{"id":"3919007c-d41d-4e06-b58a-76a4282d3937","arxiv_id":"2608.04765","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Adding an explicit rolling language memory to a hierarchical VLA policy improves long-horizon stage success and yields readable progress logs, but the evidence lacks confidence intervals and an ablation.","lead":"Robot policies that add a running text summary of progress performed better on long multi-step tasks in simulations and on a real robot. The approach makes each step's reasoning visible and reduces skipped stages, though the measured gains are modest and not yet statistically robust.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No ablation isolates the language-memory variable from the added high-level VLM and subtask conditioning, so the reported gains may not be caused by memory.","rationale":"The reader's weakest assumption focuses on the low-level bottleneck: actions depend on history only through the short subtask s_t, so memory content not verbalized in s_t cannot affect control. That is a real constraint on the mechanism, and it is related to the attribution problem. However, the more load-bearing gap is that no ablation removes memory while keeping the high-level VLM and subtask supervision. If memory is only changing subtask predictions, the bottleneck is already the intended interface; what needs to be proven is that the memory variable improves those predictions beyond what the current observation plus instruction can provide. The paper's comparisons conflate the memory module with the entire hierarchical architecture, and its own limitations section admits lack of seed variation and confidence intervals. Thus the reader's conditional verdict is appropriate, but for a slightly different reason than the reader emphasized. I see no reason to move from CONDITIONAL; the concern is addressable by a targeted ablation, so UNCHANGED is the correct verdict adjustment.","tokens_in":70,"tokens_out":4806,"duration_ms":65147,"concrete_test":"Train a memory-free ablation with the same architecture and training budget: high-level condition π_HL(s_t | o_t, l) with no m_{t-1} and no memory target, while the low-level VLA and data pipeline remain unchanged. Compare stage success rates and subtask accuracy against the memory model over at least 5 seeds with bootstrap CIs. If the memory-free model matches the 63.9% single-package result, the memory variable is not the causal driver; if memory wins with non-overlapping intervals, the central claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is causal: explicit language memory improves long-horizon VLA performance. The experiments compare native π0.5 against π0.5 + language memory, but that comparison changes at least three variables at once: it adds a separately trained high-level PaliGemma VLM, provides subtask supervision to that branch, and conditions the low-level policy on the resulting subtask string in addition to the global instruction. The rolling memory m_{t-1} is only the third change. No condition keeps the new high-level branch and subtask interface while removing memory, so the improvements in Tables 1–2 (e.g., 41.7% to 63.9% on single-package Genie Sim) could come from extra semantic modeling capacity or from giving the action policy a concise predicted subtask, rather than from cross-time memory. Equation (1) sharpens the issue: π_LL depends on o_t and s_t only, so memory can affect actions only by changing the generated subtask. If a memory-free high-level VLM conditioned on (o_t, l) produces equally good subtasks, the memory is functionally inert. The qualitative logs in Fig. 4 show memory changes the subtask, but they do not compare against a no-memory hierarchy trained on the same subtask labels. Section 5.5 also concedes there are no confidence intervals or seed variation, so the size of the measured gaps is unknown.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hierarchical vision-language-action (VLA) architecture in which a high-level PaliGemma VLM maintains a rolling natural-language memory m_t and emits a concise subtask s_t given the current observation o_t, previous memory m_{t-1}, and global instruction l; a low-level VLA then generates action chunks conditioned on (o_t, s_t). Memory and subtask targets are produced automatically from expert videos with skill timestamps via a prompted VLM. The method is evaluated against native π0.5 on a BEHAVIOR-1K radio task, Genie Sim 3.0 single/continuous package sorting, and a real XLeRobot pick-and-place task, reporting higher stage-wise success in all settings and qualitative logs showing the memory preventing premature stage transitions. Section 5.5 acknowledges the absence of confidence intervals and seed variation.","tokens_in":8170,"tokens_out":6258,"duration_ms":76267,"significance":"If the reported gains are real and attributable to the rolling language memory, the paper would offer an interpretable, compact alternative to frame buffers or latent-history states for long-horizon VLA control, together with a reusable supervision pipeline and a falsifiable stage-success comparison. The architecture also makes semantic tracking and action generation separately optimizable, and the memory logs are a genuine interpretability asset. However, the evaluation does not currently isolate the memory variable from the added high-level VLM and subtask conditioning, and the real-robot effect size is very close to the resolution of the evaluation (three additional stage successes out of 80). The central causal claim is therefore plausible but not yet