{"id":"70fbd98b-4876-4cfc-9f35-dde850133554","arxiv_id":"2608.09771","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A compact latent interaction policy trained with bidirectional masked trajectory prediction matches or exceeds large VLA and world-action-model baselines on robot manipulation benchmarks while using far less compute.","lead":"SLIM is a 0.47-billion-parameter robot policy that learns compact latent representations of how actions change what a robot sees, instead of relying on a large vision-language model or generating future images. Across simulation and real-world manipulation tests, the small policy matches or beats much larger systems while running faster and using less GPU memory.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline comparison rests on mixing SLIM’s own rollouts with published baseline numbers from different training/evaluation protocols; a single strong baseline rerun under SLIM’s exact harness could materially shift the claimed margins.","rationale":"The reader’s weakest-assumption analysis identifies the same load-bearing issue: the comparison to baselines relies on published numbers that are not produced under a common evaluation protocol. I agree with that assessment, and I do not think a different concern is more fundamental. The ablation evidence is internally consistent, the parameter counts are clearly broken down, and the method description is sufficiently detailed to reproduce; those parts of the paper are not under attack. The conditional verdict is appropriate because the central headline claim depends on baseline comparability that the paper does not establish. I would keep the verdict CONDITIONAL rather than escalating to REJECT: a concrete re-evaluation of one or two strong baselines under SLIM’s harness could settle the question, and nothing in the manuscript suggests an internal contradiction or a fabricated result. The real-world evaluation has only 10 trials per condition, which is a secondary concern; the simulated baseline comparability is the more load-bearing risk because it affects the paper’s primary quantitative claim.","tokens_in":123,"tokens_out":4507,"duration_ms":93612,"concrete_test":"Obtain the open-source OpenVLA-OFT checkpoint and evaluate it with SLIM’s exact LIBERO/LIBERO-Plus harness: same 50-rollout evaluation per LIBERO suite, same 10,030 LIBERO-Plus perturbation cases, same seed set, same success criteria, and same action-horizon setting. Recompute the LIBERO-Plus overall score and its 95% confidence interval over seeds. If the re-evaluated score moves more than 2 points from the published 69.6, or if VLA-JEPA re-evaluated under the same protocol scores at or above SLIM’s 77.45, then the claimed margin over baselines is not robust and the abstract’s “matches or exceeds” statement must be re-scoped.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper’s central claim is that SLIM “matches or exceeds” representative VLA and world-action-model baselines. That claim is supported mainly by Tables 1, 2, and 5, which combine SLIM’s in-house evaluation with baseline numbers taken from prior papers. The baselines differ in training data, seeds, number of evaluation rollouts, action horizons, flow-sampling schedules, and checkpoint selection. The paper does not retrain or re-evaluate any baseline under a common protocol, and it reports only point estimates without confidence intervals. This matters concretely: on zero-shot LIBERO-Plus, SLIM scores 77.45, while VLA-JEPA scores 79.5, so the headline claim is already not uniformly true on that axis; the question is whether the gap is a real algorithmic deficit or an artifact of different evaluation conditions. The same concern applies to CALVIN, where SLIM’s 4.556 average length is within 0.026 of FLOWER’s 4.53, a difference smaller than typical seed noise. The strongest load-bearing assumption is therefore that the published baseline numbers are directly comparable to SLIM’s own evaluation. If that assumption fails, the central efficiency-versus-performance claim loses its quantitative foundation. The method itself is clearly specified and the internal ablations support the value of Stage-1 masked trajectory prediction, so the issue is not soundness of the method but soundness of the comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes SLIM, a 0.47B-parameter latent interaction policy for language-conditioned robot manipulation. SLIM uses a Mixture-of-Transformers (MoT) backbone with two streams (observation and action) and a two-stage training procedure: Stage 1 learns action-grounded predictive latents