{"id":"d753077d-24ea-4a62-b73a-01a0b8b580d5","arxiv_id":"2608.11350","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"SHAPER evolves a frozen embodied agent's skill text and context-code harness through rollout feedback, improving VLABench and ESI-Bench performance without any parameter updates.","lead":"This paper introduces SHAPER, a method that improves a frozen embodied agent by automatically rewriting the textual skill instructions and the code that builds its context, without changing the model's weights. The approach improves success rates on two robot benchmarks compared with the unfrozen seed agent, suggesting a cheap alternative to fine-tuning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No ablation isolates the contribution of rollout-derived feedback: without a no-feedback control, the gains attributed to 'self-evolving through target-environment rollouts' could stem from the optimizer model's priors and executor description alone.","rationale":"The paper's contribution is not merely that a frozen agent can be improved by editing prompts, but that the improvement is driven by rollouts from the target environment. Without a no-feedback control, the observed gains could be produced by the optimizer model's prior knowledge about effective prompting, especially because the VLABench skill-optimizer prompt explicitly includes the VLA actor's command distribution. This is a more fundamental concern than baseline fairness: even if every baseline were perfectly controlled, the causal role of the rollout-derived textual gradient would remain unidentified. The reader's weakest assumption concerned whether the self-generated summaries are a sufficient and unbiased optimization signal; my concern is closely related but distinct: whether that signal is necessary at all. I do not think this changes the overall verdict of CONDITIONAL, because the missing ablation is an addressable empirical requirement rather than a demonstrated flaw. However, it should be an explicit condition for acceptance: without the no-feedback control, the phrase 'self-evolving through target-environment rollouts' is not established. Secondary issues such as the unfair SFT baseline, missing sampling budgets for TTS baselines, and absent error bars reinforce the conditional verdict but are not the primary load-bearing point.","tokens_in":13678,"tokens_out":7865,"duration_ms":68767,"concrete_test":"Run the identical two-stage evolution on VLABench and ESI-Bench with the optimizer prompt modified to omit Γ(c): replace episode summaries and aggregate execution statistics with a constant placeholder such as 'No feedback available', keeping the seed skill, harness contract, executor command description, beam search, and validation selection unchanged. If the final held-out success is within sampling error of the reported SHAPER numbers, the gains are not attributable to rollout-guided self-evolution; if it is substantially lower, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SHAPER improves a frozen agent specifically through target-environment rollouts, i.e., that the textual gradient Γ(c) in Eqs. (3)–(5) is the driver of artifact improvement. The paper never ablates this signal. The skill optimizer receives the current skill, read-only harness, episode summaries, and execution statistics (Fig. A8); in VLABench it also receives the VLA actor's command distribution, which is a priori interface knowledge, not rollout-derived. The same 27B model could plausibly propose a better skill from prior knowledge of good prompting and the executor description alone, without any episode summaries. If so, the gain over Seed Agent reflects offline prompt engineering by a strong model, not self-evolution from rollouts, and the central claim becomes unsupported. The case studies show evolved artifacts that address observed failures, but these are selected examples; they do not quantify how much of the gain depends on Γ(c). A no-feedback control is therefore load-bearing: without it, the mechanism that distinguishes SHAPER from ordinary prompt optimization is unverified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SHAPER, a train-free framework for adapting frozen embodied agents by evolving a textual skill and a context-code harness through target-environment rollouts. The agent is factored into a frozen VLM planner, a frozen executor, a skill, and a harness. SHAPER uses the same frozen model in two roles: as the planner during interaction, and as an optimizer that revises artifacts based on hierarchical textual diagnoses of rollouts. The method is evaluated on VLABench and ESI-Bench, where it reports higher success than the seed agent, same-data SFT, and test-time scaling baselines.","tokens_in":13921,"tokens_out":5926,"duration_ms":48696,"significance":"If the results are robust, SHAPER offers a practical, low-cost alternative to parameter updates for adapting embodied agents. The paper's controlled case studies, prompt templates, and cost accounting are strengths that support reproducibility. However, the central mechanism--rollout-derived textual