acceptodds
Under review as a conference paper at ICLR 2027

From System Performance to Decision Attribution in Agent Evaluation

Abstract

In agent evaluation, how much of an apparent statistical advantage comes from the decision policy, and how much from the human-designed system around it? An expert shortlist can remove difficult choices before the statistical policy makes a decision, while the LLM may have to propose actions on its own. End-to-end scores combine these contributions. We introduce Matched Policy Intervention (MPI), a framework for component-level decision attribution in agent evaluation. MPI holds the decision affordance (the external information, action space, budget, and execution rules) fixed and randomly preassigns which policy's recommendation is executed. Paired recommendations reveal whether similar outcomes arise from similar decisions. Across CausalGame, CARE, and a discrete-action adaptation of CausaLab, we find that some apparent end-to-end advantages disappear under matched affordances. On CARE's ChemLex task, a +5.72-point system gain accompanies a matched gate-minus-LLM effect of -0.09 points. In CausalGame, changing the affordance alone raises benchmark success from 25% to 62.5% under a fixed statistical policy. MPI distinguishes the contribution of a decision policy within an agent from the support supplied by its decision affordance.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, agent evaluation, decision attribution, statistical decision making, experimental design

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_KPFrOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Matched Policy Intervention (MPI), a module-replacement experiment under strictly controlled conditions, to address action-space privilege and better analyze the relative strengths and weaknesses of systems. The paper applies MPI to conduct experiments and analyses on CausalGame, CARE, and a discrete-action adaptation of CausaLab.

Critical strengths

None

Critical weaknesses
  1. The paper’s contribution rests on the assumption that prior scientific research involves over-attribution, as illustrated on the left side of Figure 1. However, the paper does not examine how existing studies conducted their comparisons or establish the prevalence of issues such as action-space privilege to justify its motivation.
  2. In my view, MPI is not a novel scientific method but is essentially the controlled-variable method. Furthermore, the paper provides relatively few experiments, and its analyses are limited to comparisons between LLMs and statistical methods.
Questions for authors

None

4 comments
AuthorsOpenReview ↗

Response to Reviewer KPFr

Thank you for asking us to examine the paper’s central insights more closely. Your concerns about motivation and novelty prompted us to revisit the evidence for the attribution problem and explain what MPI adds to established experimental methods. Your suggestion to broaden the evaluation also led us to test additional decision algorithms.

1. Evidence from existing agent studies

Many agent studies compare their methods with other complete systems. In these comparisons, the policy can change along with the search or planning procedure (SwiftSage; LATS; Boomerang), available information (DS-Agent; Causal-Copilot; Suspicion-Agent), or tool support (LLM-Assisted Light). The performance difference therefore cannot be attributed to the policy alone. MPI asks a more specific question: how does the final outcome change when we replace the policy at one decision stage, keeping external information, candidate-construction rules, budget, and execution rules fixed?

Many of the comparisons above do not control decision support separately from the policy. This can lead researchers to credit the policy for gains that come from better support. A major source of this ambiguity that we address is action-space privilege: one policy receives curated actions, while another must construct or search for them. Even with the same raw observations, the candidate menu can contain the results of expert judgment or statistical computation performed before selection. The complete-system comparisons in DECKARD and LLM-Assisted Light, for example, do not separate action-interface and decision-support differences, respectively, from the contribution of the policy. MPI makes candidate construction and supporting computations explicit, then compares policies under matched conditions.

The gap is in connecting a component-level explanation to an experiment that tests that component. MPI provides a formal framework for defining the boundary between policy and support and measuring a specified policy’s contribution while retaining shared support.

AuthorsOpenReview ↗

Response to Reviewer KPFr

2. Methodological contribution and additional experiments

We agree that controlled experimentation and factorial contrasts are established methods. The difficulty in applying them is that replacing an agent’s decision module can change more than its decision rule. Candidate generation, internal model state, tool outputs, and execution logic may be spread across the implementation, so the boundary between policy and support is not always clear. MPI addresses three practical difficulties.

