text string | source string |
|---|---|
arXiv:2505.20315v1 [cs.CL] 22 May 2025Arctic-Text2SQL-R1: Simple Rewards, Strong Reasoning in Text-to-SQL Zhewei Yao1,∗, Guoheng Sun2, Lukasz Borchmann1, Zheyu Shen2, Minghang Deng3, Bohan Zhai1, Hao Zhang1,3, Ang Li2, Yuxiong He1 1Snowflake AI Research,2University of Maryland, Colleage Park, 3University of California,... | https://arxiv.org/abs/2505.20315v1 |
lightweight RL formulation using execution-only rewards that generalizes well across model sizes and benchmarks, enabling stable training and strong performance without complex reward design. •State-of-the-Art Performance Across Benchmarks: Arctic-Text2SQL-R1 achieves #1 ex- ecution accuracy on the BIRD leaderboard, wi... | https://arxiv.org/abs/2505.20315v1 |
multiple plausible solutions, enabling the model to build a deeper understanding of what constitutes high-quality reasoning [ 18]. For example, GRPO has enabled models like DeepSeek-R1-Zero to develop complex reasoning skills such as multi-step chain-of- thought and self-verification simply by being rewarded for correc... | https://arxiv.org/abs/2505.20315v1 |
robust policy improvement. The GRPO objective is as follows: JGRPO(θ) =E1 NNX i=1min ( riAi,clip(ri,1−ϵ,1 +ϵ)) −βDKL(πθ∥πref), where ri=πθ(oi|Q) πθold(oi|Q)is the likelihood ratio, Aithe advantage, and DKLis a KL-divergence penalty to keep the policy close to a reference (supervised fine-tuned) model [ 33]. In practi... | https://arxiv.org/abs/2505.20315v1 |
SPIDER, Gretel-Synth-NonFiltered 64.6 86.4 14B-Inst BIRD, SPIDER, Gretel-Synth-Filtered 66.5 88.3 32B-Inst BIRD, SPIDER, Gretel-Synth-Filtered 64.9 87.7 32B-Inst BIRD, SPIDER, Gretel-Synth-Filtered, BIRD-aug-NonFiltered 62.5 86.1 32B-Inst BIRD, SPIDER, Gretel-Synth-Filtered, BIRD-aug-Filtered 64.9 86.8 Table 5: Learnin... | https://arxiv.org/abs/2505.20315v1 |
11.9 79.1 35.5 51.8 55.7 4.2 Learnings from Training Strategy Unless stated otherwise, the following rely on filtered BIRD, SPIDER, and Gretel-Synth. GRPO vs PPO. We compare GRPO and Proximal Policy Optimization (PPO). GRPO, by design, reduces memory usage of critic models and is well-suited for large-scale settings; P... | https://arxiv.org/abs/2505.20315v1 |
[ 54]. This highlights the risk of overfitting and the necessity of evaluating on multiple datasets to ensure robust generalization. 6 Table 7: Comparison between Arctic-Text2SQL-R1 and other OSS/Proprietary models. *: SQL- R1 [26] uses majority voting in evaluation. **: [35, Table 2] might use schema linking, our numb... | https://arxiv.org/abs/2505.20315v1 |
The reinforcement learning setup consists of (1) GRPO as the optimiza- tion algorithm, (2) online training, (3) initializing from OmniSQL supervised checkpoints, and (4) using a modified OmniSQL prompt for both training and evaluation. We report execution-based accuracy results across all six benchmarks detailed in Tab... | https://arxiv.org/abs/2505.20315v1 |
in [ 35, Table 2], but we are unsure; see Table 2 for the reward design comparison. Method Base Model Filtering MethodReward BIRD SPIDER SPIDER Design -dev -test -DK Reasoning-SQL Qwen2.5-Coder-14B-Inst Large Models Prompt Complex 64.21 81.43 73.03 Arctic-Text2SQL-R1 Qwen2.5-Coder-14B-Inst Empty Golden-SQL Simple 66.49... | https://arxiv.org/abs/2505.20315v1 |
model—setting a new state-of-the-art and surpassing much larger LLMs. Across six challenging benchmarks, Arctic-Text2SQL-R1 yields up to 4 points improvement over strong baselines, with notable parameter efficiency (e.g., the 7B version matches or outperforms prior ExCoT-70B models and GPT4o). Our results highlight tha... | https://arxiv.org/abs/2505.20315v1 |
Vittorio Castelli, and Bing Xiang. Importance of syn- thesizing high-quality data for text-to-SQL parsing. In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki, editors, Findings of the Association for Computational Linguistics: ACL 2023 , pages 1327–1343, Toronto, Canada, July 2023. Association for Computational Lin... | https://arxiv.org/abs/2505.20315v1 |
Kendrick Boyd, Maarten Van Segbroeck, Matthew Grossman, Piotr Mlocek, and Drew Newberry. Synthetic-Text-To- SQL: A synthetic dataset for training language models to generate sql queries from natural language prompts, April 2024. 11 [28] Mayank Mishra, Matt Stallone, Gaoyuan Zhang, Yikang Shen, Aditya Prasad, Adriana Me... | https://arxiv.org/abs/2505.20315v1 |
