{"id":"12c0b12c-80aa-4472-b2cd-717e12f8e0aa","arxiv_id":"2509.04535","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"ICPAD learns domain-agnostic prototype skills with a diffusion-based adapter and uses retrieved few-shot demonstrations to condition action generation in unseen target domains.","lead":"This paper introduces ICPAD, a framework that adapts robot policies to new environments without retraining, by learning reusable 'prototype skills' and conditioning a skill decoder on a few target-domain examples. A generalist reader might care because it is a concrete approach to few-shot cross-domain transfer, a key bottleneck for deploying learned robots.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cross-domain action consistency loss (Eq. 10) appears to supervise the adapter on generated actions paired with source-state sequences rather than states reached under target dynamics, and the encoders supplying that supervision are themselves trained by this loss, so the objective can be satisfied","rationale":"The central claim depends on cross-domain skill diffusion, and Eq. (10) is the main mechanism for action-level consistency. The paper does not specify how the state sequence needed to evaluate ΨE/ΦE on generated actions is obtained, and the obvious reading is that source states are reused. This is a train/test mismatch: the adapter is trained on generated-action/source-state pairs but deployed with target states. The circularity concern is also real because ΦE and ΨE receive gradients from the same loss that uses them as supervisors, so the objective could be minimized by representation collapse rather than genuine transfer. This does not disprove the empirical results; the ablations and qualitative embedding analysis provide some support. However, without code or the referenced appendix, the implementation of this central loss cannot be verified, so the conditional verdict remains appropriate. I do not recommend changing the reader's verdict, but this is the load-bearing technical issue to check first.","tokens_in":12909,"tokens_out":9370,"duration_ms":97909,"concrete_test":"Run an ablation where Eq. (10) is recomputed with the generated action sequence paired with the target state sequence s' from τ' (i.e., use ΦA(s'_t,d',z) and compare encodings to (s',a')), and with stop-gradients or removed gradients on ΦE and ΨE in this loss, while keeping all other hyperparameters and evaluation identical. If the Metaworld/CARLA target-domain results stay within the reported confidence intervals, the original loss was not relying on the problematic pairing/circularity; if they degrade or improve beyond those intervals, the published objective is not the one driving the claimed transfer.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (5) defines ΦE and ΨE as functions of full state-action sub-trajectories, while Eq. (6) defines ΦA: (s_t,d,z) ↦ a using only a single state. Eq. (10) then evaluates ΨE(·|s, ΦA(s_t,d',z)) and ΦE(·|s, ΦA(s_t,d',z)). Because no dynamics model or environment interaction is described in offline learning, the state sequence s paired with the generated action sequence must be taken from the source trajectory τ=(s,a). This is not the trajectory that would result from executing the generated actions under target dynamics. Under dynamics or embodiment shift, (s, generated-a) can be far out-of-distribution for ΨE/ΦE, so minimizing the KL terms may teach the adapter to make source-state/generated-action pairs look like the target domain, not to produce executable target-domain behavior. At test time the adapter is conditioned on target states, creating a train/test mismatch.\n\nAdditionally, although Eq. (10) is written as Lcross-A(ΦA), Algorithm 1 updates ΦE and ΨE from the sum of all losses; no stop-gradient is indicated on the encoders inside Eq. (10). The same encoders provide the supervision and receive gradients from this loss, so the objective is circular: it can be driven down by the encoders becoming insensitive to domain/skill differences (representation collapse) rather than by the adapter actually transferring. The ablation removing Lcross-A shows a performance drop, but does not establish the mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ICPAD, an offline skill-learning framework for in-context policy adaptation to unseen target domains without model updates. Offline, it learns domain-agnostic prototype skills with a skill encoder and a diffusion-based skill adapter, using two cross-domain consistency losses (a skill-prior KL and an action-embedding KL) plus a contrastive domain encoder; online, a dynamic prompting scheme retrieves few-shot target demonstrations and computes a domain embedding that conditions the adapter. The authors evaluate on Metaworld and CARLA with dynamics, embodiment, and horizon shifts, reporting consistent improvements over FIST, DCMRL, and SPiRL variants, alongside ablations, a data-availability study, and a language-prompt extension.","tokens_in":13379,"tokens_out":6739,"duration_ms":67805,"significance":"Should the method work as claimed, it would be a useful contribution: it is one of few frameworks attempting no-update, few-shot cross-domain policy adaptation by treating skills as a middle-tier abstraction, and it combines skill diffusion with consistency learning in a novel way. The experimental section is thorough for this scope: 5 seeds with 95% confidence intervals, multiple domain-disparity levels, three domain factors, ablation of both framework components, a data-availability study, and a