acceptodds
Under review as a conference paper at ICLR 2027

Beyond Final Answers: Evidence-Grounded Evaluation of Agent Trajectories

Abstract

Evaluating an agent trajectory is difficult when correctness cannot be determined from a final string alone: relevant evidence may be distributed across the execution trace, task-time skills, generated files, and executable behavior. We study how such evidence should be exposed to an evaluator and whether a single evaluation pipeline can adapt to different diagnostic targets through a declarative output schema. Our pipeline combines a MetaAgent, a small set of task-specific micro-judges, deterministic checks, and a structured OutcomeJudge, and supports both optional evidence retrieval and direct materialization of evidence into the shared context. On the full 203-case SkillTV split, evidence-delivery policy substantially changes evaluation behavior: materializing task-time skills yields much higher pass recall than output-only context (70.21% vs. 30.43%), while jointly materializing skills and outputs gives 59.9% accuracy among the matched Gemini-2.5-Flash configurations. A vanilla Gemini-2.5-Flash judge given the same evidence reaches only 44.6% accuracy but a higher F1 score (42.0% vs. 30.8%), showing that evaluator architecture changes how the same evidence is converted into a decision. On a 52-run collection derived from agent executions on AFTER, our Gemini-2.5-Flash pipeline raises pass recall to 83.7% (90.1% F1), compared with 30.6% (46.9% F1) for our reproduction of JudgeSkill v4 with Claude Sonnet 4.6; with the same Sonnet 4.6 backbone, our pipeline reaches 53.1% (69.3% F1). Finally, we apply the same core evaluation architecture to Who&When and TRAIL using benchmark-specific output schemas. On the handcrafted Who&When split, it matches TraceElephant in agent attribution accuracy (56.9%) while improving step accuracy from 1.72% to 12.07%, although agent attribution on the algorithm-generated split remains below both baselines; on the TRAIL GAIA split, a larger micro-judge budget raises Category F1 from 15.40% to 27.75%, exceeding the native TRAIL judge's 24.16%, although fine-grained localization remains substantially harder. The results separate three factors that are often conflated in agent evaluation: which evidence is available, how it is exposed and aggregated, and how evaluation criteria are specified across tasks. Code and data are available at https://anonymous.4open.science/r/codegen-agent-eval-6744.

Primary area:
interpretability and explainable AI
Keywords:
LLM agents, agent evaluation, artifact verification, trajectory evaluation, skill-augmented agents

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_Ldk8OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors studies how the evidence (traces, task-time skills, generated files, and executable behavior) for evaluating the agent trajectories should be exposed to the evaluator. The proposed pipeline has a MetaAgent that routes each trajectories to one or two task-specific micro-judges, and a structured OutcomeJudge that aggregates their findings into a schema-defined output. The paper reported the wide range of experimental results transparently.

Critical strengths

None

Critical weaknesses
  • The performance of judge is too limited. The balanced accuracy in Table 3 shows nearly the chance (50%), and most of results are lower than JudgeSkill. Furthermore, the experimental results only provide single-run, without variance of CI.

  • As the short reference list suggests, the paper does not include appropriate citations where the text needs them. While reading, it is often unclear whether a particular term comes from prior work or was proposed by the authors themselves, because no citation is given. Furthermore, the related work section is missing much of the comparison against similar works.

Questions for authors
  • How does the proposed pipeline differ from prior agentic evaluator, AutoJudge [1], which also route trajectories to specialized judges and aggregate their findings?

[1] Zhidkovskaya, Alina Borisovna, et al. "AutoJudge: Automatic Generation of LLM-Based Judges from Execution Traces." Workshop on Failure Modes of Agentic AI at ICML 2026.

Reviewer_ckgQOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a modular framework for evidence-grounded agent trajectory evaluation. It combines specialized judges to assess execution traces, skills, and generated artifacts. Experiments show evidence presentation significantly affects accuracy, while cross-benchmark performance remains mixed.

Critical strengths
  • Evaluating complex agent tasks beyond final answers is important and timely.
  • The paper presents explicit research questions and a modular methodology, which makes it relatively easy to follow
  • The evaluation is broad: Multiple benchmarks and baselines help assess generalizability.
Critical weaknesses
  • The related work section would benefit from a broader coverage of the literature. As a minor remark, the entries are currently not clickable, which makes it harder to check the references.
  • Novelty seems limited. Specialized judges and evidence-aware evaluation have already been explored in related work.
  • Confidence intervals, significance tests, and run-to-run variability appear to be missing.
  • Several interacting components make a causal interpretation difficult.
Questions for authors
  • Why does the proposed method underperform JudgeSkill in pass F1 and balanced accuracy despite achieving higher raw accuracy?
  • How well do the results generalize, given the strong class imbalance and inconsistent performance across benchmarks?
Reviewer_LN2HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how to evaluate an Agent's task completion and failure behaviors by utilizing its execution trajectory, skills employed during task execution, generated files, and executable verification results. The authors propose a modular Evidence-Grounded Evaluation framework, which dynamically selects Micro-Judges via MetaAgent, aggregates diverse evidence through deterministic verifiers and OutcomeJudge, and adapts to various evaluation tasks using declarative Output Schema. The article focuses on the impact of evidence acquisition methods on evaluation results, including on-demand tool retrieval and direct incorporation of evidence into evaluation context. On the SkillTV dataset, the author found that Skill and Output Evidence have different error distribution characteristics: Skill Context is more conducive to improving the recall rate of successful samples, while Output Evidence is more conducive to identifying failed samples. Furthermore, the author extended the framework to AFTER, Who&Then, and TRAIL to validate their applicability in scenarios such as task outcome judgment, failure attribution, and error localization.