Jingyi Wang, Xiaochong Lan, Jiahui Gong, Tianjian Ouyang, Fanjin Meng, Chenyang Shao, Yuwei Yan, Qinglong Yang, Yiwen Song, Sijian Ren, Xinyuan Hu, Yu Li, Jie Feng, Chen Gao, and Yong Li. Towards large reasoning models: A survey of reinforced reasoning with Large Language Models, 2025. [48] An Yang, Baosong Yang, Beich... | https://arxiv.org/abs/2505.20315v1 |
the correctness of the <SQL Query>. Therefore, please help me add one INSERT statement for each table in the <SQL Context>, with 5 sample rows per table. The inserted data should ensure that the <SQL Query> can retrieve results from the tables. Please ensure that it does not cause errors when using sqlite3 . Please do ... | https://arxiv.org/abs/2505.20315v1 |
Table information: Table name: {table name} Column name: {column name} Column description: {column desc} Sample rows: {samples} Optional(Task: {task}. The answer to the task is: {answer}.) Based on this, write 10 more complex nested SQLite SQL queries or SQLs with CTEs in sql code block format. You can use any informat... | https://arxiv.org/abs/2505.20315v1 |
the question. User: Database Engine: SQLite Database Schema: {Database Schema} This schema describes the database’s structure, including tables, columns, primary keys, foreign keys, and any relevant relationships or constraints. Question: {evidence + question} Instructions: - Make sure you only output the information t... | https://arxiv.org/abs/2505.20315v1 |
free rate, which is calculated as ‘Free Meal Count (K-12)‘ divided by ‘Enrollment (K-12)‘. The resulting query is efficient and meets the requirements of the natural language question. </answer> Figure D.1: Example I of response from Arctic-Text2SQL-R1-32B 19 Example II of response from Arctic-Text2SQL-R1-32B Question:... | https://arxiv.org/abs/2505.20315v1 |
both tables. 5. The final query needs to join the ‘frpm‘ and ‘schools‘ tables on their common ‘CDSCode‘ key, filter the rows for the specific administrator, and then compute and return the eligible free rate for students aged 5-17. </think> <answer> ```sql SELECT (T1.‘Free Meal Count (Ages 5-17)‘ / T1.‘Enrollment (Ages... | https://arxiv.org/abs/2505.20315v1 |
arXiv:2505.20318v1 [cs.CL] 23 May 2025Beyond Demonstrations: Dynamic Vector Construction from Latent Representations Wang Cai1,2*, Hsiu-Yuan Huang1,3*, Zhixiang Wang1,2, Yunfang Wu1,3† 1National Key Laboratory for Multimedia Information Processing, Peking University 2School of Software and Microelectronics, Peking Univ... | https://arxiv.org/abs/2505.20318v1 |
2024), as illustrated in Figure 1. ICV methods have shown promise in various applica- tions—such as promoting honesty (Li et al., 2023), reducing harmful outputs (Liu et al., 2024), and en- abling role-playing (Potertì et al., 2025)—all while maintaining the efficiency of zero-shot inference. While ICV presents a promi... | https://arxiv.org/abs/2505.20318v1 |
— each inference involves a full forward pass over lengthy prompts with repeated demonstra- tions, leading to high memory and compute costs, especially in resource-limited settings (Liu et al., 2022); (2) Instability — performance is highly sen- sitive to prompt design, including example order and selection (Liu et al.... | https://arxiv.org/abs/2505.20318v1 |
model where these vectors are injected. Existing approaches typically extract ICVs by running ICL with randomly sampled demonstra- tions and averaging the resulting representations. While this mitigates the model’s sensitivity to fac- tors such as demonstration position and composi- tion, it does not fully resolve the ... | https://arxiv.org/abs/2505.20318v1 |
fine-grained, task- centric control. 4.2 Dynamic Latent Segmentation and Injection To extract task-specific information in a more dy- namic manner, we re-partition the latent repre- sentation o(i)intoSsegments, where S∈ {s| dmod s= 0}denotes the set of values that evenly divide the hidden dimension dof the LLM. Weobtai... | https://arxiv.org/abs/2505.20318v1 |
corresponding labels in Dtrainare drawn directly from the original example set, allowing DyVec to operate effectively even under extremely limited supervision. In the next section, we elaborate on how to select the optimal intervention function L from a set of candidate strategies.4.3 Dynamic Vector Injection for Infer... | https://arxiv.org/abs/2505.20318v1 |