language-prompt variant. The consistent margins over the strongest baseline DCMRL (11.6-21.6% in CARLA; roughly 14% average success rate in Metaworld) are non-trivial. However, the validity of the central cross-domain consistency objective hinges on details and corrections that are currently missing or ambiguous, so the result is not yet established at the level the claims require.","major_comments":[{"comment":"Equation (10) is syntactically malformed: the second DKL term reads DKL(ΦE(·|s, ΦA(st,d′,z), ΦE(·|s,a)) with no second argument; presumably a parenthesis is misplaced and the intended term is DKL(ΦE(·|s, ΦA(st,d′,z)), ΦE(·|s,a)). More substantively, the loss is labeled Lcross-A(ΦA), but Algorithm 1 updates ΦE and ΨE from the sum of all losses. If the encoders receive gradients from this term, the KL can be minimized by making the encoders insensitive to domain/skill differences (representation collapse) rather than by improving the adapter. The ablation in Table 6 shows a performance drop without Lcross-A, but it does not establish the mechanism. Please either apply stop-gradient to the encoder outputs used inside Eq. (10) or provide representation-quality diagnostics (e.g., retrieval accuracy, skill-space separation) that rule out collapse.","section":"Cross-domain consistency learning, Eq. (10)"},{"comment":"The loss aligns embeddings of the synthetic pair (s, ΦA(st,d′,z)) with those of a real trajectory (s′,a′), where s is a state sequence taken from the source trajectory. Since no dynamics model or environment interaction is described in offline learning, (s, ΦA(...)) is not the trajectory that would result from executing the generated actions under target dynamics. Under dynamics or embodiment shift, this synthetic pair can be far out of distribution for ΦE and ΨE, so minimizing the KL may teach the adapter to make source-state/generated-action pairs look like the target domain rather than to produce executable target-domain behavior. At test time the adapter is conditioned on target states, creating a train/test mismatch. Please justify this open-loop surrogate or provide an experiment that evaluates the generated actions under target dynamics (e.g., in a simple simulated domain) to show","section":"Cross-domain consistency learning, Eq. (10)"},{"comment":"The first experiment paragraph states: 'Details of the domain disparity settings are in Appendix.' The version under review contains no appendix. The main tables (Tables 1-3) are organized around 'Low/Medium/High' domain disparity, but the operational definitions (e.g., noise variances, wind magnitudes, embodiment changes, horizon changes) are not given. The learning hyperparameters H (skill horizon), m (number of prompts), β, µ, λ, δ, and the training schedules are also absent. Without these, the reported 5-seed comparisons are not reproducible and the trend across disparity levels cannot be checked. Please include the appendix or move the essential definitions into the main text, and specify all hyperparameters.","section":"Experiments: missing appendix and hyperparameters"},{"comment":"The retriever notation is inconsistent: Eq. (12) defines the retriever as ΨR, but Eq. (13) sums over 'ˆτi ∼ ΦR(T)', where ΦR was already defined as the skill prior in Eq. (8). This is presumably a typo for ΨR(T). More importantly, the retrieval scheme assumes that states in the few-shot target demonstrations are close to test-time states; Table 4 tests the extreme case of 8% task coverage, but no analysis of retrieval quality is reported. Please fix the notation and add a quantitative analysis of retrieval precision (e.g., how often retrieved prompts come from the same task as the current state) to support the 'dynamic prompting' claim.","section":"Dynamic Domain Prompting, Eqs. (12)-(13)"}],"minor_comments":[{"comment":"The expectation notation 'Ek∼[1,K],η∼N(0,I)' should be written more explicitly, e.g., 'Ek∼Uniform{1,...,K}, η∼N(0,I)', since '[1,K]' as a subscript is not a standard distribution.","section":"Preliminary, Eq. (2)"},{"comment":"Typo: 'CALRA' should be 'CARLA'.","section":"Experiments, first paragraph"},{"comment":"Eq. (6) writes ΦA: (st,d,z) ↦ a, but the surrounding text and Eq. (7) define the adapter as a diffusion model that also takes a noised action sequence x_k and a denoising step k. Please align the notation to avoid ambiguity.","section":"Problem Formulation / Eq. (6)"},{"comment":"The normalization 'w_i^{-1} / Σ_i w_i^{-1}' is correct in intent but the subscript ranges are implicit; write the sums explicitly for clarity.","section":"Dynamic Domain Prompting, Eq. (13)"},{"comment":"Figure 4 is described as qualitative; consider adding a quantitative metric, such as silhouette score or retrieval accuracy, to substantiate the claim that skill embeddings are domain-invariant.","section":"Qualitative analysis, Figure 4"},{"comment":"The paper does not discuss limitations. Given the central empirical generalization assumption (cross-domain consistency on source domains transfers to arbitrary target domains), a brief limitations paragraph would help readers calibrate the method's applicability.","section":"Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be an AAAI-format submission with an appendix that is absent from the provided text. Before recommending acceptance, I would want to see the appendix and the authors' response on the Eq. (10) stop-gradient issue. The 'first to achieve in-context adaptation of skill-based policies' claim in Related Works is not necessary for the contribution and could be toned down."