Daily Papers

Self-evolving agents accumulate capability by distilling reusable skills from their execution trajectories, but we find this process is not monotonic: past a critical pool size, newly added skills degrade performance instead of improving it. We formalize this capability-contamination phase transition and trace it to a structural cause: once a defective skill enters the decision context, it becomes reference material for distilling later skills, forming cross-round contamination chains. We further show the contamination is structurally irreversible: removing a source skill after the fact cannot erase the flawed reasoning its descendants have already inherited, so post-hoc rollback recovers only a small fraction of the lost performance. This makes skill admission a pre-commit necessity rather than a post-hoc fix, and motivates Verifier-as-Gatekeeper (VaG): a progressive trust hierarchy whose three heterogeneous critics - structural validity, behavioral harmlessness, and semantic consistency - filter each skill individually, coupled with a marginal-gain subset selection that removes combinatorial contamination at the top tier before skills reach the runtime context. On Terminal-Bench 2, unconditional accumulation rises to a peak and then degrades, giving back most of its gains as the pool keeps growing, and post-hoc removal of the culprit skills recovers only a small part of the drop - the empirical signature of irreversibility. In contrast, VaG improves every round, reaching 72% pass@1 with a pool roughly 5x smaller, and its frozen skill pool transfers positively to four other backbones and a second benchmark without re-evolution. Ablations confirm the three critics are complementary and mutually non-substitutable, each intercepting a largely disjoint class of harmful skills.

We introduce Generative Universal Verifier, a novel concept and plugin designed for next-generation multimodal reasoning in vision-language models and unified multimodal models, providing the fundamental capability of reflection and refinement on visual outcomes during the reasoning and generation process. This work makes three main contributions: (1) We build ViVerBench, a comprehensive benchmark spanning 16 categories of critical tasks for evaluating visual outcomes in multimodal reasoning. Results show that existing VLMs consistently underperform across these tasks, underscoring a substantial gap from human-level capability in reliable visual verification. (2) We design two automated pipelines to construct large-scale visual verification data and train OmniVerifier-7B, the first omni-capable generative verifier trained for universal visual verification and achieves notable gains on ViVerBench(+8.3). Through training, we identify three atomic capabilities in visual verification and demonstrate how they generalize and interact synergistically. (3) We propose OmniVerifier-TTS, a sequential test-time scaling paradigm that leverages the universal verifier to bridge image generation and editing within unified models, enhancing the upper bound of generative ability through iterative fine-grained optimization. Beyond generation, we extend universal verifier to broader world-modeling interleaved reasoning scenarios. Empirically, OmniVerifier-TTS achieves improvements on T2I-ReasonBench(+3.7), and GenEval++(+4.3), outperforming existing parallel test-time scaling methods, such as Best-of-N. By endowing multimodal reasoning with reliable visual verification, OmniVerifier advances both reliable reflection during generation and scalable test-time refinement, marking a step toward more trustworthy and controllable next-generation reasoning systems.

Skills are becoming a reusable operational layer for LLM agents, encoding SOPs, domain rules, tool workflows, scripts, and validation routines. In realistic skill repositories, overlapping skills make reliable skill-use difficult. Final verifier success is too coarse for both evaluation and training, since an agent may pass through trial and error while selecting distractor skills, skipping required steps, composing workflows incorrectly or omitting final checks. We introduce SkillCoach, a self-evolving rubric framework for evaluating and enhancing agentic skill-use. SkillCoach derives skill-grounded process rubrics from real rollouts and evaluates trajectories along four dimensions: skill selection, skill following, skill composition, and skill-grounded reflection. It keeps the external verifier as a separate outcome signal, allowing process quality to be distinguished from accidental task success. The evolved rubrics further serve as process supervision for selecting high-quality training trajectories. Experiments show that evolved rubrics substantially improve evaluation quality, expose failures hidden by final accuracy, and provide stronger supervision signals than outcome-only filtering for enhancing agentic skill-use.

Reinforcement Learning with Verifiable Rewards (RLVR) trains policies against automated verifiers to avoid costly human labeling. To reduce vulnerability to verifier hacking, many RLVR systems collapse rewards to binary {0,1} during training. This choice carries a cost: it introduces false negatives (rejecting correct answers, FNs) and false positives (accepting incorrect ones, FPs). For instance, a rule-based checker may mark the correct fraction 12{36} as wrong when compared against the canonical 1{3} due to brittle parsing/equivalence rules (FN), while a large language model (LLM) judges can be gamed by superficial cues or even a single adversarial token, yielding inflated correctness for wrong solutions (FP). We formalize verifier unreliability by modeling the verifier as a stochastic reward channel with asymmetric noise rates. From this abstraction, we derive two correction algorithms for verifier errors. The first is a backward correction that de-biases the observed binary reward to recover an unbiased estimator of the clean policy gradient. The second is a forward correction that reweights score-function terms so that the expected update direction aligns with the clean gradient; notably, it requires only the FN rate. We implement both as lightweight hooks in a group relative policy optimization (GRPO)-based RLVR pipeline and evaluate them on math-reasoning models and benchmarks. Across models and datasets, both corrections improve over uncorrected training; the forward variant converges faster and remains stable under heavier noise. Finally, we show a practical appeal mechanism in which a lightweight LLM verifier estimates the FN rate online by rechecking rule-based negatives, obtaining outperformance compared with other state-of-the-art contenders.

Large language models have made significant progress in mathematical reasoning, which serves as an important testbed for AI and could impact scientific research if further advanced. By scaling reasoning with reinforcement learning that rewards correct final answers, LLMs have improved from poor performance to saturating quantitative reasoning competitions like AIME and HMMT in one year. However, this approach faces fundamental limitations. Pursuing higher final answer accuracy doesn't address a key issue: correct answers don't guarantee correct reasoning. Moreover, many mathematical tasks like theorem proving require rigorous step-by-step derivation rather than numerical answers, making final answer rewards inapplicable. To push the limits of deep reasoning, we believe it is necessary to verify the comprehensiveness and rigor of mathematical reasoning. Self-verification is particularly important for scaling test-time compute, especially for open problems without known solutions. Towards self-verifiable mathematical reasoning, we investigate how to train an accurate and faithful LLM-based verifier for theorem proving. We then train a proof generator using the verifier as the reward model, and incentivize the generator to identify and resolve as many issues as possible in their own proofs before finalizing them. To maintain the generation-verification gap as the generator becomes stronger, we propose to scale verification compute to automatically label new hard-to-verify proofs, creating training data to further improve the verifier. Our resulting model, DeepSeekMath-V2, demonstrates strong theorem-proving capabilities, achieving gold-level scores on IMO 2025 and CMO 2024 and a near-perfect 118/120 on Putnam 2024 with scaled test-time compute.

