PaperlayerView exact revision on arXiv

Source-linked AI summary

Not Just Reason, Not Just Scan: Reinforcement Learning for Proactive Scientific Error Verification over Academic Paper

Rongjin Li, Yuanxin Liu, Hao Zhou, Fandong Meng, Jie Zhou, Xu Sun

arXiv:2608.26596v1cs.CL

TL;DR

Scientific assistants still lack reliable issue- and evidence-absent verification across academic papers, despite the importance of paper-level judgment for autonomous research. VERA-RL addresses this gap with Reason–Verify–Scan training, matched data, and fine-grained rewards, substantially improving Qwen3-VL-8B’s verifiable reasoning and approaching larger MLLMs on Scan. The paper concludes that Scan-style verification can be systematically improved through aligned task design, data construction, and reward modeling.

  • Problem

    Existing MLLM methods usually specify both the target problem and supporting evidence, leaving limited training approaches for issue- and evidence-absent verification across full papers.

  • Method

    VERA-RL trains scientific error detection through Reason–Verify–Scan stages, VERA-13K matched chains, and rewards for reasoning completeness, evidence alignment, and error precision.

  • Results

    VERA-RL substantially improves Qwen3-VL-8B and shows consistent improvements on VERA-13K and ScholScan, with ablations confirming that task staging and reward design are necessary for stable gains.

  • Takeaways & Limitations

    The results support Scan-style verification as a trainable capability improved through aligned task design, data construction, and reward modeling.

  • Takeaways & Limitations

    The study focuses on paper-verifiable scientific errors, while broader peer-review judgments and weaknesses requiring extensive external domain knowledge may be underrepresented.

Abstract

from arXiv · show

Multimodal large language models (MLLMs) are increasingly capable scientific assistants, yet they remain far from fully autonomous research. This transition requires models to actively inspect academic papers, build global evidence views, and make traceable judgments without prespecified issues or evidence. However, existing work provides limited task paradigms or training studies for such issue- and evidence-absent verification. We study this challenge through scientific error detection, where models must determine whether errors exist and justify them with evidence-based reasoning. To fill this gap, we present VERA-RL, a reinforcement-learning formulation for scientific error detection over academic papers. Following a Reason--Verify--Scan progression, we construct VERA-13K, a 12,900-sample dataset organized into 4,300 matched chains, covering 6 scientific-error categories across the research workflow and broad natural-science domains. We further introduce fine-grained rewards for reasoning completeness, evidence alignment, and error precision. Training Qwen3-VL-8B with VERA-RL substantially improves verifiable reasoning, approaching flagship MLLMs such as Gemini 3 Pro and Qwen3-VL-235B-A22B on Scan.

1 Introduction

The paper frames issue- and evidence-absent scientific verification as a missing capability for autonomous research and proposes VERA-RL to train it through staged tasks, targeted rewards, and matched data. Experiments report improved verifiable reasoning and show that both task staging and reward design support stable gains.

  • Motivation: MLLMs must move beyond answering specified questions toward deciding what to check, gathering dispersed evidence, and judging whether paper claims are supported.Such paper-level judgment underlies research-quality assessment, idea discovery, and scientific-work refinement.
  • Motivation: Existing methods usually provide the target problem and supporting evidence in advance, while training methodologies for global Scan verification remain limited.ScholScan defines Scan as full-text verification without section-anchored evidence, but treats it primarily as an evaluated capability.
  • VERA-RL: VERA-RL turns Scan into a trainable Reason–Verify–Scan progression from evidence-specified reasoning to issue- and evidence-absent verification.The formulation couples progressive task decomposition with reinforcement-learning rewards for verifiable reasoning over academic papers.
  • VERA-13K: VERA-13K contains 12,900 filtered samples covering 6 scientific-error categories across research-process risk points and natural-science domains.The dataset uses a reusable construction pipeline and matched three-stage chains.
  • Reward design: Fine-grained rewards target reasoning completeness, evidence alignment, and error precision, aligning reinforcement learning with scientific error verification.These rewards represent core dimensions required for verifiable judgments.
  • Results: VERA-RL substantially improves Qwen3-VL-8B-Instruct and reaches performance comparable to the Qwen3-VL-235B-A22B series.Ablations show that both the staged task paradigm and reward system are essential for stable gains.

2 Related Work

Prior scientific-paper benchmarks often reduce evaluation to localized retrieval, explicit-clue QA, or static Scan assessment. The paper positions VERA-RL as a training approach for full-paper, issue- and evidence-absent verification rather than merely another evaluation setting.

  • Full-document reasoning: Earlier work often focused on isolated paragraphs or figures, while full-document methods treated papers as sparse mixtures of key passages and irrelevant text.This framing narrows evaluation toward long-context retrieval paired with localized reasoning.
  • Benchmark paradigms: Most benchmarks retain a QA paradigm with explicit clues and presupposed answer existence, diverging from real-world scientific tasks.PRISMM-Bench simulates reviewer-style understanding but still embeds these assumptions.
  • Scan verification: ScholScan introduces assumption-absent and evidence-absent Scan conditions, but treats Scan as a static capability without examining how to train it.VERA-RL builds on this task setting to study systematic training.
  • Reinforcement learning: Existing long-context reinforcement-learning designs improve grounding or retrieval-conditioned reasoning but provide limited guidance for Scan-style verification over full academic papers.LoongRL and QwenLong-L1 use paragraph concatenation, while VRAG-RL incorporates retrieval into the RL pipeline.
  • Taxonomy and data: VERA-13K extends the ScholScan taxonomy into 6 error categories organized around major failure points in the research workflow.The dataset provides a structured basis for training and evaluating the proposed staged verification capability.