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: ICPAD combines skill diffusion with cross-domain consistency and retrieval-based prompting, and reports consistent gains over DCMRL and others in Metaworld and CARLA. The adaptation protocol is genuinely no-gradient—at test time only the prompt changes, no parameter updates. That is the real contribution, and the ablations support the two main design choices.\n\nWhat's new: the specific combination is new, and the 'middle-tier' idea—domain-agnostic prototype skills plus a domain-grounded adapter—is a sensible way to think about zero-gradient cross-domain transfer. The experiments are broad: dynamics shifts, embodiment shifts, horizon shifts, limited data, even language-conditioned prompts. The gains over DCMRL are large and consistent, and the ablations show that removing the action consistency loss or the retrieval attention hurts.\n\nSoft spots, in order of concern. First, the cross-domain action consistency loss (Eq. 10) has a distribution-shift problem. It pairs generated actions with source states, not with states actually reached under target dynamics. The encoders ΨE and ΦE are then asked to recognize these synthetic (s, â) pairs as real trajectories. Under dynamics or embodiment shift, those pairs are out-of-distribution, so the loss may teach the adapter to fool the encoders rather than produce executable target behavior. The encoders are also trained with this same loss, so representation collapse is a plausible alternative explanation for the observed benefit of this loss term. The ablation shows a drop, but not the mechanism.\n\nSecond, the preprint is incomplete. The appendix is referenced but absent, so domain disparity settings, hyperparameters, and architecture details are unverifiable. Eq. (10) is malformed notationally and there are other typos. No code is released.\n\nThird, the 'first to achieve in-context adaptation of skill-based policies' claim is too strong. DCMRL already uses context inference for skill-based adaptation; the no-gradient protocol is different, but the 'first' phrasing needs a tighter comparison.\n\nThe central argument—that a middle-tier skill representation plus dynamic prompting gives no-gradient adaptation—is plausible and the evidence is suggestive, but the mechanism is not nailed down. The missing appendix and code matter for a claim this strong.\n\nThis paper is for people working on skill-based RL, offline RL, and few-shot policy transfer. It deserves a serious referee, but the referee should ask for missing details and a cleaner analysis of Eq. (10), ideally with a stop-gradient variant and a check for encoder collapse. Send it to peer review, but expect a request for major revision.","headline":"Solid empirical gains, but the mechanism in the cross-domain action consistency loss is under-specified and possibly training a shortcut; missing appendix and code make full verification impossible.","tokens_in":13783,"tokens_out":2813,"would_cite":true,"duration_ms":26803,"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":"A skill-based policy can adapt to a new environment without any gradient updates, using only a few target demonstrations.","keywords":["in-context policy adaptation","skill-based reinforcement learning","cross-domain generalization","diffusion models","few-shot adaptation","domain prompting","prototype skills","offline RL"],"falsifier":"Train ICPAD on source domains with mild dynamics variation, then evaluate on a target domain with inverted action sign or otherwise adversarial dynamics using five demonstrations; if success rate collapses to the no-adaptation baseline, the cross-domain consistency losses are not sufficient for unseen domains.","tokens_in":12829,"feed_emoji":"🤖","tokens_out":6005,"duration_ms":57538,"temperature":0.7,"pith_summary":"This paper argues that a policy trained on reusable skills can be redeployed in a new environment without updating any network weights, given only a handful of target demonstrations. The authors propose ICPAD, which learns a shared set of domain-agnostic prototype skills from offline data and a diffusion-based skill adapter that translates those skills into domain-specific actions. At test time, a retrieval-based dynamic prompt built from the few target demonstrations guides the adapter, so the same policy handles changes in dynamics, embodiment, and horizon. In robotic manipulation and autonomous driving benchmarks, this yields substantially higher returns and success rates than skill-based meta-RL baselines under the same few-shot constraints.","feed_headline":"Shared skills let a policy adapt to new domains without retraining","feed_subtitle":"In CARLA and Metaworld, five demonstrations are enough; gains over the closest baseline reach 21.6%.","key_machinery":"The framework centers on domain-agnostic prototype skills—a shared latent space of behavior primitives—and a domain-grounded skill adapter, a diffusion model that decodes a prototype skill into an action sequence conditioned on state, domain embedding, and prompt. Cross-domain consistency losses align the skill prior and the generated action sequences across source domains, while a contrastively trained domain encoder plus retrieval-based attention forms a dynamic prompt from a few target demonstrations at deployment time.","core_discovery":"The central claim is that in-context adaptation, not fine-tuning, is sufficient for skill-based policies to transfer across domains. During offline learning, ICPAD trains a domain-agnostic prototype skill space together