+      ]
+    }
+  ],
+  "page_stats": [
+    {
+      "page_id": 0,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          326
+        ],
+        [
+          "Line",
+          157
+        ],
+        [
+          "Text",
+          6
+        ],
+        [
+          "SectionHeader",
+          4
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Footnote",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 1,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          481
+        ],
+        [
+          "Line",
+          153
+        ],
+        [
+          "Text",
+          13
+        ],
+        [
+          "ListItem",
+          3
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Footnote",
+          1
+        ],
+        [
+          "Equation",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 2,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          199
+        ],
+        [
+          "Span",
+          132
+        ],
+        [
+          "Text",
+          51
+        ],
+        [
+          "Equation",
+          3
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "PageHeader",
+          2
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 3,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          214
+        ],
+        [
+          "Span",
+          152
+        ],
+        [
+          "Text",
+          66
+        ],
+        [
+          "Equation",
+          5
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Footnote",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 4,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          237
+        ],
+        [
+          "Span",
+          173
+        ],
+        [
+          "Text",
+          62
+        ],
+        [
+          "Code",
+          2
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Equation",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 5,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          589
+        ],
+        [
+          "TableCell",
+          180
+        ],
+        [
+          "Line",
+          128
+        ],
+        [
+          "Text",
+          13
+        ],
+        [
+          "PageHeader",
+          2
+        ],
+        [
+          "Equation",
+          2
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 6,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          129
+        ],
+        [
+          "Span",
+          82
+        ],
+        [
+          "TableCell",
+          32
+        ],
+        [
+          "Text",
+          8
+        ],
+        [
+          "PageHeader",
+          2
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 7,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          218
+        ],
+        [
+          "Span",
+          116
+        ],
+        [
+          "Text",
+          59
+        ],
+        [
+          "TableCell",
+          14
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 8,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          354
+        ],
+        [
+          "Line",
+          152
+        ],
+        [
+          "ListItem",
+          17
+        ],
+        [
+          "Reference",
+          17
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "Text",
+          2
+        ],
+        [
+          "ListGroup",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 9,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          287
+        ],
+        [
+          "Line",
+          120
+        ],
+        [
+          "ListItem",
+          19
+        ],
+        [
+          "Reference",
+          19
+        ],
+        [
+          "Text",
+          12
+        ],
+        [
+          "ListGroup",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 10,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          198
+        ],
+        [
+          "Line",
+          86
+        ],
+        [
+          "TableCell",
+          36
+        ],
+        [
+          "Text",
+          17
+        ],
+        [
+          "SectionHeader",
+          4
+        ],
+        [
+          "Reference",
+          3
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 11,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          251
+        ],
+        [
+          "Line",
+          99
+        ],
+        [
+          "Text",
+          26
+        ],
+        [
+          "SectionHeader",
+          6
+        ],
+        [
+          "Reference",
+          3
+        ],
+        [
+          "ListItem",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Code",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 12,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          266
+        ],
+        [
+          "Line",
+          93
+        ],
+        [
+          "ListItem",
+          11
+        ],
+        [
+          "Text",
+          9
+        ],
+        [
+          "SectionHeader",
+          7
+        ],
+        [
+          "ListGroup",
+          5
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 13,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          181
+        ],
+        [
+          "Line",
+          82
+        ],
+        [
+          "Text",
+          8
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "ListItem",
+          3
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    }
+  ],
+  "debug_data_path": "debug_data/19e76363-53a6-4f3c-8b50-844e1aea4e26"
+}

icml26/19e76363-53a6-4f3c-8b50-844e1aea4e26/model_text_v3.txt

@@ -0,0 +1,284 @@
+[p. 1 | section: Abstract | type: Text]
+Large language models have achieved strong performance on medical reasoning benchmarks, yet their deployment in clinical settings demands rigorous verification to ensure factual accuracy. While reward models offer a scalable approach for reasoning trace verification, existing methods face two limitations: they produce only scalar reward values without explicit justification, and they rely on single-pass retrieval that precludes adaptive knowledge access as verification unfolds. We introduce Med-TIV, an agentic framework that addresses these limitations by training medical reasoning verifiers to iteratively query external medical corpora during evaluation. Our approach combines tool-augmented verification with an iterative reinforcement learning paradigm that requires only trace-level supervision, alongside an adaptive curriculum mechanism that dynamically adjusts training data distribution. Across four medical reasoning benchmarks, Med-TIV achieves substantial gains over existing methods, improving MedQA accuracy by 23.5% and MedXpertQA by 32.0% relative to the base generator in particular. Crucially, Med-TIV demonstrates an 8× reduction in sampling budget requirement compared to prior reward model baselines. These findings establish that grounding verification in dynamically retrieved evidence offers a principled path toward more reliable medical reasoning systems.
+
+[p. 1 | section: 1. Introduction | type: Text]
+Large Language Models (LLMs) have demonstrated remarkable capabilities in medical reasoning, achieving competitive performance on clinical question answering, diagnostic
+
+[p. 1 | section: 1. Introduction | type: FigureGroup]
+Figure 1. Comparison of medical reasoning verification paradigms. Text-based judges rely on parametric knowledge and may validate erroneous reasoning, while tool-integrated judges dynamically retrieve evidence to ground their judgments.
+
+[p. 1 | section: 1. Introduction | type: Text]
+inference, and medical knowledge benchmarks (Ji et al., 2025a; Xiao et al., 2026) . While these advances hold significant promise for augmenting clinical decision making and democratizing access to medical expertise, the deployment of LLMs in high-stakes clinical settings demands rigorous verification mechanisms to ensure that generated reasoning is both factually accurate and logically sound (Zhang et al., 2025; Wang et al., 2025) .
+
+[p. 1 | section: 1. Introduction | type: Text]
+Reward-based judges have therefore emerged as a scalable solution for evaluating model outputs, supporting both posttraining refinement via reinforcement learning from human feedback (RLHF) and inference-time scaling through tree search (Snell et al., 2024) . These judges can be broadly categorized by the granularity of their supervision. Outcome Reward Models (ORMs) provide sparse trace-level supervision that quantifies the quality of the entire output, while Process Reward Models (PRMs) offer dense step-level feedback that scores each intermediate reasoning step, enabling fine-grained credit assignment and precise error localization within multi-step reasoning. Recent work has adapted both paradigms to the medical domain to assess complex clinical reasoning traces. In parallel, advances in generative reward modeling have extended judge models beyond
+
+[p. 2 | section: 1. Introduction | type: Text]
+scalar scoring, enabling them to produce natural-language critiques that explicitly justify their decisions (Liu et al., 2025c; Xiong et al., 2025) .
+
+[p. 2 | section: 1. Introduction | type: Text]
+Despite their effectiveness, reward-based judges exhibit fundamental limitations when applied to clinical reasoning tasks (Yun et al., 2025) . A primary concern is the prevalence of hallucinations in critique traces, where judge models generate plausible yet factually incorrect assessments (Figure 1) . This issue is particularly noticeable in the medical domain, where reliable verification demands grounding in authoritative clinical evidence and established medical knowledge. Unverified judgments could lead to the propagation of incorrect diagnostic or treatment recommendations. Existing medical reasoning verifiers typically provide only scalar reward signals, offering little or no justification for their judgments and thus limiting interpretability (Jiang et al., 2025b) . Furthermore, these methods often rely on a static Retrieval-Augmented Generation (RAG) pipeline, in which a fixed set of retrieved documents is prefixed to the context and remains unchanged throughout evaluation (Yun et al., 2025) . Such static design precludes adaptive, multi-turn evidence gathering and forces the verifier to a fixed retrieval budget, thus limiting scalability.
+
+[p. 2 | section: 1. Introduction | type: Text]
+To address these issues, we propose Med-TIV (Medical Tool-Integrated reasoning Verifier), an agentic reinforcement learning (RL) framework that trains LLMs to leverage external knowledge bases for judging medical reasoning traces 1 . Med-TIV features three key design principles: (1) a tool-augmented verification paradigm that enables dynamic, iterative knowledge retrieval during the evaluation process; (2) an iterative RL approach that progressively improves verification capabilities without requiring step-level expert annotations; and (3) an adaptive curriculum formulation strategy that adjusts the data distribution in response to the evolving capability of the model. By equipping judge models with tool-use capabilities, Med-TIV grounds evaluation decisions in external evidence rather than relying solely on parametric knowledge, thereby mitigating hallucination, improving interpretability, and overcoming the limitations of static RAG (Ji et al., 2025b; Xia et al., 2025) .
+
+[p. 2 | section: 1. Introduction | type: Text]
+To verify the effectiveness of Med-TIV, we conduct extensive experiments on common medical reasoning benchmarks. Our results demonstrate that Med-TIV trains strong medical verifiers: when guiding inference-time search for a 7B generator model, our trained verifier achieves relative improvements of 23.5% on MedQA and 32.0% on MedXpertQA compared to the generator model alone. Moreover, Med-TIV consistently outperforms existing medical reward model baselines and surpasses the performance of models that are up to 4× larger in scale. Notably, Med-TIV
+
+[p. 2 | section: 1. Introduction | type: Text]
+also demonstrates an 8× gain in sampling efficiency compared to prior reward-based approaches, achieving equivalent accuracy with substantially fewer sampled reasoning traces during test-time search.
+
+[p. 2 | section: 1. Introduction | type: Text]
+Our main contributions are summarized as follows:
+
+[p. 2 | section: 1. Introduction | type: ListGroup]
+We propose Med-TIV, a novel tool-integrated verification framework that enables dynamic, iterative knowledge retrieval during medical reasoning evaluation, providing both interpretable, fine-grained justification and improved factual grounding. We introduce an iterative RL paradigm with curriculumbased difficulty adaptation that progressively improves verification capabilities through self-bootstrapping, requiring only trace-level supervision rather than dense steplevel expert annotations. Med-TIV achieves state-of-the-art performance on four medical reasoning benchmarks, with comprehensive ablation studies that validate each component's contribution.
+
+[p. 2 | section: 2.1. Problem Setup | type: Text]
+We define medical reasoning verification as the task of assessing the correctness of a multi-step reasoning trace generated in response to a medical question. Formally, given a medical question q ∈ Q and a multi-step reasoning trace τ = (s1, s2, . . . , sm) from a generator model, a verifier model determines whether τ contains any errors. We formulate this problem as binary classification, where the verifier Vθ(q, τ ) produces a judgment ℓ ∈ {0, 1}, where ℓ = 1 indicates a error-free reasoning trace, and ℓ = 0 indicates the presence of one or more errors. Unlike scalar reward models that output continuous scores, we adopt a generative judge paradigm in which the verifier produces a discrete judgment accompanied by a detailed critique trace that provides a structured justification for the decision.
+
+[p. 2 | section: 2.2. Tool-Augmented Reasoning Verifier | type: Text]
+Following prior works (Jin et al., 2025) , we extend the verifier with access to an external search engine E that retrieves top-k documents from a curated medical corpus (See Appendix B.2 for details). Retrieved documents are appended verbatim to the verifier context. Given a verification instance (q, τ ), the verifier constructs an iterative verfication trajectory. At step k, the trajectory is represented as:
+
+[p. 2 | section: 2.2. Tool-Augmented Reasoning Verifier | type: Equation]
+\mathbf{t}_k = \{r_1, a_1, o_1, \dots, r_k, a_k, o_k\},\
+
+[p. 2 | section: 2.2. Tool-Augmented Reasoning Verifier | type: Text]
+where r i denotes a natural language reasoning step analyzing the medical content, a i is a search query formulated to retrieve relevant medical knowledge, and o i = E(ai) represents the retrieved documents. The iterative verification
+
+[p. 2 | section: 2.2. Tool-Augmented Reasoning Verifier | type: Footnote]
+1 Code is available at PittNAIL/med-tiv
+
+[p. 3 | section: 2.2. Tool-Augmented Reasoning Verifier | type: Caption]
+Figure 2. Overview of Med-TIV. Left: Tool-integrated verification iteratively analyzes reasoning traces, formulates search queries, and retrieves medical evidence before producing correctness judgments. Middle: Curriculum formulation filters trivial and impossible instances, retaining boundary cases for RL training. Right: At inference time, the verifier evaluates candidate medical reasoning traces generated by a frozen model and final answers are selected via weighted self-consistency.
+
+[p. 3 | section: 2.2. Tool-Augmented Reasoning Verifier | type: Text]
+process is defined as:
+
+[p. 3 | section: 2.2. Tool-Augmented Reasoning Verifier | type: Text]
+123124125
+
+[p. 3 | section: 2.2. Tool-Augmented Reasoning Verifier | type: Text]
+135136
+
+[p. 3 | section: 2.2. Tool-Augmented Reasoning Verifier | type: Text]
+137138
+
+[p. 3 | section: 2.2. Tool-Augmented Reasoning Verifier | type: Equation]
+(r_k, a_k) \sim V_{\theta}(q, \tau, \mathbf{t}_{k-1}), o_k = \mathcal{E}(a_k), \mathbf{t}_k = \mathbf{t}_{k-1} \oplus r_k \oplus a_k \oplus o_k,
+
+[p. 3 | section: 2.2. Tool-Augmented Reasoning Verifier | type: Text]
+where \oplus denotes sequence concatenation. This process continues until the verifier produces a final judgment \ell \sim V_{\theta}(q,\tau,\mathbf{t}_T) at the terminal step T. By allowing multiple tool executions, the verifier dynamically retrievs medical knowledge as need to verify specific claims in the reasoning trace. Table 5 in the Appendix shows the explicit instruction used in our experiments.
+
+[p. 3 | section: 2.3. Test-Time Search | type: Text]
+Test-time search strategies improve reasoning performance by leveraging reward models to evaluate and select among multiple candidate solutions (Shi et al., 2024). Given a frozen generator model \pi_{\rm gen} and a question q, we first sample N independent reasoning traces:
+
+[p. 3 | section: 2.3. Test-Time Search | type: Equation]
+\{\tau^{(j)}\}_{j=1}^N \sim \pi_{\mathrm{gen}}(\cdot \mid q).
+
+[p. 3 | section: 2.3. Test-Time Search | type: Text]
+A trained verifier V_{\theta} then scores each candidate trace, and the final output is selected based on these scores. Common selection strategies include Best-of-N sampling, which
+
+[p. 3 | section: 2.3. Test-Time Search | type: Text]
+selects the trace with the highest score:
+
+[p. 3 | section: 2.3. Test-Time Search | type: Equation]
+\hat{\tau} = \arg\max_{\tau^{(j)}} V_{\theta}(q, \tau^{(j)}),
+
+[p. 3 | section: 2.3. Test-Time Search | type: Text]
+and verification-based majority voting, where candidate traces are first filtered by the verifier and the final answer is determined by consensus among verified traces. Med-TIV trains such a plug-in verifier that provides tool-grounded assessments that can be used to augment decision-making for any frozen generator model at inference time.
+
+[p. 3 | section: 3. Tool-Integrated Medical Reasoning Verifier | type: Text]
+Med-TIV is an agentic verification framework that trains models to leverage external knowledge bases for verifying whether a given medical reasoning trace contains errors. We adopt an iterative training approach based on dynamic curriculum learning, which requires no fine-grained step-level expert supervision and trains solely through multiple rounds of reinforcement learning (Figure 2). We next describe the training procedure of Med-TIV in details.
+
+[p. 3 | section: 3.1. Tool-Integrated RL with Verifiable Rewards | type: Text]
+Data Construction. All training data across iterations is derived from the open-source Med-PRM dataset. Each original instance consists of a tuple (q, \tau, \ell_{\text{step}}, \ell_{\text{trace}}) , where q is a medical question, \tau is a multi-step reasoning trace, \ell_{\text{step}} de-
+
+[p. 4 | section: 3.1. Tool-Integrated RL with Verifiable Rewards | type: Text]
+notes step-level labels, and \ell_{trace} is a trace-level correctness label<sup>2</sup>.
+
+[p. 4 | section: 3.1. Tool-Integrated RL with Verifiable Rewards | type: Text]
+At each training iteration, we only utilize the triplet (q,\tau,\ell_{\text{trace}}) with human-annotated trace-level labels. Step-level labels \ell_{\text{step}} is intentionally excluded, as Med-TIV is designed to improve verification performance without replying on fine-grained supervision. For each training iteration, we fix the training data budget to 20K instances and enforce a balanced label distribution between correct ( \ell_{\text{trace}}=1 ) and incorrect ( \ell_{\text{trace}}=0 ) reasoning traces.
+
+[p. 4 | section: 3.1. Tool-Integrated RL with Verifiable Rewards | type: Text]
+Algorithm. We employ Dr. GRPO (Liu et al., 2025b) as the RL algorithm for training the verifier. Given a verification instance (q_i, \tau_i) , we sample a group of G verification trajectories \{\mathbf{o}_i\}_{i=1}^G from the current policy \pi_\theta . Each trajectory \mathbf{o}_i = (o_i^1, \dots, o_i^{|\mathbf{o}_i|}) consists of reasoning tokens, search queries, retrieved documents, and a final judgment. The objective is:
+
+[p. 4 | section: 3.1. Tool-Integrated RL with Verifiable Rewards | type: Equation]
+\frac{1}{G} \sum_{i=1}^{G} \sum_{t=1}^{|\mathbf{o}_i|} \left\{ \min \left[ r_i^t \hat{A}_i^t, \operatorname{clip} \left( r_i^t, 1 - \epsilon_l, 1 + \epsilon_h \right) \hat{A}_i^t \right] \right\}, \quad (1)
+
+[p. 4 | section: 3.1. Tool-Integrated RL with Verifiable Rewards | type: Text]
+where r_i^t = \frac{\pi_\theta(o_i^t|\mathbf{q},\mathbf{o}_i^{< t})}{\pi_{\theta_{\mathrm{old}}}(o_i^t|\mathbf{q},\mathbf{o}_i^{< t})} , \mathbf{q} = (q,\tau) denotes the input prompt containing the question and reasoning trace, \mathbf{o}_i^{< t} represents previously generated tokens, and \epsilon_l and \epsilon_h are the clipping parameters. The advantage term \hat{A}_i^t is defined as:
+
+[p. 4 | section: 3.1. Tool-Integrated RL with Verifiable Rewards | type: Equation]
+\hat{A}_i^t = R(\mathbf{q}, \mathbf{o}_i) - \text{mean}\left(\left\{R(\mathbf{q}, \mathbf{o}_1), \dots, R(\mathbf{q}, \mathbf{o}_G)\right\}\right).
+
+[p. 4 | section: 3.1. Tool-Integrated RL with Verifiable Rewards | type: Text]
+Reward Designs. To facilitate multi-turn RL with tool execution, we design a structured reward covering two complementary objectives, following prior practices (Jin et al., 2025):
+
+[p. 4 | section: 3.1. Tool-Integrated RL with Verifiable Rewards | type: Text]
+(i) Correctness Reward (R_c) : This component measures whether the verifier's judgment aligns with the ground-truth label. Let \mathbf{q} = (q, \tau) denote the verification prompt and \ell \in \{0, 1\} the ground-truth label. We define:
+
+[p. 4 | section: 3.1. Tool-Integrated RL with Verifiable Rewards | type: Equation]
+R_c = \mathbb{1}(\text{extract}(\mathbf{o}) = \ell),
+
+[p. 4 | section: 3.1. Tool-Integrated RL with Verifiable Rewards | type: Text]
+where \mathbb{1}(\cdot) is the indicator function and extract(o) parses the final judgment from the <answer> tags in the generated trajectory o. Intuitively, R_c=1 if the verifier's decision is correct, and R_c=0 otherwise.
+
+[p. 4 | section: 3.1. Tool-Integrated RL with Verifiable Rewards | type: Text]
+(ii) Format Reward (R_f) : To ensure reliable tool use and structured outputs, the verifier is required to adhere to a predefined format. Specifically, reasoning steps must be enclosed within <think> tags, search queries within <search> tags, and the final judgment within <answer>
+
+[p. 4 | section: 3.1. Tool-Integrated RL with Verifiable Rewards | type: Text]
+tags. To discourage degenerate outputs, we further penalize excessive tag usage. Specifically, R_f=1 if the output satisfies all formatting constraints and contains no more than 10 < answer> tag pairs; <math>R_f=0.25 if the output is correct but exhibits tag overflow; and R_f=0 otherwise.
+
+[p. 4 | section: 3.1. Tool-Integrated RL with Verifiable Rewards | type: Text]
+The final reward {\cal R} is defined as the product of the two components:
+
+[p. 4 | section: 3.1. Tool-Integrated RL with Verifiable Rewards | type: Equation]
+R = R_c \times R_f .
+
+[p. 4 | section: 3.2. Training Strategies | type: Text]
+Adaptive Curriculum Formulation. A central challenge in RL for verification is ensuring that training data remains appropriately calibrated to the evolving capability of the model. Instances that are either trivially easy or impossibly difficult yields minimal learning signal, as the resulting policy gradients approach zero. To address this issue, we adopt a model-aware curriculum formulation mechanism that dynamically adapts the task distribution at each training iteration.
+
+[p. 4 | section: 3.2. Training Strategies | type: Text]
+Concretely, before each iteration t, we perform online filtering on the sampled batch \mathcal{B}_t to construct an effective training set \mathcal{D}_t :
+
+[p. 4 | section: 3.2. Training Strategies | type: Equation]
+\mathcal{D}_t = \{(q, \tau, \ell) \in \mathcal{B}_t : \exists g, g' \in \{1, \dots, G\} \text{ s.t. } r^{(g)} \neq r^{(g')}\}.
+
+[p. 4 | section: 3.2. Training Strategies | type: Text]
+Here, for each candidate instance (q, \tau, \ell) \in \mathcal{B}_t , we sample G verification trajectories \{o^{(g)}\}_{g=1}^G from the current policy \pi_{\theta_t} . We then compute the corresponding rewards \{r^{(g)}\}_{g=1}^G . Finally, we retain only instances if any two rewards are different, i.e., reward variance is non-zero (Khatri et al., 2025).
+
+[p. 4 | section: 3.2. Training Strategies | type: Text]
+This criterion eliminates prompts where the model either consistently succeeds or consistently fails across all sampled trajectories. By filtering these zero-gradient instances, optimization is focused on decision-boundary cases where the verifier exhibits uncertainty.
+
+[p. 4 | section: 3.2. Training Strategies | type: Text]
+To maintain a fixed training budget per iteration, we iteratively resample additional instances from the labeled pool \mathcal{B} and apply the same filtering criterion until |\mathcal{D}_t|=20K . This dynamic curriculum evolves naturally across iterations as the verifier improves, eliminating the need for manually designed difficulty schedules.
+
+[p. 4 | section: 3.2. Training Strategies | type: Text]
+Iterative Training via Self-Bootstrapping. We adopt an iterative training approach that progressively improves verification capabilities through multiple rounds of RL. Unlike prior work that alternates between rejection sampling, supervised fine-tuning (SFT), and RL (Xu et al., 2025), our approach operates entirely through iterative RL, following the RL-Zero paradigm where the model reinforces its verification capabilities without requiring dense turn-level expert demonstrations for cold start.