First, a final score cannot separate policy quality from the quality of its support. A statistical routine may generate and score candidates while an LLM has to propose actions from raw observations. Alternatively, statistical computations may continue supplying recommendations after the LLM takes over. Repeating either comparison cannot separate factors that change together. MPI separates candidate construction and supporting computations from selection, then estimates policy effects by replacing the selector under shared support. In the support experiments, it changes menus while holding the policy fixed to measure the support effect.

Second, a shared dataset does not establish equal information at the decision point, and replacing a model does not establish a change in behavior. One routine may read fitted state and cached scores while another sees a textual summary. An LLM may also reproduce the statistical policy’s choices under the supplied support. Similar final scores cannot distinguish these cases. MPI checks whether the statistical rule can be reconstructed from the visible record and logs both recommendations at the same reached state. Reconstruction checks for information dependencies missing from the interface; policy disagreement measures whether the replacement changes behavior. Together, they help distinguish unequal information, identical choices, and different choices with similar outcomes.

Third, different recommendations need not produce different outcomes. LLM outputs may be repaired, retried, or replaced by a fallback, while another controller’s outputs are executed directly. Earlier choices also affect the evidence and opportunities available later. A complete-system score cannot locate the contribution of the original recommendation, the execution wrapper, or later stages. MPI fixes execution and failure rules, records recommendations separately from enactment, and preassigns execution authority across stages. The crossed design estimates each stage’s effect on the final outcome, including its downstream consequences. Confidence bounds then distinguish effects estimated to be small from effects that remain too uncertain to resolve, even when the policies behave differently.

None of these problems has to produce a software error. A reproducible benchmark gain can still answer the wrong attribution question. MPI makes the intervention auditable by specifying the boundary between policy and support, the matching requirements, and the stage assignments. This is the contribution of the protocol; the factorial estimator itself is established.

We added an exploratory acquisition experiment in a six-node CausaLab-style environment to test policies beyond the LLM and statistical pair. It compares a coverage rule, random selection, depth-two surrogate search, minimal tabular Q-learning, and fitted-Q policies using Ridge regression or a two-hidden-layer MLP. Under each support condition, all policies choose from the same four plan constructors and use the same final prediction module. The screened menu contains four distinct intervention-plan multisets. Training uses separate simulator seeds, and the MLP results average three frozen initializations. Evaluation uses 512 fresh test seeds under two support conditions, for 8,192 episodes across eight implementations.

Acquisition policyConcentrated menuScreened menuScreened − concentrated (adjusted 95% CI)
Coverage rule81.45%87.70%+6.25 pp [+1.18, +11.32]
Random plan77.54%78.32%+0.78 pp [−5.59, +7.15]
Depth-two surrogate search82.23%87.89%+5.66 pp [+0.62, +10.71]
Tabular Q-learning79.10%85.94%+6.84 pp [+1.09, +12.58]
Ridge fitted-Q81.64%87.30%+5.66 pp [+0.69, +10.64]
MLP fitted-Q, three-run mean79.49%85.09%+5.60 pp [+1.00, +10.20]

The intervals use paired seed-level contrasts with Bonferroni adjustment over 21 specified comparisons. Five policy families show positive support effects; the random-selection contrast is inconclusive. Action-space construction therefore affects rule-based, search-based, and learned policies as well as the LLM and statistical policies in our main experiments. We chose the latter as cases where decision procedures and supporting computations can be intertwined, but MPI applies beyond those two classes.

References

  1. Lin, B. Y., Fu, Y., Yang, K., Brahman, F., Huang, S., Bhagavatula, C., Ammanabrolu, P., Choi, Y., and Ren, X. 2023. SwiftSage: A Generative Agent with Fast and Slow Thinking for Complex Interactive Tasks. arXiv:2305.17390.

  2. Zhou, A., Yan, K., Shlapentokh-Rothman, M., Wang, H., and Wang, Y.-X. 2023. Language Agent Tree Search Unifies Reasoning, Acting, and Planning in Language Models. arXiv:2310.04406.

  3. Wang, M., Pang, A., Kan, Y., Pun, M.-O., Chen, C. S., and Huang, B. 2024. LLM-Assisted Light: Leveraging Large Language Model Capabilities for Human-Mimetic Traffic Signal Control in Complex Urban Environments. arXiv:2403.08337.

