{"id":"4eb7d5ea-a532-4e64-b94f-3d0ca960be5b","arxiv_id":"2509.01657","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Importance Weighted Retrieval scores prior robot data by the ratio of Gaussian kernel density estimates of the target and prior distributions, improving few-shot imitation learning.","lead":"This paper proposes a new rule for choosing which prior robot demonstrations to recycle when learning a new task: score each candidate by an estimated density ratio of target to prior data, instead of nearest-neighbor distance. The method, called IWR, reports consistent success-rate gains over prior retrieval baselines in simulation and on real Bridge robot tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (4)'s h→0 limit shrinks to d_min=0, so the claimed KDE equivalence for retrieval is not established.","rationale":"The reader's CONDITIONAL verdict is reasonable, but their identified weakest assumption—that Gaussian KDE ratios in latent space are faithful indicators of usefulness—captures an empirical limitation rather than the most load-bearing flaw in the paper's argument. The more fundamental issue is internal: Eq. (4), which is the sole formal bridge between nearest-neighbor retrieval and KDE, does not have the claimed limit. The nearest-neighbor term dominates as -d_min^2/h^4, so the threshold shrinks to a point unless ζ is rescaled with h; the fixed-radius rule of Eq. (2) is not recovered. This does not automatically invalidate the empirical claim that IWR improves retrieval in VAE spaces, but it undermines the paper's central theoretical motivation. It also connects to the reader's concern: the KDE density ratio is only a meaningful selection signal if the latent space is smooth and low-dimensional, but the derivation that makes this connection principled is currently incorrect. I therefore agree with the conditional verdict, but for a different primary reason. Fixing Eq. (4) or recharacterizing the equivalence as a ranking equivalence (top-K nearest neighbors vs. top-K KDE density) would resolve the issue without changing the experimental conclusions.","tokens_in":16279,"tokens_out":12304,"duration_ms":139724,"concrete_test":"Analytically re-derive the h→0 limit of Eq. (4) for a fixed query point and two candidate distances, e.g., d1^2=0.1 and d2^2=0.2, evaluating the LHS at h=0.1, 0.01, and 0.001. Check whether the accepted set converges to {d_min^2 < ζ} for fixed ζ, or instead to d_min=0. Independently, inspect the released code to see whether IWR actually uses Eq. (4) or directly thresholds the KDE density of Eq. (5); if the latter, the flawed equivalence equation should be corrected or removed, and the paper's theoretical claim should be reworded to describe the ranking equivalence, not the threshold equivalence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's theoretical contribution rests on the claim in §4.1 that the standard nearest-neighbor retrieval rule (Eq. 2) is recovered as the zero-bandwidth limit of a Gaussian KDE. The formal bridge is Eq. (4), which thresholds (1/h^2) log Σ_{z'} exp(-||z-z'||^2/h^2) > -ζ. As h→0, the log-sum-exp is dominated by the nearest neighbor: (1/h^2) log exp(-d_min^2/h^2) ≈ -d_min^2/h^4. The inequality therefore becomes d_min^2 < ζ h^4, which converges to d_min = 0, not the fixed-radius ball d_min^2 < ζ of Eq. (2). A correct log-sum-exp approximation would require h^2 log Σ exp(-d^2/h^2), and even then the equivalent KDE threshold would involve the log-density's dimension-dependent additive terms. As written, the derivation does not demonstrate the claimed equivalence; it only shows that a differently-scaled expression has a nearest-neighbor limit. The empirical IWR method may still work, but the paper's motivating theoretical result—that prior retrieval is the zero-bandwidth limit of KDE—is unsupported in the submitted formulation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Importance Weighted Retrieval (IWR), a modification of the scoring rule used in retrieval-based few-shot imitation learning. Instead of selecting prior data by minimum L2 distance to target demonstrations in a learned latent space, IWR scores prior points by a Gaussian KDE estimate of the density ratio p_t / p_prior, computed with a bandwidth from Scott's rule. The authors claim that the standard nearest-neighbor rule is the zero-bandwidth limit of a KDE density estimate, and that replacing it with smoothed KDE ratios addresses high variance and bias. They evaluate IWR on Robomimic Square, five LIBERO tasks, and three real Bridge V2 tasks, using Behavior Retrieval, Flow Retrieval, and SAILOR embeddings, and report improved success rates over the base retrieval methods. The paper includes ablations on the denominator p_prior, bandwidth, and retrieval threshold, as well as an analysis of retrieved data composition.","tokens_in":16483,"tokens_out":4509,"duration_ms":53133,"significance":"If