+
+[p. 4 | section: 3.2. Training Strategies | type: Footnote]
+& lt;sup>2</sup>Dataset is available at
+
+[p. 5 | section: 3.2. Training Strategies | type: Text]
+Algorithm 1 Iterative Training of Tool-Integrated Medical Reasoning Verifier
+
+[p. 5 | section: 3.2. Training Strategies | type: Text]
+249250
+
+[p. 5 | section: 3.2. Training Strategies | type: Code]
+Require: Base verifier \pi_{\theta_0}, labeled dataset pool \mathcal{D} = \{(q_i, \tau_i, \ell_i)\}_{i=1}^N, maximum iterations T_{\text{max}}, batch size B, group size G, search engine \mathcal{E} Ensure: Trained verifier \pi_{\theta^*} 1: for t = 1 to T_{\text{max}} do Sample labeled batch \mathcal{B}_t \leftarrow \text{SAMPLEBATCH}(\mathcal{D}, B) 3: 4: \mathcal{D}_t \leftarrow \emptyset ▷ Curriculum formulation 5: for each (q, \tau, \ell) \in \mathcal{B}_t do 6: Sample verification trajectories: 7: \{\hat{\ell}^{(g)}\}_{q=1}^G \sim \pi_{\theta_t}(\cdot \mid q, \tau, \mathcal{E}) Compute rewards within group: 8: r^{(\hat{g})} \leftarrow \mathbb{1}[\hat{\ell}^{(g)} = \ell], \text{ for } g \in 1, \dots, G if \exists g \neq g' such that r^{(g)} \neq r^{(g')} then 9: Add (q, \tau, \ell) to curriculum set \mathcal{D}_t 10: 11: end if 12: end for 13: ▶ RL optimization on curriculum data \pi_{\theta_{t+1}} \leftarrow \text{DR.GRPO}(\pi_{\theta_t}, \mathcal{D}_t, \mathcal{E}) 15: end for 16: Return \pi_{\theta_{T_{\max}}}
+
+[p. 5 | section: 3.2. Training Strategies | type: Text]
+Starting from the base model \pi_{\theta_0} , we perform T_{\text{max}} iterations. Each iteration consists of three stages:
+
+[p. 5 | section: 3.2. Training Strategies | type: Code]
+\begin{split} \mathcal{B}_t \leftarrow \text{SampleBatch}(\mathcal{D}, B), \\ \mathcal{D}_t \leftarrow \text{Filter}(\mathcal{B}_t, \pi_{\theta_t}), \\ \pi_{\theta_{t+1}} \leftarrow \text{RL}(\pi_{\theta_t}, \mathcal{D}_t). \end{split}
+
+[p. 5 | section: 3.2. Training Strategies | type: Text]
+Each iteration draws a fresh batch \mathcal{B}_t from the annotated pool \mathcal{D} with trace-level labels, ensuring a balanced distribution of correct and incorrect reasoning traces. The curriculum filtering then constructs the training set \mathcal{D}_t as described above, and RL optimization updates the policy based on the structured reward.
+
+[p. 5 | section: 3.2. Training Strategies | type: Text]
+The key insight underlying this iterative approach is the coevolution of model capability and training distribution. As the verifier improves, the filtering mechanism automatically removes instances that have become too easy, while the fresh sampling introduces new challenging cases. This creates a self-bootstrapping cycle: stronger models encounter harder verification tasks, which in turn drive further improvements. Since the trace-level correctness reward is deterministic and unambiguous, this self-bootstrapping process converges reliably without the instabilities that can arise from noisy synthetic step-level labels. We summarize the overall training procedure in Algorithm 1.
+
+[p. 5 | section: 4.1. Experimental Setup | type: Text]
+Evaluation benchmarks. We evaluated Med-TIV on four open-source medical question-answering benchmarks: MedQA (Jin et al., 2020), MedMCQA (Pal et al., 2022), MMLU-Med (Hendrycks et al., 2021), and MedX-pertQA (Zuo et al., 2025), using accuracy as the evaluation metric. These benchmarks collectively assess the verifier's ability to distinguish correct from erroneous reasoning across varying difficulty levels and medical subdomains. Detailed descriptions of benchmarks are in Appendix C.1.
+
+[p. 5 | section: 4.1. Experimental Setup | type: Text]
+Implementation details. We trained verifiers using two light-weight backbone models: Llama3.1-8B and Qwen2.5-7B, with Llama3.1-8B as the default for results reporting. All training was conducted using the VeRL-Tool framework (Jiang et al., 2025a). Detailed hyperparameters are shown in Appendix B.1. All experiments were conducted on 4 NVIDIA H100 GPUs with 80GB of memory. Due to computational constraints, we limit the maximum number of RL iterations to T_{\rm max}=2 and we set the group size for curriculum formulation (Section 3.2) to G=8.
+
+[p. 5 | section: 4.1. Experimental Setup | type: Text]
+For inference, we used the default sampling hyperparameters for all models. In reward-guided search experiments, unless otherwise specified, we used Qwen2.5-7B as the frozen generator and sampled up to 32 candidate reasoning traces per question. We applied Hard-Weighted Self-Consistency as the default test-time search strategy.
+
+[p. 5 | section: 4.1. Experimental Setup | type: Text]
+Baselines. We compared Med-TIV against two groups of baselines. 1): Off-the-shelf LLMs : GPT-40-mini (OpenAI et al., 2024), Gemini-2.0-Flash, DeepSeek-R1 series (Guo et al., 2025), Qwen2.5 series (Yang et al., 2025), Llama3.1 (Grattafiori et al., 2024), AlphaMed (Liu et al., 2025a), UltraMedical (Zhang et al., 2024), and HuatuoGPT-01 (Chen et al., 2024). 2): Medical domain-specialized Reward Models : MedS<sup>3</sup> (Jiang et al., 2025b) and Med-PRM (Yun et al., 2025). Detailed descriptions of each reward model baseline are shown in Appendix B.3.
+
+[p. 5 | section: 4.1. Experimental Setup | type: Text]
+Test-Time Search Strategies. We evaluated three test-time search strategies that leverage Med-TIV to improve the reasoning performance of frozen generators. Given a reasoning trace \tau=(s_1,s_2,\ldots,s_K) with K steps, our verifier assigns a confidence score r_{\tau}\in[0,1] for the entire trace, defined as the softmax probability of the 1 token over the logits of both 1 and 0 tokens.
+
+[p. 5 | section: 4.1. Experimental Setup | type: Text]
+• Best-of-N. Given a question q, we sampled N candidate traces \{\tau^{(j)}\}_{j=1}^N from the generator and selected the trace with the highest verifier confidence score:
+
+[p. 5 | section: 4.1. Experimental Setup | type: Equation]
+\hat{\tau} = \arg\max_{\tau^{(j)}} r_{\tau^{(j)}}.
+
+[p. 6 | section: 4.1. Experimental Setup | type: TableGroup]
+Table 1. Main evaluation results on medical reasoning benchmarks. We report accuracy (%) on MedQA, MedMCQA, MMLU-Med, and MedXpertQA. Bold numbers indicate the best results among the reward model group. ✓: Verifier supports external tools for judging; ✗: Verifier does not support external tools for judging. Baselines å |Train| Size MedQA MedMCQA MMLU-Med MedXpertQA Avg. Proprietary Models GPT-4o-mini - - - 79.03 68.20 87.79 17.84 63.22 Gemini-2.0-Flash - - - 87.51 72.60 92.01 20.57 68.17 General Reasoning Models DeepSeek-R1 - - 671B 90.34 78.80 94.40 37.76 75.33 R1-Distill-Qwen - - 7B 24.82 36.40 47.47 7.43 29.03 R1-Distill-Llama - - 8B 34.96 43.60 64.19 5.35 37.03 General Non-reasoning Models Qwen2.5 - - 32B 73.21 64.83 84.94 13.87 59.21 Qwen2.5 - - 7B 60.96 56.56 76.96 12.15 51.66 Llama3.1 - - 8B 70.93 61.60 78.97 13.02 56.13 Medical Reasoning Models AlphaMed - - 7B 71.01 61.46 81.16 19.16 58.20 UltraMedical - - 8B 72.66 62.60 79.61 15.25 57.53 HuatuoGPT-o1 - - 8B 72.19 63.60 75.30 16.84 56.98 Medical Reward Models MedS3 ✗ 225k 7B 64.89 58.91 80.53 12.90 54.31 Med-PRM ✓ 111k 7B 69.99 62.36 80.99 13.51 56.71 Med-TIV (Ours) ✓ 20k 7B 75.26 64.70 85.58 16.04 60.40
+
+[p. 6 | section: 4.1. Experimental Setup | type: Text]
+• Hard-Weighted Self-Consistency. We first filtered traces by the verifier's binary judgment, keeping only those labeled correct (Vθ(q, τ ) = 1). Among the filtered traces, we applied majority voting to determine the final answer:
+
+[p. 6 | section: 4.1. Experimental Setup | type: Equation]
+\hat{a} = \arg\max_{a} \sum_{j=1}^{N} \mathbb{1} \left[ V_{\theta}(q, \tau^{(j)}) = 1 \right] \cdot \mathbb{1} \left[ \operatorname{ans}(\tau^{(j)}) = a \right].
+
+[p. 6 | section: 4.1. Experimental Setup | type: Text]
+• Soft-Weighted Self-Consistency. Instead of binary filtering, we weighted each trace's vote by the verifier's confidence score:
+
+[p. 6 | section: 4.1. Experimental Setup | type: Equation]
+\hat{a} = \arg\max_{a} \sum_{j=1}^{N} r_{\tau^{(j)}} \cdot \mathbb{1} \big[ \mathrm{ans}(\tau^{(j)}) = a \big].
+
+[p. 6 | section: 4.2. Main Results | type: Text]
+Table 1 presents the main results on four medical reasoning benchmarks. Models trained with Med-TIV consistently outperform existing baselines across all benchmarks. Specifically, under guided-search using a Med-TIVtrained verifier, Qwen2.5-7B attains accuracies of 75.26%
+
+[p. 6 | section: 4.2. Main Results | type: Text]
+on MedQA, 64.70% on MedMCQA, 85.58% on MMLU-Med, and 16.04% on MedXpertQA, yielding an average accuracy of 60.40%. Notably, Med-TIV enables this 7B generator to rival substantially larger models, even surpassing the base performance of Qwen2.5-32B despite using a generator that is approximately 4× smaller. Compared to domain-specialized medical reasoning models of similar scale, Med-TIV outperforms HuatuoGPT-o1-8B and UltraMedical-8B by 3.07% and 2.60% on MedQA, respectively, demonstrating the effectiveness of our tool-integrated verification. Case analysis in Appendix D further illustrates how Med-TIV identifies subtle reasoning errors.
+
+[p. 6 | section: 4.3. Analysis | type: Text]
+We conducted a series of ablation analyses to investigate six key research questions regarding our proposed framework.
+
+[p. 6 | section: 4.3. Analysis | type: Text]
+Q1: Does Med-TIV generalize across different generator models? To evaluate the generalizability of the trained verifier, we applied Med-TIV to guide test-time search across generator models of varying sizes and capa-
+
+[p. 7 | section: 4.3. Analysis | type: FigureGroup]
+Figure 3. Test-time scaling analysis across three medical reasoning benchmarks. Each plot shows accuracy versus sampling budget N \in \{1, 2, 4, 8, 16, 32\} for four baselines. Med-TIV consistently outperforms baselines across all sampling budgets and benchmarks.
+
+[p. 7 | section: 4.3. Analysis | type: TableGroup]
+Table 2. Performance improvements from using Med-TIV as a verifier on MedQA. For each generator model, the first row indicates the accuracy over single sampled trace per question. Models MedQA Qwen2.5-7B 60.96 + Self-Consistency 66.38 (+5.42) + Best-of-N (Med-TIV) 72.35 (+11.39) + Soft Weighted SC (Med-TIV) 75.02 (+14.06) + Hard Weighted SC (Med-TIV) 75.26 (+14.30) AlphaMed-7B 71.01 + Self-Consistency 74.23 (+3.22) + Best-of-N (Med-TIV) 75.02 (+4.01) + Soft Weighted SC (Med-TIV) 75.33 (+4.32) + Hard Weighted SC (Med-TIV) 75.65 (+4.64) Qwen2.5-32B 73.21 + Self-Consistency 75.26 (+2.05) + Best-of-N (Med-TIV) 75.57 (+2.36) + Soft Weighted SC (Med-TIV) 75.96 (+2.75) + Hard Weighted SC (Med-TIV) 75.96 (+2.75)
+
+[p. 7 | section: 4.3. Analysis | type: Text]
+bilities. As shown in Table 2, when using Qwen2.5-7B as the generator, Hard-Weighted Self-Consistency yields a relative improvement of 23.5% over the base model's single-sample accuracy, substantially outperforming the 12.2% gain achieved by standard Self-Consistency. Notably, the domain-specialized AlphaMed-7B model also benefits from verifier guidance with a 6.5% relative improvement, indicating that our verifier provides complementary verification capabilities beyond domain-specific fine-tuning. The improvements extend to larger models as well: Qwen2.5-32B achieves a 3.8% relative gain during test-time search, demonstrating that a light-weight 8B verifier can effectively guide models that are significantly larger than itself. This
+
+[p. 7 | section: 4.3. Analysis | type: Text]
+cross-scale generalization suggests that Med-TIV learns transferable verification patterns rather than overfitting to specific generator characteristics.
+
+[p. 7 | section: 4.3. Analysis | type: Text]
+Q2: How do different test-time search strategies compare under Med-TIV? We then systematically compare different test-time search strategies under verifier guidance to identify the most effective approach for leveraging verification signals. As shown in Table 2, Hard-Weighted Self-Consistency consistently achieves the highest accuracy across all generators, followed by Soft-Weighted Self-Consistency and Best-of-N selection. On Qwen2.5-7B, Hard-Weighted Self-Consistency outperforms Best-of-N by 3% absolute accuracy, suggesting that majority voting among verified traces provides more robust answer selection than simply choosing the highest-confidence individual trace.
+
+[p. 7 | section: 4.3. Analysis | type: Text]
+Q3: Can Med-TIV reduce the sampling budget required to achieve state-of-the-art performance compared to existing baselines? Next, we investigated how verification performance scales with sampling budget, a critical consideration for deployment under varying computational constraints. As shown in Figure 3, Med-TIV achieves substantial efficiency advantage over existing medical reward models across all three benchmarks. In particular, Med-TIV matches the performance of baselines using only 4 samples, whereas the baselines require 32 samples, representing an 8× reduction in sampling budget. On MedOA, Med-TIV achieves 72.1% accuracy at N=4, while Med-PRM requires the full N=32 budget to reach 70.0% accuracy. Since inference cost scales approximately linearly with the number of sampled traces, this translates to equivalent performance at one-eighth the generator inference cost in practical deployment settings.
+
+[p. 7 | section: 4.3. Analysis | type: Text]
+Q4: Does Med-TIV generalize across different base models? To assess the generality of our proposed framework, we compared verification performance using two
+
+[p. 8 | section: 4.3. Analysis | type: FigureGroup]
+Figure 4. Ablation on base model selection and training iterations.
+
+[p. 8 | section: 4.3. Analysis | type: Text]
+distinct verifier backbones: Llama3.1-8B and Qwen2.5-7B. As shown in Figure 4, both backbones achieve strong performance after two training iterations. Llama3.1-8B consistently outperforms Qwen2.5-7B by approximately 3.5% absolute accuracy on MedQA, achieving 75.86% versus 72.35% after 2 iterations of training. The parallel performance gains observed across both models indicate that Med-TIV is agnostic to backbone architectures.
+
+[p. 8 | section: 4.3. Analysis | type: Text]
+Q5: What is the impact of iterative training? Figure 4 presents ablation results examining the impact of iterative training with adaptive curriculum formulation. Llama3.1-8B improves from 60.96% to 75.26% after iteration 1, with marginal gains to 75.86% at iteration 2. Qwen2.5-7B follows a similar pattern, reaching 72.35% after two iterations. The rapid convergence suggest that the majority of verification capability is acquired in the first round, with subsequent iterations refining boundary cases.
+
+[p. 8 | section: 4.3. Analysis | type: Text]
+O6 : How does RL and tool integration impact verification performance? Table 3 highlights the dual benefits of our framework across two generators. RL
+
+[p. 8 | section: 4.3. Analysis | type: TableGroup]
+Table 3. Ablation on RL and tool integration. Models MedQA Qwen2.5-7B 60.96 + Med-TIV (RL) 69.60 + Med-TIV (RL + Tool) 70.54 AlphaMed-7B 71.01 + Med-TIV (RL) 76.12 + Med-TIV (RL + Tool) 77.14
+
+[p. 8 | section: 4.3. Analysis | type: Text]
+training drives the primary gain, boosting MedQA accuracy of Qwen2.5-7B by 8.64%, confirming that the verifier effectively internalizes reasoning patterns. Tool integration provides a critical secondary boost, further elevating accuracy to 70.54%. A similar cumulative trend is observed with AlphaMed-7B. This demonstrates that while RL
+
+[p. 8 | section: 4.3. Analysis | type: Text]
+anchors logical verification, dynamic retrieval is essential for resolving knowledge-intensive boundary cases beyond the model's parametric memory.
+
+[p. 8 | section: 5. Related Work | type: Text]
+Medical Reasoning Models. The application of large language models to medical reasoning has attracted considerable attention. Early efforts focused on domain-adaptive pretraining and instruction tuning on medical corpora (Wu et al., 2023; Singhal et al., 2025; Chen et al., 2023). More recent work has explored reasoning-enhanced medical models. HuatuoGPT-o1 (Chen et al., 2024) incorporates chainof-thought reasoning with verification mechanisms, and UltraMedical (Zhang et al., 2024) combines high-quality instruction data with preference optimization. AlphaMed (Liu et al., 2025a) employs RL to improve medical reasoning capabilities. Despite these advances, most existing approaches focus on improving the generator model itself, whereas our work addresses the complementary problem of training a plug-and-play verifier that can improve any frozen generator through test-time search.
+
+[p. 8 | section: 5. Related Work | type: Text]
+Tool-Assisted Reward and Judge Models. Standard LLM-based judges typically function as passive scorers limited by parametric knowledge. Recent work addresses this through agentic reward modeling, equipping verifiers with executable tools. Themis (Li et al., 2024) established the foundational framework by enabling access to calculators, search engines, and knowledge bases through structured tool-calling traces. TIR-Judge (Xu et al., 2025) advanced this paradigm in the general domain by integrating code execution to judge paired responses. TIM-PRM (Kuang et al., 2025) introduced independent tool queries for multi-modal verification to eliminate confirmation bias. The concept has further expanded to the Agent-as-a-Judge paradigm (You et al., 2026), which employs dynamic planning, tool augmentation and multi-agent coordination to decompose complex evaluation tasks. Our work instantiates this agentic paradigm within the medical domain, moving beyond static retrieval to iterative, evidence-grounded clinical verification.
+
+[p. 8 | section: 6. Conclusion | type: Text]
+We presented Med-TIV, an agentic RL framework for medical reasoning verification. Our approach addresses key limitations of existing medical reward models by offering explicit critique traces and enabling dynamic knowledge retrieval during verification. Empirical evaluations across four medical reasoning benchmarks demonstrate that Med-TIV substantially outperforms prior approaches. More broadly, Med-TIV introduces a general paradigm for training toolaugmented verifiers that can be extended to other highstakes domains requiring evidence-grounded evaluation.
+
+[p. 9 | section: Impact Statement | type: Text]
+This paper introduces research aimed at improving the reliability of large language models for medical reasoning tasks. We believe our work contributes positively to the development of trustworthy medical AI systems by providing mechanisms to verify reasoning correctness before clinical deployment. Med-TIV holds potential to enhance the safety of LLM-assisted clinical decision support by reducing erroneous reasoning outputs through systematic verification. By grounding judgments in retrieved medical evidence, our approach offers improved transparency compared to opaque scalar reward models, enabling practitioners to better understand and audit verification decisions. The efficiency gains demonstrated by Med-TIV could democratize access to reliable medical reasoning verification, making robust verification feasible even in resource-constrained settings.

icml26/19e76363-53a6-4f3c-8b50-844e1aea4e26/paper.md

@@ -0,0 +1,899 @@
+
+
+{0}------------------------------------------------
+
+# Scaling Medical Reasoning Verification via Tool-Integrated Reinforcement Learning
+
+### Anonymous Authors<sup>1</sup>
+
+# Abstract
+
+Large language models have achieved strong performance on medical reasoning benchmarks, yet their deployment in clinical settings demands rigorous verification to ensure factual accuracy. While reward models offer a scalable approach for reasoning trace verification, existing methods face two limitations: they produce only scalar reward values without explicit justification, and they rely on single-pass retrieval that precludes adaptive knowledge access as verification unfolds. We introduce Med-TIV, an agentic framework that addresses these limitations by training medical reasoning verifiers to iteratively query external medical corpora during evaluation. Our approach combines tool-augmented verification with an iterative reinforcement learning paradigm that requires only trace-level supervision, alongside an adaptive curriculum mechanism that dynamically adjusts training data distribution. Across four medical reasoning benchmarks, Med-TIV achieves substantial gains over existing methods, improving MedQA accuracy by 23.5% and MedXpertQA by 32.0% relative to the base generator in particular. Crucially, Med-TIV demonstrates an 8× reduction in sampling budget requirement compared to prior reward model baselines. These findings establish that grounding verification in dynamically retrieved evidence offers a principled path toward more reliable medical reasoning systems.
+
+# 1. Introduction
+
+Large Language Models (LLMs) have demonstrated remarkable capabilities in medical reasoning, achieving competitive performance on clinical question answering, diagnostic
+
+Preliminary work. Under review by the International Conference on Machine Learning (ICML). Do not distribute.
+
+<span id="page-0-0"></span>![](_page_0_Figure_9.jpeg)
+
+*Figure 1.* Comparison of medical reasoning verification paradigms. Text-based judges rely on parametric knowledge and may validate erroneous reasoning, while tool-integrated judges dynamically retrieve evidence to ground their judgments.
+
+inference, and medical knowledge benchmarks [\(Ji et al.,](#page-8-0) [2025a;](#page-8-0) [Xiao et al.,](#page-9-0) [2026\)](#page-9-0). While these advances hold significant promise for augmenting clinical decision making and democratizing access to medical expertise, the deployment of LLMs in high-stakes clinical settings demands rigorous verification mechanisms to ensure that generated reasoning is both factually accurate and logically sound [\(Zhang et al.,](#page-9-1) [2025;](#page-9-1) [Wang et al.,](#page-9-2) [2025\)](#page-9-2).
+
+Reward-based judges have therefore emerged as a scalable solution for evaluating model outputs, supporting both posttraining refinement via reinforcement learning from human feedback (RLHF) and inference-time scaling through tree search [\(Snell et al.,](#page-9-3) [2024\)](#page-9-3). These judges can be broadly categorized by the granularity of their supervision. Outcome Reward Models (ORMs) provide sparse trace-level supervision that quantifies the quality of the entire output, while Process Reward Models (PRMs) offer dense step-level feedback that scores each intermediate reasoning step, enabling fine-grained credit assignment and precise error localization within multi-step reasoning. Recent work has adapted both paradigms to the medical domain to assess complex clinical reasoning traces. In parallel, advances in generative reward modeling have extended judge models beyond
+
+<sup>1</sup>Anonymous Institution, Anonymous City, Anonymous Region, Anonymous Country. Correspondence to: Anonymous Author <anon.email@domain.com>.
+
+{1}------------------------------------------------
+
+scalar scoring, enabling them to produce natural-language critiques that explicitly justify their decisions [\(Liu et al.,](#page-9-4) [2025c;](#page-9-4) [Xiong et al.,](#page-9-5) [2025\)](#page-9-5).
+
+104
+
+106
+
+108 109 Despite their effectiveness, reward-based judges exhibit fundamental limitations when applied to clinical reasoning tasks [\(Yun et al.,](#page-9-6) [2025\)](#page-9-6). A primary concern is the prevalence of hallucinations in critique traces, where judge models generate plausible yet factually incorrect assessments (Figure [1\)](#page-0-0). This issue is particularly noticeable in the medical domain, where reliable verification demands grounding in authoritative clinical evidence and established medical knowledge. Unverified judgments could lead to the propagation of incorrect diagnostic or treatment recommendations. Existing medical reasoning verifiers typically provide only scalar reward signals, offering little or no justification for their judgments and thus limiting interpretability [\(Jiang et al.,](#page-8-1) [2025b\)](#page-8-1). Furthermore, these methods often rely on a static Retrieval-Augmented Generation (RAG) pipeline, in which a fixed set of retrieved documents is prefixed to the context and remains unchanged throughout evaluation [\(Yun et al.,](#page-9-6) [2025\)](#page-9-6). Such static design precludes adaptive, multi-turn evidence gathering and forces the verifier to a fixed retrieval budget, thus limiting scalability.
