AI Paper Atlas

Research map / Large language models / Agents, code and math

In Agents We Trust, but Who Do Agents Trust? Latent Preferences Steer LLM Generations

Mohammad Aflah Khan, Mahsa Amani, Soumi Das, Bishwamittra Ghosh, Qinyuan Wu, Krishna P. Gummadi, Manish Gupta, Abhilasha Ravichander

ICLR 2026 · Large language models › Agents, code and math · approach: Agent instruction-following benchmarks

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
attach different source labels (news outlets) to semantically identical content12 LLMs · synthetic source-preference taskspreference percentageImproved
scale up model sizeLLMs (GPT-4.1-Mini, Phi-4, Qwen2.5-7B vs smaller Llama/Qwen) · source preference strengthvariance of preference percentageImproved
change paper topic context (ablation)NEJM source tag · Research paper selectionselection rate: 96% -> 19% (Health vs Computer Vision)Mixed
change product category context (ablation)BestBuy source tag · E-commerce selectionselection rate: 51% -> 97% (Grocery vs Electronics)Mixed
represent source by X handle or URL instead of brand name (ablation)GPT-4.1-Nano · Associated Press Fact Check news setpreference percentage: 80% -> 53%Hurt
swap source labels between left and right sources (ablation)LLM news agents · AllSides case studyarticle selection shareImproved
show source information vs hide it (ablation)LLM news agents · AllSides case studyarticle selection shareImproved
add 'Do Not Be Biased' instruction to prompt (ablation)LLM news agents · AllSides case studyselection bias toward left/center sourcesNo clear effect
prompt to optimize price (ablation)GPT-4.1-Nano · Amazon seller selectioncheapest seller chosen: 48.4% -> 70%Improved
prompt to optimize delivery time (ablation)GPT-4.1-Nano · Amazon seller selectionfastest-delivery seller chosen: 53.9% -> 69.3%Improved
pretraining co-occurrence frequency of source with 'journalistic standards'source preference ranking · Infinigram pretraining corporaexplanation of rankingsNo 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