{"id":"63f61c11-4ece-4bf3-991e-fa0f6b887def","arxiv_id":"2607.07076","paper_version":1,"verdict":"CONDITIONAL","confidence":"UNKNOWN","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"A lightweight primitive classifier and differentiable guidance mechanism improve pretrained diffusion and flow manipulation policies by 3–7 points at test time without retraining.","lead":"PriGo adds a plug-in module that predicts what type of robot motion (grasp, push, rotate, etc.) should happen next, then steers pretrained diffusion or flow policies toward that motion type during inference—no retraining needed. A smart generalist might read it because test-time guidance for robot policies is a practical way to make learned manipulation more robust without collecting new data or retraining large models.","discovery_kind":"new_method","skeptic_critique":{"model":"glm-5.2","headline":"The 3–5 point LIBERO improvements (Table 1) are reported without seed-level variance; diffusion policies commonly exhibit ±3–5 point run-to-run variance, so 'consistently improves' may be unsupported.","rationale":"The reader identified the seed-variance issue as point (4) in their rationale but made the taxonomy completeness their primary weakest_assumption. I think the statistical significance concern is more directly load-bearing for the central claim, which is fundamentally empirical ('consistently improves by 3–7 points'). The taxonomy concern is valid but more speculative—the paper does provide ablation evidence (Table 5) that taxonomy choices matter, and the sensitivity analysis (Fig. 3) shows robustness to thresholds. In contrast, the seed-variance concern directly challenges whether the reported improvements are real, and it is trivially testable. The reader's CONDITIONAL verdict is appropriate and my concern reinforces rather than changes it. The real-world and CALVIN results are sufficiently large-margin to be convincing despite variance; the concern is specifically about the LIBERO results where gains are smallest and most likely to be within noise. The shuffled-label control I propose would additionally distinguish whether the primitive content matters or whether any gradient-based action constraint helps equally—this goes to the mechanistic claim, not just the empirical one. If shuffled labels help as much as correct ones, the paper's framing around 'primitive guidance' would need revision, though the method might still be practically useful as a generic regularizer.","tokens_in":15896,"tokens_out":5482,"duration_ms":166063,"concrete_test":"Re-run all four LIBERO settings in Table 1 for DP and CogACT with ≥5 random seeds each, with and without PriGo. Report mean ± std and perform paired t-tests. If any configuration's improvement is not significant at p < 0.05, the 'consistently improves' claim is unsupported for that configuration. Additionally, run one control with shuffled PANet predictions (random primitive labels) to test whether correct primitive structure or generic action regularization drives the gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that PriGo 'consistently improves' success rates by 3–7 points across benchmarks. On LIBERO (Table 1), the gains are 3–5 points across four backbones (π0: +2.6 avg, SmolVLA: +2.1, DP: +3.5, CogACT: +4.5). However, each cell appears to be a single run with no seed-level variance reported. Diffusion and flow policies are well-documented to exhibit run-to-run variance of ±3–5 points due to stochastic initialization, data ordering, and inference-time noise sampling. If the LIBERO improvements fall within this variance band, the word 'consistently' is unsupported for the benchmark where gains are smallest. The real-world results (Table 4: +22 to +29 points) and CALVIN results (Table 3: +0.83 avg length) show larger margins and are more convincing despite small sample sizes. But LIBERO is the primary benchmark, the one where PANet is trained and evaluated in-domain, and the one where the gains are smallest. Additionally, the guidance step size η is never specified or sensitivity-analyzed, making it impossible to assess whether the guidance meaningfully alters actions or is negligibly small. The ablation in Table 5 shows taxonomy choices matter (removing push/pull distinction drops 2.4 points), which partially addresses whether primitives vs. generic regularization drive gains—but this ablation itself also lacks seed variance. Without seed-level analysis on the headline benchmark, the core empirical claim remains uncertain.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"PriGo introduces a test-time primitive guidance framework for diffusion and flow policies in robotic manipulation. The core idea is to train a lightweight classifier (PANet) that predicts an 8-category primitive distribution from visual observations and language instructions, then use a differentiable guidance loss (cross-entropy between PANet predictions and soft-classified actions) to refine generated actions during inference via gradient correction. The guidance is integrated into both DDPM-based diffusion policies and flow matching policies without retraining the backbone. Experiments span LIBERO, CALVIN, SIMPLER, and real-world Franka tasks across four backbones (π0, SmolVLA, DP, CogACT), with ablations on the primitive taxonomy and PANet accuracy.","tokens_in":16265,"tokens_out":1388,"duration_ms":157780,"significance":"The plug-and-play, no-retraining design is the paper's main contribution and is well-motivated: it decouples primitive planning from continuous action learning and applies to both lightweight and VLA-scale policies. The differentiable soft-classification (Def. 3.2) enabling gradient flow through the primitive taxonomy is a clean technical