feedback as the driver of improvement--is not isolated experimentally, and the lack of error bars makes the quantitative claims fragile. The work is a useful contribution to train-free agent adaptation, but the evidence currently falls short of establishing that the gains come from self-evolution rather than offline prompt optimization.","major_comments":[{"comment":"The paper's central claim is that skill-and-harness optimization is driven by target-environment rollouts, yet no experiment ablates the rollout-derived feedback Γ(c). The skill optimizer receives episode summaries, execution statistics, and (on VLABench) a description of the VLA actor's command distribution; a no-feedback control in which the optimizer sees only the current artifact and executor description would determine whether the gains arise from the model's priors rather than from self-evolution. This control is load-bearing because it separates SHAPER from ordinary offline prompt optimization.","section":"3.2-3.3, Eq. (3)-(5), Fig. A8"},{"comment":"The main quantitative claims are made without repeated runs, error bars, or significance tests. On VLABench, the overall gain of SHAPER over the seed is 6.25 points, but the difference between skill-only and full SHAPER is 1.00 point, and several split-level differences are smaller (C1 42.5 vs 40.0). On ESI-Bench, category-level differences such as Specular Reflection 20.0 to 60.0 rest on 20 questions and Action Sequencing on 5. Without confidence intervals or multiple evolution runs, the reported improvements cannot be distinguished from noise.","section":"Tables 2 and 3"},{"comment":"The same-data SFT baseline fine-tunes only the low-level VLA actor, whereas SHAPER optimizes the upper-level planner's textual skill and harness. This is not a like-for-like comparison for the claim that SHAPER is a competitive alternative to fine-tuning, because the planner remains unadapted in the SFT arm. Please add an SFT baseline that also adapts the planner (e.g., LoRA or full fine-tuning) or explicitly scope the claim to executor-level SFT.","section":"4.1 Compared methods, Table 2"},{"comment":"The method's hyperparameters (beam width K=3, branch factor 2, evolution rounds 4, minibatch size 4, 15 train / 24 validation episodes on VLABench, 10/10 on ESI-Bench) are fixed without sensitivity analysis. Since the paper proposes a generic self-evolving framework, it would be informative to report how performance varies with the rollout count and validation set size on at least one benchmark.","section":"4.1 Optimization protocol"}],"minor_comments":[{"comment":"The definition of Stats(c) is omitted; please specify what aggregate statistics are included in the textual gradient.","section":"3.2, Eq. (3)"},{"comment":"The notation TopK uses K, but the beam width is also called K in Section 4.1; clarify that they are the same quantity or rename one of them.","section":"3.3, Eq. (6)"},{"comment":"The claim that 'the improvements from artifact evolution are not reproduced by additional inference-time sampling' should be softened, because the test-time scaling baselines are only evaluated on VLABench and with fixed configurations; no error bars support the comparison.","section":"4.2"},{"comment":"The comparison with GPT-5 PS is external, as the paper notes, but the abstract and conclusion might state more explicitly that the two numbers are not directly comparable due to different evaluation sets.","section":"4.3"},{"comment":"The skill-optimizer prompt includes a description of the VLA actor's command distribution; this is a form of interface knowledge that should be disclosed in the main text, as it may already encode part of the improvement.","section":"Appendix A.3.1"},{"comment":"The term 'textual gradient' is used as a central concept but never formally defined; consider adding a definition box or a one-sentence formalization.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is plausible and the case studies are compelling, but the missing no-feedback control is the main obstacle to accepting the central claim. The editor may also consider whether the journal expects statistical rigor (error bars, repeated runs) for empirical claims of this strength."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the paper. The core idea is real: no one else, to my knowledge, co-evolves a textual skill and a context-code harness around a frozen planner/executor in embodied settings with fixed action interfaces. The two-stage schedule, validation-gated top-K selection, and hierarchical diagnosis are sensible engineering. The held-out gains are consistent: 34.5 vs 28.25 seed on VLABench, 49.8 vs 32.5 on ESI-Bench, and the gains appear across distribution shifts. That part is solid as far as it goes.