{"id":"4a522457-d5a2-45ee-a274-0479a5e4c7d1","arxiv_id":"2607.06403","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":7,"one_line_summary":"LingBot-VLA 2.0 combines 60k hours of multi-embodiment pretraining data, an expanded whole-body action space, and dual-query distillation from depth and video teachers to improve VLA performance on GM-100 and long-horizon mobile manipulation tasks.","lead":"LingBot-VLA 2.0 is a robot control model trained on 60,000 hours of multi-embodiment data that adds whole-body control (head, waist, mobile base, dexterous hands) and predictive future-scene modeling. It outperforms prior models on manipulation benchmarks, but the paper does not isolate which of its three main improvements drives the gains. A smart generalist might read it to understand how VLA systems are scaling toward real-world deployment.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"No contribution-level ablation isolates the three headline modifications (data scale, expanded action space, predictive dynamics), so the causal attribution is unsupported; per-task variance and missing statistical details compound the uncertainty.","rationale":"The reader correctly identified the most load-bearing concern: the absence of a contribution-level ablation for the three headline modifications. This is the central weakness because the paper's claim is specifically about three modifications working 'in concert,' yet provides no evidence that all three contribute. The reader also correctly noted the uncontrolled comparison with π0.5 (training data volume not matched), missing error bars/trial counts for GM-100, and unreported hyperparameters (K, Nr, λ, γ). I add the observation that per-task results in Table 5 show high variance — LingBot-VLA-2.0 loses to π0.5 on 4 of 9 Agilex tasks and 5 of 9 Galaxea tasks — making the 'best overall' claim sensitive to which tasks are included and underscoring the need for statistical reporting. The paper has genuine strengths: real-robot multi-platform evaluation, a thorough data pipeline, insightful micro-level ablations in §6.1, and the MoE scaling comparison in Fig. 7. The DINO-Video teacher is validated on LARYBench (Tab. 3). But without isolating the three headline contributions, the causal claim remains unsupported. CONDITIONAL is the right verdict: if the authors add a contribution-level ablation with statistical details, this would move toward ACCEPT; as-is, the engineering contribution is clear but the scientific claim is not established.","tokens_in":21760,"tokens_out":1776,"duration_ms":159353,"concrete_test":"Train four model variants on the same compute budget: (A) LingBot-VLA-1.0 baseline, (B) A + expanded 60k-hour data only, (C) B + whole-body action space, (D) C + dual-query distillation. Evaluate all four on the same 9 GM-100 tasks with ≥15 trials each and report per-task progress/success with confidence intervals. If variant B already captures most of the gain over LingBot-VLA-1.0 (e.g., within 2 points of D's overall average), the architectural contributions (action space + predictive dynamics) are not load-bearing and the headline claim of three coordinated modifications is unsupported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is that three coordinated modifications — (1) 60k-hour multi-embodiment pretraining data, (2) whole-body action space expansion, and (3) dual-query predictive dynamics distillation — jointly produce the best GM-100 and long-horizon results. However, no ablation isolates these three contributions. The ablations in §6 only examine micro-level action representation choices (relative vs absolute, EEF vs joint, normalization, loss function) on 4 tasks — these are orthogonal to the three headline claims. The paper never runs a baseline → +data → +action-space → +predictive-dynamics staircase, so we cannot determine whether all three contribute, one dominates, or the gains are primarily from data scale (60k hours vs LingBot-VLA-1.0's undisclosed amount and vs π0.5's undisclosed amount). This matters because Table 5 shows high per-task variance: on Agilex, LingBot-VLA-2.0 dramatically beats π0.5 on Retrieve keychain (100/100 vs 20/20) but loses on Block sorting (56.8/0.0 vs 90.4/60.0), Sort snacks (66.2/10.0 vs 82.4/30.0), Pack eggs (44.4/0.0 vs 72.4/20.0), and Tool packing (60/0 vs 66/30). The overall average of 66.2/34.4 is pulled up by a few outlier tasks, and without error bars or trial counts for GM-100 (only the long-horizon tasks specify 15 trials), the statistical robustness of the headline comparison is unclear. The MoE vs Dense comparison in Fig. 7 and the action-representation ablations in §6.1 demonstrate the authors can do rigorous ablation work, making the absence of headline-contribution ablations conspicuous.