3 Methodology

The methodology frames scientific error detection as a three-stage Reason–Verify–Scan curriculum and trains structured, evidence-based judgments with DAPO and multidimensional rewards. VERA-13K supplies matched task chains derived from errors in accepted papers and peer reviews.

  • 3.1 Task Definition: Reason assumes an error with specified evidence, Verify supplies candidate evidence without assuming an error, and Scan removes both issue and evidence cues.Each stage produces evidence points, reasoning steps, and a final judgment.
  • 3.2 Dataset Construction: VERA-13K contains 12,900 samples formed by rewriting 4,300 errors into matched Reason–Verify–Scan chains across six research-workflow error categories.The dataset combines accepted papers with review-derived objective errors and balances the three stages in each split.
  • 3.2 Dataset Construction: Accepted papers receive paragraph-level injected errors, while review-derived papers contribute filtered objective scientific errors before chain rewriting and quality control.Seed-1.6-Thinking Pass@4 filtering retains samples judged correct or partially correct.
  • 3.3 RL for Verifiable Reasoning: DAPO samples multiple trajectories for each paper-verification instance and updates the policy using token-level advantages normalized within each sampled trajectory group.The training setup uses structured outputs consisting of a question, input paper, and gold answer.
  • 3.3 RL for Verifiable Reasoning: The final reward combines reasoning completeness, evidence alignment, and error precision, with structured matching reducing dependence on unconstrained LLM-as-a-Judge preferences.Completeness measures covered reference answer points, alignment grounds judgments in evidence, and precision penalizes unsupported error claims.
  • 3.3 RL for Verifiable Reasoning: Reason and Verify emphasize completeness and precision, whereas Scan assigns evidence alignment a comparable role because evidence must be identified rather than given.The task-specific reward weights are (0.6, 0, 0.4) for Reason and Verify and (0.4, 0.4, 0.2) for Scan.

4 Experiments

Experiments evaluate VERA-RL across baselines, post-training variants, external transfer, training dynamics, reward ablations, task-framework ablations, and qualitative Scan behavior. Results show improved verifiable reasoning, but gains depend on staged supervision and multidimensional rewards.

  • Main Results: Gemini 3 Pro reaches 60.0 on Reason but only 24.3 on Scan, exposing a large gap in global evidence construction.
  • Main Results: Post-training improves Qwen3-VL-8B across rewards, with RL approaching Qwen3-VL-235B-A22B-Thinking on Rcompleteness and Ralignment while surpassing it on the composite score.SFT also raises Rprecision through format and distributional regularization.
  • External Transfer: VERA-RL transfers to ScholScan, moving Qwen3-VL-8B from near-zero performance to non-trivial scores and close to Qwen3-VL-235B-A22B-Instruct on several metrics.Absolute scores remain low in this harder cross-benchmark setting.
  • Training Dynamics: Reason and Verify show converging fine-grained reward dynamics, consistent with Reason serving as an internal-knowledge reference point.Overall rewards increase in phases with moderate oscillations during staged training.
  • Reward Ablation: The multidimensional reward configuration outperforms the Rcompleteness-only variant on Rfinal and Rcompleteness, whose sparse training signal causes instability and Scan collapse.The reward components are described as coupled rather than independent.
  • Task-Framework Ablation: Pure Scan performs worse than the main setting on Reason, Verify, and Scan, while additional Scan-only training improves slowly and fails to stabilize nearly all metrics.Removing Reason and Verify supervision weakens the evidence-conditioned reasoning and judgment needed for Scan.

5 Conclusion

The paper presents VERA-RL as a staged reinforcement-learning approach for verifiable scientific error detection, supported by the VERA-13K dataset. Experiments on VERA-13K and ScholScan show improvements, while ablations identify task staging and reward design as necessary for stable gains.

  • Conclusion: VERA-RL decomposes Scan into Reason, Verify, and Scan and rewards reasoning completeness, evidence alignment, and error precision.
  • Conclusion: VERA-13K contains 12,900 samples organized into matched three-stage chains across 6 scientific-error categories.
  • Conclusion: Experiments on VERA-13K and ScholScan show consistent improvements, while ablations confirm that task staging and reward design are necessary for stable gains.

Limitations

The study focuses on paper-verifiable scientific error detection rather than the full range of peer-review judgments. Its dataset, model training, and paper representations define additional scope boundaries.