Inference-time planning with large language models frequently breaks under partial observability: when task-critical preconditions are not specified at query time, models tend to hallucinate missing facts or produce plans that violate hard constraints. We introduce Self-Querying Bidirectional Categorical Planning (SQ-BCP), which explicitly represents precondition status (Sat/Viol/Unk) and resolves unknowns via (i) targeted self-queries to an oracle/user or (ii) bridging hypotheses that establish the missing condition through an additional action. SQ-BCP performs bidirectional search and invokes a pullback-based verifier as a categorical certificate of goal compatibility, while using distance-based scores only for ranking and pruning. We prove that when the verifier succeeds and hard constraints pass deterministic checks, accepted plans are compatible with goal requirements; under bounded branching and finite resolution depth, SQ-BCP finds an accepting plan when one exists. Across WikiHow and RecipeNLG tasks with withheld preconditions, SQ-BCP reduces resource-violation rates to 14.9\% and 5.8\% (vs.\ 26.0\% and 15.7\% for the best baseline), while maintaining competitive reference quality.

In-context watermarking (ICW) prepends an instruction to a query asking the model to embed a statistically detectable signal in its response. It thus equips LLMs with a watermarking interface that third parties can invoke without access to model internals. Its reliability hinges on the LLM following the instruction without degrading answer quality, yet how well current LLMs do so has not been measured. We introduce ICWBench, a benchmark of three verifiable ICW instruction families, each scored on both detectability and answer quality. Evaluating 14 frontier proprietary and open-source LLMs, we find that none of the evaluated LLMs achieves both objectives across all three families. To address this, we propose a self-contained two-stage training method, requiring no distillation from a stronger model, no manual annotation, and no pre-existing ICW IF ability. The first stage, self-distillation with logits perturbation (SDLP), uses the same base LLM as both teacher and student: an instruction-equivalent decoding-time logits perturbation makes the teacher follow the ICW instruction, and the student is trained to match the teacher's output distribution. The second stage applies reinforcement learning with the automatic verifier as the reward. Applied to Qwen3-14B and GPT-OSS-20B, our method raises average TPR@1%FPR across three ICW instructions from 0.100 to 0.974 and from 0.337 to 0.968, respectively, while maintaining high response quality under both perplexity evaluation and LLM-as-a-Judge.

Inference-time skill augmentation provides a lightweight way to improve data-analytic agents by injecting reusable procedural knowledge without updating model parameters. However, discovering effective skills for data analysis remains challenging, as reliable supervision is expensive and success criteria vary across analytical formats. This raises the key question of how to discover reusable data-analysis skills from unlabeled exploration alone. We propose DataCOPE, an unsupervised verifier-guided skill discovery framework for data-analytic agents. DataCOPE derives verifier signals from the exploration trajectories and uses them to characterize relative quality or aggreement among trajectories. It iteratively coordinates a Data-Analytic Agent for trajectory generation, an Unsupervised Verifier for signal extraction, and a Skill Manager for contrastive skill distillation. For report-style analysis, we instantiate the verifier as an Adaptive Checklist Verifier that derives task-specific criteria, scores reports by verifiable coverage, and iteratively refines the checklist. For reasoning-style analysis, we instantiate it as an Answer Agreement Verifier that groups trajectories by answer agreement and uses self-consistency as an auxiliary signal. We evaluate DataCOPE on report-style analysis from Deep Data Research and reasoning-style analysis from DABStep. Across both settings, DataCOPE consistently improves held-out performance over baselines. Averaged across four model settings, DataCOPE improves the mean score by 9.71% and 32.30% on report-style and reasoning-style tasks respectively.

On-policy distillation (OPD) trains a student on its own trajectories with dense per-token supervision from a stronger teacher, and often outperforms off-policy distillation and standard reinforcement learning. However, we find that its effectiveness implicitly relies on two assumptions that frequently break in practice: trajectory-level alignment between the student and the teacher, and uniform token-level reliability of the teacher's preferences. We therefore propose Sign-Gated On-Policy Distillation (SG-OPD), which uses a binary verifier as a trust signal for the teacher at two complementary granularities: phased teacher sampling mixes in verifier-endorsed teacher rollouts at cold-start, and a sign-consistency gate extrapolates the distillation update on tokens where the teacher agrees with the verifier-correct direction and interpolates it where it disagrees. Experiments on competition-level mathematical reasoning benchmarks show that SG-OPD consistently outperforms standard OPD, with average gains of 1.98 and 7.50 at the per-sample and per-question levels, respectively.

Translating natural language sentences to first-order logic (NL-FOL translation) is a longstanding challenge in the NLP and formal logic literature. This paper introduces LogicLLaMA, a LLaMA-7B model fine-tuned for NL-FOL translation using LoRA on a single GPU. LogicLLaMA is capable of directly translating natural language into FOL rules, which outperforms GPT-3.5. LogicLLaMA is also equipped to correct FOL rules predicted by GPT-3.5, and can achieve similar performance as GPT-4 with a fraction of the cost. This correction ability was achieved by a novel supervised fine-tuning (SFT) + reinforcement learning with human feedback (RLHF) framework, which initially trains on synthetically perturbed NL-FOL pairs to encourage chain-of-thought reasoning and then fine-tunes with RLHF on GPT-3.5 outputs using a FOL verifier as the reward model. To train LogicLLaMA, we present MALLS (large language Model generAted NL-FOL pairS), a dataset of 34K high-quality and diverse sentence-level NL-FOL pairs collected from GPT-4. The dataset was created by implementing a pipeline that prompts GPT-4 for pairs, and dynamically adjusts the prompts to ensure the collection of pairs with rich and diverse contexts at different levels of complexity, and verifies the validity of the generated FOL rules. Codes, weights, and data are available at https://github.com/gblackout/LogicLLaMA{{small https://github.com/gblackout/LogicLLaMA}}.

