{"id":"1b7e67b4-2dd7-4c4e-9362-336d23d08652","arxiv_id":"2512.05693","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A hierarchical Mixture-of-Experts action module for VLA policies turns negative transfer from heterogeneous robot data into positive transfer and achieves state-of-the-art results on CALVIN, LIBERO, and real xArm7/ALOHA robots.","lead":"This paper introduces HiMoE-VLA, a robot-control model whose action module uses a hierarchical mixture-of-experts to specialize on different robot action spaces while sharing general knowledge. It reports top results on standard benchmarks and real robots, and shows that training on heterogeneous action data can improve rather than hurt performance.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 7's 'positive transfer' claim is one-sided: co-training gains on CALVIN-D are reported without the corresponding ABC-side success or variance, so the headline conclusion is not yet established.","rationale":"The reader's weakest assumption points to the fixed 24D unified vector representation. That is a real limitation, but within the evaluated robots—xArm, ALOHA, CALVIN, LIBERO, and the chosen OXE subset—the representation appears adequate, so it does not directly threaten the reported central experiment. The more load-bearing gap is in Table 7 itself: the evidence for 'avoids negative transfer and improves' depends on comparing single-domain training against co-training, but the table never states what is being evaluated or whether ABC performance was preserved. Without the ABC-side numbers, the headline claim of positive transfer is not demonstrable from the paper as written. The reader's rationale already notes missing error bars and small sample sizes, which compounds the issue: the reported gain is modest and may be within noise. I am not arguing the result is false; the architecture and ablations are suggestive, and the code is public. The concern is that the central quantitative evidence is incomplete as presented, and a targeted experiment can settle it. This keeps the verdict at CONDITIONAL: the claim is plausible but requires the missing comparison and variance before it can be accepted as demonstrated.","tokens_in":20002,"tokens_out":11465,"duration_ms":123515,"concrete_test":"Using the released code, run 5 seeds of each configuration: (1) Full HiMoE trained on CALVIN-D only, (2) Full HiMoE trained on CALVIN-ABC only, (3) Full HiMoE co-trained on ABC+D. Evaluate each seed on both ABC and D separately, and report per-domain CALVIN sums plus the average across domains, with confidence intervals. If co-trained ABC performance is not within noise of the ABC-only baseline, or if the average over both domains does not improve over the single-domain models, Table 7's positive-transfer conclusion should be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that HiMoE converts negative transfer into positive transfer—rests on Table 7, which compares a D(Joint)-only training run with an ABC(EEF)+D(Joint) co-training run. The table does not state the evaluation domain. If the reported sums are CALVIN-D only, the +0.186 gain for Full HiMoE shows only that extra ABC data helps D; it does not demonstrate positive transfer on ABC. A model can improve on D while degrading on ABC, which would falsify the claim as stated. If the sums are intended to be aggregate across ABC and D, then the 'D (Joint)' column is not a matched baseline, since the co-trained column covers two environments and the single-domain column covers one. In either reading, the paper omits the ABC-only baseline and the co-trained ABC success. The comparison also appears single-seed; the Full gain (3.826→4.012) is small relative to run-to-run variation typical of CALVIN. The architecture is plausible and the code is public, but Table 7 is the direct evidence for the paper's strongest sentence and it is currently under-specified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HiMoE-VLA, a 4B vision-language-action model whose action module is organized as a hierarchy of mixture-of-experts layers: Action-Space MoE layers at the boundaries specialize computation for different action spaces (e.g., joint-angle vs. end-effector control), Heterogeneity-Balancing MoE layers in adjacent layers handle broader embodiment/sensor/scene variability, and dense Transformer blocks in the middle integrate shared representations. Two auxiliary losses are introduced: a contrastive Action-Space Regularization and a load-balancing Heterogeneity-Balancing Regularization. The model is pretrained on OXE and Aloha data, then fine-tuned and evaluated on CALVIN, LIBERO, and real xArm7 and ALOHA robots; the paper reports state-of-the-art results on these benchmarks and claims that, under heterogeneous co-training, HiMoE turns negative transfer into positive transfer. Code and models are publicly released.","tokens_in":20265,"tokens_out":7245,"duration_ms":71536,"significance":"If the reported results are correct, the paper makes a practically valuable contribution: it offers an architectural treatment of action-space and embodiment heterogeneity in VLAs, with extensive benchmark coverage, public code and models, and