via masked trajectory prediction, combining an inverse-dynamics loss (reconstructing noised action chunks from current and future observation latents) and a forward-dynamics loss (predicting future observation latents from current observations and clean actions, supervised by an EMA target); Stage 2 trains a flow-matching policy conditioned only on the current observation latent, learned future-slot embeddings, proprioception, language, and noised actions. The paper reports results on LIBERO (97.5% overall), zero-shot LIBERO-Plus (77.45%), CALVIN ABC→D (4.556 average sequence length), and a five-task real-world manipulation suite, alongside ablations showing that Stage-1 training and the EMA target improve performance. The central claim is that SLIM matches or exceeds representative vision-language-action and world-action-model baselines with fewer parameters, no additional embodied pretraining, lower inference latency, and lower GPU memory usage.","tokens_in":16815,"tokens_out":6795,"duration_ms":55463,"significance":"If the empirical comparison holds, the paper demonstrates a valuable design point: compact latent interaction policies can be competitive with much larger VLA and world-action-model baselines on standard manipulation benchmarks, with substantial efficiency gains. The method is clearly specified, the two-stage pipeline is well-described, and the ablations (especially the EMA collapse diagnostics in Table 3) provide useful insights into predictive representation learning. The controlled inference benchmark (Table 4) is a strength, as is the real-world evaluation comparing SLIM, π0.5, and Fast-WAM under a common training protocol. The primary risk is the comparability of the simulation baselines, which are taken from prior papers with different protocols and reported without error bars.","major_comments":[{"comment":"The headline claim that SLIM 'matches or exceeds representative large-scale VLA and world-action-model baselines' rests on comparing SLIM's in-house rollouts with published baseline numbers that differ in training data, seeds, evaluation rollouts, flow-sampling schedules, and hyperparameters. The reported numbers themselves do not uniformly support the claim: on zero-shot LIBERO-Plus, SLIM obtains 77.45% whereas VLA-JEPA obtains 79.5%, and the paper provides no confidence intervals or seed-level statistics. A single strong baseline (e.g., VLA-JEPA or OpenVLA-OFT) rerun under SLIM's exact evaluation harness could materially shift the claimed margins. This is a load-bearing gap for the paper's central efficiency-versus-performance claim.","section":"§4.1, Table 1"},{"comment":"On CALVIN ABC→D, SLIM's 4.556 average sequence length is within 0.026 of FLOWER's 4.53, a difference smaller than typical seed noise. With no reported variance or number of seeds, the 'matches' part of the claim is unsupported; the same issue applies to several other CALVIN entries where differences are within a few hundredths of a point. The paper should provide seed-level statistics or otherwise justify that these differences are meaningful.","section":"§4.1, Table 2"},{"comment":"The real-world comparison reports only point estimates from 10 trials per task-condition, without error bars or statistical tests. The observed differences, e.g., SLIM's average progress 67.8 vs π0.5's 56.8, and the 'close' performance under background shift (49 vs 54), could be within trial noise. This undercuts the real-world generalization claim, which is one of the paper's four main evaluation axes.","section":"§4.2, Figure 6"},{"comment":"The forward-dynamics target is produced by an EMA copy of the same online encoder, and the policy is trained in that same latent space. The no-EMA ablation (Table 3) shows collapse, but the EMA target still does not provide an externally grounded supervision signal; the Stage-1 improvement could be an auxiliary-task regularizer rather than evidence for action-grounded predictive latents. A concrete test would be to compare against a fixed pre-trained target encoder (e.g., frozen DINOv2) or to measure whether the predicted latent aligns with a genuinely independent future-state encoding. This would strengthen the conceptual contribution of the paper.","section":"§3.3, Eq. (7)"}],"minor_comments":[{"comment":"The 'Overall' column for LIBERO-Plus appears to be an average over seven perturbations; please state explicitly whether the average is equally weighted and how the per-perturbation scores are computed.","section":"§4.1, Table 1 caption"},{"comment":"The notation 'LN' is used without definition; please clarify whether