remain substantial at 8.5% and 31.5%, respectively. These results confirm DyVec’s strong generalization abil- ity in low-resource scenarios. Data Size ICL LoRA DyVec 4 51.17 (↑8.5%) 39.78 (↑39.5%) 55.51 8 56.13 (↑8.5%) 46.31 (↑31.5%) 60.93 16 65.04 (↑0.1%) 52.69 (↑23.6%) 65.11 Table 3: Average performance across models... | https://arxiv.org/abs/2505.20318v1 |
85.00 24.10 55.08 8ICL 34.50 56.00 94.90 56.80 75.90 26.40 57.42 LoRA 50.90 64.40 53.70 53.00 36.70 6.90 44.27 DyVec 53.90 50.50 92.70 60.10 80.40 25.20 60.47 16ICL 63.40 79.60 94.20 53.30 87.90 29.30 67.95 LoRA 66.70 72.60 53.70 57.70 51.30 7.80 51.63 DyVec 67.00 50.50 91.00 58.70 68.50 26.20 60.32 DeepSeek4ICL 39.60 ... | https://arxiv.org/abs/2505.20318v1 |
Results are averaged over three tasks with 8-shot data. Solid lines represent perfor- mance using different numbers of randomly constructed prompts ( N= 1,50,100), while dashed lines indicate performance using EQR strategy. As shown in Figure 4, EQR consistently achieves the highest average performance across all model... | https://arxiv.org/abs/2505.20318v1 |
performance, DyVec offers a lightweight, generalizable framework for vector- based intervention, deepening our understanding of latent task representations in LLMs. Limitations While DyVec demonstrates promising results on classification and lexical generation tasks, it still has several limitations. First, the current... | https://arxiv.org/abs/2505.20318v1 |
OxAISH-AL-LLM. 2023. Trec-6 classification dataset. https://huggingface.co/datasets/ OxAISH-AL-LLM/trec6 . Accessed: 2025-05-09. Daniele Potertì, Andrea Seveso, and Fabio Mercorio. 2025. Designing role vectors to improve llm infer- ence behaviour. Preprint , arXiv:2502.12055. Ohad Rubin, Jonathan Herzig, and Jonathan B... | https://arxiv.org/abs/2505.20318v1 |
\n A: This template is applied uniformly across all classification and generation tasks, enabling a fair comparison between few-shot and intervention- based methods. B Data Construction Details For all classification tasks used in our experiments, we construct training data with an emphasis on diversity , aiming to hel... | https://arxiv.org/abs/2505.20318v1 |
Time To further understand the practical efficiency of DyVec, we compare the inference time of different adaptation methods across LLaMA, Qwen, and DeepSeek models. All evaluations are conducted on the same hardware with a fixed test set size of 1000 examples per task, ensuring a fair comparison of decoding speed. Resu... | https://arxiv.org/abs/2505.20318v1 |
in Table 9, EQR consistently improves performance across multiple datasets and model Model Method NHSD Sarcasm SST2 ADE AG_News TREC6 Avg llamaEQR 49.00 50.40 91.50 67.40 79.20 25.40 60.48 N = 1 46.70 58.70 77.50 58.20 58.90 34.30 55.72 N = 50 49.10 50.10 57.10 54.40 57.70 24.10 48.75 N = 100 50.60 50.50 55.10 50.00 25... | https://arxiv.org/abs/2505.20318v1 |
42.03 60.81 64.04 59.34 16.12 8 49.74 58.22 91.75 67.40 74.46 18.13 16 44.24 42.03 84.49 41.99 68.39 18.26 164 58.66 60.74 84.93 59.42 53.38 21.01 8 58.02 76.89 92.22 58.67 72.51 19.13 16 63.17 82.68 89.90 70.31 74.27 20.18 Qwen44 57.50 53.55 68.67 62.32 84.81 16.06 8 43.29 45.20 79.14 58.60 73.90 20.48 16 45.32 45.09 ... | https://arxiv.org/abs/2505.20318v1 |
arXiv:2505.20320v1 [cs.CL] 23 May 2025Less Context, Same Performance: A RAG Framework for Resource-Efficient LLM-Based Clinical NLP Satya Narayana Cheetirala1, Ganesh Raut1, Dhavalkumar Patel1, Fabio Sanatana1, Robert Freeman1, Matthew Levin2,4, Girish N. Nadkarni3,4, Omar Dawkins6, Reba Miller5, Randolph M. Steinhagen... | https://arxiv.org/abs/2505.20320v1 |
Standard Transformer-based models scale quadratically in both time and memory with respect to input sequence length due to their self-attention mechanism [1], [2]. Recent variants employing optimized attention mechanisms, such as Flash Attention [3], mitigate the quadratic memory complexity but still require sub- stant... | https://arxiv.org/abs/2505.20320v1 |
medical center, tertiary care urban hospitals, and community hospitals. Prior to beginning the study, we obtained a waiver of informed consent from our Institutional Review Board (IRB) (IRB-18- 00573-MODCR001). The Department of Surgery maintains a Mortality and Mor- bidity M/&M data registry, which provides a comprehe... | https://arxiv.org/abs/2505.20320v1 |