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"The paper presents LingBot-VLA 2.0, an improved vision-language-action (VLA) model targeting the gap between laboratory benchmarks and real-world robotic deployment. The system advances three areas: (1) a curated 60,000-hour multi-embodiment pretraining dataset (50k robot + 10k egocentric), (2) an expanded 55-dimensional action space supporting whole-body DoF (head, waist, mobile base, dexterous hands), and (3) a dual-query distillation framework using LingBot-Depth and DINO-Video as teachers for predictive dynamics modeling. The architecture uses token-level sparse MoE layers in the action expert with auxiliary-loss-free load balancing. Evaluation is conducted on nine GM-100 bimanual tasks under a generalist setting on two platforms (Agilex Cobot Magic, Galaxea R1 Pro), plus two long-horizon mobile manipulation tasks. The authors report improvements over LingBot-VLA-1.0 and π0.5 on aggregate metrics.","tokens_in":22825,"tokens_out":1505,"duration_ms":285149,"significance":"The paper makes a substantial engineering contribution to the VLA field. The 60k-hour curated dataset across 20 embodiments is a significant resource, and the unified 55-D action representation is a clean design for heterogeneous platforms. The MoE scaling experiment (Fig. 7) with matched active parameters provides a fair comparison. The dual-query distillation framework is well-motivated, and the DINO-Video teacher is benchmarked on LARYBench (Tab. 3). The release of checkpoints and code is commendable and supports reproducibility. The long-horizon mobile manipulation results with 15-trial evaluation and ID/OOD settings add practical value. However, the significance of the three headline contributions is undermined by the absence of contribution-level ablations, as detailed below.","major_comments":[{"comment":"§5.2, Table 5: The central claim that three coordinated modifications (data scale, expanded action space, predictive dynamics) jointly improve performance is not supported by a contribution-level ablation. The ablations in §6.1 only examine micro-level action representation choices (relative vs absolute, EEF vs joint, normalization, loss) on four tasks. No baseline → +data → +action-space → +predictive-dynamics staircase is run. Without isolating these three contributions, the causal attribution to all three modifications working in concert is unsupported. This is the paper's central thesis.","section":null},{"comment":"§5.2, Table 5: The headline comparison with π0.5 is not controlled for training data volume. LingBot-VLA-2.0 uses 60k hours; π0.5's data volume is undisclosed. The aggregate improvements (66.2/34.4 vs 59.1/32.2 on Agilex) could largely reflect data scale rather than architectural innovation. A data-matched comparison or at minimum explicit discussion of this confound is needed to support the claim that the three modifications are responsible for the gains.","section":null},{"comment":"§5.2, Table 5: Per-task variance is high and the aggregate average may be misleading. On Agilex, LingBot-VLA-2.0 dramatically beats π0.5 on Retrieve keychain (100/100 vs 20/20) but loses on Block sorting (56.8/0.0 vs 90.4/60.0), Sort snacks (66.2/10.0 vs 82.4/30.0), Pack eggs (44.4/0.0 vs 72.4/20.0), and Tool packing (60/0 vs 66/30). No error bars or trial counts are reported for GM-100 (only long-horizon tasks specify 15 trials). The statistical robustness of the headline comparison is unclear without these details.","section":null},{"comment":"§4.2, Eqs. (9)–(10): The dual-query distillation uses LingBot-Depth [28] and DINO-Video as teachers, both of which are co-authored by the same team. While this is disclosed, the paper does not provide an ablation isolating the contribution of predictive dynamics distillation to the final action performance. Fig. 13 shows perceptual quality of the distilled queries, but no experiment links this to improved manipulation success. A simple ablation (with vs without distillation) on a few GM-100 tasks would substantially strengthen the claim.","section":null}],"minor_comments":[{"comment":"§3.1.1: The data filtering thresholds (jerk Z-score, static proportion >95%, valid-frame ratio <20%) are described qualitatively but not specified numerically. Stating the actual threshold values would aid reproducibility.","section":null},{"comment":"§4.1: The MoE hyperparameters (K, Nr, λ, γ) are mentioned in the equations but their specific values for the final model are not stated in the text. Please specify.","section":null},{"comment":"Table 1: The 'Body DoF' column includes arm DoF in some rows but not others, and the relationship between 'Body DoF' and 'Total DoF' is not always consistent (e.g., AgiBot G1: Body DoF 4, Total 20, but Arm DoF 14 + Body 4 = 18, not 20). Clarifying the definitions would help.","section":null},{"comment":"§3.1.3: The 55-D action vector includes 4 reserved dimensions. It would help to state what these are reserved for.","section":null},{"comment":"Table 5: The 'Overall average' row appears to average 9 tasks, but 10 task rows are listed for Agilex (BM-19 through BM-107). Please clarify which tasks are averaged and why one is excluded.","section":null},{"comment":"§5.3, Table 6: The long-horizon comparison only includes π0.5, not LingBot-VLA-1.0. Adding the v1.0 comparison or explaining its absence would strengthen the comparison.","section":null},{"comment":"Fig. 7: The x-axis label 'Training step (k)' is ambiguous — are these optimizer steps or epochs? Specifying would help interpretation.","section":null},{"comment":"The