acceptodds
Withdrawn from ICLR 2027

EnactOpt: Diagnosis-Guided Harness Evolution for Embodied Agents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Embodied agents built on foundation models rely on an execution harness to organize observations, guide actions, and incorporate execution feedback into subsequent decisions. Failures can arise not only from limited model capabilities, but also from how the harness presents evidence and guides its use. These model-external mechanisms provide an adaptation target without updating model parameters, but task-specific interaction experience can be costly to collect. Task-level outcomes alone do not reveal whether relevant evidence was missing from the model input or was available but inadequately used. Moreover, these two aspects are coupled: effective action guidance may require evidence that the current context omits, while exposing that evidence alone does not ensure an appropriate response. We introduce EnactOpt, which uses limited execution feedback to jointly evolve reusable textual skills and context-construction code while keeping the underlying models and runtime fixed. It diagnoses adjacent decisions by relating action feedback to the actual model input and subsequent action. Structured episode summaries guide joint revisions through prompt-level edit operators, and candidate harnesses undergo compatibility checks and validation-based retention. Across VLABench, ALFWorld, and Embodied Agent Interface (EAI), EnactOpt improves task success by approximately 17–41 percentage points over the respective Naive baselines and outperforms the evaluated GEPA, TextGrad, and Trace2Skill baselines. These gains extend to task conditions not used for harness evolution or candidate selection. The results support a complementary route to embodied adaptation: turning limited execution experience into reusable behavioral improvements by jointly refining how evidence is presented and acted upon.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied agents, harness optimization, skill evolution, context construction, execution diagnosis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_MPYEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EnactOpt, a framework that adapts embodied agents by evolving their execution harness, consisting of reusable textual skills S and context-construction code C, while keeping the foundation models and runtime fixed. For each training rollout, a diagnostic LLM examines adjacent decisions, relating the preceding action feedback to the actual model input x_t and the current action, to judge whether the feedback was presented in the input and whether it was used appropriately; missing evidence suggests revising C, and inadequate responses suggest revising S. These diagnoses are condensed into structured episode summaries, which a single optimizer LLM uses to propose joint revisions of S and C under six prompt-level edit operators (add, revise, delete, move, split, merge), and compatible candidates are retained in a top-k pool according to validation success.

The method is evaluated on VLABench (with a frozen π0 low-level policy), ALFWorld, and EAI VirtualHome action sequencing, using GPT-5.4 mini for all LLM components, 30 training and 30 validation instances per benchmark, and three evolution iterations. EnactOpt improves test success over the initial harness by 10.43, 26.64, and 23.40 percentage points, respectively, and outperforms GEPA and skill-only reimplementations of TextGrad and Trace2Skill on all three benchmarks. Gains also appear on held-out VLABench tracks (T3/T4/T6) and ALFWorld unseen scenes, and VLABench ablations indicate that removing adjacent-decision diagnosis or structured edit operators hurts generalization, while joint revision outperforms separate skill and code updates with half the optimization passes.

Critical strengths
  • The paper separates failures in which evidence was missing from the actual model input x_t from failures in which available evidence was misused, which gives a principled rationale for jointly revising skills and context code. This decomposition is the conceptual core that distinguishes the work from outcome-driven prompt optimization and is broadly applicable to agent harness design.
  • Improvements hold across a low-level policy benchmark and two high-level action benchmarks, and the VLABench ablation shows that variants with similar in-distribution success can differ substantially under shift (e.g., outcome-only feedback gains 3.75 pp in-distribution but loses 1.91 pp on generalization). This finding, together with the full release of evolved artifacts in Appendix B, is useful for how harness optimization should be evaluated.
Critical weaknesses
  • The closest prior work, SHAPER (Wang et al., 2026b), which also evolves skills and context-construction code with diagnosis and preservation-aware edits, is not compared, and TextGrad and Trace2Skill are skill-only reimplementations with C fixed. Since several evolved changes require editing C (e.g., expanding the ALFWorld history window from 2 to 50 turns), the main-table gains partly reflect a larger editable scope rather than the diagnosis or joint-revision design.
  • VLABench gains may rely on privileged information and favorable task selection. The evolved context code exposes exact entity identifiers and the track label from observation.extra to the planner, and the evolved skill mainly instructs the planner to emit short canonical commands for the frozen π0, which is consistent with much larger gains under instruction shift (T3/T4, about +16 pp) than under visual shift (T6, +5 pp). Tasks were also selected based on the checkpoint’s reported T1 performance and T1-T4 gaps rather than randomly, so it is unclear how much of the reported generalization reflects harness evolution versus a setting that favors instruction normalization.
  • All results come from single evolution runs without variance or confidence intervals, and Figure 3 explicitly uses runs different from Table 1 for some entries, implying run-to-run variation that is not quantified. With only 30 validation instances (3.3 pp per instance) driving selection and 240 T1 test episodes, small differences such as the in-distribution results of EnactOpt and its ablations (63.75% vs. 63.33% vs. 62.08%) cannot be reliably distinguished from noise.
  • The paper does not evaluate whether adjacent-decision diagnoses correctly attribute failures to presentation (C) or usage (S), and the w/o-diagnosis ablation only replaces them with terminal outcomes, so it cannot separate the specific adjacent-decision design from simply providing step-level information. In addition, the structured edit operators are unenforced prompt conventions, yet removing them drops performance below H0, which makes it hard to attribute the gains to a well-defined mechanism rather than to prompt wording.
Questions for authors
  • What are the total number of environment rollouts and LLM calls (or tokens) used during evolution, including per-candidate validation and per-decision diagnosis, and how many candidates are generated per iteration?
  • In EAI, do Naive and all baselines use the same execution feedback loop and attempt budget as EnactOpt?
  • Can you report results over multiple evolution seeds, and explain why some Figure 3 entries use runs different from Table 1?
Reviewer_ygaJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies how to improve embodied agents by revising their execution harnesses without updating foundation-model parameters. EnactOpt optimizes textual skills and context-construction code: skills guide execution and recovery, while code selects and organizes observations and history. It diagnoses adjacent decisions by relating action feedback to the actual model input and subsequent action, then produces structured episode summaries for joint revision.