rapid retrieval in high- dimensional spaces and simplifies large-scale data handling for downstream tasks. Fig. 5. RAG-based classification. The top- Nchunks from the FAISS index are concatenated (under 4,000 tokens) and fed into an LLM classifier. b) Classification with RAG (Figure 5): Once the vec- tor database is es... | https://arxiv.org/abs/2505.20320v1 |
(0.61), despite the RAG method using significantly fewer tokens. Similar patterns were observed for the LLaMA and Mistral models.TABLE I PERFORMANCE RESULTS FOR POST-OPERATIVE COMPLICATION CLASSIFICATION Experiment AUROC Precision Recall F1 PR AUC GPT4o-RAG 0.67 0.53 0.71 0.61 0.64 GPT4o-Long Context 0.66 0.46 0.90 0.6... | https://arxiv.org/abs/2505.20320v1 |
and that targeted retrieval of key segments can suffice. For healthcare decision-support systems, RAG provides a viable path to scaling NLP solutions without requiring extremely large context windows. 3) Potential for Broader Applicability: While this study focused on surgical complication identification, the RAG appro... | https://arxiv.org/abs/2505.20320v1 |
of Retrieval-Augmented Generation (RAG): Evolution, Current Landscape and Future Directions,” arXiv preprint arXiv:2410.12837 , 2024. [9] R. Qu, R. Tu, and F. Bao, “Is Semantic Chunking Worth the Computa- tional Cost?,” arXiv preprint arXiv:2410.13070 , 2024. [10] Y . Wang et al. , “A comparison of word embeddings for ... | https://arxiv.org/abs/2505.20320v1 |
searches in high- dimensional spaces. Vector EmbeddingsVector Embeddings Numerical representations of text capturing semantic meaning. Chunking Chunking Splitting lengthy documents into smaller segments (e.g., 512 tokens). Transformer- based EncodersTransformer-based Encoders Models that convert text seg- ments into ve... | https://arxiv.org/abs/2505.20320v1 |
arXiv:2505.20321v1 [cs.CL] 23 May 2025BiomedSQL: Text-to-SQL for Scientific Reasoning on Biomedical Knowledge Bases Mathew J. Koretsky1,2Maya Willey1,2Adi Asija2,3Owen Bianchi1,2 Chelsea X. Alvarado1,2Tanay Nayak2,3Nicole Kuznetsov1,2Sungwon Kim2,3 Mike A. Nalls1,2,4Daniel Khashabi2,3*Faraz Faghri1,2,4* 1Center for Alz... | https://arxiv.org/abs/2505.20321v1 |
scientific reasoning on large- scale tabular data that is required for interpreting biomedical data—such as inferring statistical significance thresholds or chaining multi-step filtering logic across ontologies that a skilled biomedical analyst would employ. To address this gap, we introduce BiomedSQL , the first bench... | https://arxiv.org/abs/2505.20321v1 |
limited by its narrow schema and templated query structure. EHRSQL [4] advanced realism by crowdsourcing 1,000+ natural language queries from clinicians across MIMIC-III and eICU, highlighting challenges in temporal logic, answerability, and data sparsity. Recent datasets have diversified query paradigms across relatio... | https://arxiv.org/abs/2505.20321v1 |
we included summary statistics from large-scale GWAS studies of Alzheimer’s disease [ 19] and Parkinson’s disease [ 20], obtained from the GWAS Catalog. We retained SNP-level data including p-values, rsIDs, allele frequencies, and nearest-gene mappings after quality control filtering. We further integrated causal infer... | https://arxiv.org/abs/2505.20321v1 |
1.Operationalizing implicit scientific conventions: Queries often invoke domain-specific concepts (e.g., "significantly associated SNPs") that imply non-obvious statistical thresholds, such as p <5×10−8for GWAS hits or directionality based on beta coefficients. These conventions are rarely explicit in schemas and must ... | https://arxiv.org/abs/2505.20321v1 |
sample table rows ( 3-rows ,5-rows ) – Adding few-shot examples ( 1-shot ,3-shot ,5-shot ) – Adding explicit domain-specific instructions (e.g., statistical thresholds via stat-instruct ) – A combined variant that includes 3-rows ,3-shot , and stat-instruct (combo ) Prompt templates are provided in the Appendix A.3. In... | https://arxiv.org/abs/2505.20321v1 |
Evaluation Results LLMs struggle with scientific reasoning in SQL generation. Table 3 shows that even top- performing models such as GPT-o3-mini (EX = 53.5%, JAC = 60.4%, RQR = 73.3%) fall short of domain expert performance (90–95%). GPT-4o performs slightly lower (EX = 46.9%, JAC = 54.7%, RQR = 71.2%). Among open mode... | https://arxiv.org/abs/2505.20321v1 |