paper mentions 'Qwen3.6-27B' [25] for annotation but the reference appears to be a future-dated preprint. Confirming the correct model version and citation would be appropriate.","section":null}],"recommendation":"major_revision","confidential_remarks":"The self-citation pattern is worth noting: LingBot-Depth [28], the dual-query distillation framework [35], and DINO-Video are all co-authored by the same team. This is disclosed but the interdependence means the novelty of the distillation component is partially internal. The contribution-level ablation gap is the most serious concern — the paper makes three headline claims but only ablates micro-level action representation choices. A staged ablation isolating the three contributions is essential for the causal claim to be credible. The per-task variance in Table 5 is also concerning: the aggregate average is pulled up by a few outlier tasks (especially Retrieve keychain at 100/100), while several tasks show regressions vs π0.5. Without error bars or trial counts for GM-100, the statistical significance is unclear."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a thorough and constructive report. The referee acknowledges the engineering contributions (dataset, unified action space, MoE scaling, dual-query distillation, reproducible release) but identifies a central concern: the absence of contribution-level ablations isolating the three headline modifications (data scale, expanded action space, predictive dynamics). Additional concerns address the uncontrolled data-volume confound in the π0.5 comparison, missing trial counts and error bars for GM-100, and the lack of a distillation ablation linking perceptual quality to manipulation success. We agree that contribution-level ablations and statistical reporting are needed and will add them. On the data-volume confound, we provide honest discussion of what can and cannot be controlled given π0.5's undisclosed training data.","responses":[{"response":"The referee is correct that the current manuscript does not include a contribution-level staircase ablation isolating the three headline modifications. The ablations in §6.1 address micro-level action representation choices, which are orthogonal to the three functional-domain contributions. We agree this is a gap relative to the paper's central thesis. In the revision, we will add a contribution-level ablation on a representative subset of GM-100 tasks, incrementally adding: (1) expanded pretraining data (from LingBot-VLA-1.0's data to the 60k-hour corpus), (2) the expanded 55-D action space, and (3) dual-query predictive dynamics distillation. We will report both progress score and success rate at each stage. We acknowledge that a fully controlled staircase on all nine GM-100 tasks across both platforms may not be feasible within the revision timeline due to compute and robot evaluation constraints, so we will be explicit about the subset and platform used. We will also temper the language in the abstract and conclusion to accurately reflect what the ablation does and does not prove about the joint contribution of all three modifications.","revision_made":"yes","referee_comment":"§5.2, Table 5: The central claim that three coordinated modifications (data scale, expanded action space, predictive dynamics) jointly improve performance is not supported by a contribution-level ablation. The ablations in §6.1 only examine micro-level action representation choices (relative vs absolute, EEF vs joint, normalization, loss) on four tasks. No baseline → +data → +action-space → +predictive-dynamics staircase is run. Without isolating these three contributions, the causal attribution to all three modifications working in concert is unsupported. This is the paper's central thesis."},{"response":"This is a fair concern. π0.5's training data volume and composition are not publicly disclosed, so we cannot run a data-matched comparison against π0.5 directly. We will add an explicit discussion of this confound in the revision, acknowledging that the aggregate improvements over π0.5 cannot be cleanly attributed to architectural innovation versus data scale. To partially address this, the contribution-level ablation described in our response to the first comment will include a LingBot-VLA-1.0 → +data-only condition, which isolates the effect of scaling from 1.0's data to 60k hours without the other two modifications. This provides an internal data-scale control, even if an external data-matched comparison with π0.5 is not possible. We will also revise the language in §5.2 to avoid implying that the gains over π0.5 are solely due to the three architectural modifications.","revision_made":"partial","referee_comment":"§5.2, Table 5: The headline comparison with π0.5 is not controlled for training data volume. LingBot-VLA-2.0 uses 60k hours; π0.5's data volume is undisclosed. The aggregate improvements (66.2/34.4 vs 59.1/32.2 on Agilex) could largely reflect data scale rather than architectural innovation. A data-matched comparison or at minimum explicit discussion of this confound is needed to support the claim that the three modifications are responsible for the gains."