+
+To address these issues, we propose Med-TIV (Medical Tool-Integrated reasoning Verifier), an agentic reinforcement learning (RL) framework that trains LLMs to leverage external knowledge bases for judging medical reasoning traces[1](#page-1-0) . Med-TIV features three key design principles: (1) a tool-augmented verification paradigm that enables dynamic, iterative knowledge retrieval during the evaluation process; (2) an iterative RL approach that progressively improves verification capabilities without requiring step-level expert annotations; and (3) an adaptive curriculum formulation strategy that adjusts the data distribution in response to the evolving capability of the model. By equipping judge models with tool-use capabilities, Med-TIV grounds evaluation decisions in external evidence rather than relying solely on parametric knowledge, thereby mitigating hallucination, improving interpretability, and overcoming the limitations of static RAG [\(Ji et al.,](#page-8-2) [2025b;](#page-8-2) [Xia et al.,](#page-9-7) [2025\)](#page-9-7).
+
+To verify the effectiveness of Med-TIV, we conduct extensive experiments on common medical reasoning benchmarks. Our results demonstrate that Med-TIV trains strong medical verifiers: when guiding inference-time search for a 7B generator model, our trained verifier achieves relative improvements of 23.5% on MedQA and 32.0% on MedXpertQA compared to the generator model alone. Moreover, Med-TIV consistently outperforms existing medical reward model baselines and surpasses the performance of models that are up to 4× larger in scale. Notably, Med-TIV
+
+also demonstrates an 8× gain in sampling efficiency compared to prior reward-based approaches, achieving equivalent accuracy with substantially fewer sampled reasoning traces during test-time search.
+
+Our main contributions are summarized as follows:
+
+- We propose Med-TIV, a novel tool-integrated verification framework that enables dynamic, iterative knowledge retrieval during medical reasoning evaluation, providing both interpretable, fine-grained justification and improved factual grounding.
+- We introduce an iterative RL paradigm with curriculumbased difficulty adaptation that progressively improves verification capabilities through self-bootstrapping, requiring only trace-level supervision rather than dense steplevel expert annotations.
+- Med-TIV achieves state-of-the-art performance on four medical reasoning benchmarks, with comprehensive ablation studies that validate each component's contribution.
+
+# 2. Preliminaries
+
+#### 2.1. Problem Setup
+
+We define *medical reasoning verification* as the task of assessing the correctness of a multi-step reasoning trace generated in response to a medical question. Formally, given a medical question q ∈ Q and a multi-step reasoning trace τ = (s1, s2, . . . , sm) from a generator model, a verifier model determines whether τ contains any errors. We formulate this problem as binary classification, where the verifier Vθ(q, τ ) produces a judgment ℓ ∈ {0, 1}, where ℓ = 1 indicates a error-free reasoning trace, and ℓ = 0 indicates the presence of one or more errors. Unlike scalar reward models that output continuous scores, we adopt a generative judge paradigm in which the verifier produces a discrete judgment accompanied by a detailed critique trace that provides a structured justification for the decision.
+
+#### 2.2. Tool-Augmented Reasoning Verifier
+
+Following prior works [\(Jin et al.,](#page-8-3) [2025\)](#page-8-3), we extend the verifier with access to an external *search engine* E that retrieves top-k documents from a curated medical corpus (See Appendix [B.2](#page-10-0) for details). Retrieved documents are appended verbatim to the verifier context. Given a verification instance (q, τ ), the verifier constructs an iterative verfication trajectory. At step k, the trajectory is represented as:
+
+$$\mathbf{t}_k = \{r_1, a_1, o_1, \dots, r_k, a_k, o_k\},\$$
+
+where r<sup>i</sup> denotes a natural language reasoning step analyzing the medical content, a<sup>i</sup> is a search query formulated to retrieve relevant medical knowledge, and o<sup>i</sup> = E(ai) represents the retrieved documents. The iterative verification
+
+<span id="page-1-0"></span><sup>1</sup>Code is available at [https://github.com/](https://github.com/PittNAIL/med-tiv) [PittNAIL/med-tiv](https://github.com/PittNAIL/med-tiv)
+
+{2}------------------------------------------------
+
+Figure 2. Overview of Med-TIV. Left: Tool-integrated verification iteratively analyzes reasoning traces, formulates search queries, and retrieves medical evidence before producing correctness judgments. Middle: Curriculum formulation filters trivial and impossible instances, retaining boundary cases for RL training. Right: At inference time, the verifier evaluates candidate medical reasoning traces generated by a frozen model and final answers are selected via weighted self-consistency.
+
+process is defined as:
+
+<span id="page-2-0"></span>110
+
+111
+
+112
+
+113
+
+114
+
+115
+
+116 117
+
+118
+
+119
+
+121
+
+122
+
+123124125
+
+127 128
+
+129
+
+130 131
+
+132
+
+134
+
+135136
+
+137138
+
+139 140
+
+141
+
+142
+
+143
+
+144
+
+145
+
+147
+
+148
+
+149
+
+150
+
+151
+
+152
+
+153
+
+154
+
+155
+
+156
+
+157
+
+158
+
+159
+
+160
+
+161
+
+162
+
+163
+
+164
+
+$$(r_k, a_k) \sim V_{\theta}(q, \tau, \mathbf{t}_{k-1}),$$
+  $o_k = \mathcal{E}(a_k),$   $\mathbf{t}_k = \mathbf{t}_{k-1} \oplus r_k \oplus a_k \oplus o_k,$ 
+
+where  $\oplus$  denotes sequence concatenation. This process continues until the verifier produces a final judgment  $\ell \sim V_{\theta}(q,\tau,\mathbf{t}_T)$  at the terminal step T. By allowing multiple tool executions, the verifier dynamically retrievs medical knowledge as need to verify specific claims in the reasoning trace. Table 5 in the Appendix shows the explicit instruction used in our experiments.
+
+#### 2.3. Test-Time Search
+
+Test-time search strategies improve reasoning performance by leveraging reward models to evaluate and select among multiple candidate solutions (Shi et al., 2024). Given a frozen generator model  $\pi_{\rm gen}$  and a question q, we first sample N independent reasoning traces:
+
+$$\{\tau^{(j)}\}_{j=1}^N \sim \pi_{\mathrm{gen}}(\cdot \mid q).$$
+
+A trained verifier  $V_{\theta}$  then scores each candidate trace, and the final output is selected based on these scores. Common selection strategies include Best-of-N sampling, which
+
+selects the trace with the highest score:
+
+$$\hat{\tau} = \arg\max_{\tau^{(j)}} V_{\theta}(q, \tau^{(j)}),$$
+
+and verification-based majority voting, where candidate traces are first filtered by the verifier and the final answer is determined by consensus among verified traces. Med-TIV trains such a plug-in verifier that provides tool-grounded assessments that can be used to augment decision-making for any frozen generator model at inference time.
+
+#### 3. Tool-Integrated Medical Reasoning Verifier
+
+Med-TIV is an agentic verification framework that trains models to leverage external knowledge bases for verifying whether a given medical reasoning trace contains errors. We adopt an iterative training approach based on dynamic curriculum learning, which requires no fine-grained step-level expert supervision and trains solely through multiple rounds of reinforcement learning (Figure 2). We next describe the training procedure of Med-TIV in details.
+
+#### 3.1. Tool-Integrated RL with Verifiable Rewards
+
+**Data Construction.** All training data across iterations is derived from the open-source Med-PRM dataset. Each original instance consists of a tuple  $(q, \tau, \ell_{\text{step}}, \ell_{\text{trace}})$ , where q is a medical question,  $\tau$  is a multi-step reasoning trace,  $\ell_{\text{step}}$  de-
+
+{3}------------------------------------------------
+
+notes step-level labels, and  $\ell_{trace}$  is a trace-level correctness label<sup>2</sup>.
+
+At each training iteration, we only utilize the triplet  $(q,\tau,\ell_{\text{trace}})$  with human-annotated trace-level labels. Step-level labels  $\ell_{\text{step}}$  is intentionally excluded, as Med-TIV is designed to improve verification performance without replying on fine-grained supervision. For each training iteration, we fix the training data budget to 20K instances and enforce a balanced label distribution between correct ( $\ell_{\text{trace}}=1$ ) and incorrect ( $\ell_{\text{trace}}=0$ ) reasoning traces.
+
+**Algorithm.** We employ Dr. GRPO (Liu et al., 2025b) as the RL algorithm for training the verifier. Given a verification instance  $(q_i, \tau_i)$ , we sample a group of G verification trajectories  $\{\mathbf{o}_i\}_{i=1}^G$  from the current policy  $\pi_\theta$ . Each trajectory  $\mathbf{o}_i = (o_i^1, \dots, o_i^{|\mathbf{o}_i|})$  consists of reasoning tokens, search queries, retrieved documents, and a final judgment. The objective is:
+
+$$\frac{1}{G} \sum_{i=1}^{G} \sum_{t=1}^{|\mathbf{o}_i|} \left\{ \min \left[ r_i^t \hat{A}_i^t, \operatorname{clip} \left( r_i^t, 1 - \epsilon_l, 1 + \epsilon_h \right) \hat{A}_i^t \right] \right\}, \quad (1)$$
+
+where  $r_i^t = \frac{\pi_\theta(o_i^t|\mathbf{q},\mathbf{o}_i^{< t})}{\pi_{\theta_{\mathrm{old}}}(o_i^t|\mathbf{q},\mathbf{o}_i^{< t})}$ ,  $\mathbf{q} = (q,\tau)$  denotes the input prompt containing the question and reasoning trace,  $\mathbf{o}_i^{< t}$  represents previously generated tokens, and  $\epsilon_l$  and  $\epsilon_h$  are the clipping parameters. The advantage term  $\hat{A}_i^t$  is defined as:
+
+$$\hat{A}_i^t = R(\mathbf{q}, \mathbf{o}_i) - \text{mean}\left(\left\{R(\mathbf{q}, \mathbf{o}_1), \dots, R(\mathbf{q}, \mathbf{o}_G)\right\}\right).$$
+
+**Reward Designs.** To facilitate multi-turn RL with tool execution, we design a structured reward covering two complementary objectives, following prior practices (Jin et al., 2025):
+
+(i) Correctness Reward  $(R_c)$ : This component measures whether the verifier's judgment aligns with the ground-truth label. Let  $\mathbf{q} = (q, \tau)$  denote the verification prompt and  $\ell \in \{0, 1\}$  the ground-truth label. We define:
+
+$$R_c = \mathbb{1}(\text{extract}(\mathbf{o}) = \ell),$$
+
+where  $\mathbb{1}(\cdot)$  is the indicator function and extract(o) parses the final judgment from the <answer> tags in the generated trajectory o. Intuitively,  $R_c=1$  if the verifier's decision is correct, and  $R_c=0$  otherwise.
+
+(ii) Format Reward  $(R_f)$ : To ensure reliable tool use and structured outputs, the verifier is required to adhere to a predefined format. Specifically, reasoning steps must be enclosed within <think> tags, search queries within <search> tags, and the final judgment within <answer>
+
+tags. To discourage degenerate outputs, we further penalize excessive tag usage. Specifically,  $R_f=1$  if the output satisfies all formatting constraints and contains no more than  $10 < answer> tag pairs; <math>R_f=0.25$  if the output is correct but exhibits tag overflow; and  $R_f=0$  otherwise.
+
+The final reward  ${\cal R}$  is defined as the product of the two components:
+
+$$R = R_c \times R_f$$
+.
+
+#### <span id="page-3-1"></span>3.2. Training Strategies
+
+Adaptive Curriculum Formulation. A central challenge in RL for verification is ensuring that training data remains appropriately calibrated to the evolving capability of the model. Instances that are either trivially easy or impossibly difficult yields minimal learning signal, as the resulting policy gradients approach zero. To address this issue, we adopt a model-aware curriculum formulation mechanism that dynamically adapts the task distribution at each training iteration.
+
+Concretely, before each iteration t, we perform online filtering on the sampled batch  $\mathcal{B}_t$  to construct an effective training set  $\mathcal{D}_t$ :
+
+$$\mathcal{D}_t = \{(q, \tau, \ell) \in \mathcal{B}_t : \exists g, g' \in \{1, \dots, G\} \text{ s.t. } r^{(g)} \neq r^{(g')}\}.$$
+
+Here, for each candidate instance  $(q, \tau, \ell) \in \mathcal{B}_t$ , we sample G verification trajectories  $\{o^{(g)}\}_{g=1}^G$  from the current policy  $\pi_{\theta_t}$ . We then compute the corresponding rewards  $\{r^{(g)}\}_{g=1}^G$ . Finally, we retain only instances if any two rewards are different, i.e., reward variance is non-zero (Khatri et al., 2025).
+
+This criterion eliminates prompts where the model either consistently succeeds or consistently fails across all sampled trajectories. By filtering these zero-gradient instances, optimization is focused on decision-boundary cases where the verifier exhibits uncertainty.
+
+To maintain a fixed training budget per iteration, we iteratively resample additional instances from the labeled pool  $\mathcal{B}$  and apply the same filtering criterion until  $|\mathcal{D}_t|=20K$ . This dynamic curriculum evolves naturally across iterations as the verifier improves, eliminating the need for manually designed difficulty schedules.
+
+Iterative Training via Self-Bootstrapping. We adopt an iterative training approach that progressively improves verification capabilities through multiple rounds of RL. Unlike prior work that alternates between rejection sampling, supervised fine-tuning (SFT), and RL (Xu et al., 2025), our approach operates entirely through iterative RL, following the RL-Zero paradigm where the model reinforces its verification capabilities without requiring dense turn-level expert demonstrations for cold start.
+
+<span id="page-3-0"></span><sup>&</sup>lt;sup>2</sup>Dataset is available at https://huggingface.co/datasets/dmis-lab/llama-3.l-medprm-reward-training-set
+
+{4}------------------------------------------------
+
+<span id="page-4-0"></span>Algorithm 1 Iterative Training of Tool-Integrated Medical Reasoning Verifier
+
+221
+
+222
+
+223
+
+224
+
+225
+
+226
+
+227
+
+228
+
+229
+
+230
+
+231
+
+232
+
+233
+
+234
+
+235
+
+236
+
+237
+
+238
+
+239
+
+240
+
+241
+
+242
+
+243
+
+244
+
+245
+
+246
+
+247
+
+248
+
+249250
+
+251
+
+252
+
+253
+
+254
+
+255
+
+256
+
+257
+
+258
+
+259
+
+260
+
+261
+
+262
+
+263
+
+264
+
+265
+
+266
+
+267
+
+268
+
+269
+
+270
+
+271
+
+272
+
+274
+
+```
+Require: Base verifier \pi_{\theta_0}, labeled dataset pool \mathcal{D} =
+      \{(q_i, \tau_i, \ell_i)\}_{i=1}^N, maximum iterations T_{\text{max}}, batch size
+      B, group size G, search engine \mathcal{E}
+Ensure: Trained verifier \pi_{\theta^*}
+ 1: for t = 1 to T_{\text{max}} do
+
+    Sample labeled batch
+
+          \mathcal{B}_t \leftarrow \text{SAMPLEBATCH}(\mathcal{D}, B)
+ 3:
+ 4:
+          \mathcal{D}_t \leftarrow \emptyset
+          ▷ Curriculum formulation
+ 5:
+          for each (q, \tau, \ell) \in \mathcal{B}_t do
+ 6:
+              Sample verification trajectories:
+ 7:
+                  \{\hat{\ell}^{(g)}\}_{q=1}^G \sim \pi_{\theta_t}(\cdot \mid q, \tau, \mathcal{E})
+              Compute rewards within group:
+ 8:
+                  r^{(\hat{g})} \leftarrow \mathbb{1}[\hat{\ell}^{(g)} = \ell], \text{ for } g \in 1, \dots, G
+              if \exists g \neq g' such that r^{(g)} \neq r^{(g')} then
+ 9:
+                  Add (q, \tau, \ell) to curriculum set \mathcal{D}_t
+10:
+11:
+              end if
+12:
+          end for
+13:
+          ▶ RL optimization on curriculum data
+          \pi_{\theta_{t+1}} \leftarrow \text{DR.GRPO}(\pi_{\theta_t}, \mathcal{D}_t, \mathcal{E})
+15: end for
+16: Return \pi_{\theta_{T_{\max}}}
+```
+
+Starting from the base model  $\pi_{\theta_0}$ , we perform  $T_{\text{max}}$  iterations. Each iteration consists of three stages:
+
+```
+\begin{split} \mathcal{B}_t \leftarrow \text{SampleBatch}(\mathcal{D}, B), \\ \mathcal{D}_t \leftarrow \text{Filter}(\mathcal{B}_t, \pi_{\theta_t}), \\ \pi_{\theta_{t+1}} \leftarrow \text{RL}(\pi_{\theta_t}, \mathcal{D}_t). \end{split}
+```
+
+Each iteration draws a fresh batch  $\mathcal{B}_t$  from the annotated pool  $\mathcal{D}$  with trace-level labels, ensuring a balanced distribution of correct and incorrect reasoning traces. The curriculum filtering then constructs the training set  $\mathcal{D}_t$  as described above, and RL optimization updates the policy based on the structured reward.
+
+The key insight underlying this iterative approach is the coevolution of model capability and training distribution. As the verifier improves, the filtering mechanism automatically removes instances that have become too easy, while the fresh sampling introduces new challenging cases. This creates a self-bootstrapping cycle: stronger models encounter harder verification tasks, which in turn drive further improvements. Since the trace-level correctness reward is deterministic and unambiguous, this self-bootstrapping process converges reliably without the instabilities that can arise from noisy synthetic step-level labels. We summarize the overall training procedure in Algorithm 1.
+
+#### 4. Experiments
+
+#### 4.1. Experimental Setup
+
+**Evaluation benchmarks.** We evaluated Med-TIV on four open-source medical question-answering benchmarks: MedQA (Jin et al., 2020), MedMCQA (Pal et al., 2022), MMLU-Med (Hendrycks et al., 2021), and MedX-pertQA (Zuo et al., 2025), using accuracy as the evaluation metric. These benchmarks collectively assess the verifier's ability to distinguish correct from erroneous reasoning across varying difficulty levels and medical subdomains. Detailed descriptions of benchmarks are in Appendix C.1.
+
+Implementation details. We trained verifiers using two light-weight backbone models: Llama3.1-8B and Qwen2.5-7B, with Llama3.1-8B as the default for results reporting. All training was conducted using the VeRL-Tool framework (Jiang et al., 2025a). Detailed hyperparameters are shown in Appendix B.1. All experiments were conducted on 4 NVIDIA H100 GPUs with 80GB of memory. Due to computational constraints, we limit the maximum number of RL iterations to  $T_{\rm max}=2$  and we set the group size for curriculum formulation (Section 3.2) to G=8.
+
+For inference, we used the default sampling hyperparameters for all models. In reward-guided search experiments, unless otherwise specified, we used Qwen2.5-7B as the frozen generator and sampled up to 32 candidate reasoning traces per question. We applied Hard-Weighted Self-Consistency as the default test-time search strategy.
+
+**Baselines.** We compared Med-TIV against two groups of baselines. 1): *Off-the-shelf LLMs*: GPT-40-mini (OpenAI et al., 2024), Gemini-2.0-Flash, DeepSeek-R1 series (Guo et al., 2025), Qwen2.5 series (Yang et al., 2025), Llama3.1 (Grattafiori et al., 2024), AlphaMed (Liu et al., 2025a), UltraMedical (Zhang et al., 2024), and HuatuoGPT-01 (Chen et al., 2024). 2): *Medical domain-specialized Reward Models*: MedS<sup>3</sup> (Jiang et al., 2025b) and Med-PRM (Yun et al., 2025). Detailed descriptions of each reward model baseline are shown in Appendix B.3.
+
+**Test-Time Search Strategies.** We evaluated three test-time search strategies that leverage Med-TIV to improve the reasoning performance of frozen generators. Given a reasoning trace  $\tau=(s_1,s_2,\ldots,s_K)$  with K steps, our verifier assigns a confidence score  $r_{\tau}\in[0,1]$  for the entire trace, defined as the softmax probability of the 1 token over the logits of both 1 and 0 tokens.
+
+• Best-of-N. Given a question q, we sampled N candidate traces  $\{\tau^{(j)}\}_{j=1}^N$  from the generator and selected the trace with the highest verifier confidence score:
+
+$$\hat{\tau} = \arg\max_{\tau^{(j)}} r_{\tau^{(j)}}.$$
+
+{5}------------------------------------------------
+
+313 314
+
+316
+
+324
+
+326
+
+328 329
+
+<span id="page-5-0"></span>*Table 1.* Main evaluation results on medical reasoning benchmarks. We report accuracy (%) on MedQA, MedMCQA, MMLU-Med, and MedXpertQA. Bold numbers indicate the best results among the reward model group. ✓: Verifier supports external tools for judging; ✗: Verifier does not support external tools for judging.
+
+| Baselines                    | å | Train | Size | MedQA | MedMCQA | MMLU-Med | MedXpertQA | Avg.  |
+|------------------------------|---|-------|------|-------|---------|----------|------------|-------|
+| Proprietary Models           |   |       |      |       |         |          |            |       |
+| GPT-4o-mini                  | - | -     | -    | 79.03 | 68.20   | 87.79    | 17.84      | 63.22 |
+| Gemini-2.0-Flash             | - | -     | -    | 87.51 | 72.60   | 92.01    | 20.57      | 68.17 |
+| General Reasoning Models     |   |       |      |       |         |          |            |       |
+| DeepSeek-R1                  | - | -     | 671B | 90.34 | 78.80   | 94.40    | 37.76      | 75.33 |
+| R1-Distill-Qwen              | - | -     | 7B   | 24.82 | 36.40   | 47.47    | 7.43       | 29.03 |
+| R1-Distill-Llama             | - | -     | 8B   | 34.96 | 43.60   | 64.19    | 5.35       | 37.03 |
+| General Non-reasoning Models |   |       |      |       |         |          |            |       |
+| Qwen2.5                      | - | -     | 32B  | 73.21 | 64.83   | 84.94    | 13.87      | 59.21 |
+| Qwen2.5                      | - | -     | 7B   | 60.96 | 56.56   | 76.96    | 12.15      | 51.66 |
+| Llama3.1                     | - | -     | 8B   | 70.93 | 61.60   | 78.97    | 13.02      | 56.13 |
+| Medical Reasoning Models     |   |       |      |       |         |          |            |       |
+| AlphaMed                     | - | -     | 7B   | 71.01 | 61.46   | 81.16    | 19.16      | 58.20 |
+| UltraMedical                 | - | -     | 8B   | 72.66 | 62.60   | 79.61    | 15.25      | 57.53 |
+| HuatuoGPT-o1                 | - | -     | 8B   | 72.19 | 63.60   | 75.30    | 16.84      | 56.98 |
+| Medical Reward Models        |   |       |      |       |         |          |            |       |
+| MedS3                        | ✗ | 225k  | 7B   | 64.89 | 58.91   | 80.53    | 12.90      | 54.31 |
+| Med-PRM                      | ✓ | 111k  | 7B   | 69.99 | 62.36   | 80.99    | 13.51      | 56.71 |
+| Med-TIV (Ours)               | ✓ | 20k   | 7B   | 75.26 | 64.70   | 85.58    | 16.04      | 60.40 |
+
+• *Hard-Weighted Self-Consistency.* We first filtered traces by the verifier's binary judgment, keeping only those labeled correct (Vθ(q, τ ) = 1). Among the filtered traces, we applied majority voting to determine the final answer:
+
+$$\hat{a} = \arg\max_{a} \sum_{j=1}^{N} \mathbb{1} \left[ V_{\theta}(q, \tau^{(j)}) = 1 \right] \cdot \mathbb{1} \left[ \operatorname{ans}(\tau^{(j)}) = a \right].$$
+
+• *Soft-Weighted Self-Consistency.* Instead of binary filtering, we weighted each trace's vote by the verifier's confidence score:
+
+$$\hat{a} = \arg\max_{a} \sum_{j=1}^{N} r_{\tau^{(j)}} \cdot \mathbb{1} \big[ \mathrm{ans}(\tau^{(j)}) = a \big].$$
+
+### 4.2. Main Results
+
+Table [1](#page-5-0) presents the main results on four medical reasoning benchmarks. Models trained with Med-TIV consistently outperform existing baselines across all benchmarks. Specifically, under guided-search using a Med-TIVtrained verifier, Qwen2.5-7B attains accuracies of 75.26%
+
+on MedQA, 64.70% on MedMCQA, 85.58% on MMLU-Med, and 16.04% on MedXpertQA, yielding an average accuracy of 60.40%. Notably, Med-TIV enables this 7B generator to rival substantially larger models, even surpassing the base performance of Qwen2.5-32B despite using a generator that is approximately 4× smaller. Compared to domain-specialized medical reasoning models of similar scale, Med-TIV outperforms HuatuoGPT-o1-8B and UltraMedical-8B by 3.07% and 2.60% on MedQA, respectively, demonstrating the effectiveness of our tool-integrated verification. Case analysis in Appendix [D](#page-12-0) further illustrates how Med-TIV identifies subtle reasoning errors.