We present Verified Multi-Agent Orchestration (VMAO), a framework that coordinates specialized LLM-based agents through a verification-driven iterative loop. Given a complex query, our system decomposes it into a directed acyclic graph (DAG) of sub-questions, executes them through domain-specific agents in parallel, verifies result completeness via LLM-based evaluation, and adaptively replans to address gaps. The key contributions are: (1) dependency-aware parallel execution over a DAG of sub-questions with automatic context propagation, (2) verification-driven adaptive replanning that uses an LLM-based verifier as an orchestration-level coordination signal, and (3) configurable stop conditions that balance answer quality against resource usage. On 25 expert-curated market research queries, VMAO improves answer completeness from 3.1 to 4.2 and source quality from 2.6 to 4.1 (1-5 scale) compared to a single-agent baseline, demonstrating that orchestration-level verification is an effective mechanism for multi-agent quality assurance.

Conversational assistants increasingly recommend follow-up edits to help users continue a task. Existing systems primarily target text-only interactions, leaving image-creation conversations underexplored. In image-creation tasks, useful follow-up edit suggestions must reflect user preferences, offer diverse directions, and remain executable on the current image. We collected 100,000 real multi-turn image-creation conversation samples from Qwen App and found that 80.1% are image-dependent, underscoring the need for multimodal recommendation. We address this setting with a three-stage framework. In Stage 1, we use real online data to build a human-reviewed table of appropriate follow-up editing intents, then create SFT targets and fine-tune a multimodal policy. In Stage 2, to align rule-guided SFT suggestions with actual user choices, we use user click feedback to optimize the policy through multi-objective reinforcement learning. In Stage 3, to reduce visual inconsistencies between suggested edits and the current image, we introduce a visual verifier as additional training supervision. Extensive experiments demonstrate that our framework significantly outperforms baselines on both automatic and human evaluations. In a live user-randomized A/B test with millions of users, our final framework reduces visual inconsistency from 3.7% to 0.9%. Furthermore, it significantly improves recommendation CTR by 32.70%, image take-away rate by 16.32%, and average conversation turns per user by 39.90% (all p<0.05).

Multi-hop QA benchmarks often reward Large Language Models (LLMs) for spurious correctness, where models reach correct answers through invalid intermediate reasoning. We propose SAFE, an LLM-as-verifier framework for evidence-grounded multi-hop QA. Rather than judging only the final answer after generation, SAFE verifies reasoning during generation by checking intermediate steps against the provided passages and previous reasoning trajectory. To make this process checkable, SAFE decomposes reasoning into atomic, evidence-grounded units represented with Knowledge Graph (KG) triples. At train-time, SAFE verifies benchmark supervision under KG-grounded constraints and constructs reliable verifier training data. At inference-time, an external verifier checks each generated step, identifies invalid reasoning, and provides correction feedback before errors propagate. Across three multi-hop QA benchmarks, SAFE improves accuracy by 8.8 pp on average. These results show that evidence-grounded multi-hop QA benefits from shifting LLM-based evaluation from post-hoc answer judgment to stepwise reasoning verification.

Large language models (LLMs) struggle with multi-step reasoning, where inference-time scaling has emerged as a promising strategy for performance improvement. Verifier-guided search outperforms repeated sampling when sample size is limited by selecting and prioritizing valid reasoning paths. However, we identify a critical limitation: scaling flaws, prevalent across different models (Mistral 7B and DeepSeekMath 7B), benchmarks (GSM8K and MATH), and verifiers (outcome value models and process reward models). As sample size increases, verifier-guided search exhibits diminishing advantages and eventually underperforms repeated sampling. Our analysis attributes this to verifier failures, where imperfect verifiers misrank candidates and erroneously prune all valid paths. These issues are further exacerbated in challenging and out-of-distribution problems, restricting search effectiveness. To mitigate verifier failures, we explore reducing reliance on verifiers and conduct preliminary investigations using two simple methods. Our findings reveal fundamental limitations in verifier-guided search and suggest future directions.

Single-cell perturbation models can reduce costly wet-lab screening by predicting how cells respond transcriptionally to interventions. While recent generative models improve population-level prediction, individual generated cells are not explicitly checked for biological consistency. We introduce PerturbCellRL, a reinforcement learning (RL) framework that post-trains a pretrained single-cell transcriptomic generator using a suite of cell-level verifiers as rewards. These verifiers define four rewards: Pearson top-k similarity, RMSE top-k proximity, DE Spearman, and Pathway activity. The Pathway activity verifier rewards cells whose pathway responses match known perturbation biology. We evaluate PerturbCellRL on multiple genetic and chemical perturbation benchmarks. Across these benchmarks, PerturbCellRL improves over the pretrained flow-matching generator on reward-aligned evaluation metrics and a held-out evaluation metric. Moreover, PerturbCellRL remains competitive with state-of-the-art methods on population-level metrics. Together, these results frame trustworthy single-cell prediction as verifier-guided generative alignment, moving beyond matching expression distributions toward predictions whose single-cell perturbation effects are explicitly checked for biological consistency.

A rule-based logic solver resolves every instance in our benchmark in under 50 microseconds with 100% accuracy; the best frontier language model reaches 65% at best and drops to 23.5% under rendering-robust evaluation (worst case over four surface renderings). We introduce DeFAb (Defeasible Abduction Benchmark), a dataset and generation pipeline that converts four decades of publicly funded knowledge bases into formally grounded instances for defeasible abduction: constructing hypotheses that explain anomalies by overriding defaults while preserving unrelated expectations. Because every hypothesis must pass polynomial-time checks for valid derivation, conservativity, and minimality, DeFAb makes logical rigor the instrument for measuring creativity and theoretical reasoning, scoring the disciplined construction of theory revisions rather than fluent but theory-destroying prose. The pipeline pairs taxonomic hierarchies (OpenCyc, YAGO, Wikidata) with behavioral property graphs (ConceptNet, UMLS) to produce 372,648+ instances across 33.75M materialized rules from 18 sources, in three levels with polynomial-time verifiable gold standards. Four frontier models do not reliably internalize defeasible reasoning: rendering-robust Level 2 accuracy is 7.8-23.5%; chain-of-thought variance (~36 pp) exceeds any inter-model gap; and a matched contamination control isolates a +19.4 pp Level 3 gap. We further release DeFAb-Hard (a 235-instance Level 3 difficulty variant; best model 53.3% vs 100% symbolic) and CONJURE (a kernel-verified transformative-creativity variant of 560 Lean 4/Mathlib instances whose gold answers are definitions the proof kernel did not previously contain, judge-free verifier; a pilot finds zero novel concepts). The same verifier doubles as an exact reward for preference optimization (DPO, RLVR/GRPO). Released under MIT at https://huggingface.co/datasets/PatrickAllenCooper/DeFAb.

