{"id":"00c56274-8d6b-4ecf-b938-8592fd107099","arxiv_id":"2603.06279","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"2D label-noise methods collapse on sparse 3D occupancy; dual-source partial-label reasoning (DPR-Occ) preserves IoU/mIoU under extreme synthetic and trailing noise on OccNL/SemanticKITTI.","lead":"This paper shows that standard 2D noisy-label tricks fail on sparse 3D occupancy voxels and introduces OccNL plus DPR-Occ, which keeps geometry and semantics usable even at 90% label noise. Safety-critical robots that train on imperfect maps need this kind of robustness.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"The 90% headline gains rest on an unmeasured dual-source coverage assumption that is least secure for sparse classes under distribution flattening.","rationale":"The reader correctly isolates the unmeasured coverage of the dual-source candidate set under extreme noise as the weakest assumption supporting the 90% gains. That assumption is load-bearing: PLL/NL/SNTD can only re-weight probability mass inside or outside a candidate set; they cannot recover a true label that was never admitted. Fig. 2 and the near-zero IoUs on sparse classes make the risk concrete rather than hypothetical. The paper’s own ablations stop at 70% and never quantify hit-rate, so the central empirical claim remains conditional on an untested intermediate quantity. No stronger internal inconsistency appears (the method is a coherent systems recipe, tables are thorough, and the domain-gap observation is well-supported). Therefore the reader’s CONDITIONAL verdict and HIGH confidence on the stated experimental numbers are unchanged; the concrete coverage diagnostic is the single check that would either solidify or weaken the headline numbers.","tokens_in":18090,"tokens_out":688,"duration_ms":7293,"concrete_test":"On the 90% OccNL split, instrument training to log, for every voxel and every epoch after E_w, whether the ground-truth class lies inside PL_v (and inside each source separately). Report mean coverage overall and per-class (especially the four sparsest dynamic classes). If overall coverage <70% or sparse-class coverage <30%, re-run Table II with an oracle that forces true-label inclusion; if the mIoU gap to baselines shrinks by more than half, the dual-source coverage assumption is load-bearing and currently unverified.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim (Table II, 90% row: DPR-Occ IoU 35.03 / mIoU 8.23 vs. best adapted baseline ~21 IoU / ~5.7 mIoU) requires that the dual-source candidate set PL_v = TopK(y_ema, K_e) ∪ TopK(S(v,·), K_e) (Eq. 4) still contains the true label often enough for PLL (Eq. 6) + NL (Eq. 7) + SNTD (Eq. 9) to recover it after warm-up. At η=90% the occupancy-asymmetric process drives the label distribution toward uniformity (Fig. 2), erasing the long-tail structure that prototypes and early EMA memory rely on; sparse dynamic classes (bicycle, person, bicyclist, motorcyclist) already sit near 0 IoU even for DPR-Occ. The paper never reports hit-rate / coverage of PL_v versus true Y, nor how often the true class is excluded when both EMA rank and prototype rank drop. Ablations (Tables VI–VII) only show that both sources help at 70%; they do not measure whether coverage itself fails at 90%. If coverage is low, the subsequent losses cannot invent the missing true label, so the reported prevention of “geometric and semantic collapse” is only partial (geometry is preserved; many minority semantics remain extinct).","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper introduces OccNL, a benchmark for 3D semantic occupancy prediction under occupancy-asymmetric category-flipping noise and real-world dynamic trailing artifacts on SemanticKITTI, and proposes DPR-Occ, a dual-source partial-label framework. After a warm-up stage that exploits memorization under noisy labels while maintaining an EMA teacher and class prototypes, the method constructs candidate sets by union of Top-K EMA predictions and Top-K feature-prototype similarities (Eq. 4), then optimizes partial-label learning, negative learning on the complementary set, and EMA-guided self-not-true distillation (Eqs. 6–11). Experiments adapt five 2D label-noise methods (AGCE, ANL, JAL, VBL, SNTD) to the voxel setting and report that DPR-Occ preserves