the empirical findings hold, IWR is a simple, broadly compatible drop-in replacement for the nearest-neighbor scoring used in several existing retrieval methods, and the paper's release of code and real-robot evaluations strengthens this contribution. The paper is also candid about limitations, notably that IWR does not help in the non-smooth BYOL latent space. However, the theoretical motivation is a central part of the paper's framing, and the claimed equivalence in Section 4.1 is not mathematically correct as written. The empirical results also do not uniformly support the abstract's claim that IWR 'consistently improves' all retrieval methods. The corrected version of the paper should be valuable, but the current manuscript overstates both the theoretical derivation and the universality of the empirical gains.","major_comments":[{"comment":"The claimed equivalence between the nearest-neighbor retrieval rule (Eq. 2) and the zero-bandwidth limit of the KDE is not established. In Eq. (4), the expression is h^{-2} log Σ exp(-||z-z'||^2/h^2). As h→0, log-sum-exp is dominated by the nearest neighbor, giving -d_min^2/h^2, and after multiplying by 1/h^2 this becomes -d_min^2/h^4. Thresholding by -ζ therefore forces d_min^2 < ζ h^4, which converges to d_min = 0, not the fixed-radius ball d_min^2 < ζ of Eq. (2). To recover Eq. (2), one would need the soft-max scaled by h^2, i.e., h^2 log Σ exp(-d^2/h^2), or an explicitly h-dependent KDE log-density threshold. As written, Eq. (4) is a different scaling and its limit is degenerate. The authors should either correct the scaling and present Eq. (4) as a soft relaxation rather than an equivalence, or remove the claim of recovering the original retrieval rule.","section":"Section 4.1, Eq. (4)"},{"comment":"The paper justifies IWR via the importance-sampling identity E_{p_prior}[p_t/p_prior log π] = E_{p_t}[log π], but the algorithm does not perform importance sampling or resampling; it thresholds the estimated density ratio and trains on retrieved data with uniform weighting, as stated in Section 4.3 ('IWR is still a biased estimate'). Thus the theoretical benefit of the importance-weight thresholding is not proven, and the claim that IWR 'mitigates the bias of previous selection rules' is an empirical heuristic. The authors should temper the theoretical language: the contribution is a smoothed density-ratio-based scoring rule, not an unbiased estimator of the target expectation.","section":"Sections 4.2-4.3, Eq. (6)"},{"comment":"The abstract states that IWR 'consistently improves performance of existing retrieval-based methods.' This is contradicted by Table 2. For Flow Retrieval, FR-IWR is worse than FR on Square (67±1.9 vs 79±5.0) and Soup-Sauce (42±3.3 vs 45±5.5). For SAILOR, SR-IWR is worse on Soup-Sauce (48±3.4 vs 51±2.2) and on the real Corn task (9/20 vs 12/20). The average improvement may be positive, and the authors later use 'generally' in Section 5.3, but the word 'consistently' in the abstract overstates the results. Please revise the claim to reflect the actual distribution of outcomes across tasks and embeddings.","section":"Abstract and Table 2"}],"minor_comments":[{"comment":"The sentence 'recovering the original retrieval rule from Eq. (3)' should refer to Eq. (2), not Eq. (3).","section":"Section 4.1"},{"comment":"The Gaussian exponent in Eq. (4) is exp(-||·||^2/h^2), while Eq. (5) has the standard 1/2 factor and covariance h^2 Σ. The relationship between the two bandwidth/scale conventions should be stated explicitly to avoid confusion.","section":"Eqs. (4) and (5)"},{"comment":"The appendix note acknowledges that the main-paper Figure 4 contains labeling errors ('Others' marked as Misc instead of Harmful, and similar ending configurations misclassified). This should be corrected in the main text before publication; a note saying 'We plan to update the main paper when possible' is not acceptable in a submitted version.","section":"Appendix D, note after Fig. 10"},{"comment":"The threshold η in Section 4.3 is introduced without a formal definition, whereas ζ is used in Eq. (2). The authors should define η consistently with the importance-weight score and explain how it relates to the retrieval percentage used in experiments.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the incorrect asymptotic derivation in Section 4.1, which is used to motivate the entire method. I believe this is fixable by rewriting the derivation as a soft approximation rather than an exact limit, and by softening the 'consistently improves' claim in the abstract. The empirical core, with code release and real-robot results, is a useful contribution. Also, the appendix note about correcting Figure 4 should be resolved before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Xie et al. propose IWR, which replaces the nearest-neighbor L2 retrieval rule in few-shot imitation learning with a density-ratio score estimated by Gaussian KDEs. The new bit is the idea of looking at p_target/p_prior instead of just distance, plus the KDE-based smoothing. That's a genuine, useful twist on existing retrieval methods, and the empirical work backs it up: three sim tasks, five LIBERO tasks, three real Bridge tasks, and ablations on the denominator, bandwidth, and threshold. The paper is honest about the BYOL failure and the biased estimator.