(±0.0) 13,317 ReAct-Gemini* 48.9 (±4.2) 56.6 (±3.8) 60.4 (±4.1) 10.2 (±2.5) 0.0 (±0.0) 13,205 Index-GPT-4o 25.5 (±3.6) 28.3 (±3.6) 44.1 (±4.2) 66.9 (±3.9) 27.5 (±3.7) 1,110 Index-GPT-o3-mini 27.1 (±3.7) 30.6 (±3.7) 44.1 (±4.1) 47.5 (±4.2) 2.0 (±0.1) 1,899 Index-Gemini* 46.1 (±4.2) 48.5 (±4.1) 54.2 (±4.2) 59.6 (±4.1) 8.... | https://arxiv.org/abs/2505.20321v1 |
they do not represent the only correct way to retrieve relevant data for a given question. Biomedical questions often permit multiple semantically valid formulations, e.g., using alternative joins, filters, or aggregations. To account for this, we evaluate models using a combination of metrics, including execution-base... | https://arxiv.org/abs/2505.20321v1 |
language using reinforcement learning. In ACL, 2017. [6]Chia-Hsuan Lee, Oleksandr Polozov, and Matthew Richardson. KaggleDBQA: Realistic evaluation of text-to-SQL parsers. arXiv preprint arXiv:2106.11455 , 2021. [7]Ping Wang, Tian Shi, and Chandan K. Reddy. Text-to-SQL generation for question answering on electronic me... | https://arxiv.org/abs/2505.20321v1 |
1474-4422. doi: https://doi.org/10.1016/S1474-4422(19)30320-5. URL https: //www.sciencedirect.com/science/article/pii/S1474442219303205 . [21] Chelsea X Alvarado, Mary B Makarious, Cory A Weller, Dan Vitale, Mathew J Koretsky, Sara Bandres-Ciga, Hirotaka Iwaki, Kristin Levine, Andrew Singleton, Faraz Faghri, et al. omi... | https://arxiv.org/abs/2505.20321v1 |
in a European population (Nalls et al., 2019). Schema: - Name: UUID | Type: STRING | Mode: REQUIRED - Name: SNP | Type: STRING | Mode: NULLABLE - Name: A1 | Type: STRING | Mode: NULLABLE - Name: A2 | Type: STRING | Mode: NULLABLE - Name: freq | Type: FLOAT | Mode: NULLABLE - Name: b | Type: FLOAT | Mode: NULLABLE - Nam... | https://arxiv.org/abs/2505.20321v1 |
p_SMR_multi | Type: FLOAT | Mode: NULLABLE - Name: p_HEIDI | Type: FLOAT | Mode: NULLABLE - Name: nsnp_HEIDI | Type: FLOAT | Mode: NULLABLE - Name: topRSID | Type: STRING | Mode: NULLABLE - Name: Omic_type | Type: STRING | Mode: NULLABLE - Name: Omic_tissue | Type: STRING | Mode: NULLABLE - Name: Disease_name | Type: S... | https://arxiv.org/abs/2505.20321v1 |
patent_no | Type: STRING | Mode: NULLABLE - Name: max_phase_for_ind | Type: FLOAT | Mode: NULLABLE - Name: mesh_id | Type: STRING | Mode: NULLABLE - Name: mesh_heading | Type: STRING | Mode: NULLABLE - Name: efo_id | Type: STRING | Mode: NULLABLE - Name: efo_term | Type: STRING | Mode: NULLABLE - Name: tradeNames_list ... | https://arxiv.org/abs/2505.20321v1 |
Significance Identifies variants that are GWAS significant for a disease (p < 5e-08). SMR Significance Identifies variants that are SMR significant for a disease (p < 2.95e-06). Functional Significance Identifies variants that are significant within a particular tissue for a disease. Effect Retrieves the effect size an... | https://arxiv.org/abs/2505.20321v1 |
drugName, drugIsApproved, newLinkedTargets_string FROM ‘card-ai-389220.bio_sql_benchmark.DrugGeneTargets_ComprehensiveAnnotations_updated‘ WHERE newLinkedTargets_string LIKE "%TUBB %" AND drugIsApproved = TRUE LIMIT 1000; Example 2: SELECT SNP, A1 AS effect_allele, freq AS effect_allele_frequency, A2 AS non_effect_alle... | https://arxiv.org/abs/2505.20321v1 |
fields: {{ "thought": "Your reasoning about how to answer the question", "action": "One of ’verify_sql’, ’execute_sql’, or ’final_answer’", "action_input": "For verify_sql/execute_sql: the SQL query; For final_answer: explanation of the results" }} IMPORTANT: - Your response must include valid JSON that can be parsed. ... | https://arxiv.org/abs/2505.20321v1 |
6: BMSQL prompt template for selecting relevant columns. You are a highly proficient BigQuery SQL generator in the biomedical domain. Database schema: {db_schema} The user asked: "{question}" Previously identified relevant columns/tables:{relevant_columns} Instructions: - Generate exactly one valid BigQuery SQL query t... | https://arxiv.org/abs/2505.20321v1 |
queries or answer contain errors, deem the answer as insufficient. 2. If you have any doubts about the SQL queries, execution results, or answer, deem the answer as insufficient. 3. If there are any inconsistencies between the SQL queries, execution results, and answer, deem the answer as insufficient. 4. Keep in mind ... | https://arxiv.org/abs/2505.20321v1 |
Answer: {gold_ans} Analyst Response: {pred_ans} Your grading Using the scoring instructions above, grade the Analyst Response. Return only the numeric score on a scale from 0.0-3.0. If the response is stating lack of knowledge or abstaining, give it -1.0. Please respond only with the score. Prompt 12: BioScore prompt t... | https://arxiv.org/abs/2505.20321v1 |
