PaperlayerView exact revision on arXiv

Source-linked AI summary

HalluPeer: A Taxonomy-driven Benchmark for Detecting Hallucinations in Scientific Peer Reviews

Tzu-Ling Lin, Dong-Ting Yao, Teng-Fang Hsiao, Wei-Chih Chen, Hong-Han Shuai

arXiv:2609.03580v1cs.AIcs.CL

TL;DR

Scientific peer-review hallucinations require verification against long technical papers, but existing benchmarks do not capture this setting. HalluPeer introduces a taxonomy-driven benchmark and controlled injection pipeline, finding that general verifiers struggle with scientific reviews while HalluPeer patterns also occur in authentic reviews. These results support source-aware verification for AI-assisted peer review.

  • Problem

    Existing hallucination benchmarks do not model scientific peer review, where review claims must be grounded in long technical papers and separated from subjective critique.

  • Method

    HalluPeer aligns paper content with human-written and hallucination-injected reviews, using a hierarchical taxonomy, aspect-compatible contexts, controlled injection, and automated filtering.

  • Results

    Existing verifiers struggle to distinguish unsupported claims from legitimate scientific critique, while domain-specific fine-tuning improves performance and HalluPeer-defined patterns occur in authentic reviews.

  • Takeaways & Limitations

    Trustworthy auditing of scientific reviews requires specialized, source-aware verification rather than off-the-shelf factuality models.

  • Takeaways & Limitations

    The dataset remains synthetic, and its injected-hallucination distribution may not perfectly reflect subtle naturally occurring review errors.

Abstract

from arXiv · show

The growing scale of academic peer review has motivated the use of Large Language Models (LLMs) as review assistants, yet LLMs can generate fluent but unsupported claims that undermine review reliability. Existing hallucination benchmarks are not designed for peer review, where verification requires grounding claims in long, technical papers. We introduce HalluPeer, a benchmark for detecting hallucinations in scientific peer reviews, providing aligned triples of paper content, human-written reviews, and hallucination-injected reviews, annotated for detection, classification, and localization. Our pipeline induces a peer-review-specific hallucination taxonomy, identifies review contexts, and injects hallucinations with automated filtering. Experiments on 12K papers and 38K reviews show that existing detectors struggle to separate hallucinations from legitimate critique, while evaluation on authentic reviews demonstrates that HalluPeer-defined hallucination patterns occur in real peer reviews, highlighting the critical need for source-aware verification. Our project page can be found in https://github.com/Lin-TzuLing/HalluPeer.git

1 Introduction

HalluPeer addresses the reliability risks of LLM-assisted peer review by defining hallucinations as paper-grounding failures and benchmarking their detection. It combines aligned review data with taxonomy-guided construction and evaluates verification across detection, classification, and localization.

  • LLM-assisted reviewing can produce fluent but unsupported claims that mislead meta-reviewers and affect editorial decisions.
  • Verifying review claims requires evidence synthesis across long technical papers while separating factual claims from subjective critique.
  • HalluPeer defines a hallucination as a review claim unsupported by or incorrect with respect to the submitted paper.
  • The benchmark pairs paper content with human-written and hallucination-injected reviews, using aspect tags, compatible templates, style-preserving injection, and automated filtering.
  • Existing verifiers struggle with scientific reviews, while domain-specific evaluation covers detection, classification, and localization and transfers to authentic review errors.

2 Related Work

Prior hallucination research addresses general-domain text and LLM-assisted reviewing, but lacks a benchmark for paper-grounded hallucination detection in scientific peer reviews. HalluPeer targets this missing evaluation setting.

  • Existing hallucination resources use general-domain text or retrieved snippets rather than long technical papers requiring cross-section evidence reasoning.
  • The growing reviewing burden motivates LLM use for review drafting, structured critique, meta-review assistance, and review-quality assessment.
  • LLM reviewers can generate inconsistent or paper-unsupported claims, establishing review hallucination as a realistic risk in scientific peer review.
  • Prior work lacks a dedicated benchmark for systematically detecting, categorizing, and localizing hallucinated review claims grounded in submitted papers.

