{"id":"79b15f84-b748-467e-b981-8cbb172724c2","arxiv_id":"2608.29937","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"AcrossWAM1.0 demonstrates that a modular latent world-action stack can retain closed-loop performance when the vision-language backbone is scaled from 2B to 0.8B parameters, with a parameter reduction of 42.4%.","lead":"AcrossWAM1.0 modularizes the LaWAM latent world-action policy stack, defining explicit adapter boundaries between the visual backbone, world decoder, and action expert. On 2,000 LIBERO episodes a compact Qwen3.5-0.8B backbone achieves 97.45% success (0.55 pp below the 2B model, p=0.266) while reducing deployable parameters by 42.4% to 1,472.6M.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Single checkpoint per backbone leaves training-seed variance unmeasured; observed gap may not be representative of the population of possible checkpoints.","rationale":"The reader's weakest assumption is exactly this: that one checkpoint per backbone is sufficient. I agree that this is the most load-bearing assumption, but the paper is transparent about it and does not overclaim (e.g., no claim of statistical equivalence or non-inferiority across seeds). The verdict of ACCEPT remains appropriate because the paper's contributions (modular design, explicit boundaries, controlled paired evaluation, and deployment verification) are solid and honestly scoped. The seed-variance issue is a known limitation, not a hidden flaw. Therefore, the verdict should not change.","tokens_in":12576,"tokens_out":2760,"duration_ms":26926,"concrete_test":"Train 3–5 independent checkpoints of each backbone (Qwen3-VL-2B and Qwen3.5-0.8B) from different random seeds, using the same training recipe and evaluation protocol. For each pair of checkpoints (one per backbone), compute the success gap and McNemar p-value on the same 2,000 paired episodes. Report the mean, standard deviation, and range of the gap across seeds. If the gap exceeds 2 pp in any seed pair, the retention criterion is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that Qwen3.5-0.8B retains performance within a two-point margin of Qwen3-VL-2B—rests on exactly one training run per backbone. The paper explicitly notes this limitation in Section 5: 'Each backbone is represented by one trained checkpoint: episode pairing controls evaluation stochasticity but not training-seed variation.' If training-seed variation is large, the observed −0.55 pp gap could be a favorable draw for the compact model or an unfavorable one for the 2B model, and would not reproduce. Because the claim is about relative retention rather than absolute performance, even a moderate seed variance (e.g., ±1 pp per backbone) could flip the conclusion of meeting the two-point criterion. The paired evaluation controls evaluation noise but does not bound training noise. This is the weakest link because the entire quantitative contribution (compact vs. large comparison) depends on the gap being small and stable, yet that stability is unmeasured.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper presents AcrossWAM1.0, a modularization and scaling study of the LaWAM latent world-action formulation. Its contributions are (1) an explicit policy-world-action contract around the LaWAM computation, including backbone adapters, query ownership, gradient/freeze boundaries, and a training-versus-deployment module split; (2) a paired closed-loop comparison on 2,000 LIBERO episodes in which Qwen3.5-0.8B achieves 97.45% success versus 98.00% for Qwen3-VL-2B, a -0.55 percentage-point gap within a prespecified two-point retention criterion and with no detected paired difference (exact McNemar p=0.266; discordant counts 46 vs 35); and (3) an auditable deployment export whose 1,472.6M unique inference-reachable parameters are 42.4% fewer than the raw 2B policy, with all 1,078 retained tensors bitwise identical to the source checkpoint. Cross-family execution is validated only by a MiniCPM-V forward/backward smoke test; closed-loop cross-family transfer, latency and peak-memory measurements, and training-seed variance are explicitly left open.","tokens_in":13110,"tokens_out":9089,"duration_ms":81343,"significance":"The paper's scoping discipline is its main strength: claims are stratified into interface compatibility, closed-loop scaling within the Qwen family, and explicitly non-claimed frozen cross-family portability (Section 3.6 and Appendix C). The central comparison is backed by matched episode keys, a prespecified two-point retention criterion, a correctly