AI Paper Atlas

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

Expert Heads: Robust Evidence Identification for Large Language Models

Qi Wu, Jianfeng Qu, Ximing Li, Zhixu Li

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
use Response-as-Source instead of Question-as-Source Expert HeadsLLaMA-3-8B-Instruct · HotpotQAPrecision@2: 88.23 -> 90.72Improved
use Response-as-Source instead of Question-as-Source Expert HeadsQwen-2.5-7B-Instruct · 2WikiMultiHopQAPrecision@2: 74.69 -> 77.94Improved
use Response-as-Source instead of Question-as-Source Expert HeadsMistral-7B-Instruct-v0.3 · 2WikiMultiHopQAPrecision@2: 70.72 -> 73.27Improved
use Expert Head voting instead of direct LLM RankMistral-7B-Instruct-v0.3 · HotpotQAPrecision@2: 36.45 -> 88.89Improved
use Expert Head voting instead of direct LLM RankLLaMA-3-8B-Instruct · HotpotQAPrecision@2: 66.31 -> 90.72Improved
use Expert Head voting instead of direct LLM RankLLaMA-3-8B-Instruct · 2WikiMultiHopQAPrecision@2: 76.49 -> 73.47Mixed
use Expert Head voting instead of direct LLM RankQwen-2.5-7B-Instruct · 2WikiMultiHopQAPrecision@2: 76.57 -> 77.94No clear effect
use LLaMA Expert Heads (R) instead of the best dense retriever BGELLaMA-3-8B-Instruct · HotpotQAPrecision@2: 75.23 -> 90.72Improved
use LLaMA Expert Heads (R) instead of BGELLaMA-3-8B-Instruct · 2WikiMultiHopQAPrecision@2: 77.12 -> 77.30No clear effect
use Mistral Expert Heads (Q) instead of BGEMistral-7B-Instruct-v0.3 · 2WikiMultiHopQAPrecision@2: 77.12 -> 70.72Hurt
use all heads of a single middle layer as expert heads vs other layers (ablation)LLaMA-3-8B-Instruct · HotpotQAPrecision@2Mixed
use stricter activation thresholds for Response-as-Source Expert Heads (ablation)LLaMA-3-8B-Instruct · HotpotQAPrecision@2Improved

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