it denotes layer normalization or a learned normalization operator.","section":"§3.3, Eq. (7)"},{"comment":"The attention maps are described as showing that SLIM 'more consistently follows the manipulated object,' but only two example sequences are shown and no quantitative attention metric is provided; this claim should be softened or supported with a quantitative analysis.","section":"§4.4, Figure 8"},{"comment":"The authors acknowledge that the experiments use a single model scale and do not establish scaling behavior; this limitation is appropriate and should be stated earlier so the headline claim is scoped as a single-scale result.","section":"§5, Conclusion"},{"comment":"There are several typographical issues, e.g., 'In F orty-third International Conference on Machine Learning' (the LaST 0 entry) and inconsistent use of spaces in 'arXiv' entries; please proofread the reference list.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the cross-paper comparison: without a common-protocol baseline rerun or confidence intervals, the headline claim is not yet solid. The paper is otherwise well-executed and the method is clearly specified. I suggest requiring the authors to provide seed-level statistics and at least one direct baseline comparison under their evaluation harness, and to address the EMA target grounding concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this paper deserves a serious referee, but the headline comparison is not as clean as the abstract suggests. The Stage-1 ablation and EMA collapse diagnostics are the strongest parts; the comparison claim needs work.\n\nWhat's actually new: the coupling of inverse and forward dynamics in one compact MoT backbone with masked trajectory prediction in latent space. That specific combination is absent from the cited prior work, including VLA-JEPA and DeFI, which either keep dynamics separate or put JEPA-style prediction on a large VLM. The method section is clearly written. The real-world evaluation is more controlled than is typical: baselines are trained on the same demonstration mixture, and OOD perturbations (background, lighting, distractor) are well defined. The latency/memory/FLOPs benchmark is carefully described with native horizons and a pooled p95. The Stage-1 gain (77.45 vs 66.82 on LIBERO-Plus) and the latent-collapse diagnostics (effective rank 61 vs 14, cosine similarity 0.071 vs 0.352 with EMA vs without) directly support the design's core mechanism. That is real evidence.\n\nSoft spots, in order of importance. First, the central comparison claim is overclaimed. On zero-shot LIBERO-Plus, VLA-JEPA scores 79.5, above SLIM's 77.45; on CALVIN, SLIM's 4.556 is within noise of FLOWER's 4.53. So the abstract should say 'competitive with, ahead on some axes, behind on others,' not 'matches or exceeds.' Second, and more seriously, all LIBERO and CALVIN baseline numbers are taken from prior papers with different training data, seeds, and evaluation protocols. The paper does not retrain or re-evaluate any baseline under SLIM's harness. A single strong baseline run could materially shift the margins, so the quantitative efficiency-versus-performance claim is only as strong as the comparability assumption. The fix is straightforward: rerun at least the closest baselines (VLA-JEPA, FLOWER) under the same evaluation script, or present the comparison as indicative rather than definitive. The real-world comparison is less vulnerable because baselines share the demo mixture, but 10 trials per condition with manually defined 0/0.5/1 milestones limits precision; minor. No code or checkpoints are released, which hurts reproducibility but doesn't invalidate the results. The conclusion's limitation paragraph is honest about the single model scale.\n\nWho gets value: researchers working on efficient robot policies, latent world models, or VLA compression. It's a useful design and a clean ablation. I'd bring it to the reading group, and I'd send it to review. The comparison weakness is condition-level, not fatal.","headline":"Solid methods paper with a genuinely new compact latent-interaction architecture; Stage-1 ablations are strong, but the headline comparison leans on unevaluated published baselines and needs a controlled rerun or softer claim.","tokens_in":17439,"tokens_out":3257,"would_cite":true,"duration_ms":27697,"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":"SLIM, a 0.47B latent interaction policy, matches or exceeds large VLA and world-action-model baselines with fewer parameters, no extra embodied pretraining, lower latency, and lower GPU