ablations of the main design choices. The layered AS-MoE/HB-MoE idea is clean and the routing visualizations are a useful sanity check. However, the paper's strongest conceptual claim, that heterogeneous co-training becomes positively transferable, rests on an under-specified comparison in Table 7, and the empirical sections generally omit uncertainty estimates. As presented, the evidence is suggestive but not yet sufficient for the strong conclusions drawn in the abstract and Section 4.3.","major_comments":[{"comment":"The headline claim that the full HiMoE model 'not only avoids negative transfer but also improves when trained with heterogeneous data' is not established by Table 7 as presented. The columns 'D (Joint)' and 'ABC (EEF) + D (Joint)' do not state the evaluation domain. If the reported sums are CALVIN-D success (the CALVIN metric from Table 1), then the +0.186 gain for Full HiMoE (3.826 → 4.012) only shows that adding ABC data helps performance on D; there is no ABC-only baseline and no co-trained ABC success rate, so the claim of positive transfer on the ABC side is unverified. If the sums are intended as an aggregate over ABC and D, then the D-only baseline is not matched to the two-environment co-trained column. Please report ABC-side success for D-only, ABC-only, and co-trained models, and include multiple seeds.","section":"§4.3, Table 7"},{"comment":"No variance, confidence intervals, or significance tests are reported anywhere. Real-world evaluations use 15–18 trials per condition (Appendix A.2), and several headline differences are comparable to plausible binomial noise, e.g., Table 4 overall 63.7% vs. 54.2% for π0, or Table 5 single-arm Sum. 67.6% vs. 55.9% for π0. Simulation results also appear to be single-seed. Adding standard errors, per-seed results, or a significance statement is necessary before 'consistently surpasses' and 'performance boost' can be accepted as quantitative claims.","section":"§4.2–§4.3, Tables 3–9"},{"comment":"The final Mixture-of-Experts configuration is ambiguous. Section 4 states N=32 experts with top-K=4 and reports that 'very high routing widths (K=8) cause instability.' Appendix B, however, states 'We set the number of experts to N=32 with a top-k routing of 8,' and refers to a table that shows this configuration outperforms alternatives. Table 9, the actual N/K ablation, contains no K=8 row at all. Since N and K are core architectural hyperparameters and the reported SOTA results depend on this choice, the paper must state the exact configuration used for the main results and make the ablation table consistent.","section":"§4 Implementation Details vs. Appendix B and Table 9"},{"comment":"The pretraining data sizes are internally inconsistent. Section 4 reports OXE 22.5M frames + Aloha 1.6M frames = 24.1M frames, while Appendix A.1 states that the Aloha datasets contribute 36.28M frames. Table 10 lists the Aloha mixture ratio as 10.4%, which matches neither 1.6/24.1 ≈ 6.6% nor 36.28/(22.5+36.28) ≈ 61.7%. These numbers must be corrected because the pretraining mixture is a central part of the experimental setup and the current text is not reproducible.","section":"§4 Pre-training Dataset; Appendix A.1"}],"minor_comments":[{"comment":"The CALVIN headline number is reported as 3.98 in the abstract, 3.94 in the text, 3.967 in Table 1, and 4.012 in Table 6(b). Please align these numbers and clarify which evaluation condition each refers to.","section":"Abstract; §4.1; Table 1; Table 6(b)"},{"comment":"The AS-Reg objective is under-specified: 'pairs of experts (i,j) assigned to the same action-space token' needs a formal definition of how i and j are selected for each token and how 'assigned' relates to the routing decisions. As written, the positive-pair selection is ambiguous.","section":"§3.3, Eq. (4)"},{"comment":"The text says that removing pretrained weights (w/o pretrain) leads to a 'more notable drop' than removing MoE re-initialization (w/o init), but the sums are 3.826 and 3.827, respectively. Either report the per-length numbers that support the qualitative claim or rephrase.","section":"§4.3, Table 6(a) text"},{"comment":"The sentence 'As shown in Table 6, this configuration consistently outperforms...' appears to refer to Table 9, since Table 6 is the ablation of initialization and heterogeneity-handling methods, not the N/K sweep.","section":"Appendix B"},{"comment":"Figure 1 labels the end-effector action space as '6D EEF Poses,' while the text and Appendix A.2 state an 8-dimensional end-effector vector (6D pose plus presumably gripper dimensions). Please make the notation consistent.