Agentic reinforcement learning requires assigning credit to environment-facing actions such as searches, clicks, edits, navigation commands, and object interactions. Standard GRPO uses the final verifier outcome as a uniform advantage over all action tokens. This outcome signal is useful but structurally incomplete: it punishes useful exploration in failed rollouts and reinforces redundant or regressive actions in successful rollouts. We propose TRIAGE, a role-typed credit assignment framework that adds a semantic role axis to outcome credit. A structured judge classifies each segment as decisive progress, useful exploration, no-progress infrastructure, or regression, and a fixed role-conditioned rule maps these labels to bounded segment-level process rewards. This keeps verifier outcomes as the source of optimization direction while correcting the two main blind spots of outcome-only credit. We further show that role-conditioned credit is the optimal segment-level correction expressible from role labels alone -- a projection of the per-segment advantage residual onto the role variable -- so that the fixed role constants reduce advantage estimation error whenever the judge is reliable, and we connect this to lower-variance policy gradients. Across ALFWorld, Search-QA, and WebShop, TRIAGE improves success rates over GRPO for two policy models and outperforms both a scalar judge-derived process reward and an outcome-supervised shared-backbone value baseline. Ablations show that the gain comes from role typing rather than merely adding dense rewards: reliable detection of regression inside successful trajectories is the dominant contributor, while exploration credit provides a consistent secondary gain; on completed ALFWorld and WebShop rollouts, TRIAGE also reduces environment-facing turns by an additional 10.4% and 14.8% relative to GRPO.

While tool-augmented Large Language Models have significantly improved multi-step reasoning in quantitative STEM tasks, a critical residual failure mode remains: intermediate reasoning steps that are syntactically well-formed, mathematically executable, and unit-consistent, yet contextually ungrounded. Current approaches either rely on formal verifiers that cannot assess semantic intent, or burden Process Reward Models (PRMs) with the dual task of checking both arithmetic and logic. In this paper, we propose a neuro-symbolic framework that cleanly decouples reasoning into two formal dimensions: Symbolic Validity (V) and Semantic Groundedness (G). We guarantee V by construction using a deterministic symbolic verifier acting as a hard filter. To assess G, we train a PRM conditionally on the verifier-accepted manifold. To train this PRM efficiently, we introduce Counterfactual Symbolic Perturbation (CSP), a novel data synthesis strategy that algorithmically generates constraint-preserving hard negatives (steps that perfectly pass the verifier but are logically flawed). At inference, we deploy a verifier-first constrained search that guarantees execution consistency for verifier-covered operations while relying on the PRM solely to rank semantic grounding. By targeting the exact residual error class of strong tool-using LLMs, our method significantly improves reasoning reliability without the sprawling heuristics of prior frameworks.

Large-scale demonstrations have driven unprecedented progress in robot learning, yet collecting robot data through teleoperation is expensive and difficult to scale to diverse environments and long-horizon tasks. Simulation offers a scalable alternative, but existing data-generation pipelines often rely on open-loop controllers, scripted skill sequences, or task-specific programs. We introduce SkillWeaver, an agentic framework that autonomously generates robot experience by exploring over Neural Interaction Skills (NIS): reusable, parameterized, closed-loop policies that expose learned physical interaction capabilities to a reasoning agent. Given a task and a simulated environment, a VLM agent reasons about what to do next, invokes and parameterizes NIS to interact with the environment, observes their outcomes, and generates verification, reflection, and memory to guide subsequent exploration. We instantiate NIS as reinforcement-learned policies for closed-loop, contact-rich manipulation and organize exploration as verifier-guided tree search, enabling the agent to discover successful long-horizon behaviors without relying on predetermined execution pipelines. SkillWeaver scales autonomously to 39.1K demonstrations across 14.1K scenes, which we distill into visuomotor policies. Across simulation benchmarks and real-world manipulation, training on SkillWeaver-generated experience substantially improves generalization to novel objects, spatial configurations, tasks, and environments, and enables zero- and few-shot sim-to-sim and sim-to-real transfer. Our results suggest agentic exploration over neural interaction skills as a scalable alternative for robot data generation.

A classical intuition holds that verifying a solution is easier than producing one. For today's coding agents, this intuition is being inverted: as foundation models develop stronger reasoning capabilities and engineering harnesses grow more sophisticated, generating complex candidate solutions is no longer difficult -- reliably verifying them has become the harder problem. Every verifier we can build is only a proxy for human intent, never the intent itself. This makes verification subject to a twofold difficulty: first, intent is underspecified by nature, making it inherently hard to faithfully check whether it has been fulfilled; second, during model training, optimization widens the gap between proxy and intent -- manifesting as reward hacking or signal saturation. To address this, we characterize the quality of verification signals along three dimensions -- scalability, faithfulness, and robustness -- and argue that achieving all three simultaneously is the central challenge. We further study four reward constructions: a test verifier for general coding tasks, a rubric verifier for frontend tasks, the user as verifier for real-world agent tasks, and an automated agent verifier for long-horizon tasks. Across different task types and policy capability levels, we conduct in-depth analysis and experiments on the core challenges of reward design and how to more effectively leverage reward signals. Experiments show that targeted verification design can effectively suppress reward hacking, improve task completion quality, and achieve significant gains across multiple internal and public benchmarks. These experiences collectively point to a core observation: no fixed reward function can remain effective as policy capability continues to grow; and verification must co-evolve with the generator.

Embodied agents are increasingly built as systems around foundation models, where performance depends not only on model weights but also on the skills, context, action interfaces, and execution harness surrounding the model. While supervised fine-tuning and reinforcement learning can adapt agents to new environments, they require additional data, rewards, and training runs; meanwhile, many train-free code-centric approaches rely on programmable robot APIs that may be unavailable in fixed-interface settings. We propose SHAPER, a self-evolving framework for train-free embodied adaptation that keeps model parameters frozen and improves the non-parametric agent system by evolving reusable skills and a context-code harness through target-environment rollouts. In SHAPER, the same frozen model can serve as both planner and optimizer, refining its external skills and context-code harness without parameter updates. We evaluate SHAPER on VLABench and ESI-Bench, covering embodied agents with different low-level action interfaces, and compare against pure execution, supervised fine-tuning, and test-time-scaling baselines such as verifier-free selection and voting. Our results suggest that skill-and-harness optimization is a practical route to self-evolving embodied agents when model training is expensive, unavailable, or undesirable.