},{"response":"The referee is correct that GM-100 trial counts and error bars are not reported in the current manuscript, and that per-task variance is substantial. We will add trial counts for all GM-100 evaluations in the revised Table 5. For error bars, we will report standard deviations or confidence intervals where the evaluation protocol supports multiple independent trials. We acknowledge that some GM-100 tasks may have been evaluated with a limited number of trials due to robot availability and time constraints; where this is the case, we will state the trial count explicitly and note the resulting statistical limitations. We will also add a brief discussion of the per-task variance pattern the referee identifies, including the observation that LingBot-VLA-2.0 underperforms π0.5 on several tasks (Block sorting, Sort snacks, Pack eggs, Tool packing on Agilex), to provide a more balanced presentation rather than relying solely on aggregate averages.","revision_made":"yes","referee_comment":"§5.2, Table 5: Per-task variance is high and the aggregate average may be misleading. On Agilex, LingBot-VLA-2.0 dramatically beats π0.5 on Retrieve keychain (100/100 vs 20/20) but loses on Block sorting (56.8/0.0 vs 90.4/60.0), Sort snacks (66.2/10.0 vs 82.4/30.0), Pack eggs (44.4/0.0 vs 72.4/20.0), and Tool packing (60/0 vs 66/30). No error bars or trial counts are reported for GM-100 (only long-horizon tasks specify 15 trials). The statistical robustness of the headline comparison is unclear without these details."},{"response":"We agree that linking the perceptual quality of distilled queries to downstream manipulation success is important and currently missing. The contribution-level ablation we propose in response to the first comment will include a with-distillation vs without-distillation comparison on a subset of GM-100 tasks, which directly addresses this concern. We will report both manipulation metrics (progress score, success rate) and, where feasible, perceptual metrics for the distilled queries. Regarding the use of co-authored teacher models: we chose LingBot-Depth and DINO-Video because they are specifically designed for robotics-relevant geometric and temporal supervision, and DINO-Video is benchmarked on LARYBench (Table 3) to demonstrate its standalone quality. However, we acknowledge that using independent teacher models would strengthen the claim that the distillation framework generalizes beyond our own models. In the revision, we will add a discussion of this limitation and, if compute permits, include a comparison using a non-co-authored depth teacher (e.g., Depth Anything) as an additional ablation point.","revision_made":"partial","referee_comment":"§4.2, Eqs. (9)–(10): The dual-query distillation uses LingBot-Depth [28] and DINO-Video as teachers, both of which are co-authored by the same team. While this is disclosed, the paper does not provide an ablation isolating the contribution of predictive dynamics distillation to the final action performance. Fig. 13 shows perceptual quality of the distilled queries, but no experiment links this to improved manipulation success. A simple ablation (with vs without distillation) on a few GM-100 tasks would substantially strengthen the claim."}],"tokens_in":21700,"tokens_out":1687,"duration_ms":236691,"standing_objections":["The data-volume confound with π0.5 cannot be fully resolved because π0.5's training data volume and composition are not publicly disclosed. We can only provide internal data-scale controls (via the LingBot-VLA-1.0 → +data-only ablation) and explicit discussion of the limitation; a true data-matched external comparison is not possible without cooperation from the π0.5 authors.","A fully controlled contribution-level staircase ablation on all nine GM-100 tasks across both platforms may not be feasible within the revision timeline due to compute and robot evaluation constraints. We will conduct the ablation on a representative subset and be transparent about this scope limitation."]},"desk_editor":{"model":"glm-5.2","letter":"Here's the bottom line: LingBot-VLA 2.0 is a serious engineering effort — 60k hours of curated multi-embodiment data, whole-body action support, dual-query distillation from depth and video teachers, and real-robot results on GM-100 and long-horizon mobile manipulation across two platforms. The system works. The question is whether the paper explains *why* it works, and there the evidence is thin.","headline":"A well-engineered VLA system with real-robot results, but the three headline contributions are never ablated against each other — the causal story is unsupported.","tokens_in":22993,"tokens_out":163,"would_cite":false,"duration_ms":92149,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Robot system handles whole-body tasks across 20 platforms","keywords":["vision-language-action","robot manipulation","mixture of experts","dual-query distillation","cross-embodiment generalization","mobile manipulation","whole-body control","VLA foundation model"],"falsifier":"If a baseline VLA trained on a comparable volume of multi-embodiment data with standard dual-arm action spaces but without the dual-query