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"I agree with the reader's main concern: Table 7 is the only direct evidence for the paper's most important claim, and it is currently under-specified. The paper is otherwise a solid empirical study with public code, but the transfer claim needs to be re-run or re-scoped with matched baselines, per-domain results, and variance information. The inconsistencies in N/K and pretraining frame counts should be treated as blocking for reproducibility, not merely cosmetic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper is better than the usual VLA architecture salad. The two-level hierarchy—Action-Space MoE at the boundaries, Heterogeneity-Balancing MoE next to them, dense transformer in the middle—is a real design idea, and the two regularizers (contrastive for action-space specialization, load-balancing for the deeper experts) are not just tacked on; the ablations show each piece earns its keep. The gains on CALVIN (3.97 vs 3.76 for π0) and LIBERO (97.8 vs 94.2 for π0) are solid, and the real-world results, while small, are consistent. Code and models are public, which makes this reproducible. I would not desk-reject this.\n\nThe soft spots are real but mostly addressable. The biggest one is Table 7, the direct evidence for the headline claim that HiMoE turns negative transfer into positive transfer. The table compares training on CALVIN-D alone with co-training on ABC+ D, but it only shows success on the D side. We don't see the ABC-side success or an ABC-only baseline. So the +0.186 for the full model shows that adding ABC data helps D—that is positive transfer in one direction—but it does not show the model is a better generalist on both action spaces. The stress-test note is right: the claim as written is one-sided. Also, there are no error bars or significance tests anywhere, and the real-world trials are 15–18 per condition. That is small but not fatal for a robotics paper; still, it needs to be stated.\n\nThere are also two numerical inconsistencies that look sloppy rather than malicious: the main text says ALOHA data is 1.6M frames while Appendix A.1 says 36.28M (a 20x difference), and Appendix B says top-k=8 while the main text and Table 9 say top-k=4, with Table 9 showing K=8 as unstable. The uncited \"GR00T-Like\" baseline should be either cited or renamed. The 24-dimensional fixed action vector with zero-padding is a pragmatic choice, but it does limit the claim of generality to robots that fit in that layout—worth a sentence in the limitations.\n\nNet: the architecture is plausible, the empirical support is decent but not as tight as the abstract implies, and the missing numbers are easy to supply. This deserves a serious referee. I'd send it out with a request for variance reporting, the ABC-side of Table 7, and a cleanup of the inconsistencies.\n\nWould I bring it to reading group? Yes, mainly to argue about whether the MoE hierarchy is the reason for the gains or just extra capacity. And I'd cite it in my own work if I were building on heterogeneous action spaces.","headline":"Genuinely new two-level MoE action module that moves CALVIN/LIBERO numbers and earns a serious referee, but the central positive-transfer claim is under-specified and the paper has a few fixable reporting inconsistencies.","tokens_in":20801,"tokens_out":3217,"would_cite":true,"duration_ms":33839,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a hierarchical mixture-of-experts action module, with action-space experts at the input/output boundaries and balancing experts at adjacent layers, converts the negative transfer caused by heterogeneous robot data into","keywords":["hierarchical mixture-of-experts","vision-language-action policy","heterogeneous robot data","negative transfer","action-space specialization","flow matching","load balancing","generalist manipulation"],"falsifier":"Co-train HiMoE-VLA on a robot with an action space that exceeds 24 dimensions—for example a 7-DoF arm with a multi-finger hand, where joint commands need more than 16 entries—and compare against a baseline with a dedicated head for that robot. If the unified-vector version does not beat the dedicated head (or at least match the positive-transfer pattern reported for 24-dim spaces), the load-bearing premise is the fixed representation, not the MoE hierarchy.","tokens_in":19873,"feed_emoji":"🤖","tokens_out":5473,"duration_ms":51374,"temperature":0.7,"pith_summary":"Most robot policies trained on large mixed datasets suffer when different robots use different action spaces: joint-angle commands versus end-effector commands, single-arm versus dual-arm. This paper argues that the conflict is architectural, not inherent, and introduces HiMoE-VLA, whose action module places specialized experts at the boundaries to absorb action-space differences, balancing experts next to them to absorb broader embodiment and observation variation, and dense transformer blocks in the middle to build shared representations. Two auxiliary losses—a contrastive action-space regularization and a load-balancing regularization—push the experts toward specialization and balanced use. On simulation benchmarks and real single-arm and dual-arm robots, the model reports the highest success rates among the compared methods; in a controlled from-scratch co-training test, it turns the negative transfer seen in baselines into a positive gain. The paper's conclusion is that heterogeneous robotic data can be unified into one policy if heterogeneity is handled at the right layers.","feed_headline":"Hierarchical experts turn mixed robot data into a gain","feed_subtitle":"Specialized edge experts plus shared middle layers let one policy co-train joint and end-effector