3 Taxonomy of Review Hallucinations

HalluPeer builds a hierarchical, review-specific hallucination taxonomy by recursively refining coarse categories into operationally defined error types. Multi-model agreement, overlap filtering, and expert validation improve taxonomy refinement.

  • Top-Down Taxonomy Construction: The taxonomy recursively refines coarse hallucination categories into fine-grained review-specific types with operational injection instructions and evaluation labels.
  • Top-Down Taxonomy Construction: Each taxonomy node stores a concept name, operational description, parent, and children, while root-to-leaf paths support analysis at multiple granularities.
  • LLM-Guided Recursive Decomposition: Taxonomy expansion uses LLM proposers to generate and define child concepts recursively until a maximum depth or no valid subdivision is reached.
  • LLM-Guided Recursive Decomposition: Prompting enforces minimally overlapping sibling categories that collectively cover major manifestations of each parent concept.
  • LLM-Guided Recursive Decomposition: A three-stage workflow combines multi-model consensus filtering, global overlap identification, and human expert validation to reduce idiosyncratic concepts and redundancy.

4 HalluPeer Construction Pipeline

HalluPeer constructs controllable hallucination examples by combining review sentences, aspect labels, and compatible taxonomy concepts. An automated injector generates fluent edits, while verification and filtering remove unsuitable outputs.

  • The dataset draws ICLR and NeurIPS records from OpenReview for paper-grounded hallucination construction and validation.
  • Review sentences receive aspect labels that constrain compatible hallucination types, such as favoring number perturbations in Evaluation sentences.
  • Templates pair review sentences with fine-grained taxonomy concepts and their injection instructions and labels.
  • Coarse screening removes inapplicable concepts before fine-grained template construction, and compatibility checks retain naturally applicable sentence-aspect pairs.
  • An LLM injector creates fluent hallucinated sentences from original sentences and taxonomy instructions, then a post-hoc verifier removes semantically equivalent outputs.

5 Experimental Results

The experiments evaluate hallucination detection, type classification, localization, and cross-venue transfer. Existing general-purpose methods struggle with technically grounded review verification, while domain-specific fine-tuning substantially improves performance and transferability.

  • Evaluation Tasks: The benchmark evaluates detection, classification, and localization at review, sentence, token, and span levels.Detection uses Accuracy, Precision, Recall, F1, and MCC; classification uses Macro-F1 and Micro-F1; localization uses Token-F1 and span-level metrics.
  • Hallucination Detection: Pretrained verifiers perform near randomly at review-level detection, with MCC scores ≤0.03.Technical review claims require multi-hop reasoning over full-length papers, unlike the datasets used to train these verifiers.
  • Hallucination Detection: MCC 0.61 and Accuracy 0.82 are achieved by sentence-level RA-LLM (KR), the strongest prompting variant.The few-shot variant outperforms prompting with explicit reasoning in the reported comparison.
  • Hallucination Detection: F1 scores of 0.90 and 0.91 are achieved by fine-tuned Qwen3-32B at review and sentence levels, with sentence-level MCC of 0.87.Fine-tuned models consistently outperform prompting-based zero-shot methods, including larger frontier models.
  • Hallucination Type Classification: Prompting-based type classification performs poorly, with the best review-level Macro-F1 reaching only 0.19.Entity and Number are relatively easier, whereas Context-based Meaning Error, Hyperbole, and Temporal remain highly challenging.
  • Hallucination Localization: GPT-5.2 reaches 0.58 Token-F1 and 0.46 Exact Span-F1 among zero-shot localization methods, but precise boundaries remain difficult.Hallucinated content is sparse and embedded within otherwise correct scientific critique, contributing to the localization challenge.
  • Cross-Venue Transferability: Cross-venue detection transfer reaches review-/sentence-level F1 scores of 0.90/0.91 and 0.86/0.93 in opposite directions.The relatively small differences across transfer directions suggest limited reliance on venue-specific writing styles or formatting.

