AI Paper Atlas

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

Topological Active Inference for Task Disambiguation

Yangbo Wei, Zhen Huang, Shaoqiang Lu, Junhong Qian, Chen Wu, Lei He

ICML 2026 · Large language models › Language, safety and interpretability · approach: Statistical frameworks for LLM evaluation

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
use persistent homology clustering instead of K-Means (ablation)TAI · Ambi-Nav and Ambi-Codesuccess rateImproved
increase sample size by 5x for Standard BED (ablation)Standard BED · Ambi-Benchsuccess rateNo clear effect
use TAI topological structure with few samples (N=20) instead of Standard BED (ablation)Standard BED · Ambi-Benchsuccess rateImproved
replace Standard BED with TAI under variable renamingAmbi-Code · Ambi-Codesuccess rate drop: -15.4% -> -0.5%Improved
replace Standard BED with TAI under hybrid syntactic attackAmbi-Code · Ambi-Codesuccess rate drop: -28.5% -> -1.2%Improved
use TAI instead of Standard BED on a Qwen-0.6B modelQwen-0.6B · Ambi-Benchsuccess rate (%): 15.2 -> 64.5Improved
use TAI instead of Standard BED on Qwen-14BQwen-14B · Ambi-Benchsuccess rate (%): 78.5 -> 99.2Improved
use TAI cluster-level soft belief update instead of candidate-level BED under combined noiseAmbi-Code · Ambi-Code, pflip=.20 punk=.20success rate (%): 38.6 -> 85.4Improved
use TAI instead of Standard BED on clean oracleAmbi-Code · Ambi-Codesuccess rate (%): 75.0 -> 100Improved

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