Skip to content

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

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Evaluating epistemic grounding and recovery in information-seeking LLM agents

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

SeekBench is the first process-level evaluation framework for LLM search agents that measures epistemic competence — not just final answer accuracy. It operationalizes three competencies (groundedness, recovery, calibration) via formal metrics computed over annotated agent traces. Applied to seven RL-trained agents on Qwen-2.5-7B across 28,493 traces, it exposes behavioral gaps that answer-only metrics miss.

Problem

Existing evaluations of LLM search agents rely almost exclusively on answer-level metrics (exact match, F1), ignoring the intermediate reasoning process captured in multi-turn traces \(T = \langle\tau_1, \tau_2, \ldots, \tau_T\rangle\). Agents can achieve high benchmark scores while hallucinating unsupported claims, failing to adapt search strategies, or answering prematurely with insufficient evidence — failures entirely invisible to final-answer metrics.

Method

SeekBench proceeds in three phases:

Phase 1 — Annotation schema. Three expert annotators coded 190 traces (1,800+ steps) across seven open-source agents through iterative Content Analysis (Krippendorff 2018), reducing 12 candidate fields to 8 well-defined features (Cohen's \(\kappa > 0.8\)). Each step is labeled with a functional type (e.g., InformationSynthesis, PlanFormation, StateAssessment for reasoning; Initial/Repeat/FollowUp/Refined for search) and a quality attribute (grounding, clarity, sufficiency).

Phase 2 — Evidence state and competency definition. Evidence state \(E_{i,t} = C_{i,t} + Q_{i,t} \in \{0,1,2\}\) encodes clarity (\(C\)) and sufficiency (\(Q\)) of retrieved evidence at turn \(t\). Three competencies are derived via latent construct inference: - Groundedness: reasoning grounded in retrieved evidence - Recovery: adaptive query reformulation to escape low-evidence states - Calibration: answering if and only if evidence is good

Phase 3 — Quantitative metrics. - Reasoning Quality Index (RQI): \(\text{RQI}_{\text{model}} = \mathbb{E}_{i}\left[\mathbb{E}_{t \in S_i}[G_{i,t}]\right]\), where \(G_{i,t}\in\{0,1\}\) is the binary grounding label per reasoning step; decomposable by evidence state and reasoning type. - Evidence Recovery Function (ERF): \(\text{ERF}(t) = \frac{1}{N}\sum_{i=1}^{N}\mathbf{1}(T_{\text{recover},i} \leq t)\), the cumulative proportion of traces achieving \(E_{i,t}=2\) or correct answer by turn \(t\). - Calibration Error (CE): \(\text{CE} = \mathbb{E}_i\left[\sum_{k=0}^{2}P(E_{i,t}=k)\left|P(\text{answer}_{i,t}=1 \mid E_{i,t}=k) - \pi^*(k)\right|\right]\), where \(\pi^*(k) = \mathbf{1}[k=2]\) is the ideal policy; \(\text{CE}=0\) is perfect calibration.

An LLM-as-judge pipeline using GPT-4.1-mini (\(\kappa=0.731\) vs. human, cost $0.0087/trace) scales annotation to 28,493 traces.

Key Contributions

  • SeekBench framework: process-level annotation schema + three operationalized epistemic metrics + automated LLM judge pipeline, all validated against human experts (\(\kappa > 0.8\) schema, \(\kappa > 0.7\) judge).
  • Formal metric definitions: RQI, ERF, and CE with mathematical grounding in evidence state \(E_{i,t}\), enabling reproducible large-scale evaluation.
  • Empirical case study: 28,493 traces, 283,950 steps across seven QA benchmarks (NQ, TriviaQA, PopQA, HotpotQA, 2Wiki, MusiQue, Bamboogle) and four RL-trained agents (Search-R1, ReSearch, ASearcher, DeepResearcher).
  • Finding: RL training fails to improve evidence-grounded reasoning; few-shot prompting achieves higher RQI than all RL-trained agents; Search-R1 shows specialized synthesis (IS-type RQI) not visible in F1 scores.

Results

  • Answer-level F1 ranking: ASearcher > Search-R1 > ReSearch > Few-shot ≈ DeepResearcher > Base.
  • Groundedness (RQI): Few-shot achieves highest RQI (0.27), outperforming all RL-trained agents — RL training does not improve reasoning groundedness.
  • Reasoning type breakdown: All models struggle most with PlanFormation and StateAssessment (lower type-level RQI) relative to InformationSynthesis; models are relatively better at synthesizing retrieved text than at planning or self-assessing knowledge gaps.
  • Search-R1 specificity: Search-R1 exhibits above-average IS-type RQI (synthesis) not captured by F1; combining Search-R1's synthesis strength with Base model's reasoning improves performance.
  • Scale: 28,493 traces across 7 benchmarks and 7 agent variants; additional evaluation of 32B ASearcher and WebSailor on GAIA with web browsing.
  • LLM judge cost-effectiveness: GPT-4.1-mini at $0.0087/trace and 2.48s achieves \(\kappa=0.731\) vs. human; GPT-5 achieves \(\kappa=0.754\) at higher cost.

Limitations

  • Evaluation is limited to Qwen-2.5-7B-based agents; generalizability to larger or differently-architected models is not fully established in the main paper.
  • Evidence state (\(E_{i,t}\)) depends on annotator/judge judgment of clarity and sufficiency, which may not generalize cleanly to highly specialized or technical domains.
  • The "ideal calibration policy" (\(\pi^*\): answer iff \(E_{i,t}=2\)) is a simplification; in practice, agents may correctly answer with partial evidence (\(E=1\)) through internal knowledge.
  • Data sanitization removes questions answerable from internal knowledge, potentially narrowing benchmark coverage and limiting cross-study comparisons to unsanitized datasets.
  • Schema development and validation relied on 190 traces — a relatively small seed set for iterative refinement, though LLM judge scaling mitigates downstream scale concerns.

Relevance to Agentic AI / LLM Agents

SeekBench directly targets the epistemic loop common to all tool-using LLM agents — identify gap → retrieve → reason → decide — making its metrics applicable beyond QA to any retrieval-augmented agent. The finding that RL training optimizes final-answer accuracy while degrading or failing to improve grounded intermediate reasoning is a critical diagnostic for the agentic AI community, which is increasingly deploying RL-trained agents in high-stakes settings. The formal decomposition of agent behavior into groundedness, recovery, and calibration provides a vocabulary and measurement toolkit for agent evaluation that complements existing benchmarks focused solely on task outcomes. This work establishes that process-level evaluation is essential for distinguishing truly capable agents from those that achieve correct outputs through epistemically unsound paths.