6 Alignment with Real Review Errors

A case study of authentic NeurIPS reviews finds that HalluPeer’s taxonomy captures error patterns occurring in real peer-review text. Manual inspection supports alignment between detected errors and the proposed hallucination categories.

  • Authentic Reviews: 13,803 human-written NeurIPS 2024 reviews were screened for hallucination patterns defined by HalluPeer.A fine-tuned detector flagged potential hallucinations, followed by manual inspection of 200 instances.
  • Authentic Reviews: Manual inspection found that identified review errors align with HalluPeer’s taxonomy-defined hallucination patterns.The study checked both whether each flagged instance was a genuine hallucination and whether its predicted type matched.

7 Conclusion

HalluPeer frames peer-review auditing as paper-grounded verification and provides a taxonomy-driven benchmark with an aspect-aware injection pipeline. Experiments show that domain-specific fine-tuning improves detection, while source-aware attribution remains challenging.

  • Conclusion: HalluPeer formulates review auditing as paper-grounded verification requiring long-context reasoning over manuscripts.The benchmark uses a hierarchical taxonomy and aspect-aware injection pipeline to generate realistic hallucinated reviews.
  • Conclusion: Existing verifiers struggle to distinguish unsupported claims from legitimate critique, while domain-specific fine-tuning substantially improves performance.Authentic-review evaluations indicate that HalluPeer-defined hallucination patterns occur in real peer reviews, but source-aware attribution remains challenging.

8 Limitations

HalluPeer acknowledges limits in coverage, realism, and representativeness. Naturally occurring hallucinations are sparse and costly to annotate, while synthetic injections and computer-science data may not capture real-world or cross-disciplinary variation.

  • Naturally occurring review hallucinations are sparse and require domain expertise, making large-scale annotation prohibitively labor-intensive.
  • The synthetic dataset may not fully reflect subtle, drift-based errors or complex reasoning failures in real reviews.
  • The data source is restricted to computer science conferences hosted on OpenReview, limiting coverage across scientific disciplines.
  • Broader sampling could improve distributional coverage but would reduce supervision reliability through noisier labels.

9 Ethical Considerations

The paper identifies dual-use and peer-review integrity risks from synthetically generating hallucinated reviews. It restricts use to defensive offline research and recommends human oversight.

  • Dual-Use Risks: The hallucination-injection pipeline could be misused to generate more convincing hallucinated reviews.
  • Dual-Use Risks: The taxonomy and generation scripts are restricted to defensive research purposes to mitigate dual-use risks.
  • Integrity of the Peer-Review Process: No hallucinated reviews were submitted to real venues or used in editorial decisions; the dataset is intended for offline training and evaluation.
  • Integrity of the Peer-Review Process: The paper advocates human-in-the-loop review systems in which AI serves as a diagnostic aid rather than an autonomous decision-maker.

A.1 HalluPeer Dataset Statistic

HalluPeer combines large-scale conference review data with a hierarchical hallucination taxonomy and deterministic LLM-assisted construction. Its taxonomy organizes broad error categories into fine-grained injection types.

  • Dataset Scale: The dataset covers ICLR 2019–2024 and NeurIPS 2021–2024, containing 12,000 papers, 38,063 reviews, and over 1.02M review sentences.
  • Taxonomy Structure: The taxonomy contains 265 nodes and 205 leaf nodes, reaches depth 3, and uses terminal leaves as fine-grained hallucination types.
  • Taxonomy Structure: The hierarchy begins with nine coarse categories and progressively decomposes them into domain-specific and fine-grained hallucination types.
  • Construction Pipeline: The construction pipeline tags review aspects, generates taxonomy descriptions and templates, and uses LLM-based filtering, checking, injection, and verification modules.
  • Taxonomy Examples: The taxonomy includes numerical discrepancies such as year discrepancies and event-year fabrication as operational hallucination categories.

