- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper introduces PHRBench, a benchmark for evaluating post-hallucination reasoning in LLMs. The benchmark augments existing reasoning tasks with truthful or synthetic hallucinated premises and evaluates not only final-answer correctness, but also how models resolve the erroneous context through hallucination compliance, avoidance, or heuristic correction. Across 18 models and 4 domains, the paper studies scaling trends, reasoning dynamics, and whether successful recovery can be predicted from prompt-level features.
Critical strengths
- The paper studies an important and relatively underexplored failure mode beyond standard hallucination detection. Separating final-answer correctness from the model’s resolution behavior is conceptually useful, since a correct answer does not necessarily imply that the hallucinated premise was properly identified and corrected.
- The distinction among hallucination compliance, avoidance, and heuristic correction, together with the definition of insightful trajectories, provides a useful framework for analyzing different recovery behaviors rather than collapsing them into a single accuracy number.
Overall I personally found this an interesting failure mode analysis where frontier LLM could keep improving.
Critical weaknesses
- I'm slightly concerned about the diversity and external validity of the benchmark. Although the paper reports 4,820 instances, these are derived from only 460 unique base questions, with many synthetic hallucinated variants generated from the same underlying tasks. Moreover, the hallucinated contexts are deliberately generated and constrained to contain a single targeted semantic error. This controlled setup is useful for analysis, but it is unclear how representative it is of naturally occurring hallucinations produced by upstream models in realistic multi-stage systems. The limited number of unique tasks may also make it easier for downstream analyses to exploit dataset-specific regularities.
- The predictive analysis may be driven substantially by surface-level or dataset-specific confounds rather than generalizable properties of hallucination recovery. E.g. strongest feature is Total Prompt Length which makes me doubt whether the analysis is valid. As it is possible that the predictor is learning task-format or difficulty proxies rather than intrinsic recoverability of hallucinated premises (Given the relatively small and structured set of unique base questions).
- An initial post training experiment for improving those behaviors would significantly strengthen the paper.
Questions for authors
- How well does the proposed benchmark transfer to naturally occurring hallucinations? And is the eval result strongly correlate to recent hallucination benchmarks?
- How much of the reported predictability of insightful trajectories is due to hallucination-specific semantics versus task-format or difficulty proxies?
- How should the model-suite results be interpreted given that many evaluated models are no longer representative of the current frontier? If the goal is primarily behavioral analysis, this may be less problematic; however, if PHRBench is intended as a benchmark of current model robustness, evaluation on more recent models would seem important.
- Does the paper provide evidence that the identified behavioral signals can actually be used to improve post-hallucination recovery? The current work is primarily diagnostic. Have the authors tried a simple training-time or inference-time intervention based on the proposed behavioral or predictive signals?
5 comments
Question2 & Weakness 2:
The predictive analysis may be driven substantially by surface-level or dataset-specific confounds ... the predictor is learning task-format or difficulty proxies rather than intrinsic recoverability of hallucinated premises (Given the relatively small and structured set of unique base questions).
How much of the reported predictability of insightful trajectories is due to hallucination-specific semantics versus task-format or difficulty proxies?
Authors’ Response:
We stratified archived prompt-level out-of-fold predictions without retraining or changing scores.
1. Task formats: constrained answer spaces (finite-set selection) versus free-form generation (answers or executable code).
| Task format | Instances | Positive instances | AUROC |
|---|---|---|---|
| Constrained answer space | 1,901 | 289 | 0.8388 |
| Free-form generation | 396 | 51 | 0.8903 |
2. Difficulty-proxy groups. We averaged within-domain percentile ranks of question-only reading length, condition count, heuristic entity count, and numerical count, then grouped questions by within-domain tertiles. Original task inputs are retained; hallucinated augmentations, candidate answers, outcome labels, and prediction scores are excluded from grouping.
| Difficulty-proxy group | Instances | Positive instances | AUROC |
|---|---|---|---|
| Low | 827 | 110 | 0.8471 |
| Medium | 662 | 121 | 0.8175 |
| High | 808 | 109 | 0.8721 |
The same within-group comparison gives 0.8470 AUROC.
| Restriction | Within-stratum AUROC | Change from overall AUROC | 95% bootstrap CI for change |
|---|---|---|---|
| Within task format | 0.8406 | −0.0067 | [−0.0189, +0.0054] |
| Within difficulty-proxy group | 0.8470 | −0.0004 | [−0.0098, +0.0055] |
Discrimination persists within these separately defined groups, so coarse format or difficulty-proxy differences alone do not explain predictability.
Question3:
How should the model-suite results be interpreted given that many evaluated models are no longer representative of the current frontier? ... evaluation on more recent models would seem important.
Authors’ Response:
We evaluated four newer models using the same PHRBench questions, truthful/hallucinated augmentations, and scoring procedures.
| Model | (%) | (%) | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| GPT-5.6 Sol | 67.3 | 58.4 (↓8.9 pp) | 92.6 | 124.8 | 0.028 | 0.041 | 0.104 | 0.162 | 0.742 | 0.881 |
| Claude Opus 4.8 | 72.1 | 68.6 (↓3.5 pp) | 146.8 | 198.3 | 0.044 | 0.049 | 0.196 | 0.237 | 1.012 | 1.146 |
| Gemini 3.1 Pro | 69.4 | 54.8 (↓14.6 pp) | 184.9 | 179.6 | 0.052 | 0.046 | 0.241 | 0.218 | 1.487 | 1.392 |
| Qwen3.7-Max | 70.8 | 69.6 (↓1.2 pp) | 121.4 | 118.9 | 0.033 | 0.030 | 0.174 | 0.168 | 0.984 | 0.951 |
We will add the results of current frontier models in the Table 1 of our paper.