  4. Guo, S., Deng, C., Wen, Y., Chen, H., Chang, Y., and Wang, J. 2024. DS-Agent: Automated Data Science by Empowering Large Language Models with Case-Based Reasoning. arXiv:2402.17453.

  5. Wang, X., Zhou, K., Wu, W., Singh, H. S., Nan, F., Jin, S., Philip, A., Patnaik, S., Zhu, H., Singh, S., Prashant, P., Shen, Q., and Huang, B. 2025. Causal-Copilot: An Autonomous Causal Analysis Agent. arXiv:2504.13263.

  6. Nottingham, K., Ammanabrolu, P., Suhr, A., Choi, Y., Hajishirzi, H., Singh, S., and Fox, R. 2023. Do Embodied Agents Dream of Pixelated Sheep?: Embodied Decision Making using Language Guided World Modelling. arXiv:2301.12050.

  7. Gonzalez-Pumariega, G., Chen, W., Kedia, K., and Choudhury, S. 2024. Query-Efficient Planning with Language Models. arXiv:2412.06162.

  8. Guo, J., Yang, B., Yoo, P., Lin, B. Y., Iwasawa, Y., and Matsuo, Y. 2023. Suspicion-Agent: Playing Imperfect Information Games with Theory of Mind Aware GPT-4. arXiv:2309.17277.

AuthorsOpenReview ↗

Response to Reviewer KPFr

Dear Reviewer KPFr, thank you again for your time and consideration. As the discussion period is nearing its end, we wanted to follow up on our response and the additional experiments. We have not yet heard back and would appreciate knowing whether our clarifications address your concerns, or whether any questions remain that we could help resolve. We would be happy to discuss them further. Thank you!

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

The paper introduces Matched Policy Intervention (MPI), which compares LLM and statistical decision policies under shared information, action menus, and execution rules. A blocked factorial design estimates stage-specific effects, while paired recommendations measure behavioral disagreement. Experiments on CausalGame, CARE, and a CausaLab adaptation show that system-level performance gaps can differ from the effects of replacing individual decision policies.

Critical strengths
  1. The paper formulates a useful attribution question and handles sequential effects carefully. Equation 1 averages over the other stages' assignments, while retaining changes in histories and menus. The estimator is justified under the given assumptions and framework. I also appreciate the use of non-inferiority bounds, thus avoiding to interpret non-significance as equivalence.
  2. The CARE results strengthen the empirical contribution. On ChemLex, a +5.72 improvement accompanies a gate-replacement effect of -0.09, with a reasonably narrow interval. On Minerva, the confirmatory gate advantage is +14.35. This shows why component-level evaluation can change our interpretation of a system improvement without presupposing that statistical policies are unnecessary.
Critical weaknesses
  1. The methodological advance is modest. Equation 5 is a standard blocked factorial contrast, and its identification argument follows from comparability of the experimental cells. Related work already suggests controlled model-harness comparisons and factorial evaluation. MPI adds a useful stage-level formulation and paired recommendations, but its contribution comprises primarily the empirical findings.
  2. The empirical comparison between policy and shared support needs further investigation. The CausalGame reconstruction rule uses exploration and exploitation scores, and CausaLab provides the weighted-average prediction as an available action. Thus, computation can remain available after control passes to the LLM. These are useful comparisons of selection policies, but their implications depend on which computations remain shared. The paper should make this dependency clearer, and easier to assess. The reconstruction diagnostic checks action agreement. Is there a way to diagnose if any remaining information discrepancy has negligible outcome consequences?
Questions for authors
  1. Which computations produce the shared CausalGame scores and CausaLab predictions, and which remain active when control passes to the LLM?
  2. What motivates the common 5% loss tolerance across the different experiments? Several conclusions depend on this choice. Please explain its practical interpretation and how the conclusions change under smaller tolerances.
2 comments
AuthorsOpenReview ↗

Response to Reviewer DKZy

Your questions about shared computations and loss tolerance helped us clarify what the policy intervention changes and what the statistical results mean. We appreciate the care you took with the sequential design and the CARE findings. We address the two questions below.

