Research map / Large language models / Agents, code and math
BeyondBench: Benchmark-Free Evaluation of Reasoning in Language Models
Gaurav Srivastava, Aafiya Shamshad Hussain, Zhenyu Bi, Swastik Roy, Priya Pitre, Meng Lu, Morteza Ziyadi, Xuan Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add explicit algorithm description (ablation) | GPT-OSS-120B · Tower of Hanoi (n=7) | accuracy: 18.4% -> 24.2% | Improved |
| add explicit algorithm description (ablation) | GPT-OSS-120B · Sudoku (9x9) | accuracy: 8.2% -> 12.1% | Improved |
| add explicit algorithm description (ablation) | GPT-OSS-120B · N-Queens (n=8) | accuracy: 22.3% -> 28.7% | Improved |
| add explicit algorithm description (ablation) | GPT-OSS-120B · Boolean SAT (5 vars) | accuracy: 38.1% -> 42.8% | Improved |
| add tools (all tools combined) (ablation) | GPT-5 · BeyondBench overall | accuracy drop when tools removed: 16.81% drop without tools | Improved |
| add tools (all tools combined) (ablation) | GPT-5-nano · BeyondBench overall | accuracy drop when tools removed: 43.95% drop without tools | Improved |
| add thinking variant | Qwen3-4B · BeyondBench overall | accuracy: 58.63% -> 60.77% | Improved |
| add thinking variant | Qwen3-30B-MOE · BeyondBench overall | accuracy: 69.00% -> 69.20% | No clear effect |
| add mathematical fine-tuning | Qwen2.5-72B · BeyondBench overall | accuracy: 53.36% -> 48.48% | Hurt |
| add mathematical fine-tuning | Qwen2.5-7B · BeyondBench overall | accuracy: 36.12% -> 31.63% | Hurt |
| scale Qwen3 from 0.6B to 1.7B parameters | Qwen3 · BeyondBench overall | accuracy: 28.21% -> 45.75% | Improved |
| scale Qwen3 from 14B to 32B parameters | Qwen3 · BeyondBench overall | accuracy: +1.79 points | No clear effect |
Closest papers by meaning
- metabench - A Sparse Benchmark of Reasoning and Knowledge in Large Language Models (ICLR 2025 · similarity 0.91)
- RE-IMAGINE: Symbolic Benchmark Synthesis for Reasoning Evaluation (ICML 2025 · similarity 0.90)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.90)
- DARG: Dynamic Evaluation of Large Language Models via Adaptive Reasoning Graph (NeurIPS 2024 · similarity 0.90)
- DyVal: Dynamic Evaluation of Large Language Models for Reasoning Tasks (ICLR 2024 · similarity 0.90)
- From Passive to Active Reasoning: Can Large Language Models Ask the Right Questions under Incomplete Information? (ICML 2025 · similarity 0.90)
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.
