Research map / Large language models / Instruction and fine-tuning
From Reasoning Traces to Reusable Modules: Understanding Compositional Generalization in Language Model Reasoning
Lingjing Kong, Xin Liu, Guangyi Chen, Martin Q. Ma, Xiangchen Song, Yuekai Sun, Mikhail Yurochkin, Taylor W. Killian, Ruslan Salakhutdinov, Kun Zhang, Eric P. Xing, Zhengzhong Liu
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 RL after SFT on compound traces | SFT on L=3 traces · unseen compositions L=4 | accuracy (%): 4.8 -> 42.6 | Improved |
| use compound traces (L=3) instead of atomic modules (L=1) for RL (ablation) | SFT+RL · OOD compound traces L=4 | accuracy (%): 14.8 -> 42.6 | Improved |
| add RL on atomic modules only | SFT · OOD compound traces L=4 | accuracy (%): 4.8 -> 14.8 | Improved |
| add RL to SFT | SFT on L=3 traces · atomic skills L=1 | accuracy (%): +43.7% | Improved |
| add RL to SFT | SFT on L=3 traces · unseen compositions across depths | accuracy (%): avg +40.3% | Improved |
| add RL | SFT, no absence · Unseen-3 | accuracy: 0.22 -> 0.55 | Improved |
| add RL | SFT, no absence · Seen-3 | accuracy: 0.83 -> 0.98 | Improved |
| include all atomic skills in SFT data (vs 4 absent) (ablation) | SFT+RL · Unseen-3 | accuracy: 0.36 -> 0.55 | Improved |
| include all atomic skills in SFT data (vs 4 absent) (ablation) | SFT+RL · Unseen-2 | accuracy: 0.52 -> 0.67 | Improved |
| include all atomic skills in SFT data (vs 4 absent) (ablation) | SFT+RL · Seen-2 | accuracy: 0.99 -> 0.99 | No clear effect |
| re-inject compositions containing held-out skill during RL (vs isolated atomic skill) (ablation) | RL with held-out skill · compositions with held-out skill, deeper L | accuracy | Improved |
| re-inject lower-depth compositions with held-out routing mechanism (ablation) | RL with held-out router · unseen L=3 and L=4 | accuracy | Improved |
Closest papers by meaning
- Compositional Reasoning in Language Models under Reinforcement Learning Post-Training (NeurIPS 2026 · similarity 0.93)
- On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models (ICML 2026 · similarity 0.91)
- RL Squeezes, SFT Expands: A Comparative Study of Reasoning LLMs (ICLR 2026 · similarity 0.91)
- Revisiting Reinforcement Learning for LLM Reasoning from A Cross-Domain Perspective (NeurIPS 2025 · similarity 0.90)
- Compositional Generalization from Learned Skills via CoT Training: A Theoretical and Structural Analysis for Reasoning (ICLR 2026 · similarity 0.90)
- How does RL Post-training Induce Skill Composition? A Case Study on Countdown (NeurIPS 2026 · 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.
