Research map / Large language models / Language, safety and interpretability
ThinkBench: Dynamic Out-of-Distribution Evaluation for Robust LLM Reasoning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces ThinkBench, a dynamic OOD evaluation framework that rephrases AIME and GPQA questions (scenario-level and attack-level semi-fact perturbations, agent-verified) to measure LLM reasoning robustness and detect data leakage, evaluating 16 LLMs and 4 PRMs.
Problem
Static benchmarks for LLM reasoning suffer from data contamination and answer leakage, so reported accuracy may reflect memorization rather than generalizable reasoning.
Approach
Dynamically generate OOD test variants of reasoning questions using LLM rephrasing agents with verifier agents (scenario-level semi-fact rewrites) plus three attack-style perturbations (TextBugger, CheckList, StressTest), and compare accuracy on original vs. OOD versions, with the AIME 2024 vs. pre-2024 AIME gap used as a leakage signal.
Key results
- Average OOD performance decay of 24.9% on AIME-500 and 11.8% on AIME 2024 across models, suggesting leakage in pre-2024 AIME data.
- o1-preview shows -3.3% gap on AIME 2024 vs -11.2% on AIME-500; o3-mini, Deepseek-R1 and o1-series achieve highest OOD accuracy.
- Human annotators found 100% of 30 AIME 2024 semi-fact questions consistent with original answers.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Evaluate on dynamically generated OOD data instead of original data | o3-mini · AIME-500 | accuracy: 0.922 -> 0.834 | Hurt |
| Evaluate on dynamically generated OOD data instead of original data | o3-mini · AIME 2024 | accuracy: 0.767 -> 0.717 | Hurt |
| Evaluate on dynamically generated OOD data instead of original data | o1-mini · AIME 2024 | accuracy: 0.567 -> 0.600 | Helped |
| Evaluate on dynamically generated OOD data instead of original data | Deepseek-R1 · GPQA Diamond | accuracy: 0.747 -> 0.682 | Hurt |
| Use Math-Shepherd PRM instead of majority voting | Qwen2.5-Math-7B-IT · AIME 2024 OOD | accuracy: 0.133 -> 0.233 | Helped |
| Use OpenR PRM instead of majority voting | Qwen2.5-Math-7B-IT · AIME 2024 OOD | accuracy: 0.133 -> 0.200 | Helped |
| Use Skywork PRM instead of majority voting | Qwen2.5-Math-7B-IT · AIME-500 OOD | accuracy: 0.464 -> 0.500 | Helped |
| Use Qwen PRM instead of majority voting | Qwen2.5-Math-7B-IT · AIME 2024 OOD | accuracy: 0.133 -> 0.300 | Helped |
| Increase test-time computation budget | Qwen2.5-Math-7B-IT with PRMs · AIME 2024 OOD and AIME-500 OOD | accuracy | Helped |
Surprising findings
- Smaller models such as LLaMA3.2-3B show higher OOD than ID accuracy on GPQA Diamond, attributed to random guessing near chance.
- GPT-4o and Llama3.1-70B answer original questions correctly but fail the rephrased versions, indicating reliance on surface form.
Limitations
- Evaluation focuses on math and science reasoning; lacks diversity of reasoning types.
- Scenario-level generation uses single-path rephrasing rather than multi-scenario branching, limiting OOD diversity.
- Single run with temperature 0.7 and pass@1 on small sets (30 AIME 2024 questions), so variance may be high.
Experimental setup
Datasets: AIME-500 AIME 2024 GPQA Diamond MMLU
Benchmarks: AIME 2024 AIME-500 GPQA Diamond MMLU 570
Metrics: accuracy pass@1 OOD gap (%) best-of-N accuracy
Baselines: o1-preview o1-mini o3-mini Deepseek-R1 GPT-4o Deepseek-V3 s1.1-32B Qwen2.5-72B-IT
Builds on: semi-factual causality TextBugger CheckList StressTest Math-Shepherd OpenR
Closest papers by meaning
- DyVal: Dynamic Evaluation of Large Language Models for Reasoning Tasks (ICLR 2024 · similarity 0.91)
- DyCodeEval: Dynamic Benchmarking of Reasoning Capabilities in Code Large Language Models Under Data Contamination (ICML 2025 · similarity 0.90)
- DARG: Dynamic Evaluation of Large Language Models via Adaptive Reasoning Graph (NeurIPS 2024 · similarity 0.89)
- metabench - A Sparse Benchmark of Reasoning and Knowledge in Large Language Models (ICLR 2025 · similarity 0.89)
- ORACLE: Optimizing Reasoning Abilities of Large Language Models via Constraint-Led Synthetic Data Elicitation (AAAI 2026 · similarity 0.89)
- A State-Transition Framework for Efficient LLM Reasoning (ICLR 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.