Repository-level software engineering (SWE) is a challenging long-horizon setting: agents must reason over extended interactions, use tools, and adapt to stateful environments. Recent work trains SWE agents with reinforcement learning methods such as Group Relative Policy Optimization (GRPO), which independently sample multiple trajectories per issue, test the resulting patches, and compare terminal rewards within a fixed group. However, this training setup does not reuse verifier feedback from failed patches as context for subsequent attempts, even though this feedback contains valuable diagnostic information about what went wrong. Training on recovery trajectories is challenging because the preceding outcome determines whether the next trajectory is generated, while the failed execution determines its conditioning context. We introduce FC-SWE, a failure-conditioned RL framework that incorporates recovery attempts into policy training. After a patch fails verification, FC-SWE restores the repository to its original task state and uses the failed patch and verifier feedback as context for a recovery trajectory. FC-SWE adapts GRPO to these chains of complete, multi-turn tool-use trajectories through two mechanisms. Trajectory-local rewards preserve each attempt's verifier outcome, preventing recovery success from rewarding an earlier failed patch. Active-set advantage estimation forms a comparison group from all initial and recovery trajectories actually executed for the same issue, so failed attempts remain in the group while unexecuted attempts are excluded. On all 500 SWE-bench Verified tasks under a verifier-assisted protocol, FC-SWE with Qwen3.5-4B and SWE-agent achieves 41.7% Resolved@1 and 52.8% Resolved@2, compared with 38.9% and 48.5% for GRPO. Although trained with at most two attempts per chain, FC-SWE reaches 70.7% Resolved@11 under an eleven-attempt test-time budget.

Scaling pre-training, post-training, and test-time compute have become the central paradigms for improving the capabilities of LLMs. In this work, we identify verification, the ability to determine the correctness of a solution, as a new scaling axis. To unlock this and demonstrate its effectiveness, we introduce LLM-as-a-Verifier, a general-purpose verification framework that provides fine-grained feedback for agentic tasks without requiring additional training. Unlike standard LM judges that prompt LLMs to produce discrete scores for candidate solutions, LLM-as-a-Verifier computes the expectation over the distribution of scoring token logits to generate continuous scores. This probabilistic formulation enables verification to scale along multiple dimensions: (1) score granularity, (2) repeated evaluation, and (3) criteria decomposition. In particular, we show that scaling the scoring granularity leads to better separation between positive and negative solutions, resulting in more calibrated comparisons. Moreover, scaling repeated evaluation and criteria decomposition consistently lead to additional gains in verification accuracy through variance and complexity reduction. We further introduce a cost-efficient ranking algorithm for selecting the best solution among candidates using the verifier's continuous scores. LLM-as-a-Verifier achieves state-of-the-art performance on Terminal-Bench V2 (86.5%), SWE-Bench Verified (78.2%), RoboRewardBench (87.4%), and MedAgentBench (73.3%). Beyond verification, the fine-grained signals from LLM-as-a-Verifier can also serve as a proxy for estimating task progress. We build an extension for Claude Code, enabling developers to monitor and improve their own agentic systems. Finally, we show that LLM-as-a-Verifier can provide dense feedback for RL, improving the sample efficiency of SAC and GRPO on robotics and mathematical reasoning benchmarks.

Large Language Models (LLMs) have demonstrated impressive capabilities in complex reasoning tasks, yet they still struggle to reliably verify the correctness of their own outputs. Existing solutions to this verification challenge often depend on separate verifier models or require multi-stage self-correction training pipelines, which limit scalability. In this paper, we propose Policy as Generative Verifier (PAG), a simple and effective framework that empowers LLMs to self-correct by alternating between policy and verifier roles within a unified multi-turn reinforcement learning (RL) paradigm. Distinct from prior approaches that always generate a second attempt regardless of model confidence, PAG introduces a selective revision mechanism: the model revises its answer only when its own generative verification step detects an error. This verify-then-revise workflow not only alleviates model collapse but also jointly enhances both reasoning and verification abilities. Extensive experiments across diverse reasoning benchmarks highlight PAG's dual advancements: as a policy, it enhances direct generation and self-correction accuracy; as a verifier, its self-verification outperforms self-consistency.

Learning alignment between language and vision is a fundamental challenge, especially as multimodal data becomes increasingly detailed and complex. Existing methods often rely on collecting human or AI preferences, which can be costly and time-intensive. We propose an alternative approach that leverages cycle consistency as a supervisory signal. Given an image and generated text, we map the text back to image space using a text-to-image model and compute the similarity between the original image and its reconstruction. Analogously, for text-to-image generation, we measure the textual similarity between an input caption and its reconstruction through the cycle. We use the cycle consistency score to rank candidates and construct a preference dataset of 866K comparison pairs. The reward model trained on our dataset outperforms state-of-the-art alignment metrics on detailed captioning, with superior inference-time scalability when used as a verifier for Best-of-N sampling. Furthermore, performing DPO and Diffusion DPO using our dataset enhances performance across a wide range of vision-language tasks and text-to-image generation. Our dataset, model, and code are at https://cyclereward.github.io

Practical video editing is not only pixel generation: an editor must turn a brief, a clip pool, music metadata, and hard constraints into an executable timeline. We study this decision layer as executable video-editing planning and introduce RefineCut, which, unlike workflow systems that wrap a prompted frontier model, trains a compact open-weight planner for it. The planner edits a typed timeline through structured patches covering clip selection, trimming, ordering, transitions, and duration and music alignment; a deterministic verifier applies each patch and checks it against an explicit constraint ledger. Because editing has no single ground-truth repair, we do not imitate teachers directly: RefineCut replays every multi-teacher branch through the verifier and keeps verifier-best repairs as supervision. A second stage, RefineCut-Evo, lets the student score its own repairs with the verifier and a task rubric and trains on high-margin preference pairs, so the final 8B planner runs in a closed verifier loop with no teacher calls at inference. On RefineCut-Bench (3{,}578 tasks, 7{,}971 captioned clips, 499 music tracks, explicit ledgers), verifier-replayed distillation lifts the planner from 0.620 to 0.858 on the protocol-specific Video-Editing Score and RefineCut-Evo reaches 0.924; the gain transfers to Llama-3.1-8B and GLM-4-9B, and in the same closed loop the 8B planner matches or exceeds its frontier teachers. Code and RefineCut-Bench are publicly released; see the Data Availability statement.