contribution. The error-bound analysis (Proposition A.1) attempts to formalize convergence, and the sensitivity analysis (Fig. 3, right) addresses threshold robustness. The breadth of benchmarks and the real-world demonstrations with limited data are commendable. Code is promised, which supports reproducibility.","major_comments":[{"comment":"Table 1 (LIBERO): The headline claim that PriGo 'consistently improves' performance by 3–5 points is reported without seed-level variance. Diffusion and flow policies are well-documented to exhibit ±3–5 point run-to-run variance from stochastic initialization, data ordering, and inference-time noise sampling. Several gains in Table 1 (e.g., π0 Spatial: +1.8, SmolVLA Goal: +1.8, DP Goal: +2.2) are within this band. Without confidence intervals or multi-seed evaluation on the primary benchmark, the word 'consistently' is unsupported for the settings where gains are smallest. The real-world (Table 4: +22–29 pts) and CALVIN (Table 3: +0.83 avg length) results show larger margins and are more convincing, but LIBERO is the in-domain benchmark where the smallest gains are reported. Please provide multi-seed results (≥3 seeds) with standard deviations for Table 1, or reframe the claim to match.","section":null},{"comment":"§3.2–3.3: The guidance step size η in Eqs. (3) and (4) is never specified anywhere in the paper, and no sensitivity analysis is provided. This is a load-bearing hyperparameter: if η is too small, the guidance is negligible and the improvements are unrelated to the proposed mechanism; if too large, it may destabilize trajectories. The sensitivity analysis in Fig. 3 (right) covers only the auto-labeling thresholds (τ_trans, τ_rot, τ_w), not η or the softmax temperature τ in Def. 3.2. Please report the η values used for each backbone and provide a sensitivity analysis, as this directly determines whether the guidance mechanism is actually responsible for the observed gains.","section":null},{"comment":"Proposition A.1 (Appendix A): The error bound assumes the guidance field G is locally strongly monotone near the target action (Eq. 9: ⟨G(a)−G(a*), a−a*⟩ ≥ C₀‖a−a*‖²). This is a strong assumption that is not verified for the cross-entropy-based guidance loss L_PG. The guidance gradient ∇_a L_PG depends on the soft-classification score functions f_k (Def. 3.2), whose monotonicity properties near arbitrary target actions are not established. The proposition thus provides a formal bound under an assumption whose applicability is unclear. Please either verify the monotonicity condition empirically (e.g., by plotting ‖∇_a L_PG‖ vs. ‖a−a*‖ near expert actions) or clearly state this as an unverified assumption and soften the theoretical contribution accordingly.","section":null}],"minor_comments":[{"comment":"§3.1, Eq. (1): The hard classification uses x_k^1 (the first component of translation) to distinguish push (x_k^1 > 0) from pull (x_k^1 < 0). It is unclear which axis this corresponds to (robot base frame? end-effector frame? world frame?). Please specify, as the push/pull distinction depends on this choice.","section":null},{"comment":"Table 2: The caption states 'Comparison of our approach PriGo-DP (CogACT with PriGo)' but the table header lists 'PriGo-DP' as a separate method row alongside CogACT. This is slightly confusing—consider labeling it 'CogACT + PriGo' for consistency with Table 1's notation.","section":null},{"comment":"§4.3: The real-world experiments report 'averaged success rates over 10 episodes per task' (Table 4). While the gains are large (+22–29 pts), 10 episodes is a small sample. Please report confidence intervals or note this limitation.","section":null},{"comment":"Appendix C: The wall-clock analysis (208 ms/step, 77 Hz effective) is mentioned only in the appendix. Given that inference overhead is practically important for a plug-and-play method, consider moving this to the main text.","section":null},{"comment":"Fig. 1: The figure caption references 'a) Structured Primitive Actions' and 'b) Unstructured Primitive Actions,' but the subfigure labels in the image appear to show specific failure modes (pull before unlock, falling) rather than a general structured vs. unstructured comparison. Consider revising the caption for clarity.","section":null},{"comment":"The manuscript states 'Codes are available on PriGo' (§1) but no URL is provided. Please include a repository link or anonymized placeholder.","section":null}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about missing seed variance on LIBERO is the most substantive issue and is well-founded—Table 1 is the primary benchmark and the gains there are within typical variance bands. The η omission is a close second: without knowing the guidance strength, it is impossible to assess whether the mechanism is actually doing work. The error-bound proposition is a minor theoretical contribution that could be dropped or moved to a footnote if the assumptions cannot be verified, but it is not load-bearing for the empirical claims. The circularity concern (primitive labels derived from actions, then used to guide actions) is not a real problem because PANet predicts primitives from observations, not from the actions being guided—but the authors could make this point more explicitly."