established.","major_comments":[{"comment":"The main comparison changes at least three factors at once: it adds a separately trained high-level PaliGemma branch, introduces subtask supervision and conditioning, and then adds the rolling memory m_{t-1}. Consequently, the stage-success improvements in Tables 1–3 (e.g., 41.7% to 63.9% in Table 2) cannot be attributed specifically to language memory; they could come from extra semantic modeling capacity or from giving the action policy a concise predicted subtask. Please include a condition that retains the high-level VLM and subtask conditioning but removes the memory, for example by resetting m_t to an empty or constant value at every decision step, and report this condition on the same tasks. This ablation is load-bearing for the paper’s central claim.","section":"§5.1, Tables 1–3"},{"comment":"Equation (1) factorizes the policy as π_LL(a_{t:t+H} | o_t, s_t) · π_HL(s_t, m_t | o_t, m_{t-1}, l). This means memory can influence actions only through the generated subtask string. If a memory-free high-level VLM conditioned on (o_t, l) produces the same subtask distribution, the memory is functionally inert for control. The paper should quantify how often the memory and memory-free high-level branches disagree in subtask predictions, particularly at stage boundaries such as PickUp→PressOn in the radio task, or otherwise demonstrate that the subtask carries decision-relevant state that cannot be recovered from (o_t, l) alone.","section":"§3.1, Eq. (1)"},{"comment":"All reported stage-success results are point estimates without confidence intervals, seed variation, or per-trial counts. The real-robot result in Table 3 is 62.5% versus 66.3% over 20 trials, which corresponds to three additional stage successes out of 80 (53/80 versus 50/80); without intervals or raw trial outcomes this does not support the claim of improved robustness. Please report per-seed statistics for the simulations (at least three seeds) and exact trial-level outcomes or confidence intervals for the real-robot experiment. The limitation statement in §5.5 acknowledges the problem, but the central quantitative claims depend on it.","section":"§5.1 and §5.5"},{"comment":"The abstract and contributions claim that language memory 'reduces skipped steps,' but the only supporting evidence is a single qualitative trajectory in Fig. 4. No quantitative measure of skipped or repeated stages is reported. The stage-success tables measure final stage completion, not the order-confusion errors the paper emphasizes. Please report per-trial stage-transition trajectories or a metric such as the number of incorrect stage transitions, together with counts across all evaluation episodes.","section":"§5.2, Fig. 4"},{"comment":"The low-level branch is described as receiving 'the target or predicted subtask s_t' during training, while deployment uses predicted subtasks from π_HL. If training uses only ground-truth subtask labels, there is a train/test distribution shift when the high-level model outputs an imperfect subtask; the paper does not describe any mechanism (e.g., scheduled sampling, DAgger, or noise injection) to make the low-level policy robust to predicted subtasks. Please clarify whether predicted subtasks are ever used during low-level training and, if not, discuss how the architecture handles the resulting mismatch.","section":"§3.4 and §3.5"}],"minor_comments":[{"comment":"Equation (4) presents the low-level objective as an L2 regression against π_LL, but π0.5 uses a flow-matching action objective. If the implementation uses conditional flow matching, the equation should be corrected; if the implementation genuinely uses an L2 loss on flow output, that should be stated explicitly.","section":"§3.4, Eq. (4)"},{"comment":"The architecture diagram contains untranslated Chinese annotations and retains source-thesis figure numbering; readers cannot verify the intended relation between the annotated components and the surrounding Chinese text. Please provide a fully labeled English diagram.","section":"Fig. 2"},{"comment":"The real-robot experiment is described as 'sim-to-real,' but no simulation-trained model is transferred to the real platform; the real task appears to be trained directly on the 25 real trajectories. Please replace the phrase with a clearer description of the real-robot evaluation and state explicitly whether any weights trained in simulation were used on the real robot.","section":"§4.1"},{"comment":"The automated annotation pipeline is reported with two routes (Gemini 3.5 Flash API and local Qwen3.5-9B), but the experiments do not state which route produced the labels used in Tables 1–3. This detail matters for reproducibility of the memory and subtask supervision.","section":"§3.3"},{"comment":"Several numeric cells are missing separators and are rendered as a single number, for example '20.040.0' in Table 1, '25.063.9' and '12.546.9' in Table 2, and '55.066.3' in Table 3; these should be spaced or placed in separate columns. In addition, the tables do not report the number of evaluation episodes N, which should be given in the captions.","section":"Tables 1–3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be derived from a thesis, and several figures/captions retain thesis-style artifacts; the editors may wish to check that the presentation meets the journal's formatting standards. The main concern is evidentiary: the central causal role of language memory is not isolated by the current experiments, and the real-robot effect is nearly indistinguishable from noise. If the authors can add the no-memory hierarchy ablation and strengthen the statistics, the paper would likely become acceptable; without those changes, the claim that 'explicit language memory improves' VLA long-horizon performance is not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe short version: this paper has a genuinely useful idea — a rolling natural-language memory maintained by a high-level VLM that feeds subtasks to a low-level VLA — plus an automated supervision pipeline that makes it trainable from expert videos and timestamps. But the experiments do not yet isolate the memory variable, so the headline gains could plausibly come from the added semantic hierarchy rather than from memory per se.