Building generalist embodied agents capable of solving complex real-world tasks remains a fundamental challenge in AI. Multimodal Large Language Models (MLLMs) have significantly advanced the reasoning capabilities of such agents through strong vision-language knowledge and chain-of-thought (CoT) reasoning, yet remain brittle when faced with challenging out-of-distribution scenarios. To address this, we propose Verifier-Guided Action Selection (VegAS), a test-time framework designed to improve the robustness of MLLM-based embodied agents through an explicit verification step. At inference time, rather than committing to a single decoded action, VeGAS samples an ensemble of candidate actions and uses a generative verifier to identify the most reliable choice, without modifying the underlying policy. Crucially, we find that using an MLLM off-the-shelf as a verifier yields no improvement, motivating our LLM-driven data synthesis strategy, which automatically constructs a diverse curriculum of failure cases to expose the verifier to a rich distribution of potential errors at training time. Across embodied reasoning benchmarks spanning the Habitat and ALFRED environments, VeGAS consistently improves generalization, achieving up to a 36% relative performance gain over strong CoT baselines on the most challenging multi-object, long-horizon tasks.

Large Language Models (LLMs) have advanced autonomous agents from deep search, which retrieves concise factual answers, to deep research, which synthesizes scattered evidence into long-form reports. However, verifiable multimodal deep research remains challenging due to open-ended synthesis without deterministic ground truth and the need to interleave textual arguments with visual evidence. We propose Ptah, a multi-agent harness for interleaved report generation. Ptah orchestrates the lifecycle from user query to rendered web report through planning, research, and writing stages, where specialized agents construct visual-aware plans, collect claim-grounded evidence, maintain source-aligned images in a Visual Working Memory, and compose reports through declarative multimodal tool use. A verifier agent serves as the harness's acceptance function, enforcing factual grounding, citation fidelity, and cross-modal consistency throughout the workflow. We further introduce PtahEval, an evaluation protocol that augments existing benchmarks with image-level and presentation-level assessments. Experiments on deep research benchmarks show that Ptah produces more reliable, visually informative, and usable human-facing multimodal reports than strong baselines.

Large Language Models frequently hallucinate in precision-critical domains such as technical diagramming and mechanical design, where outputs must satisfy strict geometric constraints. We study open-ended geometric synthesis from natural language: translating free-form descriptions into precise constructions whose entities must simultaneously satisfy dozens of interacting constraints. To make this tractable, we release PyGeoX, a programmable geometric DSL that compiles declarative constraints into a differentiable loss, and PyGeoX-Bench, a stratified suite of 300 problems with per-constraint verifiable rewards. Using PyGeoX as a verifier, we identify a failure mode we call Outlier Gradient Masking: under global-norm rewards (any scheme that aggregates residuals through a single norm, for example, exp(-MSE)), a single outlier constraint can nullify the learning signal across all others. To address this, we propose Saturating Additive Rewards (SAR), which decompose the reward into bounded per-constraint terms, preserving partial progress and ensuring consistent gradients even under severe violations. Against MSE-based rewards, the natural baseline for geometry solvers, SAR improves the hard-tier solving rate by 2.3times, and the resulting 8B model is competitive with much larger frontier systems on this benchmark. We release the engine, benchmark, and data at https://github.com/Huawei-AI4Math/PyGeoX.

Voice agents in production must handle several requests, background speech, and customers who lose patience. We introduce VAmoS Energy, a benchmark that combines these challenges in 100 calls about utility billing and payment assistance. Each caller makes two to four requests. The agent has sixteen tools backed by a stateful Stripe billing twin and the Apache Fineract loan engine, with account access blocked until caller verification succeeds. The tasks use public household electricity data and a policy based on Pennsylvania's residential billing rules. An LLM-as-a-verifier checks the agent's actions and spoken figures against explicit requirements. On a calibration run, it agrees with a code verifier on 99.1% of checks. Across fourteen voice stacks and three repeats per task, completion ranges from 17.3% to 44.7%. Grok Voice leads, and Gemini 3.8 Live and GPT-Live follow at about the same cost per call. Background television reduces pooled completion from 38.7% to 8.6%. The simulated caller often accepts an incorrect result because it hears the agent's words but cannot inspect its actions. These findings show why voice agents need evaluation across the whole call, including what they say, what they change, and how they handle competing speech.

Reasoning models have achieved remarkable performance on tasks like math and logical reasoning thanks to their ability to search during reasoning. However, they still suffer from overthinking, often performing unnecessary reasoning steps even after reaching the correct answer. This raises the question: can models evaluate the correctness of their intermediate answers during reasoning? In this work, we study whether reasoning models encode information about answer correctness through probing the model's hidden states. The resulting probe can verify intermediate answers with high accuracy and produces highly calibrated scores. Additionally, we find models' hidden states encode correctness of future answers, enabling early prediction of the correctness before the intermediate answer is fully formulated. We then use the probe as a verifier to decide whether to exit reasoning at intermediate answers during inference, reducing the number of inference tokens by 24\% without compromising performance. These findings confirm that reasoning models do encode a notion of correctness yet fail to exploit it, revealing substantial untapped potential to enhance their efficiency.

Verifiable instruction-following benchmarks emphasize output-constraint satisfaction but can overlook whether responses meaningfully answer the underlying request, creating an evaluation blind spot. We show that frontier models can receive full credit for responses with little substantive answer content. Using the same verifier-based criterion as a training reward can induce reward hacking, as publicly released checkpoints exhibit higher constraint satisfaction but lower response quality after reinforcement learning with verifiable rewards. To address this reward hacking, we introduce CALIF, a lightweight method that augments verifier rewards with a calibrated likelihood-based bonus to preserve response quality. CALIF evaluates response likelihood under the request alone using the existing frozen reference policy, requiring no additional judge or learned reward model. Across models ranging from 1.7B to 8B in reasoning and non-reasoning modes, CALIF improves constraint satisfaction on IFBench while maintaining response quality near or above pre-RL levels. It also outperforms corresponding base policies in 40 of 45 model-metric comparisons across three rubric-based instruction-following benchmarks, demonstrating gains beyond training verifiers.