(±3.1) 0.0 (±0.0) 10,951 5-rows 54.2 (±4.2) 61.9 (±3.8) 76.7 (±3.5) 15.7 (±3.0) 0.2 (±0.4) 14,312 1-shot 54.0 (±4.2) 61.7 (±3.8) 73.4 (±3.7) 32.4 (±3.9) 0.4 (±0.5) 4,058 3-shot 56.0 (±4.2) 64.2 (±3.7) 74.0 (±3.7) 33.1 (±3.9) 0.0 (±0.0) 4,099 5-shot 57.9 (±4.1) 65.6 (±3.7) 75.8 (±3.6) 23.5 (±3.5) 0.0 (±0.0) 4,566 stat-i... | https://arxiv.org/abs/2505.20321v1 |
and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. •The aut... | https://arxiv.org/abs/2505.20321v1 |
well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. •If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. •Depending on the contribution, reproducibility ca... | https://arxiv.org/abs/2505.20321v1 |
and data submission guidelines ( https: //nips.cc/public/guides/CodeSubmissionPolicy ) for more details. •The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. •The authors should provide scripts to re... | https://arxiv.org/abs/2505.20321v1 |
of execution) needed to reproduce the experiments? Answer: [Yes] Justification: We provide details on compute resources needed to reproduce the experiments in Appendix A.9. Guidelines: • The answer NA means that the paper does not include experiments. •The paper should indicate the type of compute workers CPU or GPU, i... | https://arxiv.org/abs/2505.20321v1 |
in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). 11.Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a hig... | https://arxiv.org/abs/2505.20321v1 |
research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? Answer: [NA] Justification: [NA] Guidelines: •The answer NA me... | https://arxiv.org/abs/2505.20321v1 |
Beyond Prompt Engineering: Robust Behavior Control in LLMs via Steering Target Atoms Mengru Wang1,2*, Ziwen Xu1*, Shengyu Mao1, Shumin Deng3,Zhaopeng Tu2,Huajun Chen1,Ningyu Zhang1† 1Zhejiang University,2Tencent AI Lab, 3National University of Singapore, NUS-NCS Joint Lab, Singapore Abstract Precise control over langua... | https://arxiv.org/abs/2505.20322v1 |
autoen- coders (SAEs) (Gao et al., 2024; Lan et al., 2024) offer a promising approach by decomposing LLM representations into higher-dimensional, sparser features (Lieberum et al., 2024a). This aligns with theoretical analyses of language model parameter spaces as linear projections of knowledge mani- folds, where poly... | https://arxiv.org/abs/2505.20322v1 |
(1) where Mis the model and yis the output. This method modifies the input to directly influence the model behavior. 2.2 Steering Steering strategy modifies the representations dur- ing the forward propagation to achieve the desired results without changing the model parameters . Specifically, given the hidden state at... | https://arxiv.org/abs/2505.20322v1 |
¯ai j,pos >0 (8) fneg j=1 NXN i=1I ¯ai j,neg >0 (9) ∆f=fpos−fneg (10) Then, we select target atoms abased on their amplitude and frequency in the high-dimensional representation space aj target=( ∆aj,if∆aj≥αand∆fj≥β. 0, otherwise . (11) This selection process ensures that the most rele- vant and impactful atoms are ... | https://arxiv.org/abs/2505.20322v1 |
Llama-3.1-8B. Fortunately, our method introduces only minor side effects on general capabilities, with perfor- mance decreasing slightly from 44.73% to 43.90% in Gemma-2-9B-pt and from 51.04% to 49.12% in Gemma-2-9B-it. Interestingly, we observe that steering strategies, including our STA and CAA, outperform prompting ... | https://arxiv.org/abs/2505.20322v1 |
50.47 SAE AXBENCH 90.74 98.42 94.58 5.43 70.89 72.63 49.65 STA (Ours) 95.78 99.33 97.56 5.43 70.27 71.65 49.12 Llama-3.1- 8BVanilla 59.78 58.38 59.08 4.04 58.10 43.97 35.37 Prompthand 63.70 57.30 60.50 3.62 58.10 46.78 36.17 Promptauto 61.63 60.64 61.14 4.03 58.10 41.55 34.56 CAA 68.67 72.81 70.74 3.89 57.64 44.35 35.2... | https://arxiv.org/abs/2505.20322v1 |
performance comparison. As shown in the lower panel of Fig. 5, the vec- tors obtained by converting the prompts using our method, denoted as STA prompt , significantly out- perform the original prompts. Similarly, the vectors derived from the prompts using the CAA method, denoted as CAA prompt , also significantly exce... | https://arxiv.org/abs/2505.20322v1 |