C Baseline Implementation

The baseline implementation evaluates specialized verifiers and general-purpose LLMs across detection, classification, and localization. Systems use retrieval, entailment, prompting, or supervised adaptation, with task-specific metrics.

  • Verification Baselines: Specialized baselines include faithfulness scoring, NLI verification, retrieval-then-verification, and claim-level checking frameworks.
  • LLM Baselines: General-purpose baselines include zero-shot prompting and instruction-tuned models evaluated on detection, hallucination-type classification, and localization.
  • LLM Baselines: RA-LLM retrieves relevant paper evidence for each review sentence and compares knowledge-retrieval, chain-of-thought, and contrastive prompting strategies.
  • Training Setup: Instruction-tuned baselines use supervised fine-tuning with NF4 quantization and QLoRA adapters applied across attention and MLP linear modules.
  • Reproducibility: Training and inference use deterministic temperature-zero evaluation, with experiments conducted on a single NVIDIA H100 GPU.
  • Evaluation Metrics: Classification reports Macro-F1 and Micro-F1, while localization reports token-level and span-level metrics including Exact Match Span-F1.

E.1.1 Ablation Study of RA-LLM

RA-LLM ablations show that retrieval and prompting interact: few-shot knowledge-retrieval prompting performs best at sentence level, while filtered evidence can deprive reasoning-heavy strategies of necessary context. Across ICLR results, domain-specific fine-tuning remains strongest for detection, classification, and localization.

  • Prompting and retrieval: MCC 0.61 and Accuracy 0.82 make RA-LLM (KR) the strongest sentence-level prompting configuration under BM25.
  • Prompting and retrieval: Review-level MCC for Contrast drops from 0.26 with full-paper access to 0.15 with filtered evidence.The authors attribute this degradation to missing global information needed for multi-step reasoning or contradiction analysis.
  • Detection: Specialized verifiers reach review-level MCC ≤0.04 on ICLR, indicating near-random generalization.
  • Detection: Qwen3-32B fine-tuning achieves sentence-level F1 0.94 and review-level F1 0.89 on ICLR detection.The compact Qwen2.5-3B also surpasses all zero-shot baselines.
  • Classification: Zero-shot review-level classification remains weak at Macro-F1 0.15, while fine-tuning improves difficult Hyperbole, Temporal, and Context-based Meaning Error categories.
  • Localization: Qwen3-32B fine-tuning reaches 0.93 Token-F1 and 0.90 Exact Span-F1 for hallucination localization.

G.2 Results and Analysis

On authentic NeurIPS reviews, the fine-tuned detector recovers every expert-annotated hallucination while reducing the screening pool, and cross-generator evaluations show stable detection, classification, and localization. The taxonomy is recursively refined and used to construct compatible, semantically altered hallucination examples.

  • Authentic-review evaluation: TPR = 100.0% at FPR = 22.1% lets the fine-tuned Qwen3-32B detector recover all 20 authentic hallucinations.The evaluation contains 1,161 reviews with 20 positive and 1,141 negative instances.
  • Authentic-review evaluation: The detector reduces 1,161 authentic reviews to 272 candidates while retaining all 20 expert-annotated hallucinations.The authors frame this as a high-recall screening stage before human or stronger-model adjudication.
  • Taxonomy and injection pipeline: The taxonomy is recursively decomposed from coarse concepts into fine-grained nodes, then filtered through cross-model similarity and overlap review.The refinement pipeline combines multiple LLM-generated trees, automated overlap identification, and human validation.
  • Taxonomy and injection pipeline: Compatible leaf templates are selected for each sentence before an injector rewrites it and a verifier discards semantically equivalent rewrites.
  • Cross-generator robustness: Cross-generator localization remains stable, with Token-F1 within 0.02 of the in-domain baseline for Mistral and slight ICLR improvements for Llama.
Loading 2609.03580v1…