Chain-of-thought (CoT) reasoning has been widely applied in the mathematical reasoning of Large Language Models (LLMs). Recently, the introduction of derivative process supervision on CoT trajectories has sparked discussions on enhancing scaling capabilities during test time, thereby boosting the potential of these models. However, in multimodal mathematical reasoning, the scarcity of high-quality CoT training data has hindered existing models from achieving high-precision CoT reasoning and has limited the realization of reasoning potential during test time. In this work, we propose a three-module synthesis strategy that integrates CoT distillation, trajectory-format rewriting, and format unification. It results in a high-quality CoT reasoning instruction fine-tuning dataset in multimodal mathematics, MMathCoT-1M. We comprehensively validate the state-of-the-art (SOTA) performance of the trained URSA-7B model on multiple multimodal mathematical benchmarks. For test-time scaling, we introduce a data synthesis strategy that automatically generates process annotation datasets, known as DualMath-1.1M, focusing on both interpretation and logic. By further training URSA-7B on DualMath-1.1M, we transition from CoT reasoning capabilities to robust supervision abilities. The trained URSA-RM-7B acts as a verifier, effectively enhancing the performance of URSA-7B at test time. URSA-RM-7B also demonstrates excellent out-of-distribution (OOD) verifying capabilities, showcasing its generalization. Model weights, training data and code will be open-sourced.

Recent advances in large language models (LLMs) have improved their performance on coding benchmarks. However, improvement is plateauing due to the exhaustion of readily available high-quality data. Prior work has shown the potential of synthetic self-instruct data, but naively training on a model's own outputs can cause error accumulation, especially in coding tasks, where generalization may collapse due to overly simple or erroneous training data, highlighting the need for rigorous quality checks on synthetic data. In this work, we explore an effective approach whereby the model itself verifies the correctness of its own data. We thus propose Sol-Ver, a self-play solver-verifier framework that jointly improves a single model's code and test generation capacity. By iteratively refining code (LLM-as-a-solver) and tests (LLM-as-a-verifier) together, we boost both capabilities without relying on human annotations or larger teacher models. Experiments with the Llama 3.1 8B model demonstrate substantial performance enhancements, achieving average relative improvements of 19.63% in code generation and 17.49% in test generation on MBPP and LiveCodeBench.

Terminal agents act through stochastic model generations, yet the ability to generate a useful action does not ensure its reliable execution. A poor command (e.g., wrong package install) can change the environment in ways that hinder subsequent progress, even when the model could generate a better alternative. We investigate whether allocating test-time compute at the model-harness boundary can improve action reliability and trajectory success, and what makes this allocation effective. To study these questions, we introduce Mid-Harness, which samples and verifies candidate actions before forwarding one for execution, while keeping the generator and harness unchanged. With a TMAX-9B generator, more action sampling yields little benefit under weak verification, whereas a capable verifier can exploit useful alternatives from the same generator. On TerminalBench-Lite, a GPT-5.6 Sol verifier raises Pass@1 from 50.00% for the base agent to 68.03% with 8 sampled actions. When the same TMAX-9B model serves as the verifier, pairwise verification performs best among the evaluated verification mechanisms. Distilling responses from the stronger verifier into TMAX-9B further improves Pass@1, while leaving the action generator unchanged. With TMAX-9B on TerminalBench-Lite, combining action and trajectory scaling reaches higher success at lower estimated token cost than generating more trajectories alone. Mid-Harness also improves performance across additional models, benchmarks, and harnesses. These findings identify action scaling as a promising target for test-time compute scaling in terminal agents.

Recent work in language modeling has raised the possibility of self-improvement, where a language models evaluates and refines its own generations to achieve higher performance without external feedback. It is impossible for this self-improvement to create information that is not already in the model, so why should we expect that this will lead to improved capabilities? We offer a new perspective on the capabilities of self-improvement through a lens we refer to as sharpening. Motivated by the observation that language models are often better at verifying response quality than they are at generating correct responses, we formalize self-improvement as using the model itself as a verifier during post-training in order to ``sharpen'' the model to one placing large mass on high-quality sequences, thereby amortizing the expensive inference-time computation of generating good sequences. We begin by introducing a new statistical framework for sharpening in which the learner aims to sharpen a pre-trained base policy via sample access, and establish fundamental limits. Then we analyze two natural families of self-improvement algorithms based on SFT and RLHF. We find that (i) the SFT-based approach is minimax optimal whenever the initial model has sufficient coverage, but (ii) the RLHF-based approach can improve over SFT-based self-improvement by leveraging online exploration, bypassing the need for coverage. Finally, we empirically validate the sharpening mechanism via inference-time and amortization experiments. We view these findings as a starting point toward a foundational understanding that can guide the design and evaluation of self-improvement algorithms.

Agent skills are increasingly used to equip large language model (LLM) agents with reusable procedural knowledge. Although recent work has substantially improved skill retrieval due to the increasing skill libraries, retrieving a plausible skill bundle does not guarantee that executing it is worthwhile. Since every skill-conditioned rollout is computationally expensive, deciding whether a retrieved bundle should be executed has become an increasingly important challenge. To this end, we introduce the Reward-Aware Dynamic Execution Gate (RADEG), a lightweight, retriever-agnostic decision layer between skill retrieval and agent execution. RADEG learns a low-cost surrogate model that predicts the execution utility of a query--bundle pair before the expensive rollout is launched. To obtain informative supervision while controlling for task difficulty, we locally perturb each retrieved bundle by deleting, adding, or replacing one skill, producing matched same-query rollouts that isolate the effect of bundle composition on verifier reward. During deployment, RADEG updates only a warm-started logistic head as new verifier feedback becomes available, enabling inexpensive adaptation of the execute/skip boundary without retraining either the retriever or the agent. Under a query-level held-out evaluation on 288 collected rollouts, RADEG substantially reduces unnecessary agent executions while preserving a large fraction of the downstream verifier reward. It consistently outperforms relevance-based and random gating across different execution budgets, demonstrating that execution-aware surrogate modeling provides a practical and cost-effective complement to skill retrieval.