memory.","keywords":["robot manipulation","vision-language-action models","latent world models","flow matching","self-supervised representation learning","Mixture-of-Transformers","inverse dynamics","forward dynamics"],"falsifier":"Run the strongest baselines, say OpenVLA-OFT and VLA-JEPA, under SLIM's exact LIBERO-Plus zero-shot protocol, including the same 10,030 perturbation cases, rollout counts, seeds, and flow-sampling steps; if their overall scores become statistically indistinguishable from SLIM's 77.45%, the comparative advantage collapses, even though the method itself may still work.","tokens_in":16316,"feed_emoji":"🤖","tokens_out":5468,"duration_ms":44215,"temperature":0.7,"pith_summary":"SLIM is a 0.47-billion-parameter robot manipulation policy that learns, in latent space, both which action explains a change between two observations and which future observation a given action produces. The central claim is that this bidirectional, action-grounded prediction makes a small model competitive with, and sometimes better than, much larger vision-language-action and world-action-model baselines while using less compute and memory. On LIBERO it scores 97.5%, on zero-shot LIBERO-Plus 77.45%, on CALVIN ABC→D an average sequence length of 4.556, and on real-world tasks an average progress of 67.8 against 56.8 for π0.5 and 40.0 for Fast-WAM. If the comparison holds, robot manipulation need not route every control step through a large VLM or through pixel-level future generation; a compact latent interaction process can provide the control backbone.","feed_headline":"0.5B robot policy outdoes 7B vision-language-action models","feed_subtitle":"Action-grounded latent prediction makes SLIM faster, lighter, and competitive on LIBERO, CALVIN, and real tasks.","key_machinery":"The central machinery is action-grounded masked trajectory prediction inside a compact Mixture-of-Transformers (MoT) backbone, a two-stream transformer where an observation stream and an action stream interact through shared joint attention and per-stream language cross-attention. Stage 1 masks one trajectory variable at a time: with the action chunk masked, the model performs inverse dynamics by predicting the flow velocity of the noised action from current and future observation latents; with the future latent masked, it performs forward dynamics by predicting that latent from the current observation latent and the clean action chunk, supervised by a stop-gradient EMA target. Stage 2 reuses the same backbone for flow-matching action generation without any future observation. The predictive slots learned in Stage 1 remain in the observation stream and, the paper argues, guide action generation during inference, which is why the deployed policy needs no pixel decoder or future-frame generation.","core_discovery":"The paper's discovery is that a compact policy can internalize dynamics as predictive structure in observation-latent space rather than as VLM reasoning or pixel prediction. SLIM couples an inverse-dynamics objective—reconstructing the noised action chunk from current and future observation latents—with a forward-dynamics objective—predicting the future observation latent from the current latent and clean action chunk—using a Mixture-of-Transformers backbone with separate observation and action streams. This masked trajectory prediction is trained self-supervised on robot trajectories, and the same backbone is then trained with flow matching for language-conditioned action generation. The result, the paper argues, is that action-grounded predictive latents give a 0.47B policy performance on LIBERO, LIBERO-Plus, CALVIN ABC→D, and real-world manipulation that matches or exceeds representative baselines, with 60.6 ms mean latency, 4.26 GiB peak VRAM, and no embodied pretraining beyond the target data.","pith_inferences":["Beyond the paper: scaling this latent-interaction recipe with broader cross-embodiment data may give large-VLA-level generalization at small-model efficiency, since the ablations suggest the limiting factor is the breadth of training data rather than backbone capacity.","A testable extension suggested by the EMA-target analysis is that an explicit rank or diversity regularizer could replace the EMA teacher and might transfer better to smaller data budgets, since the paper shows the EMA's value lies in preventing latent collapse.","The attention-probe result, if it holds beyond the two shown sequences, predicts that Stage-1-trained policies will be more robust to distractors and background shifts because the action stream attends to the manipulated