{"id":"d74d06df-6f2b-4def-aa83-9cd13cff8194","arxiv_id":"2607.28251","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Criticality-guided importance sampling raises training information density and cuts embodied-AI failure rates 51–67% versus random finetuning baselines.","lead":"A lightweight failure predictor steers robot training data toward rare crash-prone situations, then corrects the bias so learning stays fair. Across locomotion, manipulation, vision-language-action models, and a real arm, failure rates drop sharply where random finetuning stalls.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"Abstract’s “unbiased IS + information density” mechanism does not hold for the IL/VLA domains that supply the 8–25% SOTA gains.","rationale":"The reader correctly flags hand-engineered state features and routing dependence and lands on CONDITIONAL for the right reasons (random ablations are the strongest evidence; code absent; specialist-only often regresses). I agree the verdict should stay CONDITIONAL, but the single most load-bearing crack in the central claim as written is narrower and text-checkable: the paper asserts one unbiased IS + information-density loop across all domains, yet Stages 2–3 and the VLA protocols implement biased demo selection without weight correction precisely where the SOTA comparisons live. That is an internal mechanism mismatch, not only an external generalization risk about missing state factors. The reader’s state-representation concern remains real (Limitations; Eqs. 1–3) but is partly downstream—if C_φ is incomplete, both the RL and IL pipelines suffer; the IL pipeline fails the paper’s own unbiasedness claim even when C_φ is accurate. Credit where due: multi-domain random controls and the supp density toy still support that criticality-guided curation beats volume-matched random data in RL and, with routing, in IL. Conditioning on (i) splitting RL vs IL mechanism claims, (ii) reporting specialist-only vs routed numbers as first-class, and (iii) code/data remains the right bar—not rejection.","tokens_in":24478,"tokens_out":786,"duration_ms":77038,"concrete_test":"Re-finetune LIBERO and RoboTwin with the RL-style correction: keep the criticality-collected pool but resample demos ∝ W_i = p_unif(i)/q(i) (or reweight the BC loss by W_i), matched steps/hyperparameters, no routing. Report Finetune-only and routed failure rates on the same manifests as Tables 4–5. If unbiased reweighting removes most of the routed gain (or Finetune-only still needs τ-sweep to beat baseline), the abstract’s mechanism does not explain the VLA 8–25% claims.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim ties headline failure cuts to a single mechanism: criticality-guided collection that replaces nominal data with diverse failure-prone data, then importance-weight resampling that “preserves an unbiased learning objective under P(s)” (Abstract; Stage 2–3, Eqs. 2–4). That account is accurate only for the RL domains (Go2, ManiSkill), where buffer episodes are resampled ∝ W = P/q and the loss is unweighted. For IL/VLA (LIBERO, RoboTwin, real-robot banana)—three of five domains, and the only SOTA-VLA comparisons—Stage 3 explicitly samples expert demos from the biased q(i) (Eq. 3) and runs standard unweighted BC; “the importance bias is absorbed by q(i).” No likelihood-ratio correction is applied. The empirical signature matches biased hard-example BC, not unbiased density enrichment: Finetune-only regresses vs baseline (LIBERO 2.13% vs 1.87%; RoboTwin 16.30% vs 14.40%; Random often worse), and the reported 8–25% cuts appear only after validation-swept threshold routing (Eq. 5; τ per suite/global). Routing is therefore load-bearing for the VLA numbers, not an optional monitor. The unified information-density story and the abstract’s unbiased-objective sentence overextend the RL construction to the IL results.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The paper argues that embodied-policy finetuning plateaus because random data collection undersamples rare failures, and proposes a self-evolving loop built around a lightweight state-wise criticality model C_φ ≈ P(failure|state) trained on the policy’s own rollout outcomes. C_φ defines an importance-sampling proposal that concentrates collection on failure-prone states; the policy is then finetuned on the curated pool, and the loop iterates. At deployment, C_φ optionally routes high-criticality states to the finetuned specialist. Empirically, across Go2 locomotion, ManiSkill multi-task manipulation, LIBERO and RoboTwin VLA benchmarks, and a real banana-on-plate