while preserving fluency. We have now formalized this range in the revised manuscript to provide clearer guidance.We further investigate the changes in the token distribution for steering and prompting strategies. As shown in the Fig 7, the influence of prompting on the model’s token distribution is much smaller than t... | https://arxiv.org/abs/2505.20322v1 |
remarkable flexibil- ity in adjusting reasoning length, either extend- ing or shortening it while maintaining accuracy. Furthermore, we analyze the relationship between themultiplier coefficient and the token length of reasoning. Experimental results reveal that the multiplier coefficient can flexibly control reason- i... | https://arxiv.org/abs/2505.20322v1 |
works mainly fo- cus on toy tasks, such as entity recognition, slec-tion, and verb tense or number agreement. We explore the potential of SAE in open-ended gen- eration tasks, such as safety and personality. The most related work, AXBENCH (Wu et al., 2025), steering coarse-grained directions SAE spaces. In contrast, ou... | https://arxiv.org/abs/2505.20322v1 |
Overall, we anticipate no significant ethical or societal implications arising from our research, as our primary goal is to enhance the safety and controllability of LLMs. Acknowledgments This work was supported by the National Natu- ral Science Foundation of China (No. 62206246, No. NSFCU23B2055, No. NSFCU19B2027), th... | https://arxiv.org/abs/2505.20322v1 |
time category-wise safety steering for large language models. CoRR , abs/2410.01174. Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nicholas L Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph,... | https://arxiv.org/abs/2505.20322v1 |
Askell, Kamal Ndousse, Dawn Drain, Anna Chen, Yuntao Bai, Deep Ganguli, Liane Lovitt, Zac Hatfield-Dodds, Jackson Kernion, Tom Conerly, Shauna Kravec, Stanislav Fort, Saurav Kada- vath, Josh Jacobson, Eli TranJohnson, Jared Kaplan, Jack Clark, Tom Brown, Sam McCandlish, Dario Amodei, , and Christopher Olah. 2022a. Soft... | https://arxiv.org/abs/2505.20322v1 |
Zou, Mantas Mazeika, Dawn Song, and Jacob Stein- hardt. 2021. Measuring massive multitask language understanding. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 . OpenReview.net. Robert Huben, Hoagy Cunningham, Logan Riggs, Aidan Ewart, and Lee Sharkey. 202... | https://arxiv.org/abs/2505.20322v1 |
in natural language processing. ACM Comput. Surv. , 55(9):195:1–195:35. Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. 2022. Fantastically ordered prompts and where to find them: Overcoming few- shot prompt order sensitivity. In Proceedings of the60th Annual Meeting of the Association for ... | https://arxiv.org/abs/2505.20322v1 |
model-written evaluations. In Findings of the Association for Computational Linguistics: ACL 2023, Toronto, Canada, July 9-14, 2023 , pages 13387–13434. Association for Computational Lin- guistics. Joris Postmus and Steven Abreu. 2024. Steering large language models using conceptors: Improv- ing addition-based activati... | https://arxiv.org/abs/2505.20322v1 |
Deng, Peng Wang, Xiang Chen, Jia-Chen Gu, Yong Jiang, Pengjun Xie, Fei Huang, Huajun Chen, and Ningyu Zhang. 2024a. Knowledge mechanisms in large language models: A survey and perspective. InFindings of the Association for Computational Linguistics: EMNLP 2024, Miami, Florida, USA, November 12-16, 2024 , pages 7097–713... | https://arxiv.org/abs/2505.20322v1 |
How non-ai experts try (and fail) to design LLM prompts. In Proceedings of the 2023 CHI Con- ference on Human Factors in Computing Systems, CHI 2023, Hamburg, Germany, April 23-28, 2023 , pages 437:1–437:21. ACM. Wojciech Zaremba, Evgenia Nitishinskaya, Boaz Barak, Stephanie Lin, Sam Toyer, Yaodong Yu, Rachel Dias, Eri... | https://arxiv.org/abs/2505.20322v1 |
in the Phi-3 Mini model but is highly sensitive to hyperparameters. As noted in Table 2 of the original study, achieving effective detoxification with RefusalFeature often requires indiscriminately rejecting all queries, which signif- icantly compromises the model’s general capabili- ties. In contrast, our work aims to... | https://arxiv.org/abs/2505.20322v1 |
95.78 99.33 97.56 5.43 70.27 71.65 49.12 wo/Amplitude 95.70 99.33 97.52 5.43 70.29 71.49 49.07 wo/Frequency 90.89↓ 98.42↓ 94.65↓5.43 70.90 72.63 49.65 Table 3: The ablation study of our proposal STA. The biggest drop of detoxification performance in each column is appended ↓. E Prompting and Steering E.1 Position of Pr... | https://arxiv.org/abs/2505.20322v1 |
