AI Paper Atlas

Research map / Large language models / Retrieval and knowledge

Do LLM Agents Know How to Ground, Recover, and Assess? Evaluating Epistemic Competence in Information-Seeking Agents

Jiaqi Shao, Yuxiang Lin, Munish Prasad Lohani, Yufeng Miao, Bing Luo

ICLR 2026 · Large language models › Retrieval and knowledge · approach: Explanatory and multiple-choice QA 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
train search agents with RL (versus few-shot prompting the base model) on final-answer rewardQwen2.5-7B-Instruct search agents · Reasoning groundedness (RQI)RQI: Few-shot RQI 0.27 is higher than all RL-trained agentsHurt
RL training of search agents (versus base model) for calibrationQwen2.5-7B-Instruct · Calibration (SeekBench traces)Overconfident answer rate: 0.631 -> 0.353Improved
RL training of search agents (versus base model) on calibration errorQwen2.5-7B-Instruct · Calibration (SeekBench traces)Calibration Error: 0.329 -> 0.309Improved
RL training of search agents (versus base model) on overcautious abstentionQwen2.5-7B-Instruct · Calibration (SeekBench traces)Overcautious abstention rate: 0.030 -> 0.085Hurt
use refined and follow-up queries instead of repeated queriessearch agent recovery · Recovery from low evidencerecovery speed (Kaplan-Meier)Improved
answer with good evidence (E=2) versus without supporting evidenceRL-trained agents · Answer accuracyaccuracy (%): 8.4% -> 31.6%Improved
feed agent reasoning traces and evidence from other agents to Search-R1 for answer generationSearch-R1 · Agent synthesisaverage F1 gain: +2.61 F1Improved
pair Base model as answer generator with evidence from other agentsBase Qwen2.5-7B-Instruct · Agent synthesisaverage F1 gain: +2.42 F1Improved
add inference-time feedback from epistemic competency signalsASearcher-7B · QA benchmarksF1: 8.4% F1 improvementImproved
use more advanced LLM judge for annotation agreementLLM-as-judge · Agreement with human expertsCohen's kappa: GPT-4.1 0.693 -> GPT-4.1-mini 0.731 -> GPT-5 0.754Improved
RL training versus base on Plan Formation reasoningall search agents · RQI by reasoning typeRQI: consistently scoring below 0.2No 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