distillation or MoE architecture achieved similar GM-100 and long-horizon scores, the three headline contributions would lose their causal support.","tokens_in":21761,"feed_emoji":"🤖","tokens_out":841,"duration_ms":139917,"temperature":0.7,"pith_summary":"The paper presents LingBot-VLA 2.0, a vision-language-action (VLA) model for robotic control that argues practical deployment requires three coordinated improvements over lab-grade systems: broader multi-embodiment pretraining data (60,000 hours spanning 20 robot configurations and egocentric human video), an expanded action space covering whole-body degrees of freedom (head, waist, mobile base, dexterous hands beyond standard dual-arm), and predictive dynamics modeling via a dual-query distillation scheme where a depth teacher provides geometric supervision and a video teacher provides temporal supervision. The central claim is that these three modifications jointly bridge the lab-to-deployment gap. The paper evaluates on the GM-100 bimanual benchmark and two long-horizon mobile manipulation tasks, reporting best-overall performance against π0.5 and the prior LingBot-VLA-1.0, with gains most pronounced on tasks requiring object grounding and future-aware planning. The architecture uses token-level sparse mixture-of-experts layers in the action expert with auxiliary-loss-free load balancing, and the dual-query distillation appends learnable current and future queries to the VLM's token sequence, training them to predict depth and video features at both timesteps.","feed_headline":"Whole-body robot VLA handles 20 platforms, beats π0.5","feed_subtitle":"LingBot-VLA 2.0 pairs 60k hours of multi-embodiment data with predictive dynamics modeling to close the lab-to-deployment gap.","key_machinery":"LingBot-VLA 2.0","core_discovery":"The paper's central object is a VLA system that unifies three deployment-oriented modifications: a 55-dimensional canonical action vector mapping heterogeneous embodiment controls (arms, end-effectors, grippers, hands, waist, head, mobility) into a shared representation; a token-level sparse MoE action expert with sigmoid routing and bias-based load balancing that decouples expert assignment from expert weighting; and a dual-query distillation framework where current and future query tokens are supervised by a depth model (LingBot-Depth) for geometric structure and a causal video model (DINO-Video, built on DINOv3 with causal temporal attention and 3D rotary embeddings) for temporal dynamics","pith_inferences":["The paper does not isolate the three headline contributions (data scale, action space expansion, predictive dynamics) against each other in a controlled ablation, so the relative contribution of each remains unclear from the evidence presented.","The comparison with π0.5 is not controlled for training data volume (60k hours vs. undisclosed), so gains could partly reflect data scale rather than architectural choices.","The action-space ablations (relative vs absolute, EEF vs joint, normalization, loss) examine micro-level representation choices on four tasks but do not test the three headline modifications.","If the predictive dynamics mechanism generalizes, it could be tested as a plug-in module on other VLA architectures to see whether future-query distillation consistently improves contact-rich or long-horizon tasks."],"forward_implications":[],"fun_headline_variants":["LingBot-VLA 2.0 unifies whole-body control across 20 robot platforms","Sparse MoE action expert handles 55-DoF whole-body manipulation in VLA models","Dual-query distillation from depth and video models improves VLA temporal reasoning","60k hours of multi-embodiment data bridges lab-to-deployment gap for VLA robots","Token-level MoE with sigmoid routing decouples expert assignment from weighting"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The paper assumes the reported performance gains stem from its three proposed modifications working in concert, but it never runs an ablation isolating data scale, action-space expansion, and predictive dynamics against data-matched baselines.","fun_headline_variants_meta":{"raw":{"variants":["LingBot-VLA 2.0 unifies whole-body control across 20 robot platforms","Sparse MoE action expert handles 55-DoF whole-body manipulation in VLA models","Dual-query distillation from depth and video models improves VLA temporal reasoning","60k hours of multi-embodiment data bridges lab-to-deployment gap for VLA robots","Token-level MoE with sigmoid routing decouples expert assignment from weighting"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":739,"prompt_tokens":630,"completion_tokens":109,"prompt_tokens_details":null},"tokens_in":630,"tokens_out":109,"duration_ms":57625,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T06:48:01.650351+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If a baseline VLA trained on a comparable volume of multi-embodiment data with standard dual-arm action spaces but without the dual-query distillation or MoE architecture achieved similar GM-100 and long-horizon scores, the three headline contributions would lose their causal support.","supporting_citations":[],"review_version":1}