modular arithmetic, although the representation may vary. In standard arithmetic, the sum of two ones is always two. Answer (long thought ): Alright, so I’m trying to figure out what 1 + 1 equals. Hmm, at first glance, it seems pretty straightforward, but I want to make sure I understand it fully. Let me think about ho... | https://arxiv.org/abs/2505.20322v1 |
PMOA-TTS: Introducing the PubMed Open Access Textual Times Series Corpus Shahriar Noroozizadeh1,2∗Sayantan Kumar3∗George H. Chen2,1Jeremy C. Weiss3 1Machine Learning Department, School of Computer Science, Carnegie Mellon University 2Heinz College of Information Systems and Public Policy, Carnegie Mellon University 3Na... | https://arxiv.org/abs/2505.20323v1 |
and DeepSeek R1 [ 11]—we transform each case report into a sequence of (clinical event, time) tuples that capture key clinical events and their relative timing within the patient narrative. To our knowledge, PMOA–TTS constitutes the largest publicly available collection of clinical narratives with explicit temporal eve... | https://arxiv.org/abs/2505.20323v1 |
disease progression. While prior efforts, such as the i2b2 2012 challenge [ 7], laid foundational work, progress has been constrained by data scarcity. PMOA–TTS supports timeline extraction at scale and introduces novel evaluation metrics like AULTC, bridging the gap between event extraction and downstream temporal mod... | https://arxiv.org/abs/2505.20323v1 |
report to verify whether it described exactly one case. Our LLM-pipeline significantly outperforms PubMed metadata filtering, particularly in terms of specificity and precision (Section 4.2). 3.2 Timeline Annotation via LLM Prompting We used LLMs to extract clinical textual time series from PMOA case reports. A textual... | https://arxiv.org/abs/2505.20323v1 |
many-to-one mappings. For each reference event, the closest predicted event is selected based on the minimal distance, with event order used to break ties. Once matched, both events are removed from further consideration, and the process continues recursively with the remaining pairs. We tested several distance metrics... | https://arxiv.org/abs/2505.20323v1 |
history of (clinical_event, time) tuples, predicting whether the nextkevents will occur within specific time windows (1 hour, 1 day, 1 week). 2.Event Ordering (Concordance) : Given a history of (clinical_event, time) tuples, predicting the correct temporal ordering of the next kevents. A summary of these forecasting ta... | https://arxiv.org/abs/2505.20323v1 |
These findings highlight the significant improvements in precision and overall classification quality offered by our LLM-Pipeline approach for case report identification, ensuring that the PMOA-TTS corpus contains high-quality, single-patient case reports with minimal contamination from multi-patient or non-case report... | https://arxiv.org/abs/2505.20323v1 |
to the reference points. In contrast, LLaMA 3.3 70B (green line) produces a flatter curve, reflecting greater variation and lower precision in absolute time prediction. The right panel in Figure 2 breaks down performance across time intervals—presentation, 1 hour, 1 day, 1 week, 1 year, and beyond. All models perform w... | https://arxiv.org/abs/2505.20323v1 |
Kidney Failure, Acute DIABETES MELLITUS, NONINSULIN-DEPENDENT, 1 (disorder) DIABETES MELLITUS, INSULIN-DEPENDENT, 2 Diabetic Diet Glucose Intolerance (disease) Neoplasm Metastasis Secondary Neoplasm 01234567 log(1 + Co-occurrence Count)Clustered Diagnosis Co-occurrence Heatmap (Top 30, Log Scale) (b) Diagnoses Co-occur... | https://arxiv.org/abs/2505.20323v1 |
DeepSeek-R1-Distill-Llama-8B 0.806 ± 0.015 0.803 ± 0.011 0.809 ± 0.013 0.806 ± 0.011 0.802 ± 0.010 0.809 ± 0.011 Llama-3.1-8B-Instruct 0.815 ± 0.008 0.813 ± 0.010 0.817 ± 0.010 0.811 ± 0.007 0.808 ± 0.014 0.809 ± 0.011 Llama-3.3-70B-Instruct 0.822 ± 0.004 0.817 ± 0.011 0.821 ± 0.011 0.819 ± 0.005 0.819 ± 0.009 0.822 ± ... | https://arxiv.org/abs/2505.20323v1 |
single relative points simplifies annotation but limits expressiveness, especially in cases where durations or implicit temporal relationships are critical. Third, our one-to-one recursive event matching strategy prioritizes temporal consistency, but at the cost of reduced recall, particularly when events are expressed... | https://arxiv.org/abs/2505.20323v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.