\n\nThe catch: the theoretical motivation in §4.1 does not actually go through. Eq. (4) uses (1/h^2) log Σ exp(-||z-z'||²/h²). As h→0, the log-sum-exp is dominated by the nearest neighbor, giving -d_min²/h⁴. So the threshold becomes d_min² < ζ h⁴, which collapses to d_min=0, not the fixed-radius ball of Eq. (2). The standard soft-max approximation would be h log Σ exp(-d²/h), which does recover the max. As written, the claim that prior retrieval is the zero-bandwidth KDE limit is not established. That's a real flaw in the framing, though not in the method itself: IWR uses a finite Scott's-rule bandwidth, so the empirical results are unaffected.\n\nOther soft spots are minor. The retrieval threshold η is tuned per experiment, and 'consistently' in the abstract is a bit stronger than the tables show—some differences are within error bars or ties. The ablation in Table 3 shows the normalization helps but not dramatically. The paper's own limitations section covers the main caveat about smooth latent spaces.\n\nOverall, the core empirical claim is plausible and well-supported. The theory needs correcting or reframing. A serious referee should catch this. I'd send it out, but ask for a corrected derivation or a substantially toned-down claim about the limit. If that's fixed, it's a solid contribution to retrieval-based IL.","headline":"The method is simple and works, but the paper's headline theoretical claim—that nearest-neighbor retrieval is the zero-bandwidth limit of KDE—is wrong as written; with that fixed, it's a useful but modest empirical upgrade.","tokens_in":17031,"tokens_out":4464,"would_cite":true,"duration_ms":46563,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that scoring prior robot demonstrations by the ratio of Gaussian kernel density estimates of the target and prior distributions, instead of by minimum L2 distance, consistently improves few-shot imitation learning in simula","keywords":["few-shot imitation learning","retrieval","data selection","importance weighting","kernel density estimation","robotic manipulation"],"falsifier":"Rank the same prior samples with IWR and L2 on a fixed task; train policies on the top-IWR set and, as a control, on the bottom-IWR set. If the bottom set performs as well or better on held-out task success, the density-ratio ranking is not capturing usefulness in that latent space.","tokens_in":16111,"feed_emoji":"🤖","tokens_out":7854,"duration_ms":80231,"temperature":0.7,"pith_summary":"This paper argues that the standard nearest-neighbor retrieval rule used in few-shot imitation learning—selecting prior demonstrations by their minimum L2 distance to target demonstrations—is mathematically a crude density estimate with two avoidable flaws. It introduces Importance Weighted Retrieval, which scores prior samples by the ratio of Gaussian kernel density estimates of the target and prior distributions. The authors show that this ratio-based scoring is a direct, plug-in upgrade to several existing retrieval methods, improving success rates in simulation and on real Bridge tasks. This matters because data selection is a bottleneck for learning new robot tasks from only a handful of demonstrations, and the fix is a small change to how prior data is scored.","feed_headline":"Swap L2 for importance weights to boost few-shot imitation","feed_subtitle":"Scoring prior demos by target-vs-prior density ratios lifts success in simulation and real Bridge tasks.","key_machinery":"The central object is the importance weight p_KDE_t / p_KDE_prior, the ratio of two Gaussian kernel density estimates over latent embeddings of state-action pairs. The paper shows the classic min-L2 rule is the h->0 limit of a KDE of the target density alone, so replacing it with a smoothed density ratio corrects both the high variance of nearest-neighbor estimates and the bias from ignoring the prior distribution. The KDE bandwidth follows a multiplicative factor of Scott's rule, and the covariance is the sample covariance of the embeddings, which lets the estimate smooth over all data and account for feature correlations.","core_discovery":"The central claim is that retrieval should be viewed probabilistically: the goal is to approximate an expectation under the target task distribution using samples from a large prior dataset, which is exactly the importance-sampling identity E_prior[(p_t/p_prior) log pi] = E_t[log pi]. Existing retrieval rules only approximate the numerator p_t, by a nearest-neighbor