computed exact McNemar test with reported discordant counts (b=46, c=35, p=0.2664), and parameter accounting verified by bitwise tensor comparison of the exported checkpoint. The manuscript also ships machine-readable paired and export-verification results with the code, and its limitations section anticipates the main reviewer concerns: single checkpoint per backbone, LIBERO-only closed-loop evidence, and no frozen-decoder cross-family claim. If the results hold, the paper provides a genuinely useful systems and evaluation boundary for latent world-action policies, making 'modular' operational rather than diagrammatic. The significance is moderate rather than high, because the paper deliberately contributes no new learning algorithm and the closed-loop evidence spans one benchmark and one backbone family.","major_comments":[{"comment":"Table 1 lists AcrossWAM1.0 with RoboTwin 2.0 scores of 92.09 (Clean) and 89.12 (Randomized), yet Section 4.2 states that the RoboTwin columns are not filled with inherited LaWAM numbers or the preliminary language diagnostic 'because neither is an AcrossWAM1.0 benchmark evaluation,' and Section 5 states that closed-loop evidence covers LIBERO only. These statements are mutually inconsistent. Please either give the provenance of the two RoboTwin numbers and the evaluation details behind them, or replace them with dashes so that the table agrees with the paper's own evidence taxonomy.","section":"Table 1 / Section 4.2 / Section 5"},{"comment":"The headline retention result rests on exactly one trained checkpoint per backbone, a point the manuscript discloses in Section 4.1 ('episode pairing controls evaluation stochasticity but not training-seed variation') and Appendix C, and the conclusion is already scoped to the measured checkpoints. This is honest but leaves the central quantitative claim unhardened: the observed pooled gap of -0.55 pp, and the per-suite gaps of up to -1.6 pp, could plausibly lie within between-seed variation for either backbone. I recommend adding at least a small number of training seeds for the compact configuration (or for both) and reporting the spread of the gap; if that is infeasible, state in the abstract and conclusion that the two-point retention result is checkpoint-specific and that its seed variance is unknown.","section":"Sections 4.1, 4.3, and 5"}],"minor_comments":[{"comment":"There are typographical spacing errors in the abstract ('a-0.55percentage-point', 'McNemarp=0.266') and in the title ('ACROSSWAM1.0:A Modular'), and the LaWAM name appears with inconsistent spacing ('LaW AM' vs 'LaWAM'); please normalize these throughout.","section":"Abstract and title"},{"comment":"The abstract's '42.4% fewer than the original 2B policy' compares the compact deployment export, which has training-only state removed, against the raw 2B checkpoint; Table 3 is transparent about this, but the abstract should say 'unique inference-reachable parameters of the compact export versus the raw 2B policy checkpoint' to avoid comparing unlike artifact types.","section":"Abstract / Section 4.3 / Table 3"},{"comment":"The two-point retention criterion is called 'prespecified' without supporting evidence such as a dated protocol or project log entry; please state where and when the criterion was fixed, or rephrase it as an engineering criterion adopted before the paired evaluation.","section":"Section 4.1"},{"comment":"The compact configuration's training budget (25,000 steps) and hyperparameters are given, but the 2B model's training steps and hyperparameters are not; please confirm that the two backbones were trained with the same procedure (steps, lambda_p, lambda_d, learning rates) so that the -0.55 pp gap can be attributed to the backbone rather than to recipe differences.","section":"Section 4.1 / Section 4.3"},{"comment":"The eight latent-action queries are also called 'ACT queries' in Figure 1 and the text, and 'Stage-2' appears alongside 'Stage 2'; please define the naming convention once and use it consistently.","section":"Section 3.4 / Figure 1"}],"recommendation":"minor_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"This paper does what it says: it modularizes LaWAM's latent world–action stack, provides a controlled compact-backbone comparison, and ships a verifiable deployment export. The engineering is solid, the evaluation is paired with a McNemar test, and the limitations section is unusually candid. I appreciate that they explicitly separate interface compatibility, closed-loop scaling, and frozen cross-family transfer — most papers would have collapsed those into one grand claim.