Aider Polyglot
Harbor nativeAider Polyglot: 6种语言diff编辑
Verified71%Coding
- Samples
- 225
- Released
- 2024
- Scores
- 61
PHRBench
PHRBench: A Behavioral Evaluation of Post-Hallucination Reasoning in LLMs
A controlled benchmark with 4820 instances across 4 domains evaluating how LLMs behaviorally resolve hallucinated premises via compliance, avoidance, or correction.
How to run
ltb · phrbench
$ npm i -g lettokenburn-cli$ ltb bench info phrbench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
12 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-5.2-chat | 91.0 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| GPT-4o-mini | 89.3 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Gemini 2.5 Pro | 87.9 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Claude Sonnet 4 (2025-05-14) | 87.7 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Gemini 2.0 Flash Experimental | 84.6 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Qwen2.5-72B-Instruct | 75.7 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| GPT-4o | 72.1 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Qwen3-32B | 70.2 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Qwen2.5 32B Instruct | 67.9 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Qwen3 8B | 66.5 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Qwen2.5-7B-Instruct | 64.5 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Qwen2.5-1.5B-Instruct | 35.7 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsLoading community runs…
About
PHRBench contains 4820 controlled instances spanning Chemistry, Biomedicine, Physics, and Code Generation. It evaluates 18 LLMs by characterizing reasoning trajectories independently of final-answer correctness.
Scoring classifies behaviors as Hallucination Compliance, Avoidance, or Heuristic Correction. Key findings reveal successful recovery is rare, associated with belief updates, and predictable from prompt features (AUROC 0.847).