Research map / Large language models / Language, safety and interpretability
Reasoning Compartmentalization: Bridging the Concretization Gap via Abstraction-based Routing
Ling-I Wu, Jian Tong, Yu Sun, Xuan Gao, Xu Guo, Qipeng Guo, Kai Chen, Guoqiang Li
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 prompt-based abstraction step | Qwen3-30B-A3B · Bool 3x3 (NL) | accuracy (%): 65.6 (+34.0) | Improved |
| add abstraction-aligned RL | Qwen3-30B-A3B · Bool 3x3 (NL) | accuracy (%): 87.8 (+56.2) | Improved |
| add abstraction-aligned RL | Qwen3-30B-A3B · Bool 5x5 (NL) | accuracy (%): 53.4 (+30.2) | Improved |
| add abstraction-aligned RL | GPT-oss-20B · Bool 5x5 (NL) | accuracy (%): 24.2 (+11.2) | Improved |
| add prompt-based abstraction step | GPT-oss-20B · Bool 3x3 (NL) | accuracy (%): 81.0 (+6.8) | Improved |
| add prompt-based abstraction step | Gemini-2.5-Pro · Bool 5x5 (NL) | accuracy (%): 68.6 (-11.6) | Hurt |
| add prompt-based abstraction step | GPT-o3 · Bool 3x3 (NL) | accuracy (%): 98.0 (+1.0) | Improved |
| add abstraction-aligned RL, out of domain | Qwen3-30B-A3B · PlanBench | accuracy (%): 68.20 -> 73.20 | Improved |
| add abstraction-aligned RL, out of domain | GPT-oss-20B · PlanBench | accuracy (%): 47.40 -> 55.60 | Improved |
| add abstraction-aligned RL, out of domain | Qwen3-30B-A3B · Natural-Plan Calendar | accuracy (%): 84.80 -> 86.20 | No clear effect |
| add abstraction-aligned RL, out of domain | GPT-oss-20B · Natural-Plan Meeting | accuracy (%): 4.00 -> 9.60 | Improved |
| convert formal-language problem to natural language | Qwen3-30B-A3B · Bool 3x3 | accuracy (%): 97.6 -> 31.6 | Hurt |
Closest papers by meaning
- Calibrating Reasoning in Language Models with Internal Consistency (NeurIPS 2024 · similarity 0.91)
- Meaningful Learning: Enhancing Abstract Reasoning in Large Language Models via Generic Fact Guidance (NeurIPS 2024 · similarity 0.90)
- Benchmarking Abstract and Reasoning Abilities Through A Theoretical Perspective (ICML 2025 · similarity 0.89)
- Improving Reasoning Performance in Large Language Models via Representation Engineering (ICLR 2025 · similarity 0.89)
- Beyond Semantics: The Unreasonable Effectiveness of Reasonless Intermediate Tokens (NeurIPS 2026 · similarity 0.89)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (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.