geometric IoU and improves mIoU under 50–90% asymmetric noise and three trailing severities (Tables II–III), with ablations on warm-up length, dynamic-K, dual-source fusion, and loss terms (Tables IV–VII).","tokens_in":18532,"tokens_out":1256,"duration_ms":12671,"significance":"If the results hold, the work is a useful first systematic study of voxel-level label noise for semantic occupancy, a practically important and previously underexplored failure mode for safety-critical robotics. OccNL supplies controllable synthetic and real-world-style noise protocols plus a refined clean evaluation set (Eq. 1), and the public code/benchmark commitment is a concrete contribution. The empirical demonstration that several strong 2D robust losses collapse in sparse 3D voxel spaces (Table II, 90% row) is informative for the community. DPR-Occ’s gains in geometric integrity under extreme corruption are the most actionable finding; the dual-source partial-label construction is a reasonable, if not fully validated, design for sparse 3D settings.","major_comments":[{"comment":"The central 90% claim (abstract; Table II) rests on an unmeasured coverage assumption for the dual-source candidate set PL_v (Eq. 4). The paper never reports hit-rate of PL_v versus true labels Y, nor how often the true class is excluded when both EMA rank and prototype rank drop. At η=90% the label distribution is driven toward uniformity (Fig. 2), which undermines the long-tail structure that prototypes and early EMA memory rely on. Ablations (Tables VI–VII) only show complementary gains at 70%; they do not establish that coverage remains high enough at 90% for PLL/NL/SNTD to recover sparse classes. Without this diagnostic, the claim that the method “prevents geometric and semantic collapse” is only partially supported: geometry is largely preserved, but many minority semantics remain near zero IoU even for DPR-Occ.","section":null},{"comment":"Table II, 90% row: the headline “up to 2.57% mIoU and 13.91% IoU” gains are real relative to the adapted baselines, yet absolute mIoU is only 8.23 and several dynamic/rare classes (bicycle, motorcycle, person, bicyclist, motorcyclist, other-ground, traffic-sign) stay at or near 0. The abstract and conclusion language of preserving “sparse dynamic semantics” and providing a “reliable foundation for safety-critical” perception overstates what the per-class numbers show. The manuscript should either qualify these claims or provide additional evidence (e.g., coverage curves, class-wise recovery analysis) that minority semantics are meaningfully retained rather than merely less collapsed than the baselines.","section":null},{"comment":"All quantitative results appear to be single-run point estimates with no error bars, seeds, or multi-run statistics (Tables I–VII). Given the free parameters (Ew, Kstart/Kend/γ, τs, EMA momentum) and the known sensitivity of noisy-label methods to initialization and schedule, the reported margins—especially the 2.57 mIoU gap at 90%—cannot be assessed for stability. At least multi-seed means and standard deviations on the primary 70% and 90% settings are needed for the load-bearing comparisons.","section":null}],"minor_comments":[{"comment":"The empty-voxel flip rate is fixed at 10^{-3}η without sensitivity analysis; a short ablation would clarify whether geometric preservation is an artifact of this design choice.","section":null},{"comment":"Fig. 1 and Fig. 4 captions are dense; a clearer legend for noise type (asymmetric vs. trailing) and a consistent color key for classes would improve readability.","section":null},{"comment":"Notation for the candidate size switches between K, Ke, Kstart, Kend; a single consistent symbol with a short definition list would help.","section":null},{"comment":"Related work on uncertainty-aware occupancy and pseudo-label filtering is cited, but a brief discussion of how OccNL differs from confidence-based filtering (rather than only from 2D LNL) would strengthen positioning.","section":null},{"comment":"Implementation details state consistency with ProtoOcc; explicitly listing any ProtoOcc-specific losses that remain in L_base would aid reproducibility before code release.","section":null}],"recommendation":"major_revision","confidential_remarks":"The work is a solid first benchmark paper for a real gap; I would not reject on novelty grounds. The main risk is overclaiming on “semantic” robustness when the 90% per-class table still shows minority extinction. If the authors add coverage diagnostics and multi-seed stats and tone the abstract/conclusion accordingly, this is a clear accept for a CV/robotics venue. Scope fit is good for a perception/robotics journal; less so for a pure learning-theory venue."