Multimodal reinforcement learning with verifiable rewards (RLVR) provides reliable task-level supervision, yet offers limited guidance for localizing errors in visual reasoning. On-policy self-distillation (OPSD) complements this supervision with dense token-level corrections from a privileged teacher equipped with an evidence-centered visual view. However, the resulting corrections may conflate visual evidence with non-visual shifts induced by linguistic priors, response style, or the privileged view itself. We formalize this phenomenon as privileged-correction contamination: a token-level correction is contaminated when its update direction cannot be consistently attributed to the privileged visual evidence. Directly distilling contaminated corrections may therefore reinforce erroneous token-level updates. To address this issue, we propose reinforcement learning with visually attributable on-policy self-distillation, an evidence-consistent paradigm that augments task-level RLVR supervision with privileged token-level corrections grounded in visual evidence. We instantiate this paradigm with Visual-Attribution Gating (VAG), which uses the verifier-derived advantage as the optimization anchor and validates each privileged correction against the contrast between evidence-present and evidence-degraded teacher views. By admitting only corrections directionally supported by this visual contrast, VAG preserves visually grounded guidance while suppressing inconsistent residuals. Experiments across six visual reasoning benchmarks show that VAG consistently outperforms GRPO and OPSD methods, establishing new state-of-the-art average performance.

Recent studies have demonstrated that test-time compute scaling effectively improves the performance of small language models (sLMs). However, prior research has mainly examined test-time compute scaling with an additional larger model as a verifier, leaving self-verification by sLMs underexplored. In this work, we investigate whether sLMs can reliably self-verify their outputs under test-time scaling. We find that even with knowledge distillation from larger verifiers, sLMs struggle with verification tasks requiring memorization, such as numerical calculations and fact-checking. To address this limitation, we propose Tool-integrated self-verification (T1), which delegates memorization-heavy verification steps to external tools, such as a code interpreter. Our theoretical analysis shows that tool integration reduces memorization demands and improves test-time scaling performance. Experiments on the MATH benchmark demonstrate that, with T1, a Llama-3.2 1B model under test-time scaling outperforms the significantly larger Llama-3.1 8B model. Moreover, T1 generalizes effectively to both mathematical (MATH500) and multi-domain knowledge-intensive tasks (MMLU-Pro). Our findings highlight the potential of tool integration to substantially improve the self-verification abilities of sLMs.

Web navigation is a unique domain that can automate many repetitive real-life tasks and is challenging as it requires long-horizon sequential decision making beyond typical multimodal large language model (MLLM) tasks. Yet, specialized reward models for web navigation that can be utilized during both training and test-time have been absent until now. Despite the importance of speed and cost-effectiveness, prior works have utilized MLLMs as reward models, which poses significant constraints for real-world deployment. To address this, in this work, we propose the first process reward model (PRM) called Web-Shepherd which could assess web navigation trajectories in a step-level. To achieve this, we first construct the WebPRM Collection, a large-scale dataset with 40K step-level preference pairs and annotated checklists spanning diverse domains and difficulty levels. Next, we also introduce the WebRewardBench, the first meta-evaluation benchmark for evaluating PRMs. In our experiments, we observe that our Web-Shepherd achieves about 30 points better accuracy compared to using GPT-4o on WebRewardBench. Furthermore, when testing on WebArena-lite by using GPT-4o-mini as the policy and Web-Shepherd as the verifier, we achieve 10.9 points better performance, in 10 less cost compared to using GPT-4o-mini as the verifier. Our model, dataset, and code are publicly available at LINK.

Process Reward Models (PRMs) are a powerful mechanism for steering large language model reasoning by providing fine-grained, step-level supervision. However, this effectiveness comes at a significant cost: PRMs require expert annotations for every reasoning step, making them costly and difficult to scale. Here, we propose a method for training unsupervised PRMs (uPRM) that requires no human supervision, neither at the level of step-by-step annotations nor through ground-truth verification of final answers. The key idea behind our approach is to define a scoring function, derived from LLM next-token probabilities, that jointly assesses candidate positions of first erroneous steps across a batch of reasoning trajectories. We demonstrate the effectiveness of uPRM across diverse scenarios: (i) uPRM achieves up to 15% absolute accuracy improvements over the LLM-as-a-Judge in identifying first erroneous steps on the ProcessBench dataset; (ii) as a verifier for test-time scaling, uPRM performs comparably to supervised PRMs and outperforms the majority voting baseline by up to 6.9%, and (iii) when used as a reward signal in reinforcement learning, uPRM enables more robust policy optimization throughout training compared to a supervised PRM trained using ground-truth labels. Overall, our results open a path toward scalable reward modeling for complex reasoning tasks.

Autonomous agents for long-sequence Graphical User Interface tasks are hindered by sparse rewards and the intractable credit assignment problem. To address these challenges, we introduce GUI-Shepherd, a Process Reward Model that provides dense, step-by-step feedback to guide agents. GUI-Shepherd is trained on a diverse large-scale data set of 52k interactions that features human-annotated scores and GPT-4o generated rationales, enabling it to serve both as a reward provider for RL training and as a verifier for inference. As far as we know, we are the first to conduct a systematic study of process supervision in GUI agents, across diverse settings from online long-horizon tasks to offline single-step prediction. On the online AndroidWorld benchmark, GUI-Shepherd improves success rate by 7.7 points via multi-turn online PPO, significantly outperforming Outcome Reward Model based competitors. When used as an inference verifier, it brings 5.1 points improvements. The benefits generalize to the offline AndroidControl benchmark, with gains of 2.2 points as a reward provider and 4.3 points as a verifier. Collectively, our results establish that high-fidelity process supervision is critical for building more capable GUI agents and present a generalizable solution.

Deep research agents autonomously conduct open-ended investigations, integrating complex information retrieval with multi-step reasoning across diverse sources to solve real-world problems. To sustain this capability on long-horizon tasks, reliable verification is critical during both training and inference. A major bottleneck in existing paradigms stems from the lack of explicit verification mechanisms in QA data synthesis, trajectory construction, and test-time scaling. Errors introduced at each stage propagate downstream and degrade the overall agent performance. To address this, we present Marco DeepResearch, a deep research agent optimized with a verification-centric framework design at three levels: (1)~QA Data Synthesis: We introduce verification mechanisms to graph-based and agent-based QA synthesis to control question difficulty while ensuring answers are unique and correct; (2)~Trajectory Construction: We design a verification-driven trajectory synthesis method that injects explicit verification patterns into training trajectories; and (3)~Test-time scaling: We use Marco DeepResearch itself as a verifier at inference time and effectively improve performance on challenging questions. Extensive experimental results demonstrate that our proposed Marco DeepResearch agent significantly outperforms 8B-scale deep research agents on most challenging benchmarks, such as BrowseComp and BrowseComp-ZH. Crucially, under a maximum budget of 600 tool calls, Marco DeepResearch even surpasses or approaches several 30B-scale agents, like Tongyi DeepResearch-30B.