task, the method reports 51–67% relative failure-rate reductions versus trained baselines and 8–25% versus SOTA VLA checkpoints, while matched random-data finetuning yields negligible gain or degradation. A controlled toy experiment (Supp. §1) is used to argue that the mechanism is increased information density of distinct critical scenarios under an unbiased objective, not merely a higher critical:nominal sample ratio.","tokens_in":24818,"tokens_out":1434,"duration_ms":33759,"significance":"If the claims hold under a carefully stated mechanism, this is a practically useful systems contribution: a single cheap failure predictor that couples data selection and deployment routing, validated across RL and IL/VLA regimes including a real robot, with matched random-data ablations (Table 7), multi-round trends (Fig. 3), and an explicit information-density toy study. The work generalizes prior rare-event IS ideas (e.g., NADE-style criticality) from evaluation into iterative embodied finetuning. Strengths include multi-domain consistency, clear random-collection controls, and transparent reporting that Finetune-only can regress on VLA tasks until routing is applied. The main value is empirical and methodological rather than theoretical novelty of IS itself.","major_comments":[{"comment":"Abstract and Stage 2–3 (Eqs. 2–4) present a unified mechanism: criticality-guided collection plus importance-weight resampling that “preserves an unbiased learning objective under P(s).” This is accurate for the RL domains (Go2, ManiSkill), where episodes are resampled ∝ W = P/q and the loss is unweighted. For the IL/VLA domains (LIBERO, RoboTwin, real robot)—three of five settings and the only SOTA-VLA comparisons—Stage 3 samples expert demos from the biased q(i) (Eq. 3) and runs standard unweighted BC, with the text stating the bias is “absorbed by q(i).” No likelihood-ratio correction is applied. The abstract’s unbiased-objective sentence and the single information-density story therefore overextend the RL construction to the IL results and should be restated domain-conditionally.","section":"Abstract; Method Stage 2–3, Eqs. (2)–(4)"},{"comment":"On LIBERO and RoboTwin, Finetune-only underperforms the baseline (Table 4: 2.13% vs 1.87%; Table 5: 16.30% vs 14.40%), and the reported 8–25% SOTA-relative cuts appear only after validation-swept threshold routing (Eq. 5; per-suite/global τ). Routing is therefore load-bearing for the headline VLA numbers, not merely an optional deployment monitor as framed in Contribution 1 and the Deployment section. The paper should lead with this specialization trade-off for IL/VLA, report routed vs unrouted numbers as primary, and avoid attributing those gains solely to unbiased density enrichment.","section":"Deployment Eq. (5); Tables 4–5; Contribution 1"},{"comment":"The weakest modeling assumption is that pre-defined state features (terrain grid, force grid, padded initial-state vectors, 8-D geometric coords; Table 1 and Supp. §9) suffice for C_φ to surface residual failure modes of the improved policy. Limitations notes manual state engineering, but there is no stress test where the representation is intentionally incomplete or where post-update failure modes leave the scored support. A minimal experiment or quantitative failure analysis (e.g., modes missed by C_φ after Round k) would make the central sampling claim more falsifiable, especially for the VLA initial-state-only scores.","section":"Stage 1; Limitations; Table 1; Supp. §9"}],"minor_comments":[{"comment":"Free parameters (ε, β/κ, τ, RoboTwin α and capped-sigmoid schedule) are scattered across Method and Supp.; a single hyperparameter table with selection protocol (validation sweep vs fixed defaults) would aid reproducibility.","section":"Method; Supp. §3, §7"},{"comment":"Figure 1 is helpful but the Stage 2 panel does not visually distinguish RL trajectory-weighting from IL demo-level q(i), which is exactly where the mechanism splits.","section":"Figure 1"},{"comment":"Go2 and real-robot results use IS evaluation (Supp. §2). Main-text tables should mark IS-estimated μ explicitly so readers do not compare raw Monte Carlo rates to reweighted rates without noticing.","section":"Tables 2, 6; Supp. §2"},{"comment":"Minor prose/spacing artifacts appear throughout (e.g., “systemsroutinelyplateauduring,” “state-wisecriticalitymodel”). A full copy-edit pass is needed.","section":"Abstract; Introduction"},{"comment":"Related Work cites dense-learning / NADE training extensions appropriately; briefly clarify what is new relative to Feng et al. 2026 beyond multi-perturbation embodied settings (already hinted, but one crisp paragraph would help).","section":"Related Work"}],"recommendation":"major_revision","confidential_remarks":"The empirical package is above average for an embodied-systems submission and I would expect an accept after the mechanism is stated honestly (RL unbiased IS vs IL hard-example BC + routing). The main risk is overclaim in the abstract aligning all five domains under one unbiased information-density banner; that is fixable without new algorithms. No integrity concerns. Scope fits eess.SY / robot learning venues; novelty is incremental relative to rare-event IS literature but the multi-domain execution is the selling point."