object; this could be tested with systematic saliency evaluations.","The paper's comparative claim is more sensitive than its method to evaluation standardization: rerunning baselines under SLIM's exact seeds and rollouts could change the reported margins, and that comparison is the part most worth rechecking."],"forward_implications":["A robot control policy does not need a large vision-language backbone at its center; a 0.5B latent interaction model can reach or exceed 7B-class VLA performance on standard manipulation benchmarks.","Dynamics can be an explicit training signal without pixel reconstruction: bidirectional latent prediction grounds actions in observation changes at a fraction of the compute of video world models.","Because SLIM needs no test-time future imagination, its 60.6 ms inference latency and 4.26 GiB peak GPU memory make deployment on physically embodied robots materially cheaper than π0.5 or Fast-WAM.","Self-supervised masked trajectory prediction transfers across settings: the same LIBERO-trained checkpoint handles zero-shot LIBERO-Plus perturbations, and the same two-stage recipe works on real-world demonstrations.","The ablation results indicate that the Stage-1 latent-learning phase, not the flow-matching policy head, is what drives the gains: removing it lowers LIBERO-Plus success from 77.45% to roughly 66.8% and CALVIN average length from 4.556 to 4.382."],"supporting_citations":[{"why":"Supplies the conditional flow matching used both for noised-action interpolation in the objectives and for the final action-generation policy.","marker":"Lipman et al., 2023"},{"why":"Supplies the Mixture-of-Transformers architecture with separate observation and action streams and joint attention.","marker":"Liang et al., 2024"},{"why":"Establishes the JEPA-style masked prediction in latent space that Stage 1 extends to action-grounded trajectory prediction.","marker":"Assran et al., 2023"},{"why":"Provides the video feature-prediction principle and EMA-target setup used for the future-latent prediction objective.","marker":"Bardes et al., 2024"},{"why":"Initializes the visual encoder (DINOv2) whose output latents are the space in which dynamics are predicted.","marker":"Oquab et al., 2024"},{"why":"Defines the LIBERO benchmark and data used for training and the four-suite evaluation.","marker":"Liu et al., 2023"},{"why":"Defines the LIBERO-Plus perturbation benchmark used for the zero-shot robustness evaluation.","marker":"Fei et al., 2025"},{"why":"Defines the CALVIN ABC→D benchmark used for the long-horizon composition evaluation.","marker":"Mees et al., 2022"},{"why":"Provides the π0 vision-language-action flow-model baseline that SLIM compares against in efficiency and performance.","marker":"Black et al., 2024"},{"why":"Provides the π0.5 baseline used in the real-world multi-task comparison and latency/memory measurements.","marker":"Physical Intelligence et al., 2025"}],"fun_headline_variants":["0.5B SLIM policy rivals 7B VLA in robot control","Self-supervised latents shrink robot policy to 0.5B","Action-grounded prediction: 0.5B policy matches 7B VLA","0.5B robot policy beats 7B VLA with predictive latents","SLIM: compact 0.5B policy matches large VLA baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the published numbers for the baselines are directly comparable to SLIM's, even though the baselines were not retrained or re-evaluated under SLIM's exact seeds, rollouts, and sampling schedules.","fun_headline_variants_meta":{"raw":{"variants":["0.5B SLIM policy rivals 7B VLA in robot control","Self-supervised latents shrink robot policy to 0.5B","Action-grounded prediction: 0.5B policy matches 7B VLA","0.5B robot policy beats 7B VLA with predictive latents","SLIM: compact 0.5B policy matches large VLA baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001252,"raw_usage":{"total_tokens":5140,"prompt_tokens":962,"completion_tokens":4178,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":4083}},"tokens_in":578,"tokens_out":4178,"duration_ms":26600,"temperature":1.0,"reasoning_tokens":4083,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:03:55.703124+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the strongest baselines, say OpenVLA-OFT and VLA-JEPA, under SLIM's exact LIBERO-Plus zero-shot protocol, including the same 10,030 perturbation cases, rollout counts, seeds, and flow-sampling steps; if their overall scores become statistically indistinguishable from SLIM's 77.45%, the comparative advantage collapses, even though the method itself may still work.","supporting_citations":[],"review_version":1}