Question4 & Weakness 3:
An initial post training experiment for improving those behaviors would significantly strengthen the paper.
Does the paper provide evidence that the identified behavioral signals can actually be used to improve post-hallucination recovery? ... Have the authors tried a simple training-time or inference-time intervention based on the proposed behavioral or predictive signals?
Authors’ Response:
We tested predictive signals through an inference intervention.
1. Setup. We use 100 independent questions balanced across four domains, with one hallucinated-context variant each. Interventions share the same additional generation budget. Signal-guided conditions select identical instances; random selection matches per-domain counts. Revised answers replace initial answers without ground-truth correctness access.
2. Results. Accuracy covers all 100 questions, including unselected ones. Changes are in pp; transitions compare final answers with the same baseline answers.
| Model | Condition | Accuracy (%) | Change vs. baseline (pp) | Incorrect → correct | Correct → incorrect |
|---|---|---|---|---|---|
| Qwen2.5-7B-Instruct | Baseline | 38.0 | — | — | — |
| Qwen2.5-7B-Instruct | Random + Correction | 42.0 | +4.0 | 8 | 4 |
| Qwen2.5-7B-Instruct | Signal + General Review | 44.0 | +6.0 | 10 | 4 |
| Qwen2.5-7B-Instruct | Signal + Correction | 49.0 | +11.0 | 14 | 3 |
| Qwen3-8B | Baseline | 44.0 | — | — | — |
| Qwen3-8B | Random + Correction | 47.0 | +3.0 | 7 | 4 |
| Qwen3-8B | Signal + General Review | 50.0 | +6.0 | 9 | 3 |
| Qwen3-8B | Signal + Correction | 56.0 | +12.0 | 15 | 3 |
Signal + Correction exceeds Baseline by 11.0 and 12.0 pp, Random + Correction by 7.0 and 9.0 pp, and Signal + General Review by 5.0 and 6.0 pp, respectively. Gains mainly reflect more repaired errors, supporting the utility of predictive signals for inference recovery.
Dear Reviewer p3fR,
We sincerely appreciate your time and effort in providing such thoughtful reviews and insightful comments.
Could you please kindly let us know if our rebuttal has addressed your concerns?
If you have any further questions regarding our work, we would be delighted to address them.
If our responses have addressed your concerns, we would greatly appreciate your reconsideration of your overall assessment.
Best regards,
Authors
Response to Reviewer p3fR
Thank you for recognizing the value of post-hallucination reasoning and separating correctness from resolution behavior.
Question1 & Weakness 1:
Authors’ Response:
1. Controlled construction. Following HIVE [1], PHRBench isolates the effects of targeted erroneous premises while preserving the underlying question. The 4,820 instances derive from 460 base questions across four domains and eight source datasets; variants introduce distinct substantive errors rather than merely paraphrasing one error. This supports controlled comparisons; external validity requires further evaluation.
2. Naturally generated hallucinations. We evaluated 94 eligible questions from RAGTruth [2] and 120 from RAGBench [3], with one upstream response per question. RAGTruth provides human-annotated hallucination spans; RAGBench provides sentence-level support annotations. RAGBench questions come from PubMedQA, FinQA, and TechQA, with cross-benchmark deduplication. Five downstream models generate one response per condition:
Source documents and generation settings are fixed. Unsupported claims are not automatically treated as false; cases whose errors cannot be resolved from the evidence are excluded.
All five models have lower accuracy under naturally erroneous context, with decreases of 7.45–10.00 pp.
Behavioral annotations cover all H responses pooled across models: 470 for RAGTruth and 600 for RAGBench. Labels are independent of final-answer correctness.
All three behaviors remain observable. Correction has higher final-answer accuracy than Compliance, while Avoidance also permits correct answers. Correction does not guarantee a correct outcome. These results support transfer of selected PHRBench phenomena to naturally generated errors.
3. Model-to-model transfer. We also evaluated a two-stage system using Qwen2.5-7B-Instruct and Qwen3-8B in both roles. An upstream model generates reasoning, which a downstream model uses to answer. We use 100 questions (25 per domain) and one generation per condition, comparing naturally erroneous context (H) with a minimally corrected counterpart (T). Upstream errors occurred on 27 and 19 questions, respectively; accuracy below is within these subsets.
Consistent gains after correcting upstream errors demonstrate the framework's relevance to actual model-to-model information transfer.
References
[1] He et al., HIVE: Understanding Post-Hallucination Reasoning in Vision Language Models, ECCV 2026.
[2] Niu et al., RAGTruth: A Hallucination Corpus for Developing Trustworthy Retrieval-Augmented Language Models, ACL 2024.
[3] Friel et al., RAGBench: Explainable Benchmark for Retrieval-Augmented Generation Systems, arXiv:2407.11005, 2024.