\n\nWhat's new: the rolling-compression memory (detailed info before a milestone completes, compressed summary after) and the VLM-based frame-aligned target generation are not in the cited work as far as I can tell. The architecture is clean: Eq. (1) separates the high-level semantic branch from the low-level action branch, and the two are trained independently. The qualitative logs in Fig. 4 show exactly the phase-confusion failure the memory is supposed to fix — the baseline skips the button press, the memory model doesn't. The limitations section is refreshingly honest: it admits no confidence intervals, no seed variation, a 20-trial real-robot eval, and that PressOn stays hard.\n\nWhere it's soft. The central claim — explicit language memory improves success rate and robustness — is not yet established, because the main comparison (π0.5 vs. π0.5 + memory) changes three variables at once: it adds a trained high-level PaliGemma, provides subtask supervision, and conditions the low-level policy on the predicted subtask. The rolling memory is only the third change. The stress-test note is right: Eq. (1) has π_LL depend only on o_t and s_t, so memory can affect actions only by changing the subtask string. If a memory-free high-level VLM conditioned on (o_t, l) predicts equally good subtasks, the memory is inert. That's not a fatal flaw — memory might be the mechanism that makes the subtask predictions good — but it's an empirical claim the paper doesn't test. No ablation removes memory while keeping the hierarchy. The missing ablations and statistics are addressable, and the architecture and idea are there.\n\nOne more thing: the manuscript has formatting artifacts — captions referencing \"the source thesis,\" garbled Chinese characters in Fig. 2, an oddly placed \"10\" — suggesting it's an excerpt rather than a polished submission. A referee would want that cleaned up.\n\nBottom line: this deserves a serious referee. The idea is timely, the supervision pipeline is a real contribution, and the weaknesses are fixable with proper ablations and seed variation. It's not ready as-is, but it's a legitimate paper to engage with.\n\nRecommendation: send it to peer review, and push for an ablation that holds the hierarchy fixed and turns memory off.","headline":"Rolling language memory for VLA is a genuinely new mechanism, but the main experiments don't isolate it from the added semantic hierarchy, so the headline gains are under-supported; still worth a serious referee.","tokens_in":8730,"tokens_out":2657,"would_cite":true,"duration_ms":29939,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Giving a robot a rolling written summary of completed milestones improves multi-step task success and produces readable progress logs.","keywords":["explicit language memory","long-horizon planning","vision-language-action models","hierarchical policy","phase consistency","robot manipulation","closed-loop inference","rolling text memory"],"falsifier":"Instrument stage transitions and count, across trials, how often each policy advances before the observation supports completion of the current milestone; if the memory policy's premature-advance rate is not below the no-memory baseline's, the central mechanism is refuted.","tokens_in":7717,"feed_emoji":"🤖","tokens_out":8230,"duration_ms":87913,"temperature":0.7,"pith_summary":"The paper tries to establish that representing a robot's task history as an explicit rolling natural-language memory improves long-horizon execution in vision-language-action models, and that the improvement comes from making stage transitions conditional on observed evidence. It separates the policy into a high-level vision-language model that recursively rewrites the memory and issues a short subtask, and a low-level action model that turns the subtask and current observation into an action chunk. In a household simulation, a humanoid manipulation simulation, and a real two-arm mobile robot, average stage-wise success rises (30.0% to 40.0%, 41.7% to 63.9%, 31.3% to 46.9%, and 62.5% to 66.3%). If this is right, long-horizon robot policies can be made more reliable and inspectable without storing the full visual history, and planning errors can be separated from motor errors.","feed_headline":"Text memory lifts long-horizon robot success by 22 points","feed_subtitle":"A running verbal summary of completed milestones stops skipped stages and makes robot decisions readable.","key_machinery":"The load-bearing mechanism is the recursive language-memory loop joined to the policy factorization in Eq. (1): $\\pi(a_{t:t+H}, s_t, m_t | o_t, m_{t-1}, l) = \\pi_{\\mathrm{LL}}(a_{t:t+H} | o_t, s_t) \\cdot \\pi_{\\mathrm{HL}}(s_t, m_t | o_t, m_{t-1}, l)$. The high-level branch updates the memory with a rolling-compression rule: keep target details before a subtask is done, then collapse them into a past-tense summary and shift attention