+
+#### 4.3. Analysis
+
+We conducted a series of ablation analyses to investigate six key research questions regarding our proposed framework.
+
+Q1: Does **Med-TIV** generalize across different generator models? To evaluate the generalizability of the trained verifier, we applied Med-TIV to guide test-time search across generator models of varying sizes and capa-
+
+{6}------------------------------------------------
+
+<span id="page-6-1"></span>![](_page_6_Figure_2.jpeg)
+
+Figure 3. Test-time scaling analysis across three medical reasoning benchmarks. Each plot shows accuracy versus sampling budget  $N \in \{1, 2, 4, 8, 16, 32\}$  for four baselines. Med-TIV consistently outperforms baselines across all sampling budgets and benchmarks.
+
+<span id="page-6-0"></span>Table 2. Performance improvements from using Med-TIV as a verifier on MedQA. For each generator model, the first row indicates the accuracy over single sampled trace per question.
+
+| Models                       | MedQA                 |
+|------------------------------|-----------------------|
+| Qwen2.5-7B                   | 60.96                 |
+| + Self-Consistency           | 66.38 (+5.42)         |
+| + Best-of-N (Med-TIV)        | 72.35 (+11.39)        |
+| + Soft Weighted SC (Med-TIV) | 75.02 (+14.06)        |
+| + Hard Weighted SC (Med-TIV) | <b>75.26</b> (+14.30) |
+| AlphaMed-7B                  | 71.01                 |
+| + Self-Consistency           | 74.23 (+3.22)         |
+| + Best-of-N (Med-TIV)        | 75.02 (+4.01)         |
+| + Soft Weighted SC (Med-TIV) | 75.33 (+4.32)         |
+| + Hard Weighted SC (Med-TIV) | <b>75.65</b> (+4.64)  |
+| Qwen2.5-32B                  | 73.21                 |
+| + Self-Consistency           | 75.26 (+2.05)         |
+| + Best-of-N (Med-TIV)        | 75.57 (+2.36)         |
+| + Soft Weighted SC (Med-TIV) | 75.96 (+2.75)         |
+| + Hard Weighted SC (Med-TIV) | <b>75.96</b> (+2.75)  |
+
+bilities. As shown in Table 2, when using Qwen2.5-7B as the generator, Hard-Weighted Self-Consistency yields a relative improvement of 23.5% over the base model's single-sample accuracy, substantially outperforming the 12.2% gain achieved by standard Self-Consistency. Notably, the domain-specialized AlphaMed-7B model also benefits from verifier guidance with a 6.5% relative improvement, indicating that our verifier provides complementary verification capabilities beyond domain-specific fine-tuning. The improvements extend to larger models as well: Qwen2.5-32B achieves a 3.8% relative gain during test-time search, demonstrating that a light-weight 8B verifier can effectively guide models that are significantly larger than itself. This
+
+cross-scale generalization suggests that Med-TIV learns transferable verification patterns rather than overfitting to specific generator characteristics.
+
+Q2: How do different test-time search strategies compare under Med-TIV? We then systematically compare different test-time search strategies under verifier guidance to identify the most effective approach for leveraging verification signals. As shown in Table 2, Hard-Weighted Self-Consistency consistently achieves the highest accuracy across all generators, followed by Soft-Weighted Self-Consistency and Best-of-N selection. On Qwen2.5-7B, Hard-Weighted Self-Consistency outperforms Best-of-N by 3% absolute accuracy, suggesting that majority voting among verified traces provides more robust answer selection than simply choosing the highest-confidence individual trace.
+
+Q3: Can Med-TIV reduce the sampling budget required to achieve state-of-the-art performance compared to existing baselines? Next, we investigated how verification performance scales with sampling budget, a critical consideration for deployment under varying computational constraints. As shown in Figure 3, Med-TIV achieves substantial efficiency advantage over existing medical reward models across all three benchmarks. In particular, Med-TIV matches the performance of baselines using only 4 samples, whereas the baselines require 32 samples, representing an 8× reduction in sampling budget. On MedOA, Med-TIV achieves 72.1% accuracy at N=4, while Med-PRM requires the full N=32 budget to reach 70.0% accuracy. Since inference cost scales approximately linearly with the number of sampled traces, this translates to equivalent performance at one-eighth the generator inference cost in practical deployment settings.
+
+**Q4:** Does Med-TIV generalize across different base models? To assess the generality of our proposed framework, we compared verification performance using two
+
+{7}------------------------------------------------
+
+<span id="page-7-0"></span>![](_page_7_Figure_1.jpeg)
+
+Figure 4. Ablation on base model selection and training iterations.
+
+distinct verifier backbones: Llama3.1-8B and Qwen2.5-7B. As shown in Figure 4, both backbones achieve strong performance after two training iterations. Llama3.1-8B consistently outperforms Qwen2.5-7B by approximately 3.5% absolute accuracy on MedQA, achieving 75.86% versus 72.35% after 2 iterations of training. The parallel performance gains observed across both models indicate that Med-TIV is agnostic to backbone architectures.
+
+**Q5:** What is the impact of iterative training? Figure 4 presents ablation results examining the impact of iterative training with adaptive curriculum formulation. Llama3.1-8B improves from 60.96% to 75.26% after iteration 1, with marginal gains to 75.86% at iteration 2. Qwen2.5-7B follows a similar pattern, reaching 72.35% after two iterations. The rapid convergence suggest that the majority of verification capability is acquired in the first round, with subsequent iterations refining boundary cases.
+
+**O6**: How does RL and tool integration impact verification performance? Table 3 highlights the dual benefits of our framework across two generators. RL
+
+385
+
+387
+
+388
+
+389
+
+390
+
+394
+
+395
+
+396
+
+397
+
+398
+
+399
+
+400
+
+401
+
+402
+
+403
+
+404
+
+405
+
+406
+
+407
+
+408
+
+409
+
+410
+
+411
+
+412
+
+413
+
+414
+
+415
+
+416
+
+417
+
+418
+
+419
+
+420
+
+421
+
+422
+
+423
+
+424
+
+425
+
+426
+
+427
+
+428
+
+429
+
+430
+
+431
+
+432
+
+433
+
+434
+
+435
+
+436
+
+437
+
+438
+
+439
+
+<span id="page-7-1"></span>Table 3. Ablation on RL and tool integration.
+
+| Models                | MedQA |
+|-----------------------|-------|
+| Qwen2.5-7B            | 60.96 |
+| + Med-TIV (RL)        | 69.60 |
+| + Med-TIV (RL + Tool) | 70.54 |
+| AlphaMed-7B           | 71.01 |
+| + Med-TIV (RL)        | 76.12 |
+| + Med-TIV (RL + Tool) | 77.14 |
+
+training drives the primary gain, boosting MedQA accuracy of Qwen2.5-7B by 8.64%, confirming that the verifier effectively internalizes reasoning patterns. Tool integration provides a critical secondary boost, further elevating accuracy to 70.54%. A similar cumulative trend is observed with AlphaMed-7B. This demonstrates that while RL
+
+anchors logical verification, dynamic retrieval is essential for resolving knowledge-intensive boundary cases beyond the model's parametric memory.
+
+#### 5. Related Work
+
+Medical Reasoning Models. The application of large language models to medical reasoning has attracted considerable attention. Early efforts focused on domain-adaptive pretraining and instruction tuning on medical corpora (Wu et al., 2023; Singhal et al., 2025; Chen et al., 2023). More recent work has explored reasoning-enhanced medical models. HuatuoGPT-o1 (Chen et al., 2024) incorporates chainof-thought reasoning with verification mechanisms, and UltraMedical (Zhang et al., 2024) combines high-quality instruction data with preference optimization. AlphaMed (Liu et al., 2025a) employs RL to improve medical reasoning capabilities. Despite these advances, most existing approaches focus on improving the generator model itself, whereas our work addresses the complementary problem of training a plug-and-play verifier that can improve any frozen generator through test-time search.
+
+Tool-Assisted Reward and Judge Models. Standard LLM-based judges typically function as passive scorers limited by parametric knowledge. Recent work addresses this through agentic reward modeling, equipping verifiers with executable tools. Themis (Li et al., 2024) established the foundational framework by enabling access to calculators, search engines, and knowledge bases through structured tool-calling traces. TIR-Judge (Xu et al., 2025) advanced this paradigm in the general domain by integrating code execution to judge paired responses. TIM-PRM (Kuang et al., 2025) introduced independent tool queries for multi-modal verification to eliminate confirmation bias. The concept has further expanded to the Agent-as-a-Judge paradigm (You et al., 2026), which employs dynamic planning, tool augmentation and multi-agent coordination to decompose complex evaluation tasks. Our work instantiates this agentic paradigm within the medical domain, moving beyond static retrieval to iterative, evidence-grounded clinical verification.
+
+#### 6. Conclusion
+
+We presented Med-TIV, an agentic RL framework for medical reasoning verification. Our approach addresses key limitations of existing medical reward models by offering explicit critique traces and enabling dynamic knowledge retrieval during verification. Empirical evaluations across four medical reasoning benchmarks demonstrate that Med-TIV substantially outperforms prior approaches. More broadly, Med-TIV introduces a general paradigm for training toolaugmented verifiers that can be extended to other highstakes domains requiring evidence-grounded evaluation.
+
+{8}------------------------------------------------
+
+# Impact Statement
+
+This paper introduces research aimed at improving the reliability of large language models for medical reasoning tasks. We believe our work contributes positively to the development of trustworthy medical AI systems by providing mechanisms to verify reasoning correctness before clinical deployment. Med-TIV holds potential to enhance the safety of LLM-assisted clinical decision support by reducing erroneous reasoning outputs through systematic verification. By grounding judgments in retrieved medical evidence, our approach offers improved transparency compared to opaque scalar reward models, enabling practitioners to better understand and audit verification decisions. The efficiency gains demonstrated by Med-TIV could democratize access to reliable medical reasoning verification, making robust verification feasible even in resource-constrained settings.
+
+# References
+
+- <span id="page-8-12"></span>Chen, J., Cai, Z., Ji, K., Wang, X., Liu, W., Wang, R., Hou, J., and Wang, B. Huatuogpt-o1, towards medical complex reasoning with llms, 2024. URL [https://](https://arxiv.org/abs/2412.18925) [arxiv.org/abs/2412.18925](https://arxiv.org/abs/2412.18925).
+- <span id="page-8-13"></span>Chen, Z., Cano, A. H., Romanou, A., Bonnet, A., Matoba, K., Salvi, F., Pagliardini, M., Fan, S., Kopf, A., Mo- ¨ htashami, A., Sallinen, A., Sakhaeirad, A., Swamy, V., Krawczuk, I., Bayazit, D., Marmet, A., Montariol, S., Hartley, M.-A., Jaggi, M., and Bosselut, A. Meditron-70b: Scaling medical pretraining for large language models, 2023. URL [https://arxiv.org/abs/2311.](https://arxiv.org/abs/2311.16079) [16079](https://arxiv.org/abs/2311.16079).
+- <span id="page-8-10"></span>Grattafiori, A. et al. The llama 3 herd of models, 2024. URL <https://arxiv.org/abs/2407.21783>.
+- <span id="page-8-9"></span>Guo, D. et al. Deepseek-r1 incentivizes reasoning in llms through reinforcement learning. Nature, 645(8081), 2025. doi: 10.1038/s41586-025-09422-z.
+- <span id="page-8-7"></span>Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., and Steinhardt, J. Measuring massive multitask language understanding, 2021. URL [https:](https://arxiv.org/abs/2009.03300) [//arxiv.org/abs/2009.03300](https://arxiv.org/abs/2009.03300).
+- <span id="page-8-0"></span>Ji, Y., Ma, W., Sivarajkumar, S., et al. Mitigating the risk of health inequity exacerbated by large language models. npj Digital Medicine, 8:246, 2025a. doi: 10.1038/s41746-025-01576-4.
+- <span id="page-8-2"></span>Ji, Y., Zhang, H., and Wang, Y. Bias evaluation and mitigation in retrieval-augmented medical question-answering systems, 2025b. URL [https://arxiv.org/abs/](https://arxiv.org/abs/2503.15454) [2503.15454](https://arxiv.org/abs/2503.15454).
+
+- <span id="page-8-8"></span>Jiang, D., Lu, Y., Li, Z., Lyu, Z., Nie, P., Wang, H., Su, A., Chen, H., Zou, K., Du, C., Pang, T., and Chen, W. Verltool: Towards holistic agentic reinforcement learning with tool use, 2025a. URL [https://arxiv.org/](https://arxiv.org/abs/2509.01055) [abs/2509.01055](https://arxiv.org/abs/2509.01055).
+- <span id="page-8-1"></span>Jiang, S., Liao, Y., Chen, Z., Zhang, Y., Wang, Y., and Wang, Y. Meds<sup>3</sup> : Towards medical slow thinking with self-evolved soft dual-sided process supervision, 2025b. URL <https://arxiv.org/abs/2501.12051>.
+- <span id="page-8-3"></span>Jin, B., Zeng, H., Yue, Z., Yoon, J., Arik, S., Wang, D., Zamani, H., and Han, J. Search-r1: Training llms to reason and leverage search engines with reinforcement learning, 2025. URL [https://arxiv.org/abs/](https://arxiv.org/abs/2503.09516) [2503.09516](https://arxiv.org/abs/2503.09516).
+- <span id="page-8-6"></span>Jin, D., Pan, E., Oufattole, N., Weng, W.-H., Fang, H., and Szolovits, P. What disease does this patient have? a large-scale open domain question answering dataset from medical exams, 2020. URL [https://arxiv.org/](https://arxiv.org/abs/2009.13081) [abs/2009.13081](https://arxiv.org/abs/2009.13081).
+- <span id="page-8-16"></span>Jin, Q., Kim, W., Chen, Q., Comeau, D. C., Yeganova, L., Wilbur, W. J., and Lu, Z. Medcpt: Contrastive pre-trained transformers with large-scale pubmed search logs for zero-shot biomedical information retrieval. Bioinformatics, 39(11), November 2023. ISSN 1367-4811. doi: 10.1093/bioinformatics/ btad651. URL [http://dx.doi.org/10.1093/](http://dx.doi.org/10.1093/bioinformatics/btad651) [bioinformatics/btad651](http://dx.doi.org/10.1093/bioinformatics/btad651).
+- <span id="page-8-5"></span>Khatri, D., Madaan, L., Tiwari, R., Bansal, R., Duvvuri, S. S., Zaheer, M., Dhillon, I. S., Brandfonbrener, D., and Agarwal, R. The art of scaling reinforcement learning compute for llms, 2025. URL [https://arxiv.org/](https://arxiv.org/abs/2510.13786) [abs/2510.13786](https://arxiv.org/abs/2510.13786).
+- <span id="page-8-15"></span>Kuang, P., Wang, X., Liu, W., Dong, J., and Xu, K. Timprm: Verifying multimodal reasoning with tool-integrated prm, 2025. URL [https://arxiv.org/abs/2511.](https://arxiv.org/abs/2511.22998) [22998](https://arxiv.org/abs/2511.22998).
+- <span id="page-8-14"></span>Li, L., Chai, Y., Wang, S., Sun, Y., Tian, H., Zhang, N., and Wu, H. Tool-augmented reward modeling, 2024. URL <https://arxiv.org/abs/2310.01045>.
+- <span id="page-8-11"></span>Liu, C., Wang, H., Pan, J., Wan, Z., Dai, Y., Lin, F., Bai, W., Rueckert, D., and Arcucci, R. Beyond distillation: Pushing the limits of medical llm reasoning with minimalist rule-based rl, 2025a. URL [https:](https://arxiv.org/abs/2505.17952) [//arxiv.org/abs/2505.17952](https://arxiv.org/abs/2505.17952).
+- <span id="page-8-4"></span>Liu, Z., Chen, C., Li, W., Qi, P., Pang, T., Du, C., Lee, W. S., and Lin, M. Understanding r1-zero-like training: A critical perspective, 2025b. URL [https://arxiv.](https://arxiv.org/abs/2503.20783) [org/abs/2503.20783](https://arxiv.org/abs/2503.20783).
+
+{9}------------------------------------------------
+
+<span id="page-9-4"></span>495 496 497 498 Liu, Z., Wang, P., Xu, R., Ma, S., Ruan, C., Li, P., Liu, Y., and Wu, Y. Inference-time scaling for generalist reward modeling, 2025c. URL [https://arxiv.org/abs/](https://arxiv.org/abs/2504.02495) [2504.02495](https://arxiv.org/abs/2504.02495).
+
+499 500 501
+
+<span id="page-9-10"></span><span id="page-9-8"></span>504
+
+<span id="page-9-16"></span>513 514
+
+<span id="page-9-3"></span>516
+
+518 519 520
+
+<span id="page-9-2"></span>524 525 526
+
+<span id="page-9-15"></span><span id="page-9-7"></span>528 529 530
+
+534
+
+<span id="page-9-5"></span><span id="page-9-0"></span>536
+
+- <span id="page-9-12"></span>OpenAI et al. Gpt-4o system card, 2024. URL [https:](https://arxiv.org/abs/2410.21276) [//arxiv.org/abs/2410.21276](https://arxiv.org/abs/2410.21276).
+  - Pal, A., Umapathi, L. K., and Sankarasubbu, M. Medmcqa : A large-scale multi-subject multi-choice dataset for medical domain question answering, 2022. URL [https:](https://arxiv.org/abs/2203.14371) [//arxiv.org/abs/2203.14371](https://arxiv.org/abs/2203.14371).
+  - Shi, W., Xu, R., Zhuang, Y., Yu, Y., Sun, H., Wu, H., Yang, C., and Wang, M. D. Medadapter: Efficient test-time adaptation of large language models towards medical reasoning, 2024. URL [https://arxiv.org/abs/](https://arxiv.org/abs/2405.03000) [2405.03000](https://arxiv.org/abs/2405.03000).
+  - Singhal, K., Tu, T., Gottweis, J., et al. Toward expertlevel medical question answering with large language models. Nature Medicine, 31:943–950, 2025. doi: 10. 1038/s41591-024-03423-7.
+  - Snell, C., Lee, J., Xu, K., and Kumar, A. Scaling llm testtime compute optimally can be more effective than scaling model parameters, 2024. URL [https://arxiv.](https://arxiv.org/abs/2408.03314) [org/abs/2408.03314](https://arxiv.org/abs/2408.03314).
+  - Wang, K., Fu, Z., Xin, W., Zhou, L., and Chandrappa, S. K. Digital voices of survival: From social media disclosures to support provisions for domestic violence victims. arXiv preprint arXiv:2509.12288, 2025.
+  - Wu, C., Lin, W., Zhang, X., Zhang, Y., Wang, Y., and Xie, W. Pmc-llama: Towards building open-source language models for medicine, 2023. URL [https://arxiv.](https://arxiv.org/abs/2304.14454) [org/abs/2304.14454](https://arxiv.org/abs/2304.14454).
+  - Xia, C., Wu, Q., Tian, S., and Hao, Y. Parallelism meets adaptiveness: Scalable documents understanding in multi-agent llm systems, 2025. URL [https:](https://arxiv.org/abs/2507.17061) [//arxiv.org/abs/2507.17061](https://arxiv.org/abs/2507.17061).
+  - Xiao, W., Lian, J. J., Ouyang, K., Gu, S., Ke, Z., Wei, D., Sha, X., Wang, J., Fu, S., Qiu, M., and Xu, C. Newton downhill optimizer with application to engineering optimization and breast cancer feature selection. Biomedical Signal Processing and Control, 117:109184, 2026. ISSN 1746-8094. doi: https://doi.org/10.1016/j.bspc.2025.109184. URL [https://www.sciencedirect.com/](https://www.sciencedirect.com/science/article/pii/S1746809425016957) [science/article/pii/S1746809425016957](https://www.sciencedirect.com/science/article/pii/S1746809425016957).
+  - Xiong, W., Zhao, W., Yuan, W., Golovneva, O., Zhang, T., Weston, J., and Sukhbaatar, S. Stepwiser: Stepwise generative judges for wiser reasoning, 2025. URL <https://arxiv.org/abs/2508.19229>.
+
+- <span id="page-9-9"></span>Xu, R., Chen, J., Ye, J., Wu, Y., Yan, J., Yang, C., and Yu, H. Incentivizing agentic reasoning in llm judges via tool-integrated reinforcement learning, 2025. URL <https://arxiv.org/abs/2510.23038>.
+- <span id="page-9-13"></span>Yang, A. et al. Qwen2.5 technical report, 2025.
+- <span id="page-9-17"></span>You, R., Cai, H., Zhang, C., Xu, Q., Liu, M., Yu, T., Li, Y., and Li, W. Agent-as-a-judge, 2026. URL [https:](https://arxiv.org/abs/2601.05111) [//arxiv.org/abs/2601.05111](https://arxiv.org/abs/2601.05111).
+- <span id="page-9-6"></span>Yun, J., Sohn, J., Park, J., Kim, H., Tang, X., Shao, Y., Koo, Y., Ko, M., Chen, Q., Gerstein, M., Moor, M., and Kang, J. Med-prm: Medical reasoning models with stepwise, guideline-verified process rewards, 2025. URL [https:](https://arxiv.org/abs/2506.11474) [//arxiv.org/abs/2506.11474](https://arxiv.org/abs/2506.11474).
+- <span id="page-9-1"></span>Zhang, H., Lou, Q., and Wang, Y. Towards safe ai clinicians: A comprehensive study on large language model jailbreaking in healthcare, 2025. URL [https:](https://arxiv.org/abs/2501.18632) [//arxiv.org/abs/2501.18632](https://arxiv.org/abs/2501.18632).
+- <span id="page-9-14"></span>Zhang, K., Zeng, S., Hua, E., Ding, N., Chen, Z.-R., Ma, Z., Li, H., Cui, G., Qi, B., Zhu, X., Lv, X., Jinfang, H., Liu, Z., and Zhou, B. Ultramedical: Building specialized generalists in biomedicine, 2024. URL <https://arxiv.org/abs/2406.03949>.
+- <span id="page-9-18"></span>Zhao, X., Liu, S., Yang, S.-Y., and Miao, C. Medrag: Enhancing retrieval-augmented generation with knowledge graph-elicited reasoning for healthcare copilot, 2025. URL <https://arxiv.org/abs/2502.04413>.
+- <span id="page-9-11"></span>Zuo, Y., Qu, S., Li, Y., Chen, Z., Zhu, X., Hua, E., Zhang, K., Ding, N., and Zhou, B. Medxpertqa: Benchmarking expert-level medical reasoning and understanding, 2025. URL <https://arxiv.org/abs/2501.18362>.
