AI Paper Atlas

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

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Gradient-free adaptation and merging · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
add RL after SFT on compound tracesSFT on L=3 traces · unseen compositions L=4accuracy (%): 4.8 -> 42.6Improved
use compound traces (L=3) instead of atomic modules (L=1) for RL (ablation)SFT+RL · OOD compound traces L=4accuracy (%): 14.8 -> 42.6Improved
add RL on atomic modules onlySFT · OOD compound traces L=4accuracy (%): 4.8 -> 14.8Improved
add RL to SFTSFT on L=3 traces · atomic skills L=1accuracy (%): +43.7%Improved
add RL to SFTSFT on L=3 traces · unseen compositions across depthsaccuracy (%): avg +40.3%Improved
add RLSFT, no absence · Unseen-3accuracy: 0.22 -> 0.55Improved
add RLSFT, no absence · Seen-3accuracy: 0.83 -> 0.98Improved
include all atomic skills in SFT data (vs 4 absent) (ablation)SFT+RL · Unseen-3accuracy: 0.36 -> 0.55Improved
include all atomic skills in SFT data (vs 4 absent) (ablation)SFT+RL · Unseen-2accuracy: 0.52 -> 0.67Improved
include all atomic skills in SFT data (vs 4 absent) (ablation)SFT+RL · Seen-2accuracy: 0.99 -> 0.99No 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 LaccuracyImproved
re-inject lower-depth compositions with held-out routing mechanism (ablation)RL with held-out router · unseen L=3 and L=4accuracyImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy