{"id":"826fbfc2-c2a9-4e84-8972-f0e777e3615e","arxiv_id":"2608.03234","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"CMP learns to reweight reference motions by task context, improving task return and sample efficiency for AMP- and SMP-based humanoid control in simulation.","lead":"This paper adds a learned relevance filter to motion priors so that a simulated humanoid uses reference motions suited to the current goal instead of the whole motion library. The method improves task performance and sample efficiency in five simulation tasks, relevant to humanoid control systems trained on large and uneven motion datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The relevance signal is trained on total GAE that includes the very style reward CMP adapts; if that reward dominates, the 'context-aware' gains may be a self-referential artifact rather than task-driven improvement.","rationale":"The reader's conditional verdict is reasonable: the method is coherent and the ablations isolate the reweighting mechanism, but the signal used to learn relevance is the soft spot. My concern is not that the authors are dishonest; the limitation is acknowledged in Section 5. The issue is that the total advantage includes the very reward being adapted, so the positive labels are not an independent measurement of task relevance. The proposed experiment separates task-driven from style-driven advantage and would settle whether the mechanism is genuinely context-aware or a self-reinforcing reward artifact. I do not call for rejection because the existing qualitative alignment (Figure 5) and consistent gains provide some evidence; I recommend the verdict remain conditional, with the additional condition that this confound be tested. This agrees in spirit with the reader's weakest assumption but sharpens it to a specific reward-mixing mechanism.","tokens_in":19634,"tokens_out":13053,"duration_ms":168940,"concrete_test":"On Target Location (and, if feasible, the other four tasks), compute two versions of CMP-AMP: (1) the reported total-advantage version; (2) a task-advantage-only version in which GAE is computed using only r_task (drop the style-reward term), while keeping all hyperparameters identical. Compare final returns, samples-to-threshold, and the learned relevance weights. Also rank-correlate learned R(c,x) against an oracle task-relevance score (e.g., for Target Location, the negative distance between the reference clip's velocity/direction and the target vector). If task-only advantages yield comparable or better performance and high oracle correlation, the total-advantage contamination is not load-bearing; if the gains disappear or correlation is low, the headline claim must be qualified as dependent on the style-reward loop.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Section 4.2, Table 1) requires that the learned relevance function R_phi(c,x) identifies reference motions that are useful for the task context. The only task-specific supervision for R_phi is the positive-pair construction A_i > 0 in Section 3.2 (L_on). But A_i is the GAE of the total reward, which in the AMP instantiation is 0.5*r_task + 0.5*r_AMP (Table 7), with r_AMP(c,x) = -log(1 - sigma(l_{theta,psi})) itself generated by the adapter-augmented discriminator that CMP is training (Section 3.3). No reward-scale calibration or ablation separates task-driven advantage from style-driven advantage. If the style reward is large or has higher variance, high-advantage rollouts are selected because they please the current (adapting) prior, not because they solve the task. The relevance model then reweights references to reinforce those style preferences, and the adapter reshapes the reward, closing a loop that can persist without task improvement. The Uniform Adapter and Shuffled Relevance ablations (Section 4.4) test the weight-motion correspondence but not this reward contamination. Section 5 acknowledges 'sensitive to critic error and insufficient exploration,' but the specific confound of mixing the adapted style reward into the label is neither analyzed nor quantified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Context-Aware Motion Priors (CMP), a modular method that adapts a task-agnostic motion prior to the current task context. CMP learns a contrastive relevance function R_phi(c,x) from high-advantage policy rollouts (positive pairs) and a demonstration-positive objective that anchors context embeddings to the reference distribution. The learned relevance scores reweight the reference motion distribution and are used to train a lightweight context-conditioned adapter on top of the base prior, without modifying the base prior's own training objective. CMP is instantiated with Adversarial Motion Priors (AMP) and Score-Matching Motion Priors (SMP). Experiments on five simulated humanoid tasks report improved task return and sample efficiency for CMP over the base priors, interpretable relevance weights and motion retrieval, ablations separating adapter capacity from learned relevance, and robustness to reference-distribution imbalance. The paper also includes results on a second humanoid morphology (G1).","tokens_in":19936,"tokens_out":7232,"duration_ms":84928,"significance":"If the empirical gains hold, CMP is a worthwhile contribution: it addresses a recognized limitation of motion priors—the use of a fixed, context-independent reference distribution—without requiring skill labels, dataset partitioning, or a separate skill-discovery stage. The paper's strengths include evaluation across two distinct prior formulations, five tasks, a second humanoid morphology, and ablations that separate adapter capacity from the learned relevance signal. The reweighting equivalence in Appendix A is a useful formal framing. The main weaknesses are that validation is entirely simulated, only three seeds are used, and the advantage-based relevance labels are drawn from a total reward that includes the very style prior CMP adapts, creating a potential self-referential loop that is not quantified. Overall the idea is novel and the evidence is suggestive, but the core mechanism needs stronger validation before the central claim is fully supported.","major_comments":[{"comment":"The positive-pair construction in the online branch L_on uses the standardized GAE Â_i of the total reward. In the AMP instantiation, the total reward is 0.5*r_task + 0.5*r_AMP (Table 7), and r_AMP is computed from the adapter-augmented discriminator l_{θ,ψ} that CMP itself is training (§3.3). Thus high-advantage rollouts can be selected because they satisfy the current, adapting style prior rather than because they solve the task. The relevance model then reweights references to reinforce those style preferences, closing a self-referential loop. Section 5 acknowledges sensitivity to critic error and insufficient exploration, but the specific reward-contamination confound is neither analyzed nor ablated. A concrete test would be to train the relevance model using advantages computed only from the task reward (e.g., a critic for r_task) and compare final return and sample efficiency again","section":"§3.2 and §3.3"},{"comment":"The claim of consistent gains with SMP is not fully supported by the reported statistics. For Steering, CMP-SMP return is 512±8 vs SMP 496±10; for Dribbling, 466±2 vs 458±5. These differences are within one standard deviation with three seeds. The text states CMP 'delivers consistent gains' across all five tasks and both motion-prior formulations, which overstates the evidence for these two tasks. Please report per-seed values, effect sizes, confidence intervals, or additional seeds, or qualify the claim for the SMP cases where gains are small relative to variance.","section":"§4.2, Table 1"}],"minor_comments":[{"comment":"The notation in the online batch B_on is ambiguous: x_i is described as the 'paired motion' but it is not stated clearly whether these are policy-generated motion clips or reference motions. Given that the contrastive candidates include both, please clarify explicitly.","section":"§3.2"},{"comment":"The minibatch weight formula uses B exp(a R)/sum_j exp(a R) followed by clipping; the text's statement that 'normalization gives the weights unit mean' is true only before clipping. Please state that clipping changes the mean and mention the implications for the reweighting equivalence in Appendix A.","section":"§3.1"},{"comment":"Many entries in the relevance-weight heatmaps are at the clip boundaries (0.50 and 2.00). The paper does not discuss the extent to which the learned weights are saturated. Adding a version without clipping or stating explicitly how saturation affects interpretability would help.","section":"§4.3, Figure 5"},{"comment":"The shuffle ablation is described as 'randomly permuting' weights, but the exact permutation procedure (within batch, per context, shared across references?) is not specified. Please define it precisely for reproducibility.","section":"§4.4, Tables 4-5"},{"comment":"The samples-to-threshold metric is defined with respect to 80% of the base prior's final return. Because thresholds differ between base and CMP, a CMP method that improves early but has a lower final return could artificially appear better. Please discuss this potential bias, even if not present in the reported results.","section":"Appendix J"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: CMP is a real, modest advance, not a revolution. The idea—learn context–motion compatibility from high-advantage rollouts and use it to reweight reference supervision for a lightweight adapter—is clearly new relative to ASE/CALM/C·ASE and to adaptive-sampling work like SkillMimic-v2. The ablations are the paper's strongest asset: Uniform Adapter and Shuffled Relevance isolate the learned correspondence from extra capacity, and the imbalance experiment shows the mechanism is doing something task-directed. The results are consistent across five tasks and two prior formulations, plus a G1 morphology appendix. The prose is honest and the limitation section names real constraints.\n\nWhere I'd push back: the stress-test concern about the advantage signal is not a manufactured flaw. The positives are selected on total GAE, and in the AMP instantiation half of the reward is the adapter-influenced style reward. So the relevance model can be selecting motions because the current adapted prior likes them, not because they solve the task. The paper acknowledges critic error and exploration sensitivity but never separates task advantage from style advantage. That should be an experiment: use task-reward-only advantage for L_on, or report the reward scales and variances. This is a load-bearing soft spot for the theoretical framing, but not enough to sink the empirical claim—the qualitative relevance maps show context-appropriate retrieval (FF→run, FN→walk), which is hard to explain if the signal is purely style-driven. Still, I would not call the central mechanism fully diagnosed.\n\nMinor but real: three seeds, no sensitivity over the many hand-set hyperparameters, no code or data artifact, and the \"exact clip manifest included in the code\" is an unkept promise. The SMP gains are small enough that I would want the CMP-SMP ablations in the main text rather than supplementary.\n\nVerdict: worth refereeing. The method is clear, the evaluation is honest, and the flaw is located and fixable. I would ask for code, a task-advantage-only ablation, and variance reporting on key hyperparameters before acceptance. For a reading group, this is a good paper to discuss—the self-referential bootstrap is genuinely interesting.","headline":"A genuine, modest advance in making motion priors context-aware, with good ablations and honest limitations; the main soft spot is that the advantage-based relevance signal partly feeds on the adapter's own style reward, and the paper would benefit from a cleaner separation and code release.","tokens_in":20469,"tokens_out":2270,"would_cite":true,"duration_ms":27218,"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 motion prior that reweights references by task context improves humanoid control on five tasks.","keywords":["motion priors","humanoid control","reinforcement learning","context-aware adaptation","contrastive relevance learning","reference reweighting","adversarial motion priors","score-matching motion priors"],"falsifier":"On the Target Location task, deliberately bias the critic (for example, initialize the value function so that walking-rich states are overvalued) and compare CMP-AMP's learned weights against a well-calibrated critic. If the weights still shift toward running for far targets and the return gain persists, the advantage-positivity assumption is not decisive; if the weights track the bias instead, the critic error is the load-bearing channel. A second check: after training, hold the policy fixed and see whether reweighting references by the learned $R_\\phi$ alone improves task return when the bas","tokens_in":19476,"feed_emoji":"🤖","tokens_out":10546,"duration_ms":103336,"temperature":0.7,"pith_summary":"This paper tries to establish that a general motion prior can be adapted online to the current task context without skill labels, dataset partitioning, or a separate skill-discovery stage. Existing priors treat every reference motion as equally relevant, so heterogeneous or imbalanced motion data can mislead the policy. The proposed framework, Context-Aware Motion Priors (CMP), learns context–motion compatibility from high-advantage policy rollouts and uses the resulting weights to reweight reference supervision for a lightweight adapter. Across five humanoid tasks and two motion-prior formulations (adversarial and score-matching), CMP reports consistent gains in task return and sample efficiency. If true, this turns motion-prior learning from a static, task-agnostic step into a continuously context-conditioned mechanism.","feed_headline":"Improve humanoid control by reweighting motion references per task","feed_subtitle":"CMP learns which reference motions suit each goal, cutting training samples and lifting returns for both AMP and SMP.","key_machinery":"The load-bearing object is the contrastive relevance function $R_\\phi(c,x)$, the cosine similarity between a learned context encoder and a learned motion encoder. It is trained with two contrastive branches: an online branch that treats motion clips from rollouts with standardized advantage $>0$ as positives for the current context, weighted by the advantage magnitude, and a demonstration-positive branch that treats reference motions as positives to keep context embeddings near the reference support. The relevance scores become normalized importance weights $w_\\phi(c,x)$, which reweight the reference distribution as $q_\\phi(x\\mid c) = w_\\phi(c,x)\\,p_E(x)$. A context-conditioned residual adap","core_discovery":"The central claim is that task relevance and motion plausibility are distinct, and that the reference distribution of a motion prior should therefore be conditioned on the task context. CMP learns a relevance function $R_\\phi(c,x)$ between a task context (goal, command, upcoming trajectory, or object state) and a reference motion clip, trained with two contrastive branches: high-advantage Generalized Advantage Estimation rollouts as positive context–motion pairs, and a demonstration-positive branch that anchors the context embedding to the reference support. These scores define an effective context-conditioned reference distribution $q_\\phi(x\\mid c) = w_\\phi(c,x)\\,p_E(x)$, with weights norma","pith_inferences":["Editorial inference: the advantage-positivity signal is a generic 'relevance from experience' mechanism; the same weighting could steer which demonstrations to replay or emphasize in imitation learning and offline RL, where hand-designed priorities are common.","Editorial inference: the paper's stated sensitivity to critic error and insufficient exploration suggests a direct stress test—corrupt the value function or shrink the exploration noise and measure whether CMP's gains shrink or reverse; this would quantify how much of the benefit comes from the advantage signal versus the demonstration anchor.","Editorial inference: the smaller SMP gains suggest adapter capacity limits how much a frozen prior can be reshaped; increasing the residual scale or adapter size for SMP may narrow the gap with AMP, at the risk of overriding the pretrained denoiser.","Editorial inference: the reweighting view is a change of measure on the reference distribution, so the same mathematics could schedule a curriculum—using advantage-weighted reference probabilities to sample easier motions early and harder motions as the policy improves."],"forward_implications":["On all five tasks (Target Location, Steering, Trajectory Following, Dodgeball, Dribbling), CMP reports higher test return than the corresponding base prior, and reaches 80% of the base final return with fewer environment interactions on four of five AMP tasks and on all five SMP tasks.","The learned relevance weights evolve from near-uniform early in training to context-dependent late in training—long walking for near targets, fast running for far targets, directional runs for side and back targets—showing interpretable context–motion alignment.","Removing the demonstration-positive branch degrades retrieval quality and slows learning; the branch anchors context queries to the reference support.","When walking motions are replicated 100x in the reference set, AMP's return drops 11.5% while CMP-AMP drops at most 2.8%, indicating reduced sensitivity to reference-frequency imbalance.","Because CMP is modular and leaves the base prior objective unchanged, the same context-conditioned reweighting can be layered onto other motion-prior formulations beyond AMP and SMP."],"supporting_citations":[{"why":"Supplies the base Adversarial Motion Prior and its reference-data objective, the primary setting CMP-AMP improves.","marker":"Peng et al. 2021"},{"why":"Supplies the pretrained score-matching denoiser, evaluation timesteps, and reward construction that CMP-SMP extends.","marker":"Mu et al. 2025"},{"why":"Supplies the MimicKit framework and the shared locomotion reference dataset used across all five tasks.","marker":"Peng 2025"},{"why":"Supplies the contrastive predictive-coding formulation that CMP repurposes to learn context–motion compatibility.","marker":"van den Oord, Li, and Vinyals 2018"},{"why":"Supplies the contrastive goal-conditioned value interpretation used to justify treating high-advantage rollout motions as positives for their context.","marker":"Eysenbach et al. 2022"}],"fun_headline_variants":["Context-aware motion priors improve humanoid control","Reweight motion references per task for better humanoids","Task-adapted priors cut training sample needs for humanoids","CMP: Learn which motion clips matter for each task","Context-conditioned priors boost humanoid learning efficiency"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The method assumes that a positive advantage signal from the current policy reliably marks the motion seen in that rollout as the right reference motion for that task context; if the value estimate is biased or the policy never tries the needed motion, the learned relevance will reinforce what the policy already does rather than discovering the right motions.","fun_headline_variants_meta":{"raw":{"variants":["Context-aware motion priors improve humanoid control","Reweight motion references per task for better humanoids","Task-adapted priors cut training sample needs for humanoids","CMP: Learn which motion clips matter for each task","Context-conditioned priors boost humanoid learning efficiency"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1284,"prompt_tokens":737,"completion_tokens":547,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":479}},"tokens_in":481,"tokens_out":547,"duration_ms":6619,"temperature":1.0,"reasoning_tokens":479,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:55:23.914874+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the Target Location task, deliberately bias the critic (for example, initialize the value function so that walking-rich states are overvalued) and compare CMP-AMP's learned weights against a well-calibrated critic. If the weights still shift toward running for far targets and the return gain persists, the advantage-positivity assumption is not decisive; if the weights track the bias instead, the critic error is the load-bearing channel. A second check: after training, hold the policy fixed and see whether reweighting references by the learned $R_\\phi$ alone improves task return when the bas","supporting_citations":[{"cited_title":"2021 , publisher=","cited_arxiv_id":null,"evidence_quote":"Supplies the base Adversarial Motion Prior and its reference-data objective, the primary setting CMP-AMP improves."},{"cited_title":"2025 , publisher=","cited_arxiv_id":null,"evidence_quote":"Supplies the MimicKit framework and the shared locomotion reference dataset used across all five tasks."}],"review_version":1}