PHRBench: Post-Hallucination Reasoning

Abstract

Hallucinated information can propagate through multi-stage LLM systems and become part of the context for subsequent reasoning. Existing studies of post-hallucination reasoning (PHR) mainly characterize changes in final outcomes and aggregate reasoning dynamics, leaving how models resolve hallucinated premises at the response level insufficiently understood. In this work, we introduce PHRBench, a controlled benchmark for behaviorally structured PHR across four domains and 18 large language models. PHRBench characterizes each reasoning trajectory independently of final-answer correctness through Hallucination Compliance, Hallucination Avoidance, and Heuristic Correction, and defines an insightful trajectory as successful correction that ultimately reaches the correct answer. Across 4,820 controlled instances, we find that successful recovery remains relatively rare and is associated with more frequent belief updates along the reasoning trajectory. We further find that properties of the hallucinated prompt contain substantial predictive signal for successful recovery, with a lightweight predictor achieving an AUROC of 0.847. These findings provide a behavioral view of post-hallucination reasoning, characterizing how LLMs resolve erroneous context and when successful recovery is likely to occur.

PHRBench overview: hallucinations propagate through downstream systems; compliance, avoidance, and correction differ even when final outcomes agree; recovery shows scaling tensions and predictable signatures. — **Figure 1.** From hallucinated context to downstream recovery. PHRBench separates resolution behavior from final correctness. Click any figure to view at full resolution.

Results across Models and Domains

Heatmap and radar plots of accuracy retention across 18 models, four domains and three hallucination types. — **Accuracy retention.** Performance varies with both the domain and hallucination type. Retention is accuracy under hallucinated context relative to truthful context, expressed as a percentage; it is not absolute accuracy.

Hallucinated context reduces accuracy by 7.7 percentage points on average. Larger models can exhibit both greater vulnerability to erroneous context and more frequent corrective behavior.

Table 1: Outcome performance and reasoning trajectory statistics under the truthful augmentation (T) and hallucinated augmentation (H) settings.

Model	Acc (%)		OL		UI		BUF		BC
Model	T	H	T	H	T	H	T	H	T	H
Proprietary Models
Claude-Sonnet-4	57.5	48.4↓9.1	100.7	113.2	0.054	0.071	0.152	0.329	1.133	1.111
GPT-4o-mini	55.5	48.3↓7.2	71.2	72.4	0.072	0.059	0.130	0.186	0.752	0.720
GPT-4o	51.8	42.3↓9.5	60.0	60.8	0.061	0.053	0.113	0.214	0.659	0.618
GPT-5.2	61.6	44.7↓16.9	43.0	45.1	0.007	0.016	0.013	0.026	0.578	0.515
Gemini-2.0-Flash	56.2	45.1↓11.1	76.2	86.5	0.046	0.061	0.092	0.178	0.745	0.785
Gemini-2.5-Pro	58.5	46.4↓12.1	82.4	95.3	0.045	0.068	0.124	0.244	0.852	0.850
Open-source Models
Qwen2.5-1.5B-Instruct	24.6	20.6↓4.0	115.6	117.1	0.272	0.218	0.091	0.076	1.987	1.748
Qwen2.5-7B-Instruct	44.5	38.0↓6.5	153.2	147.1	0.319	0.151	0.148	0.121	2.265	1.939
Qwen2.5-32B-Instruct	46.8	39.6↓7.2	228.4	215.1	0.245	0.150	0.285	0.353	2.820	2.644
Qwen2.5-72B-Instruct	52.2	44.2↓8.0	276.9	262.8	0.188	0.112	0.414	0.485	3.426	3.150
Qwen3-8B	46.0	43.8↓2.2	225.8	256.6	0.217	0.072	0.576	0.392	2.855	3.470
Qwen3-32B	48.5	41.0↓7.5	296.3	288.4	0.181	0.065	0.605	0.465	3.232	3.820
Qwen3-235B-A22B	58.2	47.4↓10.8	365.0	415.2	0.085	0.045	0.843	0.753	4.155	4.828
Llama-3.1-8B-Instruct	34.2	31.9↓2.3	210.5	199.9	0.378	0.322	0.371	0.207	2.752	2.222
Llama-3.1-70B-Instruct	48.4	40.9↓7.5	285.2	275.4	0.431	0.419	0.523	0.384	3.356	2.989
Llama-3.2-3B-Instruct	31.2	29.5↓1.7	167.0	153.4	0.159	0.137	0.322	0.124	2.076	1.441
GLM-4-9B-Chat	32.3	22.4↓9.9	127.3	79.0	0.171	0.173	0.050	0.038	1.886	1.122
Mistral-7B-Instruct	21.1	15.7↓5.4	131.0	122.4	0.130	0.104	0.262	0.370	1.853	1.472

