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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| train search agents with RL (versus few-shot prompting the base model) on final-answer reward | Qwen2.5-7B-Instruct search agents · Reasoning groundedness (RQI) | RQI: Few-shot RQI 0.27 is higher than all RL-trained agents | Hurt |
| RL training of search agents (versus base model) for calibration | Qwen2.5-7B-Instruct · Calibration (SeekBench traces) | Overconfident answer rate: 0.631 -> 0.353 | Improved |
| RL training of search agents (versus base model) on calibration error | Qwen2.5-7B-Instruct · Calibration (SeekBench traces) | Calibration Error: 0.329 -> 0.309 | Improved |
| RL training of search agents (versus base model) on overcautious abstention | Qwen2.5-7B-Instruct · Calibration (SeekBench traces) | Overcautious abstention rate: 0.030 -> 0.085 | Hurt |
| use refined and follow-up queries instead of repeated queries | search agent recovery · Recovery from low evidence | recovery speed (Kaplan-Meier) | Improved |
| answer with good evidence (E=2) versus without supporting evidence | RL-trained agents · Answer accuracy | accuracy (%): 8.4% -> 31.6% | Improved |
| feed agent reasoning traces and evidence from other agents to Search-R1 for answer generation | Search-R1 · Agent synthesis | average F1 gain: +2.61 F1 | Improved |
| pair Base model as answer generator with evidence from other agents | Base Qwen2.5-7B-Instruct · Agent synthesis | average F1 gain: +2.42 F1 | Improved |
| add inference-time feedback from epistemic competency signals | ASearcher-7B · QA benchmarks | F1: 8.4% F1 improvement | Improved |
| use more advanced LLM judge for annotation agreement | LLM-as-judge · Agreement with human experts | Cohen's kappa: GPT-4.1 0.693 -> GPT-4.1-mini 0.731 -> GPT-5 0.754 | Improved |
| RL training versus base on Plan Formation reasoning | all search agents · RQI by reasoning type | RQI: consistently scoring below 0.2 | No clear effect |
Closest papers by meaning
- ReSearch: Learning to Reason with Search for LLMs via Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- Iterative Self-Incentivization Empowers Large Language Models as Agentic Searchers (NeurIPS 2025 · similarity 0.89)
- AgentBench: Evaluating LLMs as Agents (ICLR 2024 · similarity 0.89)
- MC-Search: Evaluating and Enhancing Multimodal Agentic Search with Structured Long Reasoning Chains (ICLR 2026 · similarity 0.89)
- Search More, Think Less: Rethinking Long-Horizon Agentic Search for Efficiency and Generalization (NeurIPS 2026 · similarity 0.89)
- ReSeek: A Self-Correcting Framework for Search Agents with Instructive Rewards (ICML 2026 · similarity 0.89)
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.