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"The paper worth your attention is PriGo, which applies classifier-guided diffusion (and flow matching) to robotic manipulation policies at test time, using a learned primitive classifier (PANet) to steer action generation without retraining the backbone policy. The core idea: train a lightweight classifier to predict which of 8 motion primitives (grasp, push, pull, rotation, etc.) should occur next from visual observations and language, then at inference compute a cross-entropy loss between PANet's predicted primitive distribution and a soft classification of the generated action, and use the gradient to refine the action. This is applied to both DDPM-based diffusion policies and flow matching policies via modified denoising/update steps (Eqs. 3-4). The plug-and-play framing is genuinely useful—no retraining of large VLA models is needed, and the mechanism is architecturally simple. The real-world results (Table 4: +22 to +29 points over CogACT on three table-service tasks) and CALVIN long-horizon results (Table 3: +0.83 average trajectory length over 3D Diffuser Actor) show substantial margins. The SIMPLER results under visual perturbation are also solid (+5-7 points). PANet itself achieves 94% classification accuracy, and the ablation in Table 5 confirms that distinguishing push from pull and including hybrid primitives both contribute meaningfully. The soft probabilistic formulation for guidance is a nice design choice that handles primitive transitions gracefully. Now the soft spots. The stress-test concern about seed variance on LIBERO is legitimate and I share it. The 3-5 point gains on LIBERO (Table 1) are reported as single numbers with no error bars or seed-level analysis. Diffusion policies are known to have ±3-5 point run-to-run variance, so some of these improvements could be within noise. The larger margins on CALVIN, SIMPLER, and real-world tasks are more convincing, but LIBERO is the primary in-domain benchmark where gains are smallest. The guidance step size η is never specified or sensitivity-analyzed in the main text, which is a real omission—you cannot assess whether the guidance meaningfully alters actions or is negligibly small. The error bound (Proposition A.1) relies on an unverified strong monotonicity assumption for the guidance field; it's a formal exercise rather than a proven guarantee. The 8-primitive taxonomy is hand-designed, and while the sensitivity analysis (Fig. 3) shows robustness to classification thresholds, the taxonomy itself is validated only on the tested benchmarks. Real-world results use only 10 episodes per task with high variance (±7-11 points). None of these are fatal—the method is sound in principle and the experiments are reasonably comprehensive. But the missing seed variance on the headline benchmark and the unspecified η are the two things I'd push hardest on in revision. This paper is for researchers working on diffusion/flow policies for manipulation who want test-time improvement without retraining. It deserves a serious referee who can assess whether the LIBERO gains survive proper variance analysis.","headline":"Plug-and-play test-time primitive guidance for diffusion/flow policies: real gains, but LIBERO results lack seed variance and the error bound is unverified.","tokens_in":16899,"tokens_out":696,"would_cite":true,"duration_ms":54275,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Steering robot policies with primitive-level guidance at inference time","keywords":[],"falsifier":"A manipulation domain where the eight primitives do not decompose the relevant actions — for example, tasks dominated by continuous deformations (folding cloth, pouring liquid) or tasks where the same observation legitimately requires different primitives depending on hidden state. In such domains, PANet's predictions would be unreliable, and the guidance gradient would push actions toward the wrong structural category, degrading rather than improving policy performance.","tokens_in":16164,"feed_emoji":"🤖","tokens_out":1188,"duration_ms":155974,"temperature":0.7,"pith_summary":"Robot imitation policies that generate actions via diffusion or flow matching can produce motions that look locally plausible but violate the structural intent of a manipulation task — for example, pulling before rotating a key, or grasping the wrong part of a cup. This paper claims that a lightweight classifier, trained separately to predict which of eight motion primitives (grasp, push, pull, rotate, etc.) the robot should be executing at each step, can be used as a differentiable constraint during inference to nudge the policy's output actions toward structurally consistent trajectories. The key claim is that this guidance requires no retraining of the underlying policy: at each denoising or flow-matching step, the gradient of a cross-entropy loss between the classifier's predicted primitive distribution and a soft classification of the current action proposal is used to refine the action. The authors report consistent 3–7 point success-rate improvements across four simulation benchmarks and real-world table-service tasks, for both diffusion-based and flow-based policies of varying scale.","feed_headline":"Primitives steer robot policies at inference, no retraining needed","feed_subtitle":"A lightweight classifier predicts which motion primitive the robot should execute, and its gradient nudges diffusion and flow policies 3–7%.","key_machinery":"The differentiable primitive guidance mechanism. The hard primitive classification (Eq. 1) maps a 7-DOF action to one of eight categories using thresholds on translation magnitude, rotation magnitude, and gripper-width change. The soft classification (Def. 3.2) replaces the argmax with a temperature-scaled softmax over differentiable score functions, making the mapping from action to primitive distribution smooth. The guidance loss L_PG (Eq. 2) is the cross-entropy between PANet's predicted distribution and the action-induced soft distribution. For diffusion policies, the gradient ∇_a L_PG is applied to the estimated clean action at each denoising step (Eq. 3); for flow policies, it modifies","core_discovery":"The