Critical strengths

1.The paper distinguishes constructed context from the complete model input, avoiding incorrect omission diagnoses when the runtime supplies additional observations. Treating evidence presentation and behavioral guidance as coupled components also fits practical agent systems.

2.Improvements appear across three benchmarks and are measured against the initial harness, not only Naive. VLABench separates in-distribution and generalization results, and the appendix provides the evolved skills and code, making the resulting changes concrete.

Critical weaknesses
  1. The main comparisons do not fully separate optimization method, editable scope, and feedback conditions. Section 5.1 implements TextGrad and Trace2Skill as skill-only methods, whereas EnactOpt can revise both skills and code; compute budgets are also unmatched. Table 1 therefore supports the complete system, but does not establish whether its advantage comes from diagnosis, joint search, or a larger edit space. Section 2.2 discusses SHAPER, which also combines skills, context code, and diagnosis, without a direct experimental comparison. The incremental contribution is consequently not fully located by the current results.
  2. The diagnostic ablation in Table 2 removes much of the intermediate execution evidence at the same time. The full method receives actual model inputs, adjacent actions, and feedback, while the variant retains only task descriptions and terminal outcomes. This establishes the value of detailed execution evidence, but does not isolate whether explicitly checking evidence presentation and use improves over ordinary trajectory summarization. The paper appropriately avoids treating diagnoses as causal proofs, yet the ablation also leaves the diagnostic structure itself untested.
  3. Thirty training instances and one training epoch do not establish low interaction cost. The paper explicitly states that these counts describe task coverage rather than total rollouts across harnesses, and each compatible candidate is evaluated on 30 validation instances. Candidate counts, total training and validation episodes, and diagnostic and editing calls are not reported. The 1× versus 2× optimization-pass accounting in Table 2 also does not directly translate into time, tokens, or environment interactions. These quantities matter for a method motivated by limited execution experience.
  4. Variability across independent evolution runs is not established. Selection repeatedly uses only 30 validation instances, while model-generated edits and environment execution can be stochastic. Tables 1–2 report no independent-search repetitions or uncertainty, and Figure 3 explicitly uses different runs for some entries. This does not invalidate the overall gains, but makes smaller baseline and component differences harder to assess, including the 3.73-point ALFWorld generalization advantage over GEPA. Test-episode counts do not replace replication of the optimization process.
Questions for authors
  1. How do GEPA and EnactOpt differ in edit permissions, feedback inputs, and candidate-evaluation budgets? Which distinctions from SHAPER are directly tested by existing experiments?
  2. Is there an existing ordinary-trajectory-summary control retaining the same execution evidence? Otherwise, should the diagnostic ablation primarily establish detailed feedback’s value over terminal outcomes?
  3. How many candidates, training rollouts, and validation rollouts are used per benchmark? What are the model-call or token costs of diagnosis, summarization, and editing?
  4. How many independent evolution runs underlie Table 1, Figure 3, and Table 2? Are repeated results or validation-score records available to clarify selection stability and component differences?
Reviewer_yyJ5OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents an agent harness optimization approach for LLM-based embodied agents. The harness in operation consists of skills (textual instructions) and context-construction code that composes environment states into textual descriptions for the backbone LLM. The proposed method, EnactOpt, jointly updates these two harness components and improves performance across three embodied AI benchmarks (VLABench, ALFWorld, and EAI) over harness refinement baselines.

Critical strengths
  • The writing is clear and easy to follow
  • The experiments are substantial with 3 common baselines and 3 benchmarks
Critical weaknesses
  • The paper describes yet another reflection and refinement heuristic without substantial insights into why the approach works or other significant conceptual contributions.
  • As a heuristic, the paper lacks discussions and comparisons that could highlight advantages of the proposed method, in terms of practicality and efficiency over the baselines. It could be the case that at matched compute expense, the baselines and the proposed method would perform similarly.
Questions for authors
  • Do you have any insights into why the method works, either theoretically or empirically?
  • Can you elaborate on the advantages of the method over existing approaches, e.g., in terms of ease of use, replicability, scalability, efficiency, etc.?
  • Is it possible to design experiments such that methods have matched compute expense? Would EnacOpt still outperform baselines then?

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

Related papers

Discussion (0)

Sign in to comment.