\n\nThe central result — that Qwen3.5-0.8B retains performance within 0.55 pp of Qwen3-VL-2B — is supported by 2,000 paired episodes and a prespecified retention criterion. That part is credible as far as it goes. But the stress-test note hits the real weak point: each backbone has exactly one training run, so training-seed variance is unmeasured. The paper says this, which is honest, but it means the reported gap could easily shift by a couple of points with a different seed. For a paper whose main quantitative contribution is a gap of 0.55 pp, that is a genuine fragility — not fatal, because the two-point criterion is generous, but enough that I wouldn't bet on the exact number reproducing.\n\nThe evaluation is also limited to LIBERO and Qwen backbones. The MiniCPM-V smoke test shows the adapter contract works, but it's not a closed-loop result. That's fine for a first study, but it limits the generality of the modularization claim.\n\nOverall, this is a useful piece of systems work with clear scope and honest reporting. It is not a breakthrough but it is not trying to be. The paper deserves a serious referee — the evidence is reproducible, the code is available, and the limitations are stated. I would recommend acceptance with a note asking the authors to either run one or two additional training seeds or adjust the strength of the claim to explicitly acknowledge that the observed retention is from a single seed per condition.","headline":"A clean, honest modularization study that does what it says but whose central scaling comparison rests on a single training seed per backbone — a real but acknowledged fragility.","tokens_in":13408,"tokens_out":1262,"duration_ms":13942},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"A compact 0.8B backbone holds 97.45% LIBERO success, within two points of the 2B model, while the exported policy cuts 42.4% of deployable parameters.","keywords":["latent world-action model","vision-language-action policy","modular policy stack","compact backbone scaling","LIBERO benchmark","flow matching","deployment checkpoint export","robot manipulation"],"falsifier":"Train multiple independent checkpoints of each backbone (same data, different seeds) and evaluate every one on the same 2,000 paired episodes: if the 0.8B success rates spread by much more than two points across seeds, or the mean gap exceeds the prespecified two-point retention criterion, the compact-retention claim is falsified. A second check: run the exported 1,472.6M-parameter checkpoint through the production loader on an accelerator with the LIBERO simulator — the paper states this has not yet been done — and see whether the closed-loop success rate reproduces the reported 97.45%.","tokens_in":14380,"feed_emoji":"🤖","tokens_out":9104,"duration_ms":71515,"temperature":0.7,"pith_summary":"AcrossWAM1.0 tries to establish that LaWAM's latent world-action computation — predicting an action-relevant visual subgoal in feature space instead of rendering future pixels — can be rebuilt as a modular, auditable system with explicit boundaries between the policy backbone, the world decoder, and the action expert. Its central empirical claim is that a much smaller vision-language backbone (Qwen3.5-0.8B) plugged into this stack keeps closed-loop robot success essentially intact: 97.45% versus 98.00% for the 2B model on the same 2,000 paired LIBERO episodes, a 0.55-point drop that meets a prespecified two-point retention criterion and shows no statistically detected paired difference (exact McNemar $p = 0.266$). A second claim is that the deployable artifact is genuinely compact and genuinely equivalent to training: the exported checkpoint holds 1,472.6M unique parameters (42.4% fewer), with all 1,078 retained tensors bitwise identical to the source and all training-only modules removed. The paper matters because it separates what has been demonstrated — within-family backbone scaling and exact export — from what remains open (frozen cross-family transfer, latency advantages, and seed-level robustness).","feed_headline":"Half-size robot policy holds 97.45% success, cuts 42% of its weight","feed_subtitle":"Swapping a 2B vision-language backbone for a 0.8B one costs 0.55 points across 2,000 paired LIBERO episodes.","key_machinery":"The carrying object is the latent visual subgoal: a 32-dimensional latent action $\\hat{z}_t$ extracted by the backbone's latent-action queries, compressed by a one-query cross-attention mapper $M_\\rho$, and passed through the