AI Paper Atlas

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

How Stable is the Next Token? A Geometric View of LLM Prediction Stability

Deyuan Liu, Zecheng Wang, Zhanyue Qin, Zhiying Tu, Dianhui Chu, Dianbo Sui

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Pruning and circuits in LLMs

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
dTCB-guided prompt enhancement instead of baseline promptMMLU ambiguous questions · MMLU AQ setPDR (%): 30 -> 10Improved
dTCB-guided prompt enhancement instead of baseline promptMMLU ambiguous questions · MMLU AQ setworst-case accuracy: 0.15 -> 0.30Improved
dTCB-guided prompt enhancement instead of baseline promptMMLU ambiguous questions · MMLU AQ setaverage dTCB: 1983.0 -> 2734.0Improved
dTCB-guided prompt enhancement instead of baseline promptGSM8K ambiguous questions · GSM8K AQ setaccuracy: 0.35 -> 0.65Improved
dTCB-guided prompt enhancement instead of baseline promptGSM8K very confident questions · GSM8K VCQ setaccuracy: 0.85 -> 0.92Improved
co-optimize prompts for dTCB instead of perplexity aloneprompt optimization · MMLU & GSM8Kworst-case accuracy: 0.40 -> 0.65Improved
co-optimize prompts for dTCB instead of perplexity aloneprompt optimization · MMLU & GSM8Kaverage accuracy: 0.70 -> 0.72No clear effect
co-optimize prompts for dTCB instead of perplexity aloneprompt optimization · MMLU & GSM8Kaverage PPL: 1.9 -> 2.4Hurt
add misleading clarification '(7 days)' to the questiongsm8k_811 with algebraic ICLs · GSM8K gsm8k_811accuracy: 100.0 -> 0.00Hurt
add misleading clarification '(7 days)' to the questiongsm8k_811 with algebraic ICLs · GSM8K gsm8k_811dTCB: 8.20 -> 46.97Improved
remove ICLs and use algebraic decomposition instruction zero-shotgsm8k_811 clarified question · GSM8K gsm8k_811dTCB: 46.97 -> 49450.23Improved
cluster competitor embeddings toward the top-1 token with probabilities fixed (ablation)MMLU prompts · MMLUfraction where dTCB ordering held: 90%Improved

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