AI Paper Atlas

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

BiasBusters: Uncovering and Mitigating Tool Selection Bias in Large Language Models

Thierry Blankenstein, Jialin Yu, Zixuan Li, Vassilis Plachouras, Sunando Sengupta, Philip Torr, Yarin Gal, Alasdair Paren, Adel Bibi

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Bias measurement and mitigation

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
apply subset-filter then uniform-sampling mitigationGPT-4.1 mini tool selection · bias benchmarkdelta_model (lower better): 0.380 -> 0.094Improved
apply mitigationGPT-4.1 mini · bias benchmarkdelta_pos: 0.422 -> 0.079Improved
biased continued pre-training on one endpointQwen3-8B · Language Identification clusterselection share of exposed endpoint: 0.006 -> 0.128Improved
raise temperature from 0 to 2GPT-3.5 · bias benchmarkdelta_model: 0.350 -> 0.285Improved
raise top-p from 0.7 to 1.0GPT-3.5 · bias benchmarkdelta_model: 0.346 -> 0.340No clear effect
increase model size (Qwen3 1.7B to 235B)Qwen3 · bias benchmarkdelta_modelImproved
cyclic vs random API orderingGPT-3.5 · bias benchmarkselection distributionNo clear effect
reword or restructure system promptGPT-3.5 · bias benchmarkselection distributionMixed
scramble descriptions and parameters of toolsGemini 2.5 Flash · metadata perturbationTV distance from base: 0.450Improved
swap descriptions of most- and least-selected toolsGemini 2.5 Flash · metadata perturbationTV distance from base: 0.338Improved
scramble tool names onlyGemini / GPT-3.5 · metadata perturbationselection shiftNo clear effect
add random forest instead of linear regression for predicting selectionfeature analysis · 50 APIs per modelpredictive performanceNo 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