kernel that emerges as the bandwidth of a Gaussian KDE goes to zero; they ignore the prior's own distribution and use high-variance estimates. IWR instead estimates both densities with Gaussian KDEs using Scott's rule and the sample covariance, scores prior points by p_KDE_t / p_KDE_prior, and thresholds the top-scoring points fo","pith_inferences":["Going beyond the paper's experiments, IWR's density-ratio viewpoint suggests replacing Gaussian KDE with a scalable density-ratio estimator, which could remove the low-latent-dimension restriction the authors flag as a limitation.","The same probabilistic reading implies other retrieval metrics, like dynamic-time-warping distance, may also be limit cases of a smoothed density and could be upgraded by the same ratio correction.","A practical takeaway not stated by the paper: when choosing a latent space for retrieval, one should test whether small perturbations in latent space move retrieved samples in semantically meaningful ways, because KDE-based weights inherit the space's smoothness."],"forward_implications":["IWR can be dropped into any retrieval method that produces a smooth latent space, turning L2 retrieval into density-ratio retrieval without retraining the representation.","Retrieval quality shifts from matching isolated target points toward selecting samples that sit in dense, target-relevant regions of the prior—visible as more relevant tasks and more balanced coverage across demonstration phases.","Even when the target task has a missing critical subtask in the prior, retrieval gains are capped; IWR cannot invent absent behaviors.","Threshold choice still matters: retrieving too much of the prior reintroduces harmful data and hurts performance."],"supporting_citations":[{"why":"Defines Behavior Retrieval, the VAE-latent baseline whose L2 retrieval rule IWR replaces and whose representation is used for the main IWR results.","marker":"[5]"},{"why":"Defines Flow Retrieval, a flow-based retrieval baseline whose embeddings and L2 rule IWR wraps in experiments.","marker":"[6]"},{"why":"Defines SAILOR, a skill-based retrieval method that serves as another baseline and representation for IWR.","marker":"[7]"},{"why":"Supplies the Bridge V2 real-world prior dataset used to evaluate retrieval on unseen sink tasks.","marker":"[8]"},{"why":"Supplies Scott's rule, the bandwidth formula used to set Gaussian KDE smoothing in IWR.","marker":"[26]"},{"why":"Supplies Diffusion Policy, the policy-learning backbone used to train and evaluate all retrieval methods.","marker":"[30]"}],"fun_headline_variants":["Boost few-shot imitation with importance-weighted retrieval","Why L2 retrieval fails and density ratios fix it","Retrieve demos by density ratio, not distance, for better imitation","Importance weights outperform L2 for few-shot imitation","Swap nearest-neighbor scores for KDE density ratios in retrieval"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The method's gains depend on Gaussian density estimates in the learned latent space actually tracking which prior samples help the target task; when the latent space lacks smoothness, as the authors report for BYOL, the importance weights no longer help.","fun_headline_variants_meta":{"raw":{"variants":["Boost few-shot imitation with importance-weighted retrieval","Why L2 retrieval fails and density ratios fix it","Retrieve demos by density ratio, not distance, for better imitation","Importance weights outperform L2 for few-shot imitation","Swap nearest-neighbor scores for KDE density ratios in retrieval"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00023,"raw_usage":{"total_tokens":1346,"prompt_tokens":795,"completion_tokens":551,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":471}},"tokens_in":539,"tokens_out":551,"duration_ms":5961,"temperature":1.0,"reasoning_tokens":471,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:18:11.696887+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rank the same prior samples with IWR and L2 on a fixed task; train policies on the top-IWR set and, as a control, on the bottom-IWR set. If the bottom set performs as well or better on held-out task success, the density-ratio ranking is not capturing usefulness in that latent space.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines Flow Retrieval, a flow-based retrieval baseline whose embeddings and L2 rule IWR wraps in experiments."},{"cited_title":"Nasiriany, T","cited_arxiv_id":null,"evidence_quote":"Defines SAILOR, a skill-based retrieval method that serves as another baseline and representation for IWR."},{"cited_title":"Walke, K","cited_arxiv_id":null,"evidence_quote":"Supplies the Bridge V2 real-world prior dataset used to evaluate retrieval on unseen sink tasks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies Scott's rule, the bandwidth formula used to set Gaussian KDE smoothing in IWR."}],"review_version":1}