retained latent world decoder $D_\\omega$ (twelve transformer blocks with adaptive LayerNorm modulation) to produce a predicted future feature $\\hat{h}_{t+\\Delta}$ with the same spatial token structure as the frozen DINOv3 current feature. That predicted subgoal, rather than the latent vector alone, conditions the flow-matching action expert: a conditional diffusion transformer that regresses a vector field $v_\\eta(a_\\tau, \\tau \\mid c_\\theta, h_t, \\hat{h}_{t+\\Delta}, s_t)$ toward the ground-truth velocity $u_\\tau = a_1 - a_0$ and integrates it with an ODE solver into a 50-step, 1.2-second action chunk. Around this computation the paper builds the contract it actually evaluates: a backbone adapter that returns separate transition and action contexts through a fixed interface, explicit hidden-size alignment, and a hard split between training-only teachers (the inverse-dynamics Q-Former $q_\\phi$ and the state-grounding head) and inference-reachable modules, with the stage-two loss $L = L_{\\text{flow}} + \\lambda_p L_{\\text{subgoal}} + \\lambda_d L_{\\text{distill}}$ using $\\lambda_p = \\lambda_d = 0.1$.","core_discovery":"On the paper's own terms, the discovery is an extension, not a new algorithm: the latent visual-subgoal formulation inherited from LaWAM is re-engineered into a modular stack whose inference graph contains only a policy adapter, a latent predictor, the retained latent world decoder, and a flow-matching action expert. The load-bearing empirical result is the controlled compact-backbone study: on 2,000 paired LIBERO episodes with identical seeds and episode keys, the 0.8B model succeeds in 1,949 episodes (97.45%) against 1,960 (98.00%) for the 2B model, with discordant counts 46 versus 35 and exact McNemar $p = 0.2664$; the paper explicitly states this satisfies its prespecified two-point retention criterion and explicitly does not claim statistical equivalence or non-inferiority across training seeds. The deployment claim is verified at the state-dictionary and CPU-loader level: 1,078 retained tensors bitwise identical, 321 training-only keys (411.1M parameters) removed, and 718 alias keys deduplicated. The paper also draws a three-level distinction — interface compatibility, closed-loop backbone scaling, and frozen cross-family transfer — and claims verified status only for the first two levels within the Qwen family.","pith_inferences":["Editorial inference: if the 0.55-point gap proves stable across multiple training seeds, the two-point retention criterion could be promoted from an engineering rule of thumb into a defensible statistical claim, and the 42.4% parameter reduction would likely translate into real memory and latency savings — quantities the paper does not yet measure.","Editorial inference: the paper's three-level taxonomy of portability claims (interface compatibility versus closed-loop scaling versus frozen transfer) is a workable template for the field; future portable robot-policy papers could be asked to state which level they certify.","Editorial inference: a direct testable extension is a seed-variation study in which five or more checkpoints of each backbone are trained and each is evaluated on the same paired episodes; this would settle whether the observed gap and its non-significance are typical or lucky.","Editorial inference: the conflicting-instruction diagnostic hints that the flow-matching expert may satisfy the overall success criterion while language conditioning remains under-fit; isolating the adherence-completion trade-off on more than two RoboTwin tasks would clarify whether compact policies need stronger language supervision."],"forward_implications":["A roughly 2.5-times smaller backbone can substitute for the 2B model under the same latent world-action contract without a detected closed-loop loss on LIBERO, cutting unique inference-reachable parameters by 42.4% from 2,555.2M to 1,472.6M.","The exact export procedure — bitwise-identical retained tensors, 411.1M training-only parameters removed, 718 alias keys deduplicated — gives other teams a checkable recipe for verifying that a deployed policy contains only the inference graph.","The three-level evidence distinction (interface compatibility, closed-loop scaling, frozen cross-family transfer) bounds the claim: within-family backbone scaling is verified, while the MiniCPM-V adapter execution is verified only as a forward/backward smoke test with no control result.","The language-grounding diagnostic shows aggregate success can conceal weak instruction following (27.4% adherence