+
+{10}------------------------------------------------
+
+# A. Limitation
+
+550
+
+554
+
+556
+
+574
+
+576
+
+594
+
+596
+
+While Med-TIV demonstrates substantial improvements over existing medical reasoning verification approaches, several limitations warrant discussion and suggest directions for future research.
+
+Process Supervision. Our current training paradigm relies solely on trace-level outcome rewards, providing no supervision on intermediate verification behaviors such as when to search, what queries to formulate, or how to integrate retrieved evidence. While this design eliminates the need for costly step-level annotations, it may lead to suboptimal search patterns or redundant retrieval operations. Future work could explore supervision for the verification task itself, or leverage techniques such as search behavior cloning from stronger models to provide denser optimization signals.
+
+Retrieval Corpus Coverage. Med-TIV's verification accuracy is inherently bounded by the coverage and quality of the underlying medical corpus. Our retrieval system indexes documents from PubMed abstracts and medical textbooks, which provides broad coverage of established medical knowledge but may lack recent findings, rare disease information, or region-specific clinical guidelines. Verification of reasoning traces involving cutting-edge treatments or highly specialized subspecialties may be limited by corpus gaps.
+
+Language and Domain Scope. All training and evaluation are conducted on English-language medical reasoning benchmarks. The generalization of Med-TIV to multilingual medical content or non-Western medical traditions remains unexplored. Additionally, while our benchmarks span multiple medical subdomains, certain specialized areas such as genomics, radiology interpretation, and surgical planning may require domain-adapted retrieval corpora for optimal verification performance.
+
+# B. Additional Implementation Details
+
+#### <span id="page-10-1"></span>B.1. Hyperparameter Settings
+
+<span id="page-10-2"></span>Table [4](#page-10-2) provides comprehensive hyperparameter configurations for Med-TIV training across both iterations. We maintain mostly consistent settings between iterations to isolate the effect of iterative training from hyperparameter tuning.
+
+| Hyperparameters              | Iteration 1 | Iteration 2 |
+|------------------------------|-------------|-------------|
+| RL Algorithm                 | Dr.GRPO     | Dr.GRPO     |
+| Clip ratio (low / high)      | 0.2 / 0.3   | 0.2 / 0.3   |
+| Learning rate                | 1e-6        | 1e-6        |
+| Warmup steps                 | 10          | 10          |
+| Training epochs              | 5           | 5           |
+| Global batch size            | 256         | 256         |
+| Mini-batch size              | 256         | 256         |
+| Group size (G)               | 5           | 8           |
+| Rollout sampling temperature | 1.0         | 1.0         |
+| Rollout top-p                | 0.95        | 0.95        |
+| Curriculum filtering         | Enabled     | Enabled     |
+
+*Table 4.* Hyperparameter configurations for Med-TIV training across iterations.
+
+#### <span id="page-10-0"></span>B.2. Retrieval Setup
+
+We construct our retrieval infrastructure using a dense retrieval architecture optimized for medical domain queries. The corpus is derived from the MedRAG [\(Zhao et al.,](#page-9-18) [2025\)](#page-9-18) collection, specifically combining the PubMed and Textbooks subcorpora into a unified index. The PubMed subset contains approximately 23.9 million biomedical abstracts covering research publications, while the Textbooks subset includes content from standard medical textbooks spanning clinical medicine, pharmacology, pathology, and related disciplines. After deduplication and quality filtering, the combined corpus contains approximately 24 million snippets.
+
+We employ MedCPT [\(Jin et al.,](#page-8-16) [2023\)](#page-8-16) as our dense retrieval encoder, specifically the query encoder variant for encoding
+
+{11}------------------------------------------------
+
+*Table 5.* Prompt template.
+
+606 607
+
+<span id="page-11-0"></span>605
+
+### User Prompt:
+
+You are a reasoning validator for medical problems. Your task is to think step by step and evaluate whether the given reasoning trace of a medical problem contains errors.
+
+First, you must always perform a step-by-step analysis to examine the entire reasoning process. Then, based on your analysis, you will make a definitive judgment.
+
+- Use 1 if the reasoning trace is free of errors.
+- Use 0 if the reasoning trace contains one or more errors.
+
+#### Output Instruction:
+
+You must conduct your step-by-step analysis inside <think> and </think> first every time you get new information. After reasoning, if you find you lack some knowledge, you can call a search engine by <search> query </search> and it will return the top searched results between <information> and </information>. You can search as many times as you want. If you find no further external knowledge needed, you can directly provide the answer inside <answer> and </answer>, without detailed illustrations.
+
+```
+Medical Problem:
+{The full Medical Problem on one or more lines.}
+Reasoning Trace:
+{The full Reasoning Trace on one or more lines.}
+```
+
+630 631 search queries and article encoder for encoding corpus snippets. Document embeddings are pre-computed and stored in a FAISS index using the Flat configuration for maximum retrieval accuracy, distributed across multiple GPUs using FAISS's GPU sharding capability to enable parallel similarity search. For each search query, we retrieve the top-3 most relevant documents for both training and inference.
+
+632 633 634
+
+### <span id="page-11-2"></span>B.3. Baseline Setup
+
+639
+
+We describe the configuration of reward model baselines used in our experiments. For Med-PRM, which employs static retrieval-augmented generation, we equip it with the same retrieval corpus, encoder, and top-k setting as our framework to ensure a controlled comparison. MedS<sup>3</sup> does not support external tool invocation and is therefore evaluated without retrieval augmentation. For confidence score extraction and inference hyperparameter settings, we follow the configurations specified in each baseline's original publication.
+
+640 641 642
+
+#### B.4. Prompt Template
+
+643 644 645
+
+We design a structured prompt template that guides the verifier through systematic reasoning with explicit tool invocation syntax. The complete prompt is shown in Table [5.](#page-11-0)
+
+646 647
+
+# <span id="page-11-1"></span>C. Benchmarks and Baselines
+
+648 649 650
+
+### C.1. Benchmarks
+
+We evaluate Med-TIV on four established medical reasoning benchmarks that collectively assess verification capability across varying difficulty levels and medical subdomains.
+
+651 652 653
+
+• MedQA [\(Jin et al.,](#page-8-6) [2020\)](#page-8-6): A dataset of multiple-choice questions derived from the United States Medical Licensing Examination (USMLE), designed to evaluate clinical reasoning and medical knowledge integration across diverse specialties.
+
+654 655 656
+
+• MedMCQA [\(Pal et al.,](#page-9-10) [2022\)](#page-9-10): A large-scale multi-subject benchmark sourced from Indian medical entrance examinations (AIIMS and NEET-PG), covering 21 medical subjects with emphasis on factual knowledge and clinical application.
+
+657 658 659
+
+• MMLU-Med [\(Hendrycks et al.,](#page-8-7) [2021\)](#page-8-7): An aggregation of medical-related subsets from the Massive Multitask Language Understanding benchmark, encompassing anatomy, clinical knowledge, college biology, college medicine, medical 
+
+{12}------------------------------------------------
+
+genetics, and professional medicine.
+
+• MedXpertQA [\(Zuo et al.,](#page-9-11) [2025\)](#page-9-11): An expert-level benchmark featuring challenging questions that require multi-step clinical reasoning, differential diagnosis, and treatment planning at the level expected of practicing physicians.
+
+#### C.2. Baselines
+
+We compare Med-TIV against comprehensive baselines spanning proprietary systems, general-purpose models, and domainspecialized approaches.
+
+### Proprietary Models.
+
+- GPT-4o-mini [\(OpenAI et al.,](#page-9-12) [2024\)](#page-9-12): A compact variant of OpenAI's GPT-4o optimized for efficiency while maintaining strong reasoning capabilities across diverse tasks.
+- Gemini-2.0-Flash: Google's efficient multimodal model designed for fast inference with competitive performance on knowledge-intensive benchmarks.
+
+#### General Reasoning Models.
+
+- DeepSeek-R1 [\(Guo et al.,](#page-8-9) [2025\)](#page-8-9): A 671B parameter reasoning model trained with RL, representing the current frontier of open-weight reasoning capabilities.
+- R1-Distill-Qwen / R1-Distill-Llama: Distilled variants of DeepSeek-R1 at 7B and 8B scales respectively, designed to transfer reasoning capabilities to smaller architectures.
+
+#### General Foundation Models.
+
+- Qwen2.5 [\(Yang et al.,](#page-9-13) [2025\)](#page-9-13): A family of open-weight language models with strong multilingual and reasoning capabilities, evaluated at 7B and 32B parameter scales.
+- Llama3.1 [\(Grattafiori et al.,](#page-8-10) [2024\)](#page-8-10): Meta's open-source foundation model demonstrating competitive performance across diverse benchmarks, evaluated at the 8B scale.
+
+## Medical Domain Models.
+
+- AlphaMed [\(Liu et al.,](#page-8-11) [2025a\)](#page-8-11): A medical reasoning model that employs RL with rule-based rewards to enhance clinical reasoning without reliance on distillation from larger models.
+- UltraMedical [\(Zhang et al.,](#page-9-14) [2024\)](#page-9-14): A specialized medical model combining high-quality instruction tuning on curated biomedical corpora with preference optimization for improved clinical accuracy.
+- HuatuoGPT-o1 [\(Chen et al.,](#page-8-12) [2024\)](#page-8-12): A medical reasoning model incorporating chain-of-thought reasoning with internal verification mechanisms to improve diagnostic accuracy.
+
+#### Medical Reward Models.
+
+- MedS<sup>3</sup> [\(Jiang et al.,](#page-8-1) [2025b\)](#page-8-1): A self-evolved soft dual-sided process supervision framework for medical reasoning that generates training signals through iterative self-improvement without external annotations.
+- Med-PRM [\(Yun et al.,](#page-9-6) [2025\)](#page-9-6): A process reward model for medical reasoning verification that provides step-level supervision using static retrieval-augmented generation with guideline-based verification.
+
+# <span id="page-12-0"></span>D. Case Analysis
+
+704
+
+706
+
+713 714 Table [6](#page-13-0) presents a complete verification example illustrating how a Med-TIV trained verifier identifies reasoning errors through dynamic evidence retrieval. The case involves a patient with bladder cancer who develops ototoxicity following chemotherapy. The generator's reasoning trace incorrectly attributes the symptoms to taxanes based on their known association with ototoxicity, concluding with answer (B). However, the model retrieves evidence establishing that cisplatin—the standard neoadjuvant therapy for transitional cell carcinoma—is the causative agent, and its mechanism involves DNA cross-linking rather than microtubule hyperstabilization. Through iterative search and reasoning, verifier correctly identifies the error, demonstrating the value of tool augmentation for catching subtle medical reasoning mistakes.
+
+{13}------------------------------------------------
+
+769
+
+<span id="page-13-0"></span>*Table 6.* Complete verification demonstration. Given a medical problem and a reasoning trace, the verifier retrieves relevant evidence and correctly identifies the reasoning error within the trace.
+
+#### Medical Problem:
+
+Question: A 67-year-old man with transitional cell carcinoma of the bladder comes to the physician because of a 2-day history of ringing sensation in his ear. He received this first course of neoadjuvant chemotherapy 1 week ago. Pure tone audiometry shows a sensorineural hearing loss of 45 dB. The expected beneficial effect of the drug that caused this patient's symptoms is most likely due to which of the following actions?
+
+Options: (A) Inhibition of proteasome (B) Hyperstabilization of microtubules (C) Generation of free radicals (D) Cross-linking of DNA
+
+### Generator Reasoning Trace:
+
+The patient's presentation of ringing in the ear, known as tinnitus, and a sensorineural hearing loss of 45 dB, suggest ototoxicity. Ototoxicity is a well-known side effect of certain chemotherapy drugs. Neoadjuvant chemotherapy refers to the use of chemotherapy before other treatments such as surgery. The patient has received a course of neoadjuvant chemotherapy 1 week ago, which aligns with the recent onset of ototoxic symptoms. To determine the cause of ototoxicity from the provided options, we need to identify which of the options is likely to cause sensorineural hearing loss. Option (A), Inhibition of proteasome, is associated with drugs like bortezomib (Velcade), which are less commonly associated with ototoxicity. Option (B), Hyperstabilization of microtubules, is a known mechanism of taxanes (e.g., paclitaxel, docetaxel). Taxanes are known to cause ototoxicity, making this a potential candidate. Option (C), Generation of free radicals, is a general mechanism that can cause cellular damage but is not specifically known to be associated with ototoxicity. Option (D), Cross-linking of DNA, is a known mechanism of alkylating agents like cyclophosphamide. While alkylating agents can cause ototoxicity, this mechanism is not as commonly linked to ototoxicity as other options. Considering the options, the one most commonly associated with ototoxicity is hyperstabilization of microtubules. Therefore, the drug that most likely caused this patient's symptoms is one that hyperstabilizes microtubules. The answer is (B).
+
+# Verifier Judgment:
+
+<think> Let's check the common chemotherapeutic agents used for bladder cancer and their known side effects. Transitional cell carcinoma of the bladder is often treated with cisplatin-based chemotherapy as neoadjuvant therapy. We need to determine if any of the options match the known side effects of these drugs. Let's search to confirm the side effects of cisplatin and other common bladder cancer chemotherapy drugs. </think>
+
+- <search> common side effects of cisplatin chemotherapy </search>
+- <information> [Search Results regarding Cisplatin toxicity and mechanisms...] </information>
+- <think> From the information, cisplatin is known to cause ototoxicity, which aligns with the patient's symptoms of sensorineural hearing loss. Now let's check the mechanisms of action for the options: Option (A): Inhibition of proteasome - This is associated with drugs like bortezomib and is not used for bladder cancer. Option (B): Hyperstabilization of microtubules - This is the mechanism of taxanes (e.g., paclitaxel, docetaxel). Option (C): Generation of free radicals - This is not specific to a single mechanism and is not commonly associated with ototoxicity. Option (D): Cross-linking of DNA - This is the mechanism of alkylating agents (e.g., cyclophosphamide). Cisplatin is an alkylating agent, and its mechanism involves cross-linking of DNA. Therefore, the correct answer should be (D). </think>
+
+<answer>0</answer>

icml26/19e76363-53a6-4f3c-8b50-844e1aea4e26/parse_report.json

@@ -0,0 +1,75 @@
+{
+  "paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26",
+  "pipeline": "marker_non_llm_v3",
+  "parser": "marker_single",
+  "formats": [
+    "markdown",
+    "chunks"
+  ],
+  "llm_enabled": false,
+  "pdf_path": "/network/scratch/j/jianan.zhao/ReviewAgent/data/processed_papers/icml26_20260429_1952_duequeue/raw/19e76363-53a6-4f3c-8b50-844e1aea4e26.pdf",
+  "pdf_sha256": "165b769dbdab634a8ec57eb3aa702459c749ee10eaacb5db44dc41e22c4a793b",
+  "bytes": 749273,
+  "source": "https://koala.science/storage/pdfs/19e76363-53a6-4f3c-8b50-844e1aea4e26.pdf",
+  "page_count": 14,
+  "ok": true,
+  "elapsed_seconds": 151.2,
+  "paper2markdown_v3": {
+    "paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26",
+    "pipeline": "Paper2Markdown-V3",
+    "ok": true,
+    "page_count": 14,
+    "chunk_count": 137,
+    "main_body_chunk_count": 95,
+    "appendix_chunk_count": 37,
+    "reference_chunk_count": 5,
+    "model_text_chars": 39229,
+    "raw_markdown_chars": 64704,
+    "sanitized_chars": 58960,
+    "page_provenance": {
+      "min_page": 1,
+      "max_page": 14,
+      "invalid_count": 0
+    },
+    "marker_block_type_counts": {
+      "Caption": 1,
+      "Code": 3,
+      "Equation": 12,
+      "FigureGroup": 3,
+      "Footnote": 3,
+      "ListGroup": 12,
+      "PageFooter": 14,
+      "PageHeader": 17,
+      "SectionHeader": 39,
+      "TableGroup": 4,
+      "Text": 352
+    },
+    "asset_count_raw": 3,
+    "asset_count_model_kept": 3,
+    "asset_count_rejected": 0,
+    "asset_reject_reasons": {
+      "kept": 3
+    },
+    "artifact_leak_audit": {
+      "ok": true,
+      "hits": {
+        "Anonymous Authors": [],
+        "ACKNOWLEDGMENT": [],
+        "OpenReview": [],
+        "\"accept_label\"": [],
+        "\"decision\"": [],
+        "\"decision_tier\"": [],
+        "\"source_status\"": [],
+        "Meta-review": [],
+        "Official Review": [],
+        "official_reviews": [],
+        "meta_reviews": [],
+        "suggested_verdict_score": []
+      },
+      "artifact_count": 2
+    },
+    "default_model_input": "model_text_v3.txt",
+    "appendix_input": "appendix_text_v3.txt",
+    "reference_input": "reference_text_v3.txt"
+  }
+}

icml26/19e76363-53a6-4f3c-8b50-844e1aea4e26/reference_chunks.jsonl

@@ -0,0 +1,5 @@
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0095", "section": "References", "page_start": 9, "page_end": 9, "type": "ListGroup", "text": "Chen, J., Cai, Z., Ji, K., Wang, X., Liu, W., Wang, R., Hou, J., and Wang, B. Huatuogpt-o1, towards medical complex reasoning with llms, 2024. URL https:// arxiv.org/abs/2412.18925 . Chen, Z., Cano, A. H., Romanou, A., Bonnet, A., Matoba, K., Salvi, F., Pagliardini, M., Fan, S., Kopf, A., Mo- ¨ htashami, A., Sallinen, A., Sakhaeirad, A., Swamy, V., Krawczuk, I., Bayazit, D., Marmet, A., Montariol, S., Hartley, M.-A., Jaggi, M., and Bosselut, A. Meditron-70b: Scaling medical pretraining for large language models, 2023. URL 16079 . Grattafiori, A. et al. The llama 3 herd of models, 2024. URL . Guo, D. et al. Deepseek-r1 incentivizes reasoning in llms through reinforcement learning. Nature, 645(8081), 2025. doi: 10.1038/s41586-025-09422-z. Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., and Steinhardt, J. Measuring massive multitask language understanding, 2021. URL https: //arxiv.org/abs/2009.03300 . Ji, Y., Ma, W., Sivarajkumar, S., et al. Mitigating the risk of health inequity exacerbated by large language models. npj Digital Medicine, 8:246, 2025a. doi: 10.1038/s41746-025-01576-4. Ji, Y., Zhang, H., and Wang, Y. Bias evaluation and mitigation in retrieval-augmented medical question-answering systems, 2025b. URL 2503.15454 .", "source": "marker_v2", "marker_block_id": "/page/8/ListGroup/529"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0096", "section": "References", "page_start": 9, "page_end": 9, "type": "ListGroup", "text": "Jiang, D., Lu, Y., Li, Z., Lyu, Z., Nie, P., Wang, H., Su, A., Chen, H., Zou, K., Du, C., Pang, T., and Chen, W. Verltool: Towards holistic agentic reinforcement learning with tool use, 2025a. URL abs/2509.01055 . Jiang, S., Liao, Y., Chen, Z., Zhang, Y., Wang, Y., and Wang, Y. Meds 3 : Towards medical slow thinking with self-evolved soft dual-sided process supervision, 2025b. URL . Jin, B., Zeng, H., Yue, Z., Yoon, J., Arik, S., Wang, D., Zamani, H., and Han, J. Search-r1: Training llms to reason and leverage search engines with reinforcement learning, 2025. URL 2503.09516 . Jin, D., Pan, E., Oufattole, N., Weng, W.-H., Fang, H., and Szolovits, P. What disease does this patient have? a large-scale open domain question answering dataset from medical exams, 2020. URL abs/2009.13081 . Jin, Q., Kim, W., Chen, Q., Comeau, D. C., Yeganova, L., Wilbur, W. J., and Lu, Z. Medcpt: Contrastive pre-trained transformers with large-scale pubmed search logs for zero-shot biomedical information retrieval. Bioinformatics, 39(11), November 2023. ISSN 1367-4811. doi: 10.1093/bioinformatics/ btad651. URL bioinformatics/btad651 . Khatri, D., Madaan, L., Tiwari, R., Bansal, R., Duvvuri, S. S., Zaheer, M., Dhillon, I. S., Brandfonbrener, D., and Agarwal, R. The art of scaling reinforcement learning compute for llms, 2025. URL abs/2510.13786 . Kuang, P., Wang, X., Liu, W., Dong, J., and Xu, K. Timprm: Verifying multimodal reasoning with tool-integrated prm, 2025. URL 22998 . Li, L., Chai, Y., Wang, S., Sun, Y., Tian, H., Zhang, N., and Wu, H. Tool-augmented reward modeling, 2024. URL . Liu, C., Wang, H., Pan, J., Wan, Z., Dai, Y., Lin, F., Bai, W., Rueckert, D., and Arcucci, R. Beyond distillation: Pushing the limits of medical llm reasoning with minimalist rule-based rl, 2025a. URL https: //arxiv.org/abs/2505.17952 . Liu, Z., Chen, C., Li, W., Qi, P., Pang, T., Du, C., Lee, W. S., and Lin, M. Understanding r1-zero-like training: A critical perspective, 2025b. URL org/abs/2503.20783 .", "source": "marker_v2", "marker_block_id": "/page/8/ListGroup/530"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0097", "section": "References", "page_start": 10, "page_end": 10, "type": "Text", "text": "495 496 497 498 Liu, Z., Wang, P., Xu, R., Ma, S., Ruan, C., Li, P., Liu, Y., and Wu, Y. Inference-time scaling for generalist reward modeling, 2025c. URL 2504.02495 .", "source": "marker_v2", "marker_block_id": "/page/9/Text/441"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0098", "section": "References", "page_start": 10, "page_end": 10, "type": "ListGroup", "text": "OpenAI et al. Gpt-4o system card, 2024. URL https: //arxiv.org/abs/2410.21276 . Pal, A., Umapathi, L. K., and Sankarasubbu, M. Medmcqa : A large-scale multi-subject multi-choice dataset for medical domain question answering, 2022. URL https: //arxiv.org/abs/2203.14371 . Shi, W., Xu, R., Zhuang, Y., Yu, Y., Sun, H., Wu, H., Yang, C., and Wang, M. D. Medadapter: Efficient test-time adaptation of large language models towards medical reasoning, 2024. URL 2405.03000 . Singhal, K., Tu, T., Gottweis, J., et al. Toward expertlevel medical question answering with large language models. Nature Medicine, 31:943–950, 2025. doi: 10. 1038/s41591-024-03423-7. Snell, C., Lee, J., Xu, K., and Kumar, A. Scaling llm testtime compute optimally can be more effective than scaling model parameters, 2024. URL org/abs/2408.03314 . Wang, K., Fu, Z., Xin, W., Zhou, L., and Chandrappa, S. K. Digital voices of survival: From social media disclosures to support provisions for domestic violence victims. arXiv preprint arXiv:2509.12288, 2025. Wu, C., Lin, W., Zhang, X., Zhang, Y., Wang, Y., and Xie, W. Pmc-llama: Towards building open-source language models for medicine, 2023. URL org/abs/2304.14454 . Xia, C., Wu, Q., Tian, S., and Hao, Y. Parallelism meets adaptiveness: Scalable documents understanding in multi-agent llm systems, 2025. URL https: //arxiv.org/abs/2507.17061 . Xiao, W., Lian, J. J., Ouyang, K., Gu, S., Ke, Z., Wei, D., Sha, X., Wang, J., Fu, S., Qiu, M., and Xu, C. Newton downhill optimizer with application to engineering optimization and breast cancer feature selection. Biomedical Signal Processing and Control, 117:109184, 2026. ISSN 1746-8094. doi: URL science/article/pii/S1746809425016957 . Xiong, W., Zhao, W., Yuan, W., Golovneva, O., Zhang, T., Weston, J., and Sukhbaatar, S. Stepwiser: Stepwise generative judges for wiser reasoning, 2025. URL .", "source": "marker_v2", "marker_block_id": "/page/9/ListGroup/439"}
+{"paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26", "chunk_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26:0099", "section": "References", "page_start": 10, "page_end": 10, "type": "ListGroup", "text": "Xu, R., Chen, J., Ye, J., Wu, Y., Yan, J., Yang, C., and Yu, H. Incentivizing agentic reasoning in llm judges via tool-integrated reinforcement learning, 2025. URL . Yang, A. et al. Qwen2.5 technical report, 2025. You, R., Cai, H., Zhang, C., Xu, Q., Liu, M., Yu, T., Li, Y., and Li, W. Agent-as-a-judge, 2026. URL https: //arxiv.org/abs/2601.05111 . Yun, J., Sohn, J., Park, J., Kim, H., Tang, X., Shao, Y., Koo, Y., Ko, M., Chen, Q., Gerstein, M., Moor, M., and Kang, J. Med-prm: Medical reasoning models with stepwise, guideline-verified process rewards, 2025. URL https: //arxiv.org/abs/2506.11474 . Zhang, H., Lou, Q., and Wang, Y. Towards safe ai clinicians: A comprehensive study on large language model jailbreaking in healthcare, 2025. URL https: //arxiv.org/abs/2501.18632 . Zhang, K., Zeng, S., Hua, E., Ding, N., Chen, Z.-R., Ma, Z., Li, H., Cui, G., Qi, B., Zhu, X., Lv, X., Jinfang, H., Liu, Z., and Zhou, B. Ultramedical: Building specialized generalists in biomedicine, 2024. URL . Zhao, X., Liu, S., Yang, S.-Y., and Miao, C. Medrag: Enhancing retrieval-augmented generation with knowledge graph-elicited reasoning for healthcare copilot, 2025. URL . Zuo, Y., Qu, S., Li, Y., Chen, Z., Zhu, X., Hua, E., Zhang, K., Ding, N., and Zhou, B. Medxpertqa: Benchmarking expert-level medical reasoning and understanding, 2025. URL .", "source": "marker_v2", "marker_block_id": "/page/9/ListGroup/440"}

icml26/19e76363-53a6-4f3c-8b50-844e1aea4e26/reference_text_v3.txt

@@ -0,0 +1,14 @@
+[p. 9 | section: References | type: ListGroup]
+Chen, J., Cai, Z., Ji, K., Wang, X., Liu, W., Wang, R., Hou, J., and Wang, B. Huatuogpt-o1, towards medical complex reasoning with llms, 2024. URL https:// arxiv.org/abs/2412.18925 . Chen, Z., Cano, A. H., Romanou, A., Bonnet, A., Matoba, K., Salvi, F., Pagliardini, M., Fan, S., Kopf, A., Mo- ¨ htashami, A., Sallinen, A., Sakhaeirad, A., Swamy, V., Krawczuk, I., Bayazit, D., Marmet, A., Montariol, S., Hartley, M.-A., Jaggi, M., and Bosselut, A. Meditron-70b: Scaling medical pretraining for large language models, 2023. URL 16079 . Grattafiori, A. et al. The llama 3 herd of models, 2024. URL . Guo, D. et al. Deepseek-r1 incentivizes reasoning in llms through reinforcement learning. Nature, 645(8081), 2025. doi: 10.1038/s41586-025-09422-z. Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., and Steinhardt, J. Measuring massive multitask language understanding, 2021. URL https: //arxiv.org/abs/2009.03300 . Ji, Y., Ma, W., Sivarajkumar, S., et al. Mitigating the risk of health inequity exacerbated by large language models. npj Digital Medicine, 8:246, 2025a. doi: 10.1038/s41746-025-01576-4. Ji, Y., Zhang, H., and Wang, Y. Bias evaluation and mitigation in retrieval-augmented medical question-answering systems, 2025b. URL 2503.15454 .
