Research map / Large language models / Retrieval and knowledge
Demystifying Scientific Problem-Solving in LLMs by Probing Knowledge and Reasoning
Alan Li, Yixin Liu, Arpan Sarkar, Doug Downey, Arman Cohan
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 DeepSeek-R1 knowledge ingredients in context (ablation) | Qwen2.5-7B-Instruct · GPQA | accuracy: 35.27 -> 47.19 | Improved |
| add DeepSeek-R1 knowledge ingredients in context (ablation) | Llama-3.1-8B-Instruct · GPQA | accuracy: 28.13 -> 43.57 | Improved |
| add the model's own knowledge ingredients in context (ablation) | Qwen2.5-7B-Instruct · GPQA | accuracy: 35.27 -> 34.24 | No clear effect |
| add R1 KIs on top of reasoning fine-tuning (Qwen-BOTH) (ablation) | Qwen-BOTH · GPQA | accuracy: 40.81 -> 54.46 | Improved |
| add reasoning fine-tuning on math plus STEM traces (ablation) | Qwen2.5-7B-Instruct · SciReas-Pro | accuracy: 13.97 -> 21.11 | Improved |
| add reasoning fine-tuning on math-only traces (ablation) | Llama-3.1-8B-Instruct · SciReas | accuracy: 31.25 -> 35.49 | Improved |
| use KIs from Qwen-Math instead of base Qwen (ablation) | Qwen2.5-7B-Instruct · MMLU-Pro* | accuracy: 59.03 -> 63.66 | Improved |
| use KIs from Llama-Math instead of base Llama (ablation) | Llama-3.1-8B-Instruct · GPQA | accuracy: 29.06 -> 29.69 | No clear effect |
| raise reasoning effort from low to high | o3-mini · SciReas (avg) | accuracy: 56.9 -> 63.7 | Improved |
| raise reasoning effort from low to high | Gemini-2.5-Pro · SciReas (avg) | accuracy: 66.6 -> 67.0 | No clear effect |
| measure on SciReas-Pro instead of SciReas | GPT-5 low vs high effort · SciReas-Pro | accuracy gap: 3.01 -> 12.22 | Improved |
| fine-tune base models on math reasoning traces (no science data) (ablation) | Qwen/Llama math variants · knowledge recall of KIs | accuracy: 72.30 -> 73.02 | No clear effect |
Closest papers by meaning
- SciReason: A Controllable Benchmark for Scientific Reasoning in LLMs (NeurIPS 2026 · similarity 0.91)
- SciBench: Evaluating College-Level Scientific Problem-Solving Abilities of Large Language Models (ICML 2024 · similarity 0.91)
- CURIE: Evaluating LLMs on Multitask Scientific Long-Context Understanding and Reasoning (ICLR 2025 · similarity 0.89)
- Scientists' First Exam: Probing Cognitive Abilities of MLLM via Perception, Understanding, and Reasoning (NeurIPS 2025 · similarity 0.88)
- SciInstruct: a Self-Reflective Instruction Annotated Dataset for Training Scientific Language Models (NeurIPS 2024 · similarity 0.88)
- The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 2025 · similarity 0.88)
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.