  • Scope: The work does not cover peer-review judgments such as novelty, significance, writing quality, or broader research impact.These judgments often depend on community context and subjective assessment.
  • Dataset Boundary: VERA-13K emphasizes errors verifiable from the paper itself, so implicit weaknesses and errors requiring extensive external domain knowledge may be underrepresented.
  • Evaluation Boundary: Training experiments mainly use Qwen3-VL-8B, leaving larger-scale RL training, broader model families, and alternative paper representations for future study.
  • Domain Coverage: Beyond computer science, VERA-13K includes 655 papers and 1,605 samples from broader scientific domains.
  • Data Composition: VERA-13K combines controlled-edit samples with review-derived samples, whose proportions remain similar across SFT, RL, and test splits.Review-derived samples account for 27.2%, 32.8%, and 30.7% of the SFT, RL, and test sets, respectively.

B.1 Supplementary Results

Supplementary analyses examine metrics, generalization, data distributions, and training choices. They report preserved related capabilities, paper-disjoint gains, broad dataset coverage, and stronger performance from the selected training checkpoints.

  • Metrics: For Reason and Verify, supplementary results report Rcompleteness, Rprecision, and Rfinal, while evidence alignment is disabled in these evidence-specified settings.All scores are scaled by 100.
  • External Benchmarks: VERA-RL maintains or modestly improves performance across three external benchmarks covering long-document understanding and scientific reasoning.The results suggest no cost to the related general capabilities measured there.
  • Generalization: Paper-disjoint evaluation leaves results nearly unchanged, with SFT and RL gains persisting after removing overlapping source papers.
  • Dataset Analysis: VERA-13K reports distributions across broad scientific domains and major AI/ML subfields, with category statistics covering six scientific-error types.
  • Training Checkpoints: The 1-epoch SFT checkpoint achieves stronger Scan completeness and a higher overall Scan score than the step-200 checkpoint, so it initializes RL training.Step 200 nevertheless improves over the Instruct model.

C Metrics Definition in ScholScan

ScholScan evaluates structured answers across target-error detection, evidence localization, reasoning faithfulness, and unrelated-error penalties. VERA-RL weights completeness, alignment, and precision according to whether issue and evidence cues are specified.

  • Metrics Definition: ScholScan parses each model answer into predicted evidence, reasoning, and error-related indicators before scoring it.The variables include Iexist, Icontain, predicted evidence ˆE, predicted reasoning ˆR, and unrelated-claim count n.
  • Metrics Definition: The detection score requires the answer to identify the annotated target error.
  • Metrics Definition: The evidence-location score measures overlap with gold evidence while penalizing overreporting.
  • Metrics Definition: The reasoning score uses the matched prefix length between predicted and gold reasoning chains.
  • Reward Design: For Reason and Verify, reward weights are (0.6, 0, 0.4), while Scan uses (0.4, 0.4, 0.2) for completeness, alignment, and precision.Alignment is disabled when evidence is specified and becomes central when evidence must be discovered.
  • Reward Design: Reward-variance analysis supports completeness as the primary signal in Reason and Verify, and completeness plus alignment as the main Scan objectives.Precision remains comparatively stable in Scan and functions as a constraint against excessive error reporting.
  • Robustness Analysis: Across post-hoc weight variants, Qwen3-VL-8B consistently improves over its base and SFT variants, while Scan remains relatively harder.

E Case Study

Case studies show VERA-RL changing Scan behavior from broad surface-level judgments to targeted evidence selection and cross-section verification. The examples include pipeline leakage and contradictions between ablation results and textual conclusions.

  • Case Study: VERA-RL makes Qwen3-VL-8B-Instruct more likely to identify specific verification evidence than the original instruction-tuned model.The original model often gives broad no-error judgments based on superficial consistency checks.
  • Pipeline Distortion: In a pipeline-distortion case, the model traces retrieval across text, equation, algorithm, and appendix to identify information leakage from answer entities.
  • Inference and Conclusions: In an inference-and-conclusions case, the model detects a contradiction between ablation results and the claim that current-state information is necessary.
  • Case Study: The cases indicate gains in evidence selection, cross-section consistency checking, and converting paper-level evidence into verifiable error judgments.

F Reliability Analysis

Independent rescoring by Qwen3-27B and Gemini 2.5 Flash agrees closely with the original reward scores across Reason, Verify, and Scan.

  • Cross-Evaluator Agreement: High cross-evaluator correlation across Reason, Verify, and Scan suggests the reward signals are not specific to one evaluator.The analysis independently rescores step-30 RL rollouts using Qwen3-27B and Gemini 2.5 Flash.

G Reproducibility and Ethics Statement

VERA-13K uses newly constructed samples from open or publicly available sources, and the code and data are publicly released. Supplementary materials document reward statistics, reweighting, evaluator correlations, and qualitative cases.

  • Reproducibility and Ethics: VERA-13K does not reuse samples from previously released benchmarks.
  • Reproducibility and Ethics: Conference samples were crawled from OpenReview, while Nature Communications and PNAS papers came entirely from open-access sources.
  • Reproducibility and Ethics: The authors state that these sources avoid privacy, ethical, and conflict-of-interest concerns.
  • Reproducibility and Ethics: The code and data are publicly available in the VERA-RL GitHub repository.
Loading 2608.26596v1…