control.","key_machinery":"Hierarchical Mixture-of-Experts (HiMoE) action module: a layered action network in which boundary layers use Action-Space MoE to specialize per action space, adjacent layers use Heterogeneity-Balancing MoE to spread broader variation across experts, and middle dense transformer blocks consolidate shared representations. Routing uses top-4 selection among 32 experts, with a contrastive Action-Space Regularization aligning experts routed to the same action-space token and a load-balancing regularization aligning expected gating scores with realized routing frequencies. Training uses a flow-matching objective, and all states and actions are projected into a fixed 24-dimensional vector with vali","core_discovery":"The central claim is that the main obstacle to training one vision-language-action policy on diverse robot data is not the data itself but the way a shared dense action module mixes it. When action spaces differ—joint angles versus end-effector deltas, single-arm versus dual-arm—the authors show that co-training degrades performance, with a baseline dropping from 3.81 to 3.55 on a long-horizon benchmark. Their HiMoE-VLA replaces the dense module with a hierarchical mixture of experts: Action-Space MoE layers at the input and output boundaries specialize computation for each action space; Heterogeneity-Balancing MoE layers in adjacent layers absorb residual variation in embodiments, sensors,","pith_inferences":["One extension the paper leaves implicit: the same boundary-specialize/middle-share pattern could be applied to other multi-domain sequence models, where the contrastive regularization would act as an automatic classifier of action or control modes.","The fixed 24-dimensional vector suggests a scaling path: as new embodiments are added, the vector must grow and the validity masks must be re-learned; an implicit, testable prediction is that the number of experts should grow with the number of distinct action spaces rather than with dataset size.","The load-balancing objective could be replaced by an adaptive capacity-allocation rule that lets the router discover how many action spaces actually exist, removing the need to pre-specify a fixed top-k.","A direct test of the representation's ceiling would be to feed the same hierarchy with a learned per-embodiment embedding instead of the fixed vector; if that variant matches or exceeds HiMoE, the unified vector, not the MoE structure, is the operative limit."],"forward_implications":["If correct, one policy can be pre-trained on a large mixture of single-arm and dual-arm, joint-space and end-effector-space demonstrations without needing dataset-specific heads or manual embodiment labels.","Co-training on heterogeneous action spaces becomes a source of improvement rather than a cause of degradation, at least within the range of the fixed 24-dimensional action/state representation.","The hierarchy gives a concrete recipe: specialize at the boundaries, balance in adjacent layers, share in the middle; ablations show removing either MoE stage or either regularization erodes the gain.","Performance gains on long-horizon tasks and unseen objects suggest the design supports generalization, not just fitting to the training distribution."],"fun_headline_variants":["Hierarchical experts flip negative transfer into gains","Specialized MoE layers master mixed robot action spaces","One policy, many robots: Mixture-of-Experts handles diverse data","HiMoE-VLA: Turning data mix-ups into cross-embodiment wins","Divide and conquer: Experts tame heterogeneous robot training"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Everything rests on the fixed 24-dimensional unified vector: if a robot's true action space cannot be squeezed into 8 end-effector plus 16 joint-angle numbers without losing information, the experts will route on a distorted input and the transfer gains may vanish.","fun_headline_variants_meta":{"raw":{"variants":["Hierarchical experts flip negative transfer into gains","Specialized MoE layers master mixed robot action spaces","One policy, many robots: Mixture-of-Experts handles diverse data","HiMoE-VLA: Turning data mix-ups into cross-embodiment wins","Divide and conquer: Experts tame heterogeneous robot training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000916,"raw_usage":{"total_tokens":3788,"prompt_tokens":780,"completion_tokens":3008,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":2925}},"tokens_in":524,"tokens_out":3008,"duration_ms":19320,"temperature":1.0,"reasoning_tokens":2925,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T18:18:37.197444+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Co-train HiMoE-VLA on a robot with an action space that exceeds 24 dimensions—for example a 7-DoF arm with a multi-finger hand, where joint commands need more than 16 entries—and compare against a baseline with a dedicated head for that robot. If the unified-vector version does not beat the dedicated head (or at least match the positive-transfer pattern reported for 24-dim spaces), the load-bearing premise is the fixed representation, not the MoE hierarchy.","supporting_citations":[],"review_version":1}