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple: this is the first controlled look at how occupancy labels break under occupancy-asymmetric flips and real trailing, and the five adapted 2D robust losses (AGCE, ANL, JAL, VBL, SNTD) really do fall apart on SemanticKITTI voxels. That domain-gap result is the paper’s main contribution. DPR-Occ is a practical dual-source partial-label recipe (EMA Top-K union prototype affinity, dynamic-K, PLL+NL+SNTD) that keeps geometric IoU around 35% at 90% noise while the baselines drop to ~20% or lower. Tables II–III and the ablations are thorough enough to trust the ranking on this backbone and dataset.\n\nWhat is new is the benchmark itself (asymmetric flip protocol that protects empty voxels, three trailing severities, refined clean eval via Eq. 1) plus the empirical demonstration that image-domain noise methods do not transfer. The method recombines known pieces from ProtoOcc and SNTD, but the combination is sensible and the gains at extreme noise are large enough to matter for anyone who actually trains occupancy models.\n\nThe soft spot the stress-test flags is real but not fatal: they never report hit-rate of the dual-source candidate set PL_v against true labels, especially once Fig. 2 flattens the distribution at 90%. Sparse classes (bicycle, person, bicyclist, motorcyclist) stay near zero even for DPR-Occ, so “prevents semantic collapse” is overstated for the long tail; geometry is what is actually preserved. Ablations stop at 70% for the dual-source and loss terms. Single backbone, single dataset, no multi-seed error bars, code only promised. Those are ordinary first-paper limits, not hidden circularity.\n\nThis is for people who ship occupancy or care about label quality in 3D perception. It is not a theory paper and does not need to be. I would send it to peer review; the benchmark and the collapse finding alone justify referee time. Once OccNL is public I would cite the numbers and probably reuse the noise protocols. Worth a reading-group slot if anyone in the group is working on SSC or noisy 3D labels.","headline":"Solid first stress-test for occupancy under label noise: OccNL shows 2D robust methods collapse on sparse voxels, and DPR-Occ keeps geometry usable at 90% noise even if minority classes stay weak.","tokens_in":19118,"tokens_out":562,"would_cite":true,"duration_ms":5973,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"3D occupancy models can keep usable geometry and semantics even when 90% of voxel labels are wrong, if candidates are built from model memory and feature prototypes rather than hard noisy labels.","keywords":["3D semantic occupancy","label noise","partial label learning","SemanticKITTI","dynamic trailing","prototype affinity","EMA teacher","robotic perception"],"falsifier":"On the 90% asymmetric-noise OccNL split, ablate the prototype branch or force a fixed tiny K so ground-truth coverage of sparse dynamic classes falls; if geometric IoU and rare-class mIoU then collapse to the level of the pure 2D baselines, the recovery claim fails.","tokens_in":18979,"feed_emoji":"🧭","tokens_out":898,"duration_ms":8440,"temperature":0.7,"pith_summary":"Real robot maps are built from voxel labels that are often wrong: category flips from sparse or misaligned sensors, and ghost trails left by moving cars and people. This paper asks whether autonomous systems can still trust that supervision, and shows that the usual 2D noisy-label tricks fail badly once the data become sparse 3D grids. It releases OccNL, a benchmark that injects both synthetic occupancy-asymmetric flips and real dynamic trailing artifacts into SemanticKITTI, then measures how far models collapse. The authors' answer is DPR-Occ: after a short warm-up that lets the network memorize clean patterns, it builds a small candidate label set for every voxel by uniting the top-K predictions of a slow-moving teacher with