acceptodds
Under review as a conference paper at ICLR 2027

EAV: Evidence Anchoring and Verification for Reliable Agentic Search

35 views

Abstract

Agentic search systems can return incorrect answers even after retrieving relevant evidence, a pattern we trace to three failure modes: evidence blindness, where retrieved documents go uninspected or are used ineffectively; verification failure, where self-review reconsiders existing reasoning without new falsifying evidence; and error propagation, where erroneous intermediate conclusions are reused downstream. Existing remedies either re-examine the agent's own reasoning or seek agreement across multiple complete runs; the former lacks new evidence, and the latter multiplies cost. We propose the Evidence Anchoring–Verification (EAV) framework. During execution, EAV archives raw tool observations and anchors the claims adopted by the executor, together with their sources and conflicts, into a compact evidence chain; before an answer is delivered, an independent tool-calling verifier re-checks high-risk premises against fresh evidence. A rejected candidate triggers diagnosis-guided re-execution that inherits the evidence memory; EAV-Flash and EAV-MAS apply this loop at the task level and at subtask handoffs, respectively. Across nine matched-backbone settings on GAIA (text-only validation), xBench-DeepSearch, and BrowseComp-ZH, EAV-Flash ranks among the leading methods in every setting and attains an average success rate of 82.64%, 3.69 points above MiroFlow, the previous state-of-the-art system, while using 56.1–72.4% fewer tokens; EAV-MAS reaches 80.91%, a 1.96-point gain over the same baseline. With DeepSeek-V4-Flash, EAV-Flash sets the open-source state of the art on xBench-DeepSearch (89.00%) and the state of the art on BrowseComp-Plus (92.89%). On a BrowseComp-Plus subset, archiving, anchoring, and verification contribute gains of 5.8, 7.1, and 7.8 points; anchoring mainly raises error-interception recall, and verification-driven sparse retries offer a better performance–cost trade-off than majority voting and can be combined with voting for further gains.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Agentic Search, Large Language Models, Evidence Verification, Multi-Agent Systems

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_R4xXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies why search agents can answer incorrectly even after retrieving relevant evidence. EAV archives tool observations and maintains a compact evidence chain of the claims, sources, and conflicts adopted during execution. An independent tool-using verifier checks high-risk premises in a fresh context, and rejected candidates trigger diagnostic re-execution with the existing evidence memory. EAV-Flash verifies complete tasks, while EAV-MAS verifies subtask handoffs. Experiments on four search benchmarks with multiple backbones report a favorable accuracy-token trade-off. Across nine matched settings, EAV-Flash achieves 82.64% average success, compared with 78.95% for MiroFlow.

Critical strengths
  1. The design is well motivated by concrete failure analysis. The observed failures distinguish retrieving evidence from inspecting and using it, and highlight the propagation of incorrect intermediate premises. Archiving, claim anchoring, and independent verification directly address these issues, making the system design coherent and practically relevant.
  2. The empirical evaluation covers multiple backbones, four benchmarks, token consumption, and paired component analyses. EAV-Flash improves average success by 3.69 points over MiroFlow while substantially reducing reported token use on the open-web benchmarks. The ablations and error analysis provide useful evidence for the value of anchoring and verification.
Critical weaknesses
  • The distinct contribution relative to existing verification and fact-memory systems, including Co-Sight, FineVerify, and ArcticSwarm, needs clearer differentiation.
  • Verification also triggers re-execution, so its measured gain combines checking with additional execution. A clearer comparison of capabilities and controls under comparable total compute would better attribute the gains to the proposed evidence-chain mechanism.
  • AI Use Statement is missing. ICLR requires this. Please add this to avoid desk rejection.
Questions for authors
  • Nearly all the references are not correct. Check differences among \cite{} \citep{} and \citet{}.
Reviewer_jwt9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EAV, an Evidence Anchoring and Verification framework for improving the reliability of agentic search. The method combines evidence archiving, structured evidence anchoring, and independent verification with diagnosis-guided re-execution to reduce reasoning errors. Experiments across four benchmarks demonstrate improvements in task success rates and token efficiency over existing methods.

Critical strengths
  1. The paper is well-written and well-organized, with clear motivation and detailed explanations of the proposed framework.
  2. Extensive experiments are conducted across multiple benchmarks and LLM backbones, including ablation studies and efficiency analyses, demonstrating the effectiveness of the proposed method.
Critical weaknesses
  1. Inconsistent improvements across benchmarks. EAV achieves notable gains on GAIA but shows limited improvements on xBench-DeepSearch and fails to outperform the strongest baselines on BrowseComp-ZH. The paper lacks an analysis of these discrepancies and benchmark-specific ablation studies.
  2. Insufficient component analysis. The ablation studies do not fully isolate the contributions of individual components. For example, verification is coupled with diagnosis-guided re-execution, making it difficult to distinguish their respective benefits.
  3. Limited reliability of evidence anchoring. The evidence chain relies on claims extracted from the executor's reasoning, which may omit or misinterpret critical information. The impact of such errors on verification reliability is not sufficiently investigated.
  4. Incomplete efficiency evaluation. Although token consumption and tool calls are reported, end-to-end latency comparisons are missing, leaving the actual computational overhead unclear.
Questions for authors

See above weaknesses.

Reviewer_hY92OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents an Evidence Anchoring–Verification (EAV) which solves the problem of evidence verification and error propagation in agentic search. The idea is to keep a compact record of claims, sources, and conflicts; independently check high-risk claims against fresh evidence; and retry rejected answers using the saved evidence. EAV-Flash applies this process to whole tasks, while EAV-MAS applies it across subtasks.

Critical strengths

The evidence anchoring and verification idea is interesting. This manuscript shows broad evaluation across various benchmarks and shows very good quantitative performance enhancements.

Critical weaknesses

W1. Motivation of the paper needs to be clearer. W2. No qualitative assessment of proposed methods W3. Reproducibility is limited by the lack of code and unclear implementation details. W4. There is a lack of theoretical analysis with respect to objectives that EAV claims to address.

Questions for authors

What are some of the failure cases for the current approach? What is time cost/time complexity of proposed anchoring and verification methods?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.