+
+[p. 9 | section: References | type: ListGroup]
+Jiang, D., Lu, Y., Li, Z., Lyu, Z., Nie, P., Wang, H., Su, A., Chen, H., Zou, K., Du, C., Pang, T., and Chen, W. Verltool: Towards holistic agentic reinforcement learning with tool use, 2025a. URL abs/2509.01055 . Jiang, S., Liao, Y., Chen, Z., Zhang, Y., Wang, Y., and Wang, Y. Meds 3 : Towards medical slow thinking with self-evolved soft dual-sided process supervision, 2025b. URL . Jin, B., Zeng, H., Yue, Z., Yoon, J., Arik, S., Wang, D., Zamani, H., and Han, J. Search-r1: Training llms to reason and leverage search engines with reinforcement learning, 2025. URL 2503.09516 . Jin, D., Pan, E., Oufattole, N., Weng, W.-H., Fang, H., and Szolovits, P. What disease does this patient have? a large-scale open domain question answering dataset from medical exams, 2020. URL abs/2009.13081 . Jin, Q., Kim, W., Chen, Q., Comeau, D. C., Yeganova, L., Wilbur, W. J., and Lu, Z. Medcpt: Contrastive pre-trained transformers with large-scale pubmed search logs for zero-shot biomedical information retrieval. Bioinformatics, 39(11), November 2023. ISSN 1367-4811. doi: 10.1093/bioinformatics/ btad651. URL bioinformatics/btad651 . Khatri, D., Madaan, L., Tiwari, R., Bansal, R., Duvvuri, S. S., Zaheer, M., Dhillon, I. S., Brandfonbrener, D., and Agarwal, R. The art of scaling reinforcement learning compute for llms, 2025. URL abs/2510.13786 . Kuang, P., Wang, X., Liu, W., Dong, J., and Xu, K. Timprm: Verifying multimodal reasoning with tool-integrated prm, 2025. URL 22998 . Li, L., Chai, Y., Wang, S., Sun, Y., Tian, H., Zhang, N., and Wu, H. Tool-augmented reward modeling, 2024. URL . Liu, C., Wang, H., Pan, J., Wan, Z., Dai, Y., Lin, F., Bai, W., Rueckert, D., and Arcucci, R. Beyond distillation: Pushing the limits of medical llm reasoning with minimalist rule-based rl, 2025a. URL https: //arxiv.org/abs/2505.17952 . Liu, Z., Chen, C., Li, W., Qi, P., Pang, T., Du, C., Lee, W. S., and Lin, M. Understanding r1-zero-like training: A critical perspective, 2025b. URL org/abs/2503.20783 .
+
+[p. 10 | section: References | type: Text]
+495 496 497 498 Liu, Z., Wang, P., Xu, R., Ma, S., Ruan, C., Li, P., Liu, Y., and Wu, Y. Inference-time scaling for generalist reward modeling, 2025c. URL 2504.02495 .
+
+[p. 10 | section: References | type: ListGroup]
+OpenAI et al. Gpt-4o system card, 2024. URL https: //arxiv.org/abs/2410.21276 . Pal, A., Umapathi, L. K., and Sankarasubbu, M. Medmcqa : A large-scale multi-subject multi-choice dataset for medical domain question answering, 2022. URL https: //arxiv.org/abs/2203.14371 . Shi, W., Xu, R., Zhuang, Y., Yu, Y., Sun, H., Wu, H., Yang, C., and Wang, M. D. Medadapter: Efficient test-time adaptation of large language models towards medical reasoning, 2024. URL 2405.03000 . Singhal, K., Tu, T., Gottweis, J., et al. Toward expertlevel medical question answering with large language models. Nature Medicine, 31:943–950, 2025. doi: 10. 1038/s41591-024-03423-7. Snell, C., Lee, J., Xu, K., and Kumar, A. Scaling llm testtime compute optimally can be more effective than scaling model parameters, 2024. URL org/abs/2408.03314 . Wang, K., Fu, Z., Xin, W., Zhou, L., and Chandrappa, S. K. Digital voices of survival: From social media disclosures to support provisions for domestic violence victims. arXiv preprint arXiv:2509.12288, 2025. Wu, C., Lin, W., Zhang, X., Zhang, Y., Wang, Y., and Xie, W. Pmc-llama: Towards building open-source language models for medicine, 2023. URL org/abs/2304.14454 . Xia, C., Wu, Q., Tian, S., and Hao, Y. Parallelism meets adaptiveness: Scalable documents understanding in multi-agent llm systems, 2025. URL https: //arxiv.org/abs/2507.17061 . Xiao, W., Lian, J. J., Ouyang, K., Gu, S., Ke, Z., Wei, D., Sha, X., Wang, J., Fu, S., Qiu, M., and Xu, C. Newton downhill optimizer with application to engineering optimization and breast cancer feature selection. Biomedical Signal Processing and Control, 117:109184, 2026. ISSN 1746-8094. doi: URL science/article/pii/S1746809425016957 . Xiong, W., Zhao, W., Yuan, W., Golovneva, O., Zhang, T., Weston, J., and Sukhbaatar, S. Stepwiser: Stepwise generative judges for wiser reasoning, 2025. URL .
+
+[p. 10 | section: References | type: ListGroup]
+Xu, R., Chen, J., Ye, J., Wu, Y., Yan, J., Yang, C., and Yu, H. Incentivizing agentic reasoning in llm judges via tool-integrated reinforcement learning, 2025. URL . Yang, A. et al. Qwen2.5 technical report, 2025. You, R., Cai, H., Zhang, C., Xu, Q., Liu, M., Yu, T., Li, Y., and Li, W. Agent-as-a-judge, 2026. URL https: //arxiv.org/abs/2601.05111 . Yun, J., Sohn, J., Park, J., Kim, H., Tang, X., Shao, Y., Koo, Y., Ko, M., Chen, Q., Gerstein, M., Moor, M., and Kang, J. Med-prm: Medical reasoning models with stepwise, guideline-verified process rewards, 2025. URL https: //arxiv.org/abs/2506.11474 . Zhang, H., Lou, Q., and Wang, Y. Towards safe ai clinicians: A comprehensive study on large language model jailbreaking in healthcare, 2025. URL https: //arxiv.org/abs/2501.18632 . Zhang, K., Zeng, S., Hua, E., Ding, N., Chen, Z.-R., Ma, Z., Li, H., Cui, G., Qi, B., Zhu, X., Lv, X., Jinfang, H., Liu, Z., and Zhou, B. Ultramedical: Building specialized generalists in biomedicine, 2024. URL . Zhao, X., Liu, S., Yang, S.-Y., and Miao, C. Medrag: Enhancing retrieval-augmented generation with knowledge graph-elicited reasoning for healthcare copilot, 2025. URL . Zuo, Y., Qu, S., Li, Y., Chen, Z., Zhu, X., Hua, E., Zhang, K., Ding, N., and Zhou, B. Medxpertqa: Benchmarking expert-level medical reasoning and understanding, 2025. URL .

icml26/19e76363-53a6-4f3c-8b50-844e1aea4e26/sanitization_report.json

@@ -0,0 +1,58 @@
+{
+  "paper_id": "19e76363-53a6-4f3c-8b50-844e1aea4e26",
+  "pipeline": "Paper2Markdown-V3",
+  "ok": true,
+  "page_count": 14,
+  "chunk_count": 137,
+  "main_body_chunk_count": 95,
+  "appendix_chunk_count": 37,
+  "reference_chunk_count": 5,
+  "model_text_chars": 39229,
+  "raw_markdown_chars": 64704,
+  "sanitized_chars": 58960,
+  "page_provenance": {
+    "min_page": 1,
+    "max_page": 14,
+    "invalid_count": 0
+  },
+  "marker_block_type_counts": {
+    "Caption": 1,
+    "Code": 3,
+    "Equation": 12,
+    "FigureGroup": 3,
+    "Footnote": 3,
+    "ListGroup": 12,
+    "PageFooter": 14,
+    "PageHeader": 17,
+    "SectionHeader": 39,
+    "TableGroup": 4,
+    "Text": 352
+  },
+  "asset_count_raw": 3,
+  "asset_count_model_kept": 3,
+  "asset_count_rejected": 0,
+  "asset_reject_reasons": {
+    "kept": 3
+  },
+  "artifact_leak_audit": {
+    "ok": true,
+    "hits": {
+      "Anonymous Authors": [],
+      "ACKNOWLEDGMENT": [],
+      "OpenReview": [],
+      "\"accept_label\"": [],
+      "\"decision\"": [],
+      "\"decision_tier\"": [],
+      "\"source_status\"": [],
+      "Meta-review": [],
+      "Official Review": [],
+      "official_reviews": [],
+      "meta_reviews": [],
+      "suggested_verdict_score": []
+    },
+    "artifact_count": 2
+  },
+  "default_model_input": "model_text_v3.txt",
+  "appendix_input": "appendix_text_v3.txt",
+  "reference_input": "reference_text_v3.txt"
+}

icml26/19e76363-53a6-4f3c-8b50-844e1aea4e26/sanitized_v3.txt

@@ -0,0 +1,343 @@
+{0}
+# Abstract
+Large language models have achieved strong performance on medical reasoning benchmarks, yet their deployment in clinical settings demands rigorous verification to ensure factual accuracy. While reward models offer a scalable approach for reasoning trace verification, existing methods face two limitations: they produce only scalar reward values without explicit justification, and they rely on single-pass retrieval that precludes adaptive knowledge access as verification unfolds. We introduce Med-TIV, an agentic framework that addresses these limitations by training medical reasoning verifiers to iteratively query external medical corpora during evaluation. Our approach combines tool-augmented verification with an iterative reinforcement learning paradigm that requires only trace-level supervision, alongside an adaptive curriculum mechanism that dynamically adjusts training data distribution. Across four medical reasoning benchmarks, Med-TIV achieves substantial gains over existing methods, improving MedQA accuracy by 23.5% and MedXpertQA by 32.0% relative to the base generator in particular. Crucially, Med-TIV demonstrates an 8× reduction in sampling budget requirement compared to prior reward model baselines. These findings establish that grounding verification in dynamically retrieved evidence offers a principled path toward more reliable medical reasoning systems.
+# 1. Introduction
+Large Language Models (LLMs) have demonstrated remarkable capabilities in medical reasoning, achieving competitive performance on clinical question answering, diagnostic
+<span id="page-0-0"></span>![](_page_0_Figure_9.jpeg)
+*Figure 1.* Comparison of medical reasoning verification paradigms. Text-based judges rely on parametric knowledge and may validate erroneous reasoning, while tool-integrated judges dynamically retrieve evidence to ground their judgments.
+inference, and medical knowledge benchmarks [\(Ji et al.,](#page-8-0) [2025a;](#page-8-0) [Xiao et al.,](#page-9-0) [2026\)](#page-9-0). While these advances hold significant promise for augmenting clinical decision making and democratizing access to medical expertise, the deployment of LLMs in high-stakes clinical settings demands rigorous verification mechanisms to ensure that generated reasoning is both factually accurate and logically sound [\(Zhang et al.,](#page-9-1) [2025;](#page-9-1) [Wang et al.,](#page-9-2) [2025\)](#page-9-2).
+Reward-based judges have therefore emerged as a scalable solution for evaluating model outputs, supporting both posttraining refinement via reinforcement learning from human feedback (RLHF) and inference-time scaling through tree search [\(Snell et al.,](#page-9-3) [2024\)](#page-9-3). These judges can be broadly categorized by the granularity of their supervision. Outcome Reward Models (ORMs) provide sparse trace-level supervision that quantifies the quality of the entire output, while Process Reward Models (PRMs) offer dense step-level feedback that scores each intermediate reasoning step, enabling fine-grained credit assignment and precise error localization within multi-step reasoning. Recent work has adapted both paradigms to the medical domain to assess complex clinical reasoning traces. In parallel, advances in generative reward modeling have extended judge models beyond
+{1}------------------------------------------------
+scalar scoring, enabling them to produce natural-language critiques that explicitly justify their decisions [\(Liu et al.,](#page-9-4) [2025c;](#page-9-4) [Xiong et al.,](#page-9-5) [2025\)](#page-9-5).
+108 109 Despite their effectiveness, reward-based judges exhibit fundamental limitations when applied to clinical reasoning tasks [\(Yun et al.,](#page-9-6) [2025\)](#page-9-6). A primary concern is the prevalence of hallucinations in critique traces, where judge models generate plausible yet factually incorrect assessments (Figure [1\)](#page-0-0). This issue is particularly noticeable in the medical domain, where reliable verification demands grounding in authoritative clinical evidence and established medical knowledge. Unverified judgments could lead to the propagation of incorrect diagnostic or treatment recommendations. Existing medical reasoning verifiers typically provide only scalar reward signals, offering little or no justification for their judgments and thus limiting interpretability [\(Jiang et al.,](#page-8-1) [2025b\)](#page-8-1). Furthermore, these methods often rely on a static Retrieval-Augmented Generation (RAG) pipeline, in which a fixed set of retrieved documents is prefixed to the context and remains unchanged throughout evaluation [\(Yun et al.,](#page-9-6) [2025\)](#page-9-6). Such static design precludes adaptive, multi-turn evidence gathering and forces the verifier to a fixed retrieval budget, thus limiting scalability.
+To address these issues, we propose Med-TIV (Medical Tool-Integrated reasoning Verifier), an agentic reinforcement learning (RL) framework that trains LLMs to leverage external knowledge bases for judging medical reasoning traces[1](#page-1-0) . Med-TIV features three key design principles: (1) a tool-augmented verification paradigm that enables dynamic, iterative knowledge retrieval during the evaluation process; (2) an iterative RL approach that progressively improves verification capabilities without requiring step-level expert annotations; and (3) an adaptive curriculum formulation strategy that adjusts the data distribution in response to the evolving capability of the model. By equipping judge models with tool-use capabilities, Med-TIV grounds evaluation decisions in external evidence rather than relying solely on parametric knowledge, thereby mitigating hallucination, improving interpretability, and overcoming the limitations of static RAG [\(Ji et al.,](#page-8-2) [2025b;](#page-8-2) [Xia et al.,](#page-9-7) [2025\)](#page-9-7).
+To verify the effectiveness of Med-TIV, we conduct extensive experiments on common medical reasoning benchmarks. Our results demonstrate that Med-TIV trains strong medical verifiers: when guiding inference-time search for a 7B generator model, our trained verifier achieves relative improvements of 23.5% on MedQA and 32.0% on MedXpertQA compared to the generator model alone. Moreover, Med-TIV consistently outperforms existing medical reward model baselines and surpasses the performance of models that are up to 4× larger in scale. Notably, Med-TIV
+also demonstrates an 8× gain in sampling efficiency compared to prior reward-based approaches, achieving equivalent accuracy with substantially fewer sampled reasoning traces during test-time search.
+Our main contributions are summarized as follows:
+- We propose Med-TIV, a novel tool-integrated verification framework that enables dynamic, iterative knowledge retrieval during medical reasoning evaluation, providing both interpretable, fine-grained justification and improved factual grounding.
+- We introduce an iterative RL paradigm with curriculumbased difficulty adaptation that progressively improves verification capabilities through self-bootstrapping, requiring only trace-level supervision rather than dense steplevel expert annotations.
+- Med-TIV achieves state-of-the-art performance on four medical reasoning benchmarks, with comprehensive ablation studies that validate each component's contribution.
+# 2. Preliminaries
+#### 2.1. Problem Setup
+We define *medical reasoning verification* as the task of assessing the correctness of a multi-step reasoning trace generated in response to a medical question. Formally, given a medical question q ∈ Q and a multi-step reasoning trace τ = (s1, s2, . . . , sm) from a generator model, a verifier model determines whether τ contains any errors. We formulate this problem as binary classification, where the verifier Vθ(q, τ ) produces a judgment ℓ ∈ {0, 1}, where ℓ = 1 indicates a error-free reasoning trace, and ℓ = 0 indicates the presence of one or more errors. Unlike scalar reward models that output continuous scores, we adopt a generative judge paradigm in which the verifier produces a discrete judgment accompanied by a detailed critique trace that provides a structured justification for the decision.
+#### 2.2. Tool-Augmented Reasoning Verifier
+Following prior works [\(Jin et al.,](#page-8-3) [2025\)](#page-8-3), we extend the verifier with access to an external *search engine* E that retrieves top-k documents from a curated medical corpus (See Appendix [B.2](#page-10-0) for details). Retrieved documents are appended verbatim to the verifier context. Given a verification instance (q, τ ), the verifier constructs an iterative verfication trajectory. At step k, the trajectory is represented as:
+$$\mathbf{t}_k = \{r_1, a_1, o_1, \dots, r_k, a_k, o_k\},\$$
+where r<sup>i</sup> denotes a natural language reasoning step analyzing the medical content, a<sup>i</sup> is a search query formulated to retrieve relevant medical knowledge, and o<sup>i</sup> = E(ai) represents the retrieved documents. The iterative verification
+<span id="page-1-0"></span><sup>1</sup>Code is available at [ [PittNAIL/med-tiv](
+{2}------------------------------------------------
+Figure 2. Overview of Med-TIV. Left: Tool-integrated verification iteratively analyzes reasoning traces, formulates search queries, and retrieves medical evidence before producing correctness judgments. Middle: Curriculum formulation filters trivial and impossible instances, retaining boundary cases for RL training. Right: At inference time, the verifier evaluates candidate medical reasoning traces generated by a frozen model and final answers are selected via weighted self-consistency.
+process is defined as:
+<span id="page-2-0"></span>110
+123124125
+135136
+137138
+$$(r_k, a_k) \sim V_{\theta}(q, \tau, \mathbf{t}_{k-1}),$$
+$o_k = \mathcal{E}(a_k),$ $\mathbf{t}_k = \mathbf{t}_{k-1} \oplus r_k \oplus a_k \oplus o_k,$
+where $\oplus$ denotes sequence concatenation. This process continues until the verifier produces a final judgment $\ell \sim V_{\theta}(q,\tau,\mathbf{t}_T)$ at the terminal step T. By allowing multiple tool executions, the verifier dynamically retrievs medical knowledge as need to verify specific claims in the reasoning trace. Table 5 in the Appendix shows the explicit instruction used in our experiments.
+#### 2.3. Test-Time Search
+Test-time search strategies improve reasoning performance by leveraging reward models to evaluate and select among multiple candidate solutions (Shi et al., 2024). Given a frozen generator model $\pi_{\rm gen}$ and a question q, we first sample N independent reasoning traces:
+$$\{\tau^{(j)}\}_{j=1}^N \sim \pi_{\mathrm{gen}}(\cdot \mid q).$$
+A trained verifier $V_{\theta}$ then scores each candidate trace, and the final output is selected based on these scores. Common selection strategies include Best-of-N sampling, which
+selects the trace with the highest score:
+$$\hat{\tau} = \arg\max_{\tau^{(j)}} V_{\theta}(q, \tau^{(j)}),$$
+and verification-based majority voting, where candidate traces are first filtered by the verifier and the final answer is determined by consensus among verified traces. Med-TIV trains such a plug-in verifier that provides tool-grounded assessments that can be used to augment decision-making for any frozen generator model at inference time.
+#### 3. Tool-Integrated Medical Reasoning Verifier
+Med-TIV is an agentic verification framework that trains models to leverage external knowledge bases for verifying whether a given medical reasoning trace contains errors. We adopt an iterative training approach based on dynamic curriculum learning, which requires no fine-grained step-level expert supervision and trains solely through multiple rounds of reinforcement learning (Figure 2). We next describe the training procedure of Med-TIV in details.
+#### 3.1. Tool-Integrated RL with Verifiable Rewards
+**Data Construction.** All training data across iterations is derived from the open-source Med-PRM dataset. Each original instance consists of a tuple $(q, \tau, \ell_{\text{step}}, \ell_{\text{trace}})$ , where q is a medical question, $\tau$ is a multi-step reasoning trace, $\ell_{\text{step}}$ de-
+{3}------------------------------------------------
+notes step-level labels, and $\ell_{trace}$ is a trace-level correctness label<sup>2</sup>.