1. Shared computations, information visibility, and the reconstruction diagnostic

MPI distinguishes the policies by their implemented decision procedures. An LLM may reason statistically, and the comparison does not require us to rule this out or determine how it reasons internally. We control the external evidence, available actions, and execution rules under which each policy recommends an action.

In CausalGame, the shared pipeline fits the hypothesis ensemble, constructs candidate menus, and computes predictive win probabilities and expected gains from additional observations. These computations continue under LLM control. The statistical policy applies its screening and exploration/exploitation rule; the LLM selects from the scored menu using the supplied record. The assignment determines which recommendation is executed.

In CausaLab, the task is to predict a held-out reactor frequency from observational and interventional evidence. Our discrete-action adaptation uses shared regression fitting and BIC-based weighting to produce five candidate predictions and their weighted average as a sixth option. We keep this prediction-construction pipeline under both policies so that the comparison isolates selection of the final prediction. Changing the pipeline as well would mix prediction construction with selection. The statistical submission policy selects the weighted average, while the LLM may select any of the six options. Both use the same index-to-prediction execution mapping. At acquisition, the statistical policy follows a deterministic coverage rule.

The effect we measure is the contribution of selection given this shared support. We will state more clearly which computations are shared, which decisions belong to each policy, and what information each interface provides.

We inspected the 12 reconstruction mismatches among 1,280 confirmatory acquisition decisions. They occurred in 10 of 128 episodes, all in the final two acquisition rounds. In 11 of the 12 cases, the two actions had the same displayed selection score: ten had equal win scores and one had equal exploration-gain scores.

You are right that action agreement alone cannot establish the downstream impact of these discrepancies. The logs describe where reconstruction differs, but do not contain counterfactual executions of the reconstructed choices. We will acknowledge this limitation in the main text and distinguish approximate action reconstruction from a guarantee of negligible outcome effects.

2. Choice of loss tolerance and sensitivity

Users should set the loss tolerance before confirmatory evaluation, based on the performance loss their application can accept. MPI does not prescribe a universal threshold. We use five percentage points as a common reference across the reported comparisons. The statistical error level also needs to be set in advance, as it determines the evidence required for a decision.

The CausaLab protocol specified a ten-point primary tolerance and a five-point sensitivity tolerance. The manuscript reports the stricter five-point results alongside the other sequential studies.

Our sensitivity analysis confirms that the tolerance changes the non-inferiority assessment: stricter tolerances leave more comparisons inconclusive. The five-percentage-point setting should therefore be interpreted in relation to the loss an application can accept.

We agree that this dependence needs to be explicit and will include the sensitivity results. The tolerance affects the replacement conclusions, while the estimated policy effects remain unchanged. It also leaves the attribution question open: a complete-system gain alone does not establish an individual policy’s contribution. MPI tests that contribution without assuming which policy should perform better.

Thank you for the helpful clarifications and revisions. The explanation of shared computations addresses my first question, and the tolerance sensitivity analysis is useful. The outcome impact of reconstruction discrepancies remains unverified, but I appreciate the acknowledgment of this limitation. My assessment of the modest methodological novelty remains unchanged, and I maintain my weak acceptance recommendation.

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

This paper introduces Matched Policy Intervention (MPI), a framework for separating the contribution of a decision policy from that of the system supporting it. By giving different policies the same information, choices, budget, and execution rules, the authors show that some apparent performance advantages disappear when the comparison is fair. The results also show that system design alone can substantially affect benchmark performance.

Critical strengths

The paper addresses an important practical issue in agent evaluation: end-to-end performance differences do not necessarily reflect the quality of the underlying decision policies. MPI provides a clear framework for separating policy effects from system design choices, while recording both policies’ recommendations offers useful behavioral insights. The evaluation across three tasks, including a controlled experiment, further demonstrates that affordance design can substantially affect benchmark performance even when the decision policy remains unchanged.

Critical weaknesses
  1. Evaluation covers three scientific or causal-reasoning settings, including a discrete-action adaptation of CausaLab. The extent to which MPI generalises to other domains, particularly agents that construct actions in open-ended settings, remains unclear.