to the next step. This yields a bounded textual state from which the low-level branch receives only the subtask, making the semantic state inspectable. The high-level branch is supervised with frame-aligned memory and subtask targets extracted from expert videos and skill timestamps.","core_discovery":"On its own terms, the paper's central claim is that explicit language memory improves phase consistency: the agent should advance to the next subtask only when the observation supports completion of the current milestone. The high-level branch maintains a rolling textual memory recording completed milestones, relevant object attributes, and next intent; the low-level branch generates continuous actions from the current observation and the subtask alone. The paper reports that this design reduces skipped or repeated stages, for example not moving to place the radio before pressing its button, and provides readable decision logs that localize failures. It explicitly limits the claim: language memory resolves what to do and which stage is active, not how to execute contact-rich motions, and the low button-pressing success rate remains a gap.","pith_inferences":["Because the memory is a separate textual variable, the same architecture could be adapted to a new task by changing the global instruction and skill annotations alone, possibly without retraining the high-level segment; the paper does not test this transfer.","A direct ablation would be to edit the memory string while holding the image fixed and observe whether the issued subtask and action chunk change; this would isolate how much behavior is driven by language state versus perception.","The paper's own suggestion of tactile and force feedback implies a testable extension: appending a single force-triggered sentence to memory could close the button-pressing gap that image-only memory leaves open."],"forward_implications":["Stage-wise success on the household radio task rises from 30.0% to 40.0%, with the largest gain at the pickup stage.","Single-package sorting rises from 41.7% to 63.9%, and the memory policy reaches all four upright placements in 12.5% of continuous-sorting evaluations while the baseline never reaches the third.","Real-robot pick-and-place average rises from 62.5% to 66.3%, showing that memory helps less when the task has few stages.","Failed grasps can be recorded as text and followed by a retry subtask without a separate recovery controller, an in-context adaptation behavior.","Readable memory and subtask outputs let failures be attributed to semantic planning versus motor execution rather than treated as a single end-to-end outcome."],"supporting_citations":[{"why":"Supplies the baseline and the two-stage pattern of semantic subtask prediction followed by a flow-matching action expert that the architecture extends.","marker":"[14]"},{"why":"Supplies the flow-matching action model used as the low-level action expert.","marker":"[3]"},{"why":"Supplies the vision-language backbone used for the high-level semantic branch.","marker":"[2]"},{"why":"Provides the household radio task and its stage-wise evaluation.","marker":"[10]"},{"why":"Provides the humanoid manipulation simulation and package-sorting tasks.","marker":"[19]"},{"why":"Provides the real low-cost dual-arm mobile robot used for sim-to-real evaluation.","marker":"[17]"},{"why":"Provides the software stack for real-robot teleoperation and action execution in the pick-and-place evaluation.","marker":"[6]"}],"fun_headline_variants":["Explicit memory log cuts robot stage skips","22-point robot gain from language memory","Readable memory for long-horizon robot planning","Text memory prevents robot subtask skipping","Phase-consistency memory lifts robot success 22pts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything the robot needs to remember must survive in the short subtask string passed to the low-level controller; if a decision-relevant detail such as a failed grasp, an object's color, or collision evidence is absent from that string, the language memory cannot influence the generated action.","fun_headline_variants_meta":{"raw":{"variants":["Explicit memory log cuts robot stage skips","22-point robot gain from language memory","Readable memory for long-horizon robot planning","Text memory prevents robot subtask skipping","Phase-consistency memory lifts robot success 22pts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000838,"raw_usage":{"total_tokens":3663,"prompt_tokens":964,"completion_tokens":2699,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":2629}},"tokens_in":580,"tokens_out":2699,"duration_ms":25268,"temperature":1.0,"reasoning_tokens":2629,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:08:42.443642+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Instrument stage transitions and count, across trials, how often each policy advances before the observation supports completion of the current milestone; if the memory policy's premature-advance rate is not below the no-memory baseline's, the central mechanism is refuted.","supporting_citations":[{"cited_title":"XLeRobot: A practical low-cost household dual-arm mobile robot design for general manipulation","cited_arxiv_id":null,"evidence_quote":"Provides the real low-cost dual-arm mobile robot used for sim-to-real evaluation."},{"cited_title":"Lerobot: State-of-the-art machine learning for real-world robotics in pytorch","cited_arxiv_id":null,"evidence_quote":"Provides the software stack for real-robot teleoperation and action execution in the pick-and-place evaluation."}],"review_version":1}