with a domain-grounded diffusion skill adapter, enforcing cross-domain consistency in both the skill prior and the generated action sequences. During deployment, a dynamic domain prompt is computed from few-shot target demonstrations, and the frozen policy produces actions through the prompted adapter without model updates. The paper reports consistent gains over the strongest baseline, DCMRL, including 11.6% to 21.6% higher normalized returns in CARLA and an average success-","pith_inferences":["If the middle-tier skill layer is truly domain-agnostic, the same scheme could be applied to sim-to-real transfer on a physical robot: train in one simulator, provide a handful of real-robot demonstrations, and deploy without fine-tuning; the paper tests embodiment changes in CARLA but not the real world.","A quantitative measure of domain-agnosticism—for example, the KL divergence between prototype skill embeddings for matched tasks across held-out domains—would turn the qualitative embedding figure into a testable benchmark; the paper does not report such a metric.","Since language instructions can substitute for retrieved trajectory prompts in the paper's extension, combining both modalities into a single prompt may improve adaptation in tasks where demonstrations are scarce but text is available."],"forward_implications":["A single set of trained skill-based policies can be redeployed to a new domain by swapping only the prompt, not the network weights.","Domain-wise adaptation scales across tasks: one shared skill layer replaces per-task fine-tuning, which matters in multi-task, long-horizon settings.","Few target trajectories suffice—five per task—and even 8% task coverage loses only 6.8% success in multi-stage Metaworld, so deployment cost is low.","The prompt interface is not limited to trajectories: language instructions can drive adaptation with comparable performance.","Performance degrades gracefully as domain disparity grows, with a smaller drop than skill-based meta-RL baselines."],"supporting_citations":[{"why":"Supplies the skill-prior RL formulation that ICPAD builds on for policy learning and regularization.","marker":"Pertsch, Lee, and Lim 2021"},{"why":"Supplies the skill extraction approach using diffusion models that ICPAD adapts to learn prototype skills.","marker":"Nam et al. 2022"},{"why":"Supplies the denoising diffusion probabilistic model used as the skill adapter.","marker":"HO, Jain, and Abbeel 2020"},{"why":"Defines DCMRL, the strongest baseline whose normalized returns ICPAD improves by 11.6–21.6% in CARLA and success rate by about 14% in Metaworld.","marker":"He et al. 2024"},{"why":"Defines FIST, a skill-based few-shot imitation baseline that ICPAD compares against.","marker":"Hakhamaneshi et al. 2022"},{"why":"Provides the Metaworld benchmark used for the robotic manipulation experiments.","marker":"Yu et al. 2020"},{"why":"Provides the CARLA simulator used for the autonomous driving experiments.","marker":"Dosovitskiy et al. 2017"},{"why":"Provides the contrastive embedding approach used to train the domain encoder.","marker":"Schroff, Kalenichenko, and Philbin 2015"}],"fun_headline_variants":["Adapt to a new domain with five demos—no model updates","In-context skill diffusion: no retraining, just a few demos","Frozen policy adapts to new domains via cross-domain skill diffusion","Five demos in a new domain: policy adapts without any updates","Cross-domain skill diffusion lets a policy adapt in-context"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The result depends on the consistency losses learned from source domains transferring to any unseen target domain that supplies a few demonstrations reasonably close to the states the policy encounters.","fun_headline_variants_meta":{"raw":{"variants":["Adapt to a new domain with five demos—no model updates","In-context skill diffusion: no retraining, just a few demos","Frozen policy adapts to new domains via cross-domain skill diffusion","Five demos in a new domain: policy adapts without any updates","Cross-domain skill diffusion lets a policy adapt in-context"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000207,"raw_usage":{"total_tokens":1223,"prompt_tokens":715,"completion_tokens":508,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":459,"completion_tokens_details":{"reasoning_tokens":416}},"tokens_in":459,"tokens_out":508,"duration_ms":5202,"temperature":1.0,"reasoning_tokens":416,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:30:26.202570+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train ICPAD on source domains with mild dynamics variation, then evaluate on a target domain with inverted action sign or otherwise adversarial dynamics using five demonstrations; if success rate collapses to the no-adaptation baseline, the cross-domain consistency losses are not sufficient for unseen domains.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the skill-prior RL formulation that ICPAD builds on for policy learning and regularization."},{"cited_title":"J.; and Lim, J","cited_arxiv_id":null,"evidence_quote":"Supplies the skill extraction approach using diffusion models that ICPAD adapts to learn prototype skills."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines FIST, a skill-based few-shot imitation baseline that ICPAD compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CARLA simulator used for the autonomous driving experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the contrastive embedding approach used to train the domain encoder."}],"review_version":1}