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is practical, not foundational: a lightweight failure predictor trained on the policy’s own rollouts can steer data collection toward residual failure modes, and matched random-data controls across five domains show that extra training alone does not explain the gains. That is worth knowing if you care about post-pretrain plateaus under rare failures.\n\nWhat is actually new is the packaging and breadth. The criticality/IS core is carried from the NADE and dense-learning AV line (same corresponding-author thread). Here they generalize it to mixed RL and IL/VLA settings, multi-perturbation spaces, iterative self-evolution, and optional deployment routing, with a clean supplementary toy study that isolates information density (distinct critical scenarios) from critical:nominal sample ratio. Empirically they do the right controls: Table 7 and the multi-round plots make the random-collection ablation hard to dismiss. Go2 and ManiSkill look especially clean under the stated IS construction.\n\nThe soft spot is framing, and the stress-test is mostly right. Abstract and Stage 2–3 sell one story—criticality collection plus importance-weight resampling that preserves an unbiased objective under P(s). That matches the RL domains (resample ∝ W, unweighted loss). For LIBERO, RoboTwin, and the real robot—the domains that supply the 8–25% SOTA-VLA numbers—they sample expert demos from biased q(i) and run ordinary unweighted BC. Finetune-only often regresses on nominal cases; the headline cuts appear after validation-swept threshold routing. Routing is load-bearing there, not a free monitor. Call it biased hard-example BC plus a risk switch, not the same unbiased density story as Go2. Minor other cautions: hand-engineered criticality features, small real-robot N, free knobs (ε, β, τ), no code.\n\nWho it is for: people stuck finetuning locomotion or VLA policies when failures are rare. Serious referee material—important bottleneck, honest enough empirics, fixable claim hygiene. I would engage, push them to split RL vs IL mechanisms in the abstract, and report specialist-only vs routed numbers side by side everywhere.","headline":"Solid cross-domain transfer of criticality-guided sampling with real ablations; abstract overclaims a single unbiased-IS mechanism that only fully holds in the RL domains.","tokens_in":25496,"tokens_out":559,"would_cite":true,"duration_ms":19418,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A failure-probability model trained on a policy’s own rollouts can steer data collection so finetuning finally escapes the plateau of random sampling.","keywords":["embodied AI","criticality model","importance sampling","self-evolving learning","vision-language-action","finetuning plateau","failure-prone sampling","policy routing"],"falsifier":"Run the same multi-round loop with a deliberately incomplete state input to C_φ (drop terrain, force, or object-pose channels) and check whether round-over-round failure-rate gains collapse to the random-data control while validation PR-AUC of C_φ stays high on the incomplete features.","tokens_in":25319,"feed_emoji":"🦾","tokens_out":965,"duration_ms":20608,"temperature":0.7,"pith_summary":"Embodied policies often stop improving once they are mostly competent, because random data collection almost never hits the rare failures that still matter. This paper argues that the missing ingredient is not more capacity or more raw data, but higher information density: replace redundant easy scenarios with diverse failure-prone ones, then correct the bias with importance weights so training stays unbiased. A lightweight criticality model learns P(failure|state) from the policy’s own successes and failures and drives that sampling loop; the same model later routes high-risk states to the finetuned specialist at deployment. Across quadruped locomotion, multi-task