the top-K nearest class prototypes in feature space, then trains with partial-label, negative, and not-true distillation losses. On the hardest 90% noise setting the method still reports roughly 35% geometric IoU and 8% mIoU, gains of up to 13.91 IoU and 2.57 mIoU over five adapted 2D baselines that largely lose structure and rare classes.","feed_headline":"90% wrong voxel labels, maps still usable","feed_subtitle":"Dual-source partial labels keep geometry and rare classes when 2D noise tricks collapse in 3D.","key_machinery":"DPR-Occ dual-source partial label: for each voxel the candidate set is the union of the Top-K EMA-teacher predictions and the Top-K feature-prototype similarities (Eq. 4), then optimized by partial-label learning, negative learning on the complement, and EMA-guided self-not-true distillation under a decaying K schedule.","core_discovery":"State-of-the-art 2D label-noise methods collapse in sparse 3D occupancy spaces, yet a dual-source partial-label construction that fuses temporal model memory with prototype affinity can preserve both geometric completeness and sparse semantic classes even under 90% occupancy-asymmetric corruption and real dynamic trailing noise.","pith_inferences":["The same dual-source candidate idea should transfer to other sparse voxel tasks such as LiDAR semantic segmentation under map-projection noise.","If the warm-up window is too short for a new backbone, the method may need an automatic early-stop on teacher-student agreement rather than a fixed epoch count.","Closed-loop driving simulators could inject OccNL-style trailing noise to quantify how much map corruption a planner can absorb before collision rates rise.","Long-tailed open-vocabulary occupancy may benefit from the same prototype-affinity term once language embeddings replace fixed class prototypes."],"forward_implications":["OccNL becomes a required stress test for any new 3D occupancy network that claims robustness.","Training pipelines can deliberately tolerate cheaper, noisier multi-frame labels instead of demanding perfectly cleaned voxels.","Sparse dynamic classes (bicyclists, motorcyclists) remain learnable under extreme corruption when candidates are structurally constrained.","Safety-critical planners receive maps that keep road and free-space geometry even when annotation quality is catastrophic.","Hypothesis-space regularization, not stronger loss reweighting, is the more transferable principle from 2D to sparse 3D."],"fun_headline_variants":["90% voxel noise no match for dual-source partial labels","2D label noise collapses in 3D voxels, DPR-Occ holds firm","Temporal memory and affinity save occupancy maps under 90% noise","Sparse 3D spaces break 2D noise methods but not dual partial labels","Usable 3D maps persist despite extreme occupancy label corruption"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The dual-source candidate set, after warm-up, still contains the true class often enough for partial-label learning to recover it, even when 90% of labels are flipped and the class distribution is driven toward uniformity.","fun_headline_variants_meta":{"raw":{"variants":["90% voxel noise no match for dual-source partial labels","2D label noise collapses in 3D voxels, DPR-Occ holds firm","Temporal memory and affinity save occupancy maps under 90% noise","Sparse 3D spaces break 2D noise methods but not dual partial labels","Usable 3D maps persist despite extreme occupancy label corruption"]},"model":"grok-4.5","effort":"low","cost_usd":0.008678,"raw_usage":{"total_tokens":2033,"prompt_tokens":832,"num_sources_used":0,"completion_tokens":77,"cost_in_usd_ticks":86780000,"prompt_tokens_details":{"text_tokens":832,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1124,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":832,"tokens_out":77,"duration_ms":9341,"temperature":1.0,"reasoning_tokens":1124,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-15T13:53:45.912191+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On the 90% asymmetric-noise OccNL split, ablate the prototype branch or force a fixed tiny K so ground-truth coverage of sparse dynamic classes falls; if geometric IoU and rare-class mIoU then collapse to the level of the pure 2D baselines, the recovery claim fails.","supporting_citations":[],"review_version":1}