+At each training iteration, we only utilize the triplet $(q,\tau,\ell_{\text{trace}})$ with human-annotated trace-level labels. Step-level labels $\ell_{\text{step}}$ is intentionally excluded, as Med-TIV is designed to improve verification performance without replying on fine-grained supervision. For each training iteration, we fix the training data budget to 20K instances and enforce a balanced label distribution between correct ( $\ell_{\text{trace}}=1$ ) and incorrect ( $\ell_{\text{trace}}=0$ ) reasoning traces.
+**Algorithm.** We employ Dr. GRPO (Liu et al., 2025b) as the RL algorithm for training the verifier. Given a verification instance $(q_i, \tau_i)$ , we sample a group of G verification trajectories $\{\mathbf{o}_i\}_{i=1}^G$ from the current policy $\pi_\theta$ . Each trajectory $\mathbf{o}_i = (o_i^1, \dots, o_i^{|\mathbf{o}_i|})$ consists of reasoning tokens, search queries, retrieved documents, and a final judgment. The objective is:
+$$\frac{1}{G} \sum_{i=1}^{G} \sum_{t=1}^{|\mathbf{o}_i|} \left\{ \min \left[ r_i^t \hat{A}_i^t, \operatorname{clip} \left( r_i^t, 1 - \epsilon_l, 1 + \epsilon_h \right) \hat{A}_i^t \right] \right\}, \quad (1)$$
+where $r_i^t = \frac{\pi_\theta(o_i^t|\mathbf{q},\mathbf{o}_i^{< t})}{\pi_{\theta_{\mathrm{old}}}(o_i^t|\mathbf{q},\mathbf{o}_i^{< t})}$ , $\mathbf{q} = (q,\tau)$ denotes the input prompt containing the question and reasoning trace, $\mathbf{o}_i^{< t}$ represents previously generated tokens, and $\epsilon_l$ and $\epsilon_h$ are the clipping parameters. The advantage term $\hat{A}_i^t$ is defined as:
+$$\hat{A}_i^t = R(\mathbf{q}, \mathbf{o}_i) - \text{mean}\left(\left\{R(\mathbf{q}, \mathbf{o}_1), \dots, R(\mathbf{q}, \mathbf{o}_G)\right\}\right).$$
+**Reward Designs.** To facilitate multi-turn RL with tool execution, we design a structured reward covering two complementary objectives, following prior practices (Jin et al., 2025):
+(i) Correctness Reward $(R_c)$ : This component measures whether the verifier's judgment aligns with the ground-truth label. Let $\mathbf{q} = (q, \tau)$ denote the verification prompt and $\ell \in \{0, 1\}$ the ground-truth label. We define:
+$$R_c = \mathbb{1}(\text{extract}(\mathbf{o}) = \ell),$$
+where $\mathbb{1}(\cdot)$ is the indicator function and extract(o) parses the final judgment from the <answer> tags in the generated trajectory o. Intuitively, $R_c=1$ if the verifier's decision is correct, and $R_c=0$ otherwise.
+(ii) Format Reward $(R_f)$ : To ensure reliable tool use and structured outputs, the verifier is required to adhere to a predefined format. Specifically, reasoning steps must be enclosed within <think> tags, search queries within <search> tags, and the final judgment within <answer>
+tags. To discourage degenerate outputs, we further penalize excessive tag usage. Specifically, $R_f=1$ if the output satisfies all formatting constraints and contains no more than $10 < answer> tag pairs; <math>R_f=0.25$ if the output is correct but exhibits tag overflow; and $R_f=0$ otherwise.
+The final reward ${\cal R}$ is defined as the product of the two components:
+$$R = R_c \times R_f$$
+.
+#### <span id="page-3-1"></span>3.2. Training Strategies
+Adaptive Curriculum Formulation. A central challenge in RL for verification is ensuring that training data remains appropriately calibrated to the evolving capability of the model. Instances that are either trivially easy or impossibly difficult yields minimal learning signal, as the resulting policy gradients approach zero. To address this issue, we adopt a model-aware curriculum formulation mechanism that dynamically adapts the task distribution at each training iteration.
+Concretely, before each iteration t, we perform online filtering on the sampled batch $\mathcal{B}_t$ to construct an effective training set $\mathcal{D}_t$ :
+$$\mathcal{D}_t = \{(q, \tau, \ell) \in \mathcal{B}_t : \exists g, g' \in \{1, \dots, G\} \text{ s.t. } r^{(g)} \neq r^{(g')}\}.$$
+Here, for each candidate instance $(q, \tau, \ell) \in \mathcal{B}_t$ , we sample G verification trajectories $\{o^{(g)}\}_{g=1}^G$ from the current policy $\pi_{\theta_t}$ . We then compute the corresponding rewards $\{r^{(g)}\}_{g=1}^G$ . Finally, we retain only instances if any two rewards are different, i.e., reward variance is non-zero (Khatri et al., 2025).
+This criterion eliminates prompts where the model either consistently succeeds or consistently fails across all sampled trajectories. By filtering these zero-gradient instances, optimization is focused on decision-boundary cases where the verifier exhibits uncertainty.
+To maintain a fixed training budget per iteration, we iteratively resample additional instances from the labeled pool $\mathcal{B}$ and apply the same filtering criterion until $|\mathcal{D}_t|=20K$ . This dynamic curriculum evolves naturally across iterations as the verifier improves, eliminating the need for manually designed difficulty schedules.
+Iterative Training via Self-Bootstrapping. We adopt an iterative training approach that progressively improves verification capabilities through multiple rounds of RL. Unlike prior work that alternates between rejection sampling, supervised fine-tuning (SFT), and RL (Xu et al., 2025), our approach operates entirely through iterative RL, following the RL-Zero paradigm where the model reinforces its verification capabilities without requiring dense turn-level expert demonstrations for cold start.
+<span id="page-3-0"></span><sup>&</sup>lt;sup>2</sup>Dataset is available at
+{4}------------------------------------------------
+<span id="page-4-0"></span>Algorithm 1 Iterative Training of Tool-Integrated Medical Reasoning Verifier
+249250
+```
+Require: Base verifier \pi_{\theta_0}, labeled dataset pool \mathcal{D} =
+\{(q_i, \tau_i, \ell_i)\}_{i=1}^N, maximum iterations T_{\text{max}}, batch size
+B, group size G, search engine \mathcal{E}
+Ensure: Trained verifier \pi_{\theta^*}
+1: for t = 1 to T_{\text{max}} do
+Sample labeled batch
+\mathcal{B}_t \leftarrow \text{SAMPLEBATCH}(\mathcal{D}, B)
+3:
+4:
+\mathcal{D}_t \leftarrow \emptyset
+▷ Curriculum formulation
+5:
+for each (q, \tau, \ell) \in \mathcal{B}_t do
+6:
+Sample verification trajectories:
+7:
+\{\hat{\ell}^{(g)}\}_{q=1}^G \sim \pi_{\theta_t}(\cdot \mid q, \tau, \mathcal{E})
+Compute rewards within group:
+8:
+r^{(\hat{g})} \leftarrow \mathbb{1}[\hat{\ell}^{(g)} = \ell], \text{ for } g \in 1, \dots, G
+if \exists g \neq g' such that r^{(g)} \neq r^{(g')} then
+9:
+Add (q, \tau, \ell) to curriculum set \mathcal{D}_t
+10:
+11:
+end if
+12:
+end for
+13:
+▶ RL optimization on curriculum data
+\pi_{\theta_{t+1}} \leftarrow \text{DR.GRPO}(\pi_{\theta_t}, \mathcal{D}_t, \mathcal{E})
+15: end for
+16: Return \pi_{\theta_{T_{\max}}}
+```
+Starting from the base model $\pi_{\theta_0}$ , we perform $T_{\text{max}}$ iterations. Each iteration consists of three stages:
+```
+\begin{split} \mathcal{B}_t \leftarrow \text{SampleBatch}(\mathcal{D}, B), \\ \mathcal{D}_t \leftarrow \text{Filter}(\mathcal{B}_t, \pi_{\theta_t}), \\ \pi_{\theta_{t+1}} \leftarrow \text{RL}(\pi_{\theta_t}, \mathcal{D}_t). \end{split}
+```
+Each iteration draws a fresh batch $\mathcal{B}_t$ from the annotated pool $\mathcal{D}$ with trace-level labels, ensuring a balanced distribution of correct and incorrect reasoning traces. The curriculum filtering then constructs the training set $\mathcal{D}_t$ as described above, and RL optimization updates the policy based on the structured reward.
+The key insight underlying this iterative approach is the coevolution of model capability and training distribution. As the verifier improves, the filtering mechanism automatically removes instances that have become too easy, while the fresh sampling introduces new challenging cases. This creates a self-bootstrapping cycle: stronger models encounter harder verification tasks, which in turn drive further improvements. Since the trace-level correctness reward is deterministic and unambiguous, this self-bootstrapping process converges reliably without the instabilities that can arise from noisy synthetic step-level labels. We summarize the overall training procedure in Algorithm 1.
+#### 4. Experiments
+#### 4.1. Experimental Setup
+**Evaluation benchmarks.** We evaluated Med-TIV on four open-source medical question-answering benchmarks: MedQA (Jin et al., 2020), MedMCQA (Pal et al., 2022), MMLU-Med (Hendrycks et al., 2021), and MedX-pertQA (Zuo et al., 2025), using accuracy as the evaluation metric. These benchmarks collectively assess the verifier's ability to distinguish correct from erroneous reasoning across varying difficulty levels and medical subdomains. Detailed descriptions of benchmarks are in Appendix C.1.
+Implementation details. We trained verifiers using two light-weight backbone models: Llama3.1-8B and Qwen2.5-7B, with Llama3.1-8B as the default for results reporting. All training was conducted using the VeRL-Tool framework (Jiang et al., 2025a). Detailed hyperparameters are shown in Appendix B.1. All experiments were conducted on 4 NVIDIA H100 GPUs with 80GB of memory. Due to computational constraints, we limit the maximum number of RL iterations to $T_{\rm max}=2$ and we set the group size for curriculum formulation (Section 3.2) to G=8.
+For inference, we used the default sampling hyperparameters for all models. In reward-guided search experiments, unless otherwise specified, we used Qwen2.5-7B as the frozen generator and sampled up to 32 candidate reasoning traces per question. We applied Hard-Weighted Self-Consistency as the default test-time search strategy.
+**Baselines.** We compared Med-TIV against two groups of baselines. 1): *Off-the-shelf LLMs*: GPT-40-mini (OpenAI et al., 2024), Gemini-2.0-Flash, DeepSeek-R1 series (Guo et al., 2025), Qwen2.5 series (Yang et al., 2025), Llama3.1 (Grattafiori et al., 2024), AlphaMed (Liu et al., 2025a), UltraMedical (Zhang et al., 2024), and HuatuoGPT-01 (Chen et al., 2024). 2): *Medical domain-specialized Reward Models*: MedS<sup>3</sup> (Jiang et al., 2025b) and Med-PRM (Yun et al., 2025). Detailed descriptions of each reward model baseline are shown in Appendix B.3.
+**Test-Time Search Strategies.** We evaluated three test-time search strategies that leverage Med-TIV to improve the reasoning performance of frozen generators. Given a reasoning trace $\tau=(s_1,s_2,\ldots,s_K)$ with K steps, our verifier assigns a confidence score $r_{\tau}\in[0,1]$ for the entire trace, defined as the softmax probability of the 1 token over the logits of both 1 and 0 tokens.
+• Best-of-N. Given a question q, we sampled N candidate traces $\{\tau^{(j)}\}_{j=1}^N$ from the generator and selected the trace with the highest verifier confidence score:
+$$\hat{\tau} = \arg\max_{\tau^{(j)}} r_{\tau^{(j)}}.$$
+{5}------------------------------------------------
+<span id="page-5-0"></span>*Table 1.* Main evaluation results on medical reasoning benchmarks. We report accuracy (%) on MedQA, MedMCQA, MMLU-Med, and MedXpertQA. Bold numbers indicate the best results among the reward model group. ✓: Verifier supports external tools for judging; ✗: Verifier does not support external tools for judging.
+| Baselines | å | Train | Size | MedQA | MedMCQA | MMLU-Med | MedXpertQA | Avg. |
+|------------------------------|---|-------|------|-------|---------|----------|------------|-------|
+| Proprietary Models | | | | | | | | |
+| GPT-4o-mini | - | - | - | 79.03 | 68.20 | 87.79 | 17.84 | 63.22 |
+| Gemini-2.0-Flash | - | - | - | 87.51 | 72.60 | 92.01 | 20.57 | 68.17 |
+| General Reasoning Models | | | | | | | | |
+| DeepSeek-R1 | - | - | 671B | 90.34 | 78.80 | 94.40 | 37.76 | 75.33 |
+| R1-Distill-Qwen | - | - | 7B | 24.82 | 36.40 | 47.47 | 7.43 | 29.03 |
+| R1-Distill-Llama | - | - | 8B | 34.96 | 43.60 | 64.19 | 5.35 | 37.03 |
+| General Non-reasoning Models | | | | | | | | |
+| Qwen2.5 | - | - | 32B | 73.21 | 64.83 | 84.94 | 13.87 | 59.21 |
+| Qwen2.5 | - | - | 7B | 60.96 | 56.56 | 76.96 | 12.15 | 51.66 |
+| Llama3.1 | - | - | 8B | 70.93 | 61.60 | 78.97 | 13.02 | 56.13 |
+| Medical Reasoning Models | | | | | | | | |
+| AlphaMed | - | - | 7B | 71.01 | 61.46 | 81.16 | 19.16 | 58.20 |
+| UltraMedical | - | - | 8B | 72.66 | 62.60 | 79.61 | 15.25 | 57.53 |
+| HuatuoGPT-o1 | - | - | 8B | 72.19 | 63.60 | 75.30 | 16.84 | 56.98 |
+| Medical Reward Models | | | | | | | | |
+| MedS3 | ✗ | 225k | 7B | 64.89 | 58.91 | 80.53 | 12.90 | 54.31 |
+| Med-PRM | ✓ | 111k | 7B | 69.99 | 62.36 | 80.99 | 13.51 | 56.71 |
+| Med-TIV (Ours) | ✓ | 20k | 7B | 75.26 | 64.70 | 85.58 | 16.04 | 60.40 |
+• *Hard-Weighted Self-Consistency.* We first filtered traces by the verifier's binary judgment, keeping only those labeled correct (Vθ(q, τ ) = 1). Among the filtered traces, we applied majority voting to determine the final answer:
+$$\hat{a} = \arg\max_{a} \sum_{j=1}^{N} \mathbb{1} \left[ V_{\theta}(q, \tau^{(j)}) = 1 \right] \cdot \mathbb{1} \left[ \operatorname{ans}(\tau^{(j)}) = a \right].$$
+• *Soft-Weighted Self-Consistency.* Instead of binary filtering, we weighted each trace's vote by the verifier's confidence score:
+$$\hat{a} = \arg\max_{a} \sum_{j=1}^{N} r_{\tau^{(j)}} \cdot \mathbb{1} \big[ \mathrm{ans}(\tau^{(j)}) = a \big].$$
+### 4.2. Main Results
+Table [1](#page-5-0) presents the main results on four medical reasoning benchmarks. Models trained with Med-TIV consistently outperform existing baselines across all benchmarks. Specifically, under guided-search using a Med-TIVtrained verifier, Qwen2.5-7B attains accuracies of 75.26%
+on MedQA, 64.70% on MedMCQA, 85.58% on MMLU-Med, and 16.04% on MedXpertQA, yielding an average accuracy of 60.40%. Notably, Med-TIV enables this 7B generator to rival substantially larger models, even surpassing the base performance of Qwen2.5-32B despite using a generator that is approximately 4× smaller. Compared to domain-specialized medical reasoning models of similar scale, Med-TIV outperforms HuatuoGPT-o1-8B and UltraMedical-8B by 3.07% and 2.60% on MedQA, respectively, demonstrating the effectiveness of our tool-integrated verification. Case analysis in Appendix [D](#page-12-0) further illustrates how Med-TIV identifies subtle reasoning errors.
+#### 4.3. Analysis
+We conducted a series of ablation analyses to investigate six key research questions regarding our proposed framework.
+Q1: Does **Med-TIV** generalize across different generator models? To evaluate the generalizability of the trained verifier, we applied Med-TIV to guide test-time search across generator models of varying sizes and capa-
+{6}------------------------------------------------
+<span id="page-6-1"></span>![](_page_6_Figure_2.jpeg)
+Figure 3. Test-time scaling analysis across three medical reasoning benchmarks. Each plot shows accuracy versus sampling budget $N \in \{1, 2, 4, 8, 16, 32\}$ for four baselines. Med-TIV consistently outperforms baselines across all sampling budgets and benchmarks.
+<span id="page-6-0"></span>Table 2. Performance improvements from using Med-TIV as a verifier on MedQA. For each generator model, the first row indicates the accuracy over single sampled trace per question.
+| Models | MedQA |
+|------------------------------|-----------------------|
+| Qwen2.5-7B | 60.96 |
+| + Self-Consistency | 66.38 (+5.42) |
+| + Best-of-N (Med-TIV) | 72.35 (+11.39) |
+| + Soft Weighted SC (Med-TIV) | 75.02 (+14.06) |
+| + Hard Weighted SC (Med-TIV) | <b>75.26</b> (+14.30) |
+| AlphaMed-7B | 71.01 |
+| + Self-Consistency | 74.23 (+3.22) |
+| + Best-of-N (Med-TIV) | 75.02 (+4.01) |
+| + Soft Weighted SC (Med-TIV) | 75.33 (+4.32) |
+| + Hard Weighted SC (Med-TIV) | <b>75.65</b> (+4.64) |
+| Qwen2.5-32B | 73.21 |
+| + Self-Consistency | 75.26 (+2.05) |
+| + Best-of-N (Med-TIV) | 75.57 (+2.36) |
+| + Soft Weighted SC (Med-TIV) | 75.96 (+2.75) |
+| + Hard Weighted SC (Med-TIV) | <b>75.96</b> (+2.75) |
+bilities. As shown in Table 2, when using Qwen2.5-7B as the generator, Hard-Weighted Self-Consistency yields a relative improvement of 23.5% over the base model's single-sample accuracy, substantially outperforming the 12.2% gain achieved by standard Self-Consistency. Notably, the domain-specialized AlphaMed-7B model also benefits from verifier guidance with a 6.5% relative improvement, indicating that our verifier provides complementary verification capabilities beyond domain-specific fine-tuning. The improvements extend to larger models as well: Qwen2.5-32B achieves a 3.8% relative gain during test-time search, demonstrating that a light-weight 8B verifier can effectively guide models that are significantly larger than itself. This
+cross-scale generalization suggests that Med-TIV learns transferable verification patterns rather than overfitting to specific generator characteristics.
+Q2: How do different test-time search strategies compare under Med-TIV? We then systematically compare different test-time search strategies under verifier guidance to identify the most effective approach for leveraging verification signals. As shown in Table 2, Hard-Weighted Self-Consistency consistently achieves the highest accuracy across all generators, followed by Soft-Weighted Self-Consistency and Best-of-N selection. On Qwen2.5-7B, Hard-Weighted Self-Consistency outperforms Best-of-N by 3% absolute accuracy, suggesting that majority voting among verified traces provides more robust answer selection than simply choosing the highest-confidence individual trace.
+Q3: Can Med-TIV reduce the sampling budget required to achieve state-of-the-art performance compared to existing baselines? Next, we investigated how verification performance scales with sampling budget, a critical consideration for deployment under varying computational constraints. As shown in Figure 3, Med-TIV achieves substantial efficiency advantage over existing medical reward models across all three benchmarks. In particular, Med-TIV matches the performance of baselines using only 4 samples, whereas the baselines require 32 samples, representing an 8× reduction in sampling budget. On MedOA, Med-TIV achieves 72.1% accuracy at N=4, while Med-PRM requires the full N=32 budget to reach 70.0% accuracy. Since inference cost scales approximately linearly with the number of sampled traces, this translates to equivalent performance at one-eighth the generator inference cost in practical deployment settings.
+**Q4:** Does Med-TIV generalize across different base models? To assess the generality of our proposed framework, we compared verification performance using two
+{7}------------------------------------------------
+<span id="page-7-0"></span>![](_page_7_Figure_1.jpeg)
+Figure 4. Ablation on base model selection and training iterations.
+distinct verifier backbones: Llama3.1-8B and Qwen2.5-7B. As shown in Figure 4, both backbones achieve strong performance after two training iterations. Llama3.1-8B consistently outperforms Qwen2.5-7B by approximately 3.5% absolute accuracy on MedQA, achieving 75.86% versus 72.35% after 2 iterations of training. The parallel performance gains observed across both models indicate that Med-TIV is agnostic to backbone architectures.
+**Q5:** What is the impact of iterative training? Figure 4 presents ablation results examining the impact of iterative training with adaptive curriculum formulation. Llama3.1-8B improves from 60.96% to 75.26% after iteration 1, with marginal gains to 75.86% at iteration 2. Qwen2.5-7B follows a similar pattern, reaching 72.35% after two iterations. The rapid convergence suggest that the majority of verification capability is acquired in the first round, with subsequent iterations refining boundary cases.
+**O6**: How does RL and tool integration impact verification performance? Table 3 highlights the dual benefits of our framework across two generators. RL
+<span id="page-7-1"></span>Table 3. Ablation on RL and tool integration.
+| Models | MedQA |
+|-----------------------|-------|
+| Qwen2.5-7B | 60.96 |
+| + Med-TIV (RL) | 69.60 |
+| + Med-TIV (RL + Tool) | 70.54 |
+| AlphaMed-7B | 71.01 |
+| + Med-TIV (RL) | 76.12 |
+| + Med-TIV (RL + Tool) | 77.14 |
+training drives the primary gain, boosting MedQA accuracy of Qwen2.5-7B by 8.64%, confirming that the verifier effectively internalizes reasoning patterns. Tool integration provides a critical secondary boost, further elevating accuracy to 70.54%. A similar cumulative trend is observed with AlphaMed-7B. This demonstrates that while RL
+anchors logical verification, dynamic retrieval is essential for resolving knowledge-intensive boundary cases beyond the model's parametric memory.
+#### 5. Related Work
+Medical Reasoning Models. The application of large language models to medical reasoning has attracted considerable attention. Early efforts focused on domain-adaptive pretraining and instruction tuning on medical corpora (Wu et al., 2023; Singhal et al., 2025; Chen et al., 2023). More recent work has explored reasoning-enhanced medical models. HuatuoGPT-o1 (Chen et al., 2024) incorporates chainof-thought reasoning with verification mechanisms, and UltraMedical (Zhang et al., 2024) combines high-quality instruction data with preference optimization. AlphaMed (Liu et al., 2025a) employs RL to improve medical reasoning capabilities. Despite these advances, most existing approaches focus on improving the generator model itself, whereas our work addresses the complementary problem of training a plug-and-play verifier that can improve any frozen generator through test-time search.
+Tool-Assisted Reward and Judge Models. Standard LLM-based judges typically function as passive scorers limited by parametric knowledge. Recent work addresses this through agentic reward modeling, equipping verifiers with executable tools. Themis (Li et al., 2024) established the foundational framework by enabling access to calculators, search engines, and knowledge bases through structured tool-calling traces. TIR-Judge (Xu et al., 2025) advanced this paradigm in the general domain by integrating code execution to judge paired responses. TIM-PRM (Kuang et al., 2025) introduced independent tool queries for multi-modal verification to eliminate confirmation bias. The concept has further expanded to the Agent-as-a-Judge paradigm (You et al., 2026), which employs dynamic planning, tool augmentation and multi-agent coordination to decompose complex evaluation tasks. Our work instantiates this agentic paradigm within the medical domain, moving beyond static retrieval to iterative, evidence-grounded clinical verification.
+#### 6. Conclusion
+We presented Med-TIV, an agentic RL framework for medical reasoning verification. Our approach addresses key limitations of existing medical reward models by offering explicit critique traces and enabling dynamic knowledge retrieval during verification. Empirical evaluations across four medical reasoning benchmarks demonstrate that Med-TIV substantially outperforms prior approaches. More broadly, Med-TIV introduces a general paradigm for training toolaugmented verifiers that can be extended to other highstakes domains requiring evidence-grounded evaluation.