central mechanism is a test-time correction loop. A pretrained primitive classifier (PANet) maps the current visual observation and language instruction to a probability distribution over eight manipulation primitives. Separately, each candidate action generated by a frozen diffusion or flow policy is converted into a soft probability distribution over the same eight categories using differentiable score functions on the action's translational, rotational, and gripper-width components. The cross-entropy between these two distributions serves as a guidance loss, and its gradient with respect to the action is applied as a correction term at each inference step — modifying the denoising or,","pith_inferences":["If the primitive taxonomy is domain-specific and manually designed, the approach may transfer poorly to manipulation domains with fundamentally different action structures — for instance, in-hand manipulation or deformable-object manipulation where the relevant primitives (e.g., pivoting, rolling, stretching) differ from the eight proposed here.","The guidance gradient could interact poorly with the policy's learned distribution in cases where the policy intentionally produces actions that do not match any single primitive — for example, smooth transitional motions or bimanual coordinated actions that blend primitives simultaneously. The soft classification mitigates this, but the taxonomy's discreteness may still introduce bias.","A natural extension would be to learn the primitive taxonomy itself from data rather than designing it manually, potentially via clustering in action-embedding space, which could make the guidance more adaptive to unseen task distributions.","The 15% inference-time overhead reported (181 ms → 208 ms per step) is modest for the benchmarks tested, but the overhead scales with the number of denoising or flow steps and the frequency of guidance application, which could become significant for policies with many refinement steps or high-frequency control loops."],"forward_implications":["Any policy that generates actions as iterative refinement of noisy samples (diffusion, flow matching, or related generative processes) could in principle receive test-time guidance from an external classifier that predicts a structurally meaningful property of the action, not just primitives — for example, contact state, affordance, or object-relative pose.","The plug-and-play design means that organizations deploying pretrained manipulation policies could improve robustness without access to the original training data or policy weights, applying only an inference-time gradient correction.","The primitive taxonomy and its automatic labeling procedure (Eq. 1) could be extended or adapted to new manipulation domains by redefining the primitive set and retraining only the lightweight classifier, leaving the policy untouched.","The error-bound analysis (Proposition A.1) suggests that guidance fields with a fixed point at the target action and local strong monotonicity guarantee convergence improvement, providing a theoretical template for analyzing other test-time guidance schemes."],"fun_headline_variants":["Test-time primitive guidance steers diffusion and flow robot policies without retraining","Primitive gradients correct robot actions at inference, no policy retraining","Robotic manipulation policies get test-time guidance from learned primitives","Frozen diffusion and flow policies improve via primitive-guided inference","Eight motion primitives guide robot policies at test time without retraining"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The eight-category primitive taxonomy and its fixed classification thresholds capture manipulation structure that is both predictable from visual observations by a small classifier and useful as a gradient-based constraint for action refinement. If the taxonomy is too coarse, too fine, or misaligned with the task's actual motion structure, the guidance gradient could be uninformative or counterproductive.","fun_headline_variants_meta":{"raw":{"variants":["Test-time primitive guidance steers diffusion and flow robot policies without retraining","Primitive gradients correct robot actions at inference, no policy retraining","Robotic manipulation policies get test-time guidance from learned primitives","Frozen diffusion and flow policies improve via primitive-guided inference","Eight motion primitives guide robot policies at test time without retraining","Differentiable primitive loss nudges frozen robot policies toward correct behaviors","Primitive classifier guides diffusion and flow policies during inference","Test-time correction loop improves robot policy generalization and robustness","Primitive distributions steer frozen robot policies without any retraining","Gradient from primitive classifier adapts robot policies at inference time"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":943,"prompt_tokens":502,"completion_tokens":441,"prompt_tokens_details":null},"tokens_in":502,"tokens_out":441,"duration_ms":10708,"temperature":1.0,"reasoning_tokens":375,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T20:28:20.351758+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"A manipulation domain where the eight primitives do not decompose the relevant actions — for example, tasks dominated by continuous deformations (folding cloth, pouring liquid) or tasks where the same observation legitimately requires different primitives depending on hidden state. In such domains, PANet's predictions would be unreliable, and the guidance gradient would push actions toward the wrong structural category, degrading rather than improving policy performance.","supporting_citations":[],"review_version":1}