AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Beyond Scalars: Evaluating and Understanding LLM Reasoning via Geometric Progress and Stability

Xinyan Jiang, Ninghao Liu, Di Wang, Lijie Hu

ICML 2026 · Large language models › Language, safety and interpretability · approach: Mechanistic analysis of attention and representations · 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 curvature to displacement-only features (ablation)TRACED · GPQAAUROC: 0.7812 -> 0.8300Improved
add displacement to curvature-only features (ablation)TRACED · GPQAAUROC: 0.7244 -> 0.8300Improved
add both signatures vs displacement only (ablation)TRACED · GSM8KAUROC: 0.7634 -> 0.8061Improved
add reasoning-quality subspace projection (ablation)TRACED · GPQA (DeepSeek-R1-Llama-8B)AUPR: 0.624 -> 0.661Improved
add reasoning-quality subspace projection (ablation)TRACED · GSM8K (DeepSeek-R1-Llama-8B)AUPR: 0.813 -> 0.828Improved
add TRACED geometric score to GRPO rewardGRPO Qwen2.5-0.5B · GSM8Kaccuracy (%): 39.65 -> 43.97Improved
add TRACED geometric score to GRPO rewardGRPO Qwen2.5-0.5B · MATHaccuracy (%): 28.33 -> 30.65Improved
add TRACED geometric score to GRPO rewardGRPO Qwen2.5-1.5B · GSM8Kaccuracy (%): 67.83 -> 71.22Improved
add TRACED geometric score to GRPO rewardGRPO Qwen2.5-1.5B · MATHaccuracy (%): 47.20 -> 50.10Improved
apply centroid alignment before cross-domain transferTRACED · cross-domain transferAUPRImproved
increase subspace dimension k up to 8 (ablation)TRACED · four modelsAUROCImproved
vary reasoning difficulty (Easy to Hard) (ablation)TRACED · DeepSeek-R1-Llama-8B, stratifiedAUROC: 0.775 / 0.748 / 0.766No clear effect

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