+{8}------------------------------------------------
+# Impact Statement
+This paper introduces research aimed at improving the reliability of large language models for medical reasoning tasks. We believe our work contributes positively to the development of trustworthy medical AI systems by providing mechanisms to verify reasoning correctness before clinical deployment. Med-TIV holds potential to enhance the safety of LLM-assisted clinical decision support by reducing erroneous reasoning outputs through systematic verification. By grounding judgments in retrieved medical evidence, our approach offers improved transparency compared to opaque scalar reward models, enabling practitioners to better understand and audit verification decisions. The efficiency gains demonstrated by Med-TIV could democratize access to reliable medical reasoning verification, making robust verification feasible even in resource-constrained settings.
+# References
+- <span id="page-8-12"></span>Chen, J., Cai, Z., Ji, K., Wang, X., Liu, W., Wang, R., Hou, J., and Wang, B. Huatuogpt-o1, towards medical complex reasoning with llms, 2024. URL [ [arxiv.org/abs/2412.18925](
+- <span id="page-8-13"></span>Chen, Z., Cano, A. H., Romanou, A., Bonnet, A., Matoba, K., Salvi, F., Pagliardini, M., Fan, S., Kopf, A., Mo- ¨ htashami, A., Sallinen, A., Sakhaeirad, A., Swamy, V., Krawczuk, I., Bayazit, D., Marmet, A., Montariol, S., Hartley, M.-A., Jaggi, M., and Bosselut, A. Meditron-70b: Scaling medical pretraining for large language models, 2023. URL [ [16079](
+- <span id="page-8-10"></span>Grattafiori, A. et al. The llama 3 herd of models, 2024. URL <
+- <span id="page-8-9"></span>Guo, D. et al. Deepseek-r1 incentivizes reasoning in llms through reinforcement learning. Nature, 645(8081), 2025. doi: 10.1038/s41586-025-09422-z.
+- <span id="page-8-7"></span>Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., and Steinhardt, J. Measuring massive multitask language understanding, 2021. URL [https:]( [//arxiv.org/abs/2009.03300](
+- <span id="page-8-0"></span>Ji, Y., Ma, W., Sivarajkumar, S., et al. Mitigating the risk of health inequity exacerbated by large language models. npj Digital Medicine, 8:246, 2025a. doi: 10.1038/s41746-025-01576-4.
+- <span id="page-8-2"></span>Ji, Y., Zhang, H., and Wang, Y. Bias evaluation and mitigation in retrieval-augmented medical question-answering systems, 2025b. URL [ [2503.15454](
+- <span id="page-8-8"></span>Jiang, D., Lu, Y., Li, Z., Lyu, Z., Nie, P., Wang, H., Su, A., Chen, H., Zou, K., Du, C., Pang, T., and Chen, W. Verltool: Towards holistic agentic reinforcement learning with tool use, 2025a. URL [ [abs/2509.01055](
+- <span id="page-8-1"></span>Jiang, S., Liao, Y., Chen, Z., Zhang, Y., Wang, Y., and Wang, Y. Meds<sup>3</sup> : Towards medical slow thinking with self-evolved soft dual-sided process supervision, 2025b. URL <
+- <span id="page-8-3"></span>Jin, B., Zeng, H., Yue, Z., Yoon, J., Arik, S., Wang, D., Zamani, H., and Han, J. Search-r1: Training llms to reason and leverage search engines with reinforcement learning, 2025. URL [ [2503.09516](
+- <span id="page-8-6"></span>Jin, D., Pan, E., Oufattole, N., Weng, W.-H., Fang, H., and Szolovits, P. What disease does this patient have? a large-scale open domain question answering dataset from medical exams, 2020. URL [ [abs/2009.13081](
+- <span id="page-8-16"></span>Jin, Q., Kim, W., Chen, Q., Comeau, D. C., Yeganova, L., Wilbur, W. J., and Lu, Z. Medcpt: Contrastive pre-trained transformers with large-scale pubmed search logs for zero-shot biomedical information retrieval. Bioinformatics, 39(11), November 2023. ISSN 1367-4811. doi: 10.1093/bioinformatics/ btad651. URL [ [bioinformatics/btad651](
+- <span id="page-8-5"></span>Khatri, D., Madaan, L., Tiwari, R., Bansal, R., Duvvuri, S. S., Zaheer, M., Dhillon, I. S., Brandfonbrener, D., and Agarwal, R. The art of scaling reinforcement learning compute for llms, 2025. URL [ [abs/2510.13786](
+- <span id="page-8-15"></span>Kuang, P., Wang, X., Liu, W., Dong, J., and Xu, K. Timprm: Verifying multimodal reasoning with tool-integrated prm, 2025. URL [ [22998](
+- <span id="page-8-14"></span>Li, L., Chai, Y., Wang, S., Sun, Y., Tian, H., Zhang, N., and Wu, H. Tool-augmented reward modeling, 2024. URL <
+- <span id="page-8-11"></span>Liu, C., Wang, H., Pan, J., Wan, Z., Dai, Y., Lin, F., Bai, W., Rueckert, D., and Arcucci, R. Beyond distillation: Pushing the limits of medical llm reasoning with minimalist rule-based rl, 2025a. URL [https:]( [//arxiv.org/abs/2505.17952](
+- <span id="page-8-4"></span>Liu, Z., Chen, C., Li, W., Qi, P., Pang, T., Du, C., Lee, W. S., and Lin, M. Understanding r1-zero-like training: A critical perspective, 2025b. URL [ [org/abs/2503.20783](
+{9}------------------------------------------------
+<span id="page-9-4"></span>495 496 497 498 Liu, Z., Wang, P., Xu, R., Ma, S., Ruan, C., Li, P., Liu, Y., and Wu, Y. Inference-time scaling for generalist reward modeling, 2025c. URL [ [2504.02495](
+<span id="page-9-10"></span><span id="page-9-8"></span>504
+<span id="page-9-16"></span>513 514
+<span id="page-9-3"></span>516
+<span id="page-9-2"></span>524 525 526
+<span id="page-9-15"></span><span id="page-9-7"></span>528 529 530
+<span id="page-9-5"></span><span id="page-9-0"></span>536
+- <span id="page-9-12"></span>OpenAI et al. Gpt-4o system card, 2024. URL [https:]( [//arxiv.org/abs/2410.21276](
+- Pal, A., Umapathi, L. K., and Sankarasubbu, M. Medmcqa : A large-scale multi-subject multi-choice dataset for medical domain question answering, 2022. URL [https:]( [//arxiv.org/abs/2203.14371](
+- Shi, W., Xu, R., Zhuang, Y., Yu, Y., Sun, H., Wu, H., Yang, C., and Wang, M. D. Medadapter: Efficient test-time adaptation of large language models towards medical reasoning, 2024. URL [ [2405.03000](
+- Singhal, K., Tu, T., Gottweis, J., et al. Toward expertlevel medical question answering with large language models. Nature Medicine, 31:943–950, 2025. doi: 10. 1038/s41591-024-03423-7.
+- Snell, C., Lee, J., Xu, K., and Kumar, A. Scaling llm testtime compute optimally can be more effective than scaling model parameters, 2024. URL [ [org/abs/2408.03314](
+- Wang, K., Fu, Z., Xin, W., Zhou, L., and Chandrappa, S. K. Digital voices of survival: From social media disclosures to support provisions for domestic violence victims. arXiv preprint arXiv:2509.12288, 2025.
+- Wu, C., Lin, W., Zhang, X., Zhang, Y., Wang, Y., and Xie, W. Pmc-llama: Towards building open-source language models for medicine, 2023. URL [ [org/abs/2304.14454](
+- Xia, C., Wu, Q., Tian, S., and Hao, Y. Parallelism meets adaptiveness: Scalable documents understanding in multi-agent llm systems, 2025. URL [https:]( [//arxiv.org/abs/2507.17061](
+- Xiao, W., Lian, J. J., Ouyang, K., Gu, S., Ke, Z., Wei, D., Sha, X., Wang, J., Fu, S., Qiu, M., and Xu, C. Newton downhill optimizer with application to engineering optimization and breast cancer feature selection. Biomedical Signal Processing and Control, 117:109184, 2026. ISSN 1746-8094. doi: URL [ [science/article/pii/S1746809425016957](
+- Xiong, W., Zhao, W., Yuan, W., Golovneva, O., Zhang, T., Weston, J., and Sukhbaatar, S. Stepwiser: Stepwise generative judges for wiser reasoning, 2025. URL <
+- <span id="page-9-9"></span>Xu, R., Chen, J., Ye, J., Wu, Y., Yan, J., Yang, C., and Yu, H. Incentivizing agentic reasoning in llm judges via tool-integrated reinforcement learning, 2025. URL <
+- <span id="page-9-13"></span>Yang, A. et al. Qwen2.5 technical report, 2025.
+- <span id="page-9-17"></span>You, R., Cai, H., Zhang, C., Xu, Q., Liu, M., Yu, T., Li, Y., and Li, W. Agent-as-a-judge, 2026. URL [https:]( [//arxiv.org/abs/2601.05111](
+- <span id="page-9-6"></span>Yun, J., Sohn, J., Park, J., Kim, H., Tang, X., Shao, Y., Koo, Y., Ko, M., Chen, Q., Gerstein, M., Moor, M., and Kang, J. Med-prm: Medical reasoning models with stepwise, guideline-verified process rewards, 2025. URL [https:]( [//arxiv.org/abs/2506.11474](
+- <span id="page-9-1"></span>Zhang, H., Lou, Q., and Wang, Y. Towards safe ai clinicians: A comprehensive study on large language model jailbreaking in healthcare, 2025. URL [https:]( [//arxiv.org/abs/2501.18632](
+- <span id="page-9-14"></span>Zhang, K., Zeng, S., Hua, E., Ding, N., Chen, Z.-R., Ma, Z., Li, H., Cui, G., Qi, B., Zhu, X., Lv, X., Jinfang, H., Liu, Z., and Zhou, B. Ultramedical: Building specialized generalists in biomedicine, 2024. URL <
+- <span id="page-9-18"></span>Zhao, X., Liu, S., Yang, S.-Y., and Miao, C. Medrag: Enhancing retrieval-augmented generation with knowledge graph-elicited reasoning for healthcare copilot, 2025. URL <
+- <span id="page-9-11"></span>Zuo, Y., Qu, S., Li, Y., Chen, Z., Zhu, X., Hua, E., Zhang, K., Ding, N., and Zhou, B. Medxpertqa: Benchmarking expert-level medical reasoning and understanding, 2025. URL <
+{10}------------------------------------------------
+# A. Limitation
+While Med-TIV demonstrates substantial improvements over existing medical reasoning verification approaches, several limitations warrant discussion and suggest directions for future research.
+Process Supervision. Our current training paradigm relies solely on trace-level outcome rewards, providing no supervision on intermediate verification behaviors such as when to search, what queries to formulate, or how to integrate retrieved evidence. While this design eliminates the need for costly step-level annotations, it may lead to suboptimal search patterns or redundant retrieval operations. Future work could explore supervision for the verification task itself, or leverage techniques such as search behavior cloning from stronger models to provide denser optimization signals.
+Retrieval Corpus Coverage. Med-TIV's verification accuracy is inherently bounded by the coverage and quality of the underlying medical corpus. Our retrieval system indexes documents from PubMed abstracts and medical textbooks, which provides broad coverage of established medical knowledge but may lack recent findings, rare disease information, or region-specific clinical guidelines. Verification of reasoning traces involving cutting-edge treatments or highly specialized subspecialties may be limited by corpus gaps.
+Language and Domain Scope. All training and evaluation are conducted on English-language medical reasoning benchmarks. The generalization of Med-TIV to multilingual medical content or non-Western medical traditions remains unexplored. Additionally, while our benchmarks span multiple medical subdomains, certain specialized areas such as genomics, radiology interpretation, and surgical planning may require domain-adapted retrieval corpora for optimal verification performance.
+# B. Additional Implementation Details
+#### <span id="page-10-1"></span>B.1. Hyperparameter Settings
+<span id="page-10-2"></span>Table [4](#page-10-2) provides comprehensive hyperparameter configurations for Med-TIV training across both iterations. We maintain mostly consistent settings between iterations to isolate the effect of iterative training from hyperparameter tuning.
+| Hyperparameters | Iteration 1 | Iteration 2 |
+|------------------------------|-------------|-------------|
+| RL Algorithm | Dr.GRPO | Dr.GRPO |
+| Clip ratio (low / high) | 0.2 / 0.3 | 0.2 / 0.3 |
+| Learning rate | 1e-6 | 1e-6 |
+| Warmup steps | 10 | 10 |
+| Training epochs | 5 | 5 |
+| Global batch size | 256 | 256 |
+| Mini-batch size | 256 | 256 |
+| Group size (G) | 5 | 8 |
+| Rollout sampling temperature | 1.0 | 1.0 |
+| Rollout top-p | 0.95 | 0.95 |
+| Curriculum filtering | Enabled | Enabled |
+*Table 4.* Hyperparameter configurations for Med-TIV training across iterations.
+#### <span id="page-10-0"></span>B.2. Retrieval Setup
+We construct our retrieval infrastructure using a dense retrieval architecture optimized for medical domain queries. The corpus is derived from the MedRAG [\(Zhao et al.,](#page-9-18) [2025\)](#page-9-18) collection, specifically combining the PubMed and Textbooks subcorpora into a unified index. The PubMed subset contains approximately 23.9 million biomedical abstracts covering research publications, while the Textbooks subset includes content from standard medical textbooks spanning clinical medicine, pharmacology, pathology, and related disciplines. After deduplication and quality filtering, the combined corpus contains approximately 24 million snippets.
+We employ MedCPT [\(Jin et al.,](#page-8-16) [2023\)](#page-8-16) as our dense retrieval encoder, specifically the query encoder variant for encoding
+{11}------------------------------------------------
+*Table 5.* Prompt template.
+<span id="page-11-0"></span>605
+### User Prompt:
+You are a reasoning validator for medical problems. Your task is to think step by step and evaluate whether the given reasoning trace of a medical problem contains errors.
+First, you must always perform a step-by-step analysis to examine the entire reasoning process. Then, based on your analysis, you will make a definitive judgment.
+- Use 1 if the reasoning trace is free of errors.
+- Use 0 if the reasoning trace contains one or more errors.
+#### Output Instruction:
+You must conduct your step-by-step analysis inside <think> and </think> first every time you get new information. After reasoning, if you find you lack some knowledge, you can call a search engine by <search> query </search> and it will return the top searched results between <information> and </information>. You can search as many times as you want. If you find no further external knowledge needed, you can directly provide the answer inside <answer> and </answer>, without detailed illustrations.
+```
+Medical Problem:
+{The full Medical Problem on one or more lines.}
+Reasoning Trace:
+{The full Reasoning Trace on one or more lines.}
+```
+630 631 search queries and article encoder for encoding corpus snippets. Document embeddings are pre-computed and stored in a FAISS index using the Flat configuration for maximum retrieval accuracy, distributed across multiple GPUs using FAISS's GPU sharding capability to enable parallel similarity search. For each search query, we retrieve the top-3 most relevant documents for both training and inference.
+### <span id="page-11-2"></span>B.3. Baseline Setup
+We describe the configuration of reward model baselines used in our experiments. For Med-PRM, which employs static retrieval-augmented generation, we equip it with the same retrieval corpus, encoder, and top-k setting as our framework to ensure a controlled comparison. MedS<sup>3</sup> does not support external tool invocation and is therefore evaluated without retrieval augmentation. For confidence score extraction and inference hyperparameter settings, we follow the configurations specified in each baseline's original publication.
+#### B.4. Prompt Template
+We design a structured prompt template that guides the verifier through systematic reasoning with explicit tool invocation syntax. The complete prompt is shown in Table [5.](#page-11-0)
+# <span id="page-11-1"></span>C. Benchmarks and Baselines
+### C.1. Benchmarks
+We evaluate Med-TIV on four established medical reasoning benchmarks that collectively assess verification capability across varying difficulty levels and medical subdomains.
+• MedQA [\(Jin et al.,](#page-8-6) [2020\)](#page-8-6): A dataset of multiple-choice questions derived from the United States Medical Licensing Examination (USMLE), designed to evaluate clinical reasoning and medical knowledge integration across diverse specialties.
+• MedMCQA [\(Pal et al.,](#page-9-10) [2022\)](#page-9-10): A large-scale multi-subject benchmark sourced from Indian medical entrance examinations (AIIMS and NEET-PG), covering 21 medical subjects with emphasis on factual knowledge and clinical application.
+• MMLU-Med [\(Hendrycks et al.,](#page-8-7) [2021\)](#page-8-7): An aggregation of medical-related subsets from the Massive Multitask Language Understanding benchmark, encompassing anatomy, clinical knowledge, college biology, college medicine, medical
+{12}------------------------------------------------
+genetics, and professional medicine.
+• MedXpertQA [\(Zuo et al.,](#page-9-11) [2025\)](#page-9-11): An expert-level benchmark featuring challenging questions that require multi-step clinical reasoning, differential diagnosis, and treatment planning at the level expected of practicing physicians.
+#### C.2. Baselines
+We compare Med-TIV against comprehensive baselines spanning proprietary systems, general-purpose models, and domainspecialized approaches.
+### Proprietary Models.
+- GPT-4o-mini [\(OpenAI et al.,](#page-9-12) [2024\)](#page-9-12): A compact variant of OpenAI's GPT-4o optimized for efficiency while maintaining strong reasoning capabilities across diverse tasks.
+- Gemini-2.0-Flash: Google's efficient multimodal model designed for fast inference with competitive performance on knowledge-intensive benchmarks.
+#### General Reasoning Models.
+- DeepSeek-R1 [\(Guo et al.,](#page-8-9) [2025\)](#page-8-9): A 671B parameter reasoning model trained with RL, representing the current frontier of open-weight reasoning capabilities.
+- R1-Distill-Qwen / R1-Distill-Llama: Distilled variants of DeepSeek-R1 at 7B and 8B scales respectively, designed to transfer reasoning capabilities to smaller architectures.
+#### General Foundation Models.
+- Qwen2.5 [\(Yang et al.,](#page-9-13) [2025\)](#page-9-13): A family of open-weight language models with strong multilingual and reasoning capabilities, evaluated at 7B and 32B parameter scales.
+- Llama3.1 [\(Grattafiori et al.,](#page-8-10) [2024\)](#page-8-10): Meta's open-source foundation model demonstrating competitive performance across diverse benchmarks, evaluated at the 8B scale.
+## Medical Domain Models.
+- AlphaMed [\(Liu et al.,](#page-8-11) [2025a\)](#page-8-11): A medical reasoning model that employs RL with rule-based rewards to enhance clinical reasoning without reliance on distillation from larger models.
+- UltraMedical [\(Zhang et al.,](#page-9-14) [2024\)](#page-9-14): A specialized medical model combining high-quality instruction tuning on curated biomedical corpora with preference optimization for improved clinical accuracy.
+- HuatuoGPT-o1 [\(Chen et al.,](#page-8-12) [2024\)](#page-8-12): A medical reasoning model incorporating chain-of-thought reasoning with internal verification mechanisms to improve diagnostic accuracy.
+#### Medical Reward Models.
+- MedS<sup>3</sup> [\(Jiang et al.,](#page-8-1) [2025b\)](#page-8-1): A self-evolved soft dual-sided process supervision framework for medical reasoning that generates training signals through iterative self-improvement without external annotations.
+- Med-PRM [\(Yun et al.,](#page-9-6) [2025\)](#page-9-6): A process reward model for medical reasoning verification that provides step-level supervision using static retrieval-augmented generation with guideline-based verification.
+# <span id="page-12-0"></span>D. Case Analysis
+713 714 Table [6](#page-13-0) presents a complete verification example illustrating how a Med-TIV trained verifier identifies reasoning errors through dynamic evidence retrieval. The case involves a patient with bladder cancer who develops ototoxicity following chemotherapy. The generator's reasoning trace incorrectly attributes the symptoms to taxanes based on their known association with ototoxicity, concluding with answer (B). However, the model retrieves evidence establishing that cisplatin—the standard neoadjuvant therapy for transitional cell carcinoma—is the causative agent, and its mechanism involves DNA cross-linking rather than microtubule hyperstabilization. Through iterative search and reasoning, verifier correctly identifies the error, demonstrating the value of tool augmentation for catching subtle medical reasoning mistakes.
+{13}------------------------------------------------
+<span id="page-13-0"></span>*Table 6.* Complete verification demonstration. Given a medical problem and a reasoning trace, the verifier retrieves relevant evidence and correctly identifies the reasoning error within the trace.
+#### Medical Problem:
+Question: A 67-year-old man with transitional cell carcinoma of the bladder comes to the physician because of a 2-day history of ringing sensation in his ear. He received this first course of neoadjuvant chemotherapy 1 week ago. Pure tone audiometry shows a sensorineural hearing loss of 45 dB. The expected beneficial effect of the drug that caused this patient's symptoms is most likely due to which of the following actions?
+Options: (A) Inhibition of proteasome (B) Hyperstabilization of microtubules (C) Generation of free radicals (D) Cross-linking of DNA
+### Generator Reasoning Trace:
+The patient's presentation of ringing in the ear, known as tinnitus, and a sensorineural hearing loss of 45 dB, suggest ototoxicity. Ototoxicity is a well-known side effect of certain chemotherapy drugs. Neoadjuvant chemotherapy refers to the use of chemotherapy before other treatments such as surgery. The patient has received a course of neoadjuvant chemotherapy 1 week ago, which aligns with the recent onset of ototoxic symptoms. To determine the cause of ototoxicity from the provided options, we need to identify which of the options is likely to cause sensorineural hearing loss. Option (A), Inhibition of proteasome, is associated with drugs like bortezomib (Velcade), which are less commonly associated with ototoxicity. Option (B), Hyperstabilization of microtubules, is a known mechanism of taxanes (e.g., paclitaxel, docetaxel). Taxanes are known to cause ototoxicity, making this a potential candidate. Option (C), Generation of free radicals, is a general mechanism that can cause cellular damage but is not specifically known to be associated with ototoxicity. Option (D), Cross-linking of DNA, is a known mechanism of alkylating agents like cyclophosphamide. While alkylating agents can cause ototoxicity, this mechanism is not as commonly linked to ototoxicity as other options. Considering the options, the one most commonly associated with ototoxicity is hyperstabilization of microtubules. Therefore, the drug that most likely caused this patient's symptoms is one that hyperstabilizes microtubules. The answer is (B).
+# Verifier Judgment:
+<think> Let's check the common chemotherapeutic agents used for bladder cancer and their known side effects. Transitional cell carcinoma of the bladder is often treated with cisplatin-based chemotherapy as neoadjuvant therapy. We need to determine if any of the options match the known side effects of these drugs. Let's search to confirm the side effects of cisplatin and other common bladder cancer chemotherapy drugs. </think>
+- <search> common side effects of cisplatin chemotherapy </search>
+- <information> [Search Results regarding Cisplatin toxicity and mechanisms...] </information>
+- <think> From the information, cisplatin is known to cause ototoxicity, which aligns with the patient's symptoms of sensorineural hearing loss. Now let's check the mechanisms of action for the options: Option (A): Inhibition of proteasome - This is associated with drugs like bortezomib and is not used for bladder cancer. Option (B): Hyperstabilization of microtubules - This is the mechanism of taxanes (e.g., paclitaxel, docetaxel). Option (C): Generation of free radicals - This is not specific to a single mechanism and is not commonly associated with ototoxicity. Option (D): Cross-linking of DNA - This is the mechanism of alkylating agents (e.g., cyclophosphamide). Cisplatin is an alkylating agent, and its mechanism involves cross-linking of DNA. Therefore, the correct answer should be (D). </think>
+<answer>0</answer>