\n\nThe soft spot is the mechanism. The paper claims self-evolution through target-environment rollouts, but it never ablates the rollout-derived feedback. The optimizer receives the current skill, harness description, episode summaries, and execution statistics; in VLABench it also gets the VLA actor's command distribution. A strong model could propose a better skill from that prior knowledge alone, without any episode summaries. Without a no-feedback control, you cannot distinguish 'the agent learns from its failures' from 'the optimizer does offline prompt engineering.' The case studies are illustrative, but they do not quantify the contribution of the textual gradient. This is load-bearing for the central claim, and it is fixable with one control condition.\n\nOther issues are secondary but worth noting. The SFT baseline fine-tunes only the VLA actor, not the full planner, so it is not a matched comparison. The TTS baselines do not report sampling budgets. There are no error bars or repeated seeds. No code or data release. The GPT-5 PS comparison is unpaired, but they label it as external, so that is minor. The same model acting as judge, summarizer, and optimizer is a real circularity worry, though the validation-gated selection gives some protection.\n\nWho is this for? Anyone working on adapting frozen foundation models for embodied or agentic tasks. It deserves a serious referee, but I would not accept it as-is. The missing control is a genuine gap, not a style complaint. I would push for major revision requiring the no-feedback control, a matched SFT baseline, and error bars or repeated runs. If you work in this area, cite it for the method and the combination, but wait for the control before relying on the self-evolution interpretation.","headline":"Useful method paper with consistent held-out gains, but the missing no-feedback control undercuts the specific self-evolution claim.","tokens_in":14459,"tokens_out":2672,"would_cite":true,"duration_ms":26211,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SHAPER shows that a frozen embodied agent can be adapted by evolving the textual skill and context-code harness around it, without any weight updates.","keywords":["embodied agents","train-free adaptation","skill evolution","harness optimization","frozen foundation models","textual gradient","VLABench","ESI-Bench"],"falsifier":"Take the same evolution protocol and re-run it on a second, disjoint batch of training episodes of equal size; if the selected skill and harness do not reproduce the reported held-out gains (or if artifacts evolved on one batch fail on another), the textual gradient is overfitting to particular episodes rather than capturing generalizable failures. A second check is to apply the evolved artifacts to a different executor of comparable ability and see whether the gains persist when the planner's command interface changes.","tokens_in":13499,"feed_emoji":"🤖","tokens_out":8446,"duration_ms":90636,"temperature":0.7,"pith_summary":"SHAPER claims that an embodied agent can improve itself without touching its model weights: the same frozen vision-language model that plans actions can, in a separate optimizer role, revise two pieces of external structure—a reusable textual skill and a context-building code harness—based on a handful of rollouts in the target environment. The paper argues this is a practical middle path between pure execution, which stagnates, and fine-tuning, which needs new data, rewards, and training runs. On VLABench, evolving both artifacts raises success from 28.25% for the seed agent to 34.50%, above same-data SFT at 24.00%; on ESI-Bench, it raises micro-accuracy from 32.5% to 49.8%. If correct, this means adaptation can happen entirely through model-external artifacts when weights are unavailable, expensive, or locked.","feed_headline":"Frozen agent evolves its own skills and beats fine-tuning","feed_subtitle":"Skill-harness evolution lifts frozen embodied agents on two benchmarks with no weight updates","key_machinery":"The load-bearing mechanism is the artifact pair $(s,h)$: a persistent textual skill $s$ that tells the frozen planner how to inspect, decompose, act, and recover, and a context-code harness $h$, a Python context builder that selects and formats trajectory history before each planner call. Around this pair, SHAPER builds a closed loop: rollouts are judged round-by-round on before/after observations, condensed into episode summaries plus aggregate statistics to form a textual gradient $\\Gamma(c)$, and fed to the same frozen model prompted as an optimizer $O$, which proposes replacement skills and harnesses. The two-stage schedule (skill first with the harness fixed, then harness with the skill fixed), sandboxed validation, and top-K beam search turn this into a discrete optimization over artifacts while the planner and executor weights stay frozen.","core_discovery":"The central discovery is that the bottleneck for a frozen embodied agent is not only the model but the procedural guidance and context it is given, and that this non-parametric part can be optimized by the model itself. SHAPER factorizes an agent into a frozen planner, a frozen executor, a reusable textual skill $s$, and a context-code harness $h$, then solves an optimization problem for $(s,h)$ over expected rollout reward using target-environment episodes. The same frozen vision-language model is prompted as an artifact