Critical strengths
  • The article not only considers the final output of the Agent but also incorporates execution trajectories, Skills, and Artifacts into the evaluation process. The SkillTV experiment demonstrates that different evidence types can significantly alter the error distribution of evaluators, providing valuable empirical insights for understanding Evidence-Grounded Agent Evaluation.
  • The authors reported metrics such as Accuracy, Balanced Accuracy, Precision, Recall, F1, and confusion matrices on SkillTV, rather than relying solely on final accuracy. For instance, Forced Skill achieved a Pass F1 of 40.24%, while Forced Skill + Output increased Accuracy to 59.90% but reduced Pass F1 to 30.77%. These results clarify the strengths and weaknesses of different evaluation strategies and prevent overly optimistic conclusions based solely on Accuracy.
  • The author applies the same core architecture to SkillTV, AFTER, Who&Then, and TRAIL, covering tasks such as binary classification judgment, failure attribution, and error localization. Especially in Who&Then and TRAIL, the middle Micro Judge Findings contain significantly more correct diagnostic information than the final output, which has reference value for subsequent research on evidence aggregation and trajectory diagnosis mechanisms.
Critical weaknesses
  • Although the complete framework can improve Raw Accuracy on SkillTV, its Pass F1 and Balanced Accuracy are still lower than Refined Judge Skill. For example, the accuracy of Forced Skill+Output is 59.90%, slightly higher than Judge Skill's 58.60%, but Pass F1 is only 30.77%, significantly lower than the latter's 44.70%. Under the same evidence input, Vanilla Gemini Judge also achieved a higher Pass F1 score (42.0% vs. 30.8%).
  • Although the method proposed in this article provides additional value in detecting failed samples, it is difficult to fully validate the robustness of the evaluator based solely on these three failed samples. The paper also lacks multiple runs, statistical confidence intervals, or significance analysis, resulting in uncertainty in the reliability of some small performance differences.
  • The article mainly compares the performance of different evidence strategies through numerical tables, but lacks specific agent trajectory cases to explain why Skill Evidence can improve Pass Recall, while Output Evidence is easier to identify failed samples. In addition, there is a lack of intuitive error cases to explain how information is lost during the aggregation stage regarding the significant gap between the intermediate diagnostic results and the final output of Who&Then and TRAIL.
Questions for authors
  • Can a fully matched control experiment be added to enable Optional Retrieval and Forced Materialization to use the same evidence content, Artifact Serialization, and visible range, only changing whether evidence needs to be actively obtained? In addition, if Forced Context is combined with Optional Tools, can a higher Pass Recall and True Negative Rate be retained simultaneously?
  • Suggest removing MetaAgent Routing, Micro Audit Specialization, Deterministic Validators, and OutcomeJudge Aggregation separately under the same Backbone, evidence input, and computational budget, and comparing them with single-stage LLM Judge. Especially, why can the current framework improve accuracy while reducing Pass F1? Is this phenomenon caused by the conservative judgment of the validator, the quality of Micro Judge Findings, or the final aggregation mechanism?
  • For AFTER derived Collection, it is recommended to supplement more failed samples and report Fail Class Recall, Balanced Accuracy, and confidence intervals, rather than just relying on Pass Class as the main indicator. Can we further analyze the performance of Who&Then and TRAIL under different error categories and trajectory lengths, and explain why Schema Driven Adaptation is effective in some tasks but significantly lags behind specially designed evaluators in others?
Reviewer_SQGiOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes an agentic evaluation pipeline consisting of a MetaAgent, specialized micro-judges, evidence-retrieval tools, and an OutcomeJudge. The authors study how evidence delivery affects evaluation performance and test the framework on SkillTV, AFTER, Who&When, and TRAIL for outcome verification and failure diagnosis.

Critical strengths

The experiments on evidence delivery are interesting, particularly the comparison between optional evidence retrieval and directly providing skills and artifacts to the judges.

Critical weaknesses
  1. Unclear novelty: The proposed method appears to be an agentic evaluation harness with specialized sub-agents and tools. It is unclear what is fundamentally new compared to existing Agent-as-a-Judge approaches (Zhuge et al., ICML 2025). The paper is more of a new harness for evaluation

  2. Insufficient baselines: Although the authors compare against JudgeSkill and several LLM-based judges, they do not compare against general-purpose agent harnesses such as OpenCode or OpenHands configured for evaluation. These frameworks already support tool use, sub-agents, and execution environments. A comparison using the same models, tools, and compute budgets would better establish the advantage of the proposed architecture. The paper should also consider broader evaluation settings such as AJ-Bench (Shi et al., ACL 2026).

  3. Limited evidence of generalization: Most experiments use Gemini-2.5-Flash. The authors should evaluate more models across different harnesses. I strongly encourage the authors to make the results much broader by including multi turn benchmarks.

References

Questions for authors
  1. What is the main architectural novelty compared to existing Agent-as-a-Judge frameworks?

  2. Can the authors compare against general-purpose agent harnesses such as OpenCode or OpenHands, configured as evaluators with the same tools and model backbones?

  3. How well does the proposed architecture generalize across different LLM backbones, and what is the benefit of schema-driven transfer beyond changing the evaluation prompt and output format?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.