AI Paper Atlas

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

Not All Answers Are Contextually Persuadable: Inference Dynamics in Large Language Models under Contextual Influence

Zongye Hu, Weiqing Luo, Yanjie Fu, Yu Gan, Haofeng Zhang, Ziyi Huang

ICML 2026 · Large language models › Language, safety and interpretability · approach: Reasoning mechanisms and failure modes

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
tighten convergence tolerance from 0.1 to 0.01Qwen2.5-1.5B attention layers · OpenBookQA, N=1000% converged layers: 93.0 -> 47.6Hurt
tighten convergence tolerance from 0.1 to 0.01Falcon3-3B-Base attention layers · MINTAKA, N=1000% converged layers: 100 -> 93.3Hurt
tighten convergence tolerance from 0.1 to 0.01Qwen3-4B attention layers · Simple QA, N=1000% converged layers: 100 -> 92.7Hurt
increase repetition length N from 500 to 1000answer-shift sign prediction, Transfer regime · Transfer (mean of 6 models)accuracy: 85.5 -> 86.1No clear effect
increase repetition length N from 500 to 1000answer-shift sign prediction, Mislead regime · Mislead (mean of 6 models)accuracy: 82.6 -> 80.0Hurt
increase repetition length N from 500 to 1000answer-shift sign prediction, Correct regime · Correct (mean of 6 models)F1: 78.9 -> 89.8Improved
compare predicted shift magnitude with Amp across modelsQwen2.5-1.5B · SYCONstability agreement: 4.25 -> 0.13Mixed

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