AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

HE-SNR: Uncovering Latent Logic via Entropy for Guiding Mid-Training on SWE-bench

Yueyang Wang, Jiawei Fu, Baolong Bi, Xili Wang, Xiaoqing Liu

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning for reasoning data · 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
compute the metric on Action tokens instead of Thinking tokens (ablation)HE-SNR · SWE-bench Verified correlationPearson r: 0.5581 -> 0.9666Improved
compute the metric on Action tokens instead of Thinking tokens (ablation)HE-SNR · SWE-bench Verified correlationKendall tau: 0.5192 -> 0.9440Improved
add XML tag removal to Action filtering (ablation)HE-SNR · SWE-bench Verified correlationPearson r: 0.9666 -> 0.9526Hurt
add XML tag removal to Action filtering (ablation)HE-SNR · SWE-bench Verified correlationKendall tau: 0.9440 -> 0.9558Improved
add AST-based comment removal (full pipeline) (ablation)HE-SNR · SWE-bench Verified correlationKendall tau: 0.9676 -> 0.9794Improved
add AST-based comment removal (full pipeline) (ablation)HE-SNR · SWE-bench Verified correlationPearson r: 0.9520 -> 0.9649Mixed
filter out Thought tokens and stylistic noise before computing HE-PPL (ablation)HE-PPL · SWE-bench Verified correlationlinearity vs Pass@1Improved
use HE-SNR instead of standard PPL across Long-Context TaxMoE-A26B mid-training · 128K extension, SWE-bench Verifiedagreement with SWE-bench Pass@1Improved
vary entropy threshold epsilon over ln t, t in [2,5] (ablation)HE-SNR · SWE-bench Verified correlationPearson r / Kendall tauNo clear effect
apply SFT to base checkpointsMoE-A3B / MoE-A26B · high-entropy token setHE-PPL and HE-SNRHurt
extend context window via RoPE scaling (Long-Context Tax)MoE-A26B · 128K extension step 200PPL and Top-1 accuracyHurt

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