  2. The policy effects are often small, with relatively wide confidence intervals in some experiments. Although the paper uses non-inferiority tests, the strength of the conclusions seems to depend on the chosen loss tolerance and the precision of the estimates.

  3. The matched-affordance setting may limit complete agent capabilities, while it isolates action selection from action construction. The practical implications for choosing between LLM-based and statistical systems could be discussed more explicitly.

Questions for authors

See Critical Weaknesses for details.

2 comments
AuthorsOpenReview ↗

Response to Reviewer HmaA

Thank you for the thoughtful review. Your comments on loss tolerance, estimation precision, and the scope of the conclusions have helped us make the paper more rigorous and explain what the evidence supports. We also appreciate your recognition of the attribution question and the value of paired recommendations.

1. Loss tolerance and estimation precision

We agree that loss tolerance and estimation precision affect the strength of the non-inferiority conclusions. MPI distinguishes the estimated policy effect from the criterion used to interpret it: δ specifies the acceptable performance loss, and α controls the statistical error level. Users can choose both for their application. The five-percentage-point tolerance is the reference setting for our evaluation, not a requirement of the framework.

We examined smaller tolerances using the existing confidence bounds. The 16 reported sequential stage-level comparisons give the following classifications:

Loss toleranceSupports LLM non-inferioritySupports a statistical advantage exceeding the toleranceInconclusive
1 percentage point1114
2 percentage points5110
3 percentage points709
4 percentage points1006
5 percentage points1402

The sensitivity analysis supports your concern: stricter tolerances leave more replacement assessments inconclusive. We will make this dependence explicit. The estimated effects remain the same, and a complete-system improvement still does not identify which component produced it.

For example, ChemLex has a +5.72-point complete-system gain and a gate-replacement effect of −0.09 points [−2.38, +2.20]. A lower tolerance changes whether the replacement meets the loss criterion; it does not attribute the complete-system gain to the deterministic gate. The confirmatory Minerva comparison, by contrast, identifies a substantial statistical-gate advantage.

Our claim concerns the source of a performance difference, without asserting a universal ranking of LLM and statistical policies. We will distinguish this attribution claim from the tolerance-dependent replacement conclusions and include the sensitivity results.

2. Structured decisions and generalization

MPI studies policy attribution at structured decision interfaces. It requires a common decision problem for the policies being compared; the application need not involve causal reasoning.

We chose scientific and sequential tasks because they require policies to select informative experiments under limited budgets. Early decisions affect later evidence and final outcomes, allowing us to study attribution beyond isolated, immediate-response decisions.

Structured selection is also used in other applications. TypeSafe’s Jev supports choosing among supplied options, such as routing a support request to a team. Claude’s strict tool use restricts calls to valid tools and schema-compliant arguments, including enumerated parameter choices. Both illustrate the practical use of designed decision interfaces, although schema-constrained tool use does not necessarily imply a fully finite action space. Jev documentation, Claude documentation

We agree that our evaluation does not cover unrestricted action generation. To study settings where action construction is itself an intervention, we would need to define that component and its comparison explicitly. We will clarify this scope and explain that structured selection extends beyond the domains tested here.

3. Attribution as guidance for system design

We want to help developers identify which parts of a system deserve further design effort. An improvement in the complete system can encourage investment in a decision rule even when candidate construction, information provision, or other support accounts for most of the gain. Separating these contributions helps direct that effort.

Matching the action interface isolates the contribution of selecting an action with the supplied support. Action construction is a separate contribution that can also be evaluated. The labels “delegate,” “retain,” and “inconclusive” summarize this component-level assessment; they do not recommend which complete system to deploy.

AuthorsOpenReview ↗

Response to Reviewer HmaA

Dear Reviewer HmaA, thank you again for your constructive feedback, which helped us improve the rigor and clarity of the paper. We wanted to ask whether our response and revisions have clarified our contribution and addressed your concerns about the loss tolerance, estimation precision, and scope of the evaluation. If any questions remain, we would be happy to discuss them further. Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.