T: truthful augmentation; H: hallucinated augmentation. Red annotations indicate the decrease in accuracy (percentage points). OL: Output Length; UI: Uncertainty Index; BUF: Belief Update Frequency; BC: Branching Complexity.

Insightful Trajectory Rates

Overview

PHRBench contains 460 base questions, 460 truthful augmentations, and 3,900 hallucinated augmentations, yielding 4,820 controlled instances.
It spans Chemistry, Biomedicine, Physics, and Code Generation, with base questions drawn from eight established benchmarks.
Scientific questions have ten hallucinated variants each; coding problems have three. Each augmentation introduces a targeted semantic error while preserving the task.

460 base questions: 120 Chemistry, 120 Biomedicine, 120 Physics, and 100 Code Generation. Scientific hallucinations are predominantly rule contradictions; code hallucinations are state distortions. — **Dataset composition.** Base-question reasoning categories and hallucinated augmentation types. Chemistry, Biomedicine, and Physics each include 120 base questions; Code Generation includes 100.

Hallucinated Augmentation Types

Type	Description
Rule Contradiction	A premise conflicts with an established rule, principle, or domain fact.
State Distortion	A statement is incorrect under the specific conditions of the current problem.
Pseudoscientific Entanglement	An outdated, discredited, or pseudoscientific theory is introduced as a premise.

Behavioral Resolution

Behavior	Description
Hallucination Compliance	The model accepts the hallucinated premise and reasons from it.
Hallucination Avoidance	The model bypasses the premise without explicitly correcting it.
Heuristic Correction	The model identifies and corrects the hallucinated premise.

An insightful trajectory performs Heuristic Correction and reaches the correct final answer. Behavior is classified independently of final correctness.

Evaluation Metrics

The paper reports final-answer accuracy alongside Output Length (OL), Uncertainty Index (UI), Belief Update Frequency (BUF), and Branching Complexity (BC). Each model is sampled ten times per prompt at temperature 0.7. See Section 4.2 and Appendix B for definitions and configurations.

Qualitative Examples

Qualitative ion-identification example: a compliant trajectory predicts manganese, while a corrective trajectory updates its belief and correctly identifies cadmium. — **Qualitative example.** The supplied case contrasts compliance with a successful corrective trajectory. The belief update is the key behavioral distinction.

Predicting Successful Recovery

Prompt-level structural and semantic features provide predictive signal before generation. The lightweight predictor achieves an AUROC of 0.847. SHAP attributions characterize predictive associations, rather than causal effects.

Global Feature Contributions

SHAP feature contributions led by total prompt length at 20.7%, with candidate and question length at 8.8% each. — **Global feature contributions.** Total prompt length accounts for 20.7% of normalized feature attribution. These are predictive associations, not causal effects.

BibTeX

Citation with the author list provided by the research team; publication details will be updated with the paper release.

@misc{phrbench,
  title = {PHRBench: A Behavioral Evaluation of Post-Hallucination Reasoning in LLMs},
  author = {Linghao Meng and Feng He and Xuan Yang and Junyuan Mao and Pinze Ren and Deqing Mu and Hesen Yang and Qiankun Li},
  note = {Manuscript under review}
}

PHRBench: A Behavioral Evaluation of Post-Hallucination Reasoning in LLMs