optimizer that consumes a 'textual gradient'—round-level critiques and episode summaries—and proposes candidate skills and harnesses, which are sandboxed, validated on a held-out set, and kept in a top-K beam. The paper reports that this two-stage skill-then-harness evolution improves performance over the seed agent, over training on the same data, and over test-time sampling or voting baselines in two embodied environments with different action interfaces.","pith_inferences":["Beyond the paper: if the self-generated textual gradient is unbiased, the loop should scale with planner capability: a stronger frozen model should produce sharper diagnoses and larger gains, which is directly testable by swapping the frozen planner while keeping the rest of the pipeline fixed.","Beyond the paper: evolved skills and harnesses are portable artifacts, so one testable extension is transferring an artifact evolved on one benchmark or executor to another environment without re-optimization; the paper leaves cross-embodiment transfer as future work.","Beyond the paper: the same artifact-optimization loop could be applied to non-embodied agents whose behavior is also shaped by textual procedures and context construction, though the paper only evaluates embodied settings."],"forward_implications":["Skill-and-harness evolution can improve a frozen planner both when the executor is a vision-language-action model and when the agent acts through a fixed action API, so the route does not depend on programmable robot interfaces.","Gains persist under distribution shift: on VLABench the largest improvements over the seed appear when the target category or task form is unseen (+6.0, +10.0, and +6.5 points on splits C2, C3, and C4).","The evolved artifacts are acquired once and reused across all held-out episodes, so the adaptation cost (about $2.25 on VLABench and $2.83 on ESI-Bench here) does not grow with the number of deployments.","Sampling or voting at inference time does not replicate the gains, which suggests artifact evolution changes something test-time scaling does not."],"supporting_citations":[{"why":"Supplies the VLABench benchmark and its split design that the VLABench experiments evaluate.","marker":"Zhang et al. [2024]"},{"why":"Supplies the ESI-Bench benchmark and the 231-question subset used for evaluation.","marker":"Hong et al. [2026]"},{"why":"Defines the verifier-free test-time selection baseline that SHAPER compares against.","marker":"Jang et al. [2025]"},{"why":"Defines the trajectory voting baseline that SHAPER compares against.","marker":"Lin et al. [2025]"},{"why":"Establishes the prior skill-only artifact-optimization approach that SHAPER extends by adding harness evolution.","marker":"Yang et al. [2026]"},{"why":"Establishes prior embodied skill-evolution work that SHAPER contrasts with for not co-evolving the harness.","marker":"Ju et al. [2026]"},{"why":"Establishes prior code-harness synthesis that SHAPER contrasts with for not co-optimizing reusable skills.","marker":"Lou et al. [2026]"},{"why":"Provides the frozen VLA executor used in VLABench rollouts and evaluation.","marker":"Black et al. [2024]"}],"fun_headline_variants":["Frozen agents forge their own skills, no training needed","Train-free self-evolution: frozen model upgrades its own harness","No weights changed: agent evolves skills to beat fine-tuning","Skill-harness evolution lets frozen agents self-improve","Self-evolving agents: only skills and harness change, not parameters"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the same frozen model that makes planning mistakes can, from a few rollouts, write accurate critiques of those mistakes and propose skill and harness changes that generalize beyond the handful of training episodes.","fun_headline_variants_meta":{"raw":{"variants":["Frozen agents forge their own skills, no training needed","Train-free self-evolution: frozen model upgrades its own harness","No weights changed: agent evolves skills to beat fine-tuning","Skill-harness evolution lets frozen agents self-improve","Self-evolving agents: only skills and harness change, not parameters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000278,"raw_usage":{"total_tokens":1650,"prompt_tokens":938,"completion_tokens":712,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":629}},"tokens_in":554,"tokens_out":712,"duration_ms":12915,"temperature":1.0,"reasoning_tokens":629,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:12:16.718313+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same evolution protocol and re-run it on a second, disjoint batch of training episodes of equal size; if the selected skill and harness do not reproduce the reported held-out gains (or if artifacts evolved on one batch fail on another), the textual gradient is overfitting to particular episodes rather than capturing generalizable failures. A second check is to apply the evolved artifacts to a different executor of comparable ability and see whether the gains persist when the planner's command interface changes.","supporting_citations":[],"review_version":1}