on conflicting left/right instructions), so the compact-retention result should not be read as evidence for language grounding quality.","Because the compact run fine-tunes the latent world decoder, the result supports backbone scaling of the modular stack rather than a frozen world interface that transfers across backbone families."],"supporting_citations":[{"why":"LaWAM supplies the latent visual-subgoal formulation, the stage-one inverse-dynamics teacher, the frozen DINOv3 encoder, and the published 98.6% LIBERO baseline that AcrossWAM1.0 modularizes rather than re-claims.","marker":"(Chen et al., 2026)"},{"why":"LIBERO defines the benchmark: four suites, ten tasks each, 50 trials, giving the 2,000 paired episodes on which the compact-versus-full comparison is measured.","marker":"(Liu et al., 2023)"},{"why":"The Qwen3-VL technical report documents the 2B backbone family and its image-call and token conventions, which the adapters must reproduce for the closed-loop comparison.","marker":"(Bai et al., 2025)"},{"why":"DINOv3 is the frozen patch encoder whose feature space defines the current representation $h_t$ and the prediction target $\\hat{h}_{t+\\Delta}$ in both stages.","marker":"(Siméoni et al., 2025)"},{"why":"Flow matching provides the training objective the action expert regresses, with conditional vector field $u_\\tau = a_1 - a_0$ and ODE integration at test time.","marker":"(Lipman et al., 2023)"},{"why":"The DiT conditional diffusion transformer is the architecture used for the action expert's self- and cross-attention conditioning on context and predicted subgoal.","marker":"(Peebles & Xie, 2023)"},{"why":"π0 is the prior demonstration that a flow-matching action expert can sit atop a pretrained VLM, motivating the action-expert design and serving as a Table 1 baseline.","marker":"(Black et al., 2025)"},{"why":"ACT contributes the action-chunking and query-style formulation inherited in the latent-action and flow-query layout, including the 50-step chunk prediction.","marker":"(Zhao et al., 2023)"}],"fun_headline_variants":["Robot policy shrinks 42% while losing only 0.55 points","Compact robot policy: 0.8B model achieves 97.45% on 2,000 tasks","Half-size vision backbone runs robot with 97.45% success rate","Bit-identical weights: robot policy cuts 42% parameters, retains performance","Latent world model modularizes robot policy, keeps 97.45% success"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The comparison rests on a single trained checkpoint per backbone: the paper states that episode pairing controls evaluation stochasticity but not training-seed variation, so if different random seeds produce meaningfully different policies, the observed 0.55-point gap and its non-significant paired test may not represent the typical difference.","fun_headline_variants_meta":{"raw":{"variants":["Robot policy shrinks 42% while losing only 0.55 points","Compact robot policy: 0.8B model achieves 97.45% on 2,000 tasks","Half-size vision backbone runs robot with 97.45% success rate","Bit-identical weights: robot policy cuts 42% parameters, retains performance","Latent world model modularizes robot policy, keeps 97.45% success"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000462,"raw_usage":{"cost":0.000462028,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000462028,"upstream_inference_prompt_cost":0.000131348,"upstream_inference_completions_cost":0.00033068},"total_tokens":2324,"prompt_tokens":1143,"completion_tokens":1181,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":1072}},"tokens_in":1143,"tokens_out":1181,"cost_source":"openrouter_usage","duration_ms":11142,"temperature":1.0,"reasoning_tokens":1072,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-09-09T05:05:38.896768+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-alltime-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Train multiple independent checkpoints of each backbone (same data, different seeds) and evaluate every one on the same 2,000 paired episodes: if the 0.8B success rates spread by much more than two points across seeds, or the mean gap exceeds the prespecified two-point retention criterion, the compact-retention claim is falsified. A second check: run the exported 1,472.6M-parameter checkpoint through the production loader on an accelerator with the LIBERO simulator — the paper states this has not yet been done — and see whether the closed-loop success rate reproduces the reported 97.45%.","supporting_citations":[],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}