manipulation, vision-language-action benchmarks, and a real banana-on-plate task, the method cuts failure rates by roughly half to two-thirds versus matched trained baselines, while identical training on randomly collected data barely moves the needle. The sympathetic takeaway is that self-judged rarity, not volume, is what breaks the finetuning plateau.","feed_headline":"Failure predictor steers robot data, cuts errors up to 67%","feed_subtitle":"Random finetuning plateaus; criticality-guided sampling raises information density and keeps improving.","key_machinery":"The criticality model C_φ: a lightweight network that scores P(failure|state) from rollout labels, builds the ε-mixture proposal q(s) ∝ κ(C_φ(s)), and supplies both collection weights and deployment routing thresholds.","core_discovery":"The paper’s central claim is that a state-wise criticality model C_φ ≈ P(failure|state), trained only on a policy’s own execution outcomes, can define an importance-sampling proposal that concentrates collection on failure-prone regions; after importance-weight correction (and optional threshold routing between baseline and finetuned policies), the training objective remains unbiased under the natural state distribution while the pool’s information density rises, producing large, consistent failure-rate reductions that matched random-data finetuning does not achieve.","pith_inferences":["If criticality were learned in a vision latent space instead of engineered states, the same loop could transfer across embodiments without per-domain feature design.","Coupling C_φ’s per-step scores as dense process rewards inside online VLA RL could close the loop the paper leaves open under flow-matching instability.","Domains outside robotics that also suffer a curse of rarity (rare safety events, rare medical outcomes) may admit the same evaluate-to-train transition once a cheap failure predictor exists."],"forward_implications":["Default random finetuning pipelines for locomotion, manipulation, and VLA policies will plateau even with more steps; criticality-guided collection is required to keep cutting residual failures.","The same P(failure|s) monitor can double as a deployment risk switch that hands high-criticality states to a specialist policy without retraining the base model.","Failure-rate gains of 51–67% vs trained baselines and 8–25% vs strong VLA checkpoints become available without changing the original RL or imitation loss, only the data proposal.","Information density—more distinct failure modes in the pool, not a higher critical-to-nominal batch ratio—is the operative lever, as the paper’s controlled toy classification experiment isolates."],"fun_headline_variants":["Criticality model steers robot data to failures, cuts errors 51-67%","Self-evolving sampler targets failure states, lowers robot error rates","State-wise failure predictor raises data density in embodied finetuning","Criticality-guided importance sampling breaks robot finetuning plateaus","Policy-owned criticality model cuts embodied AI failures up to 67%"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That a compact model on hand-chosen state features can keep finding the policy’s remaining failure modes as the policy improves; if those features miss the true rare causes, guided sampling aims at the wrong slice of rarity.","fun_headline_variants_meta":{"raw":{"variants":["Criticality model steers robot data to failures, cuts errors 51-67%","Self-evolving sampler targets failure states, lowers robot error rates","State-wise failure predictor raises data density in embodied finetuning","Criticality-guided importance sampling breaks robot finetuning plateaus","Policy-owned criticality model cuts embodied AI failures up to 67%"]},"model":"grok-4.5","effort":"low","cost_usd":0.003782,"raw_usage":{"total_tokens":1186,"prompt_tokens":779,"num_sources_used":0,"completion_tokens":80,"cost_in_usd_ticks":37824000,"prompt_tokens_details":{"text_tokens":779,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":327,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":779,"tokens_out":80,"duration_ms":7516,"temperature":1.0,"reasoning_tokens":327,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T13:14:55.318580+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the same multi-round loop with a deliberately incomplete state input to C_φ (drop terrain, force, or object-pose channels) and check whether round-over-round failure-rate gains collapse to the random-data control while validation PR-AUC of C_φ stays high on the incomplete features.","supporting_citations":[],"review_version":1}