{"id":"1347e6ec-55ff-4d73-b729-b43e6dd9e1db","arxiv_id":"2607.07608","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":5,"one_line_summary":"LaMem-VLA reconstructs robotic history into dual short-term and long-term latent memory tokens that are woven directly into a VLA model's reasoning sequence to improve long-horizon manipulation.","lead":"This paper introduces LaMem-VLA, a robotic control architecture that stores, retrieves, and consumes historical experience as latent memory tokens directly within the vision-language-action model's reasoning sequence. A smart generalist might read it to understand how integrating memory into the native embedding space of a model improves performance on long-horizon robotic manipulation tasks.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The claimed superiority of latent-native memory over policy-side conditioning rests on 1–2 point margins that are within statistical noise for the evaluation protocols used (24 trials/task on SimplerEnv, 50 rollouts/task on LIBERO); no confidence intervals or significance tests are reported.","rationale":"The reader correctly identifies the condenser's information preservation as a potential weakness, and the simulation-only limitation is valid. However, the more load-bearing concern is statistical: the improvement margins that support the paper's central architectural claim are small enough to be indistinguishable from evaluation noise given the trial counts used. The reader's verdict of CONDITIONAL is appropriate, but the reasoning should foreground the absence of significance testing rather than focusing primarily on the condenser's compression fidelity. The paper's ablations (Tables 3–5) are well-structured and the architecture is clearly motivated, which is why I don't recommend REJECT — but the evidence base needs stronger statistical support before the core claim can be accepted. The reader's concerns about code release and real-world experiments are secondary but also valid; they would partially address the statistical concern if real-world experiments used larger sample sizes with proper error bars.","tokens_in":17287,"tokens_out":1832,"duration_ms":88802,"concrete_test":"Re-run the SimplerEnv-Bridge evaluation with at least 100 trials per task (instead of 24) and report bootstrap 95% confidence intervals for the key comparisons in Tables 1 and 4 — specifically LaMem-VLA vs. MemoryVLA (Table 1) and LaMem-VLA vs. Policy-side Memory vs. Raw Retrieval Conditioning (Table 4). If the confidence intervals for these pairs overlap after the increased sample size, the central claim that latent-native integration outperforms policy-side conditioning is not supported by the current evidence.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central architectural claim of LaMem-VLA — that weaving memory into the VLA latent reasoning sequence outperforms policy-side memory conditioning — is supported by very small margins. On LIBERO, the headline 97.6% vs MemoryVLA's 96.5% is a 1.1-point difference. On SimplerEnv-Bridge, 73.9% vs 71.9% is 2.0 points. The critical Table 4 ablation comparing latent-native integration (73.9%) to policy-side memory (71.9%) and raw retrieval conditioning (69.8%) shows differences of 2.0 and 4.1 points respectively on SimplerEnv, where each task is evaluated over only 24 trials (4 tasks × 24 = 96 total trials). For a binomial proportion at p≈0.72 with n=96, the standard error is ≈4.6 percentage points, meaning a 2.0-point difference is well within one standard error. Even the 4.1-point gap to raw retrieval conditioning is within the 95% confidence interval. On LIBERO, the per-suite differences against MemoryVLA (e.g., 2.4 points on Long-10 with 10 tasks × 50 rollouts, 1.4 points on Long-90) are similarly small relative to binomial noise. The paper also selects the best checkpoint by validation success, which introduces optimistic bias that compounds with the small margins. Without confidence intervals, bootstrap tests, or larger sample sizes, it is unclear whether the architectural innovation (latent-native weaving vs. policy-side conditioning) produces a real performance difference or whether the reported gains reflect evaluation noise and checkpoint selection bias. This is more fundamental than the reader's concern about the condenser preserving information: even if the condenser works perfectly, the evidence that the entire approach is superior to simpler alternatives is statistically inconclusive.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This paper introduces LaMem-VLA, a memory-augmented VLA framework that organizes historical experience into dual-scale (short-term visual, long-term semantic) memory vaults, compresses retrieved evidence into fixed-length latent memory tokens via a condenser, and weaves these tokens directly into the VLA reasoning sequence before action query resolution. The central architectural claim is that integrating memory in the native latent embedding space of the VLA (rather than as auxiliary policy-side context) yields better long-horizon manipulation performance. The system is evaluated on SimplerEnv-Bridge (73.9% average success) and LIBERO (97.6% average success), with ablations isolating the dual-memory design, the latent-native integration path, and key hyperparameters (K, L_s, L_l).","tokens_in":18325,"tokens_out":1449,"duration_ms":326873,"significance":"The paper addresses a well-motivated problem: the temporal short-horizon bias of Markovian VLA models in long-horizon manipulation. The architectural design — particularly the condenser mechanism (Eq. 8) that maps variable-length retrieved evidence into fixed-length latent tokens compatible with the VLA embedding space — is a concrete and falsifiable contribution. The ablation in Table 4, which compares latent-native integration against policy-side conditioning and raw retrieval, directly tests the central claim. The dual-scale memory factorization (short-term visual vs. long-term semantic) is a reasonable design choice, and Table 3 provides evidence for complementarity. However, the empirical margins supporting the headline claim of 'superiority' are small (1.1–2.0 points over the closest baseline), and the evaluation is exclusively in simulation with no confidence intervals or significance tests reported. These issues are load-bearing for the paper's central contribution and are detailed below.","major_comments":[{"comment":"§4.4, Table 4: The paper's central architectural claim — that latent-native memory weaving outperforms policy-side memory conditioning — rests on a 2.0-point margin on SimplerEnv (73.9% vs. 71.9%) and a 2.2-point margin on LIBERO-90 (97.0% vs. 94.8%). On SimplerEnv, each of the 4 tasks is evaluated over 24 trials (96 total). For a binomial proportion at p≈0.73 with n=96, the standard error is approximately 4.5 percentage points, meaning the 2.0-point difference is well within one standard error. No confidence intervals, bootstrap tests, or multiple-seed runs are reported. The paper should either (a) report confidence intervals or significance tests for all pairwise comparisons in Tables 1–4, or (b) increase the number of evaluation trials to a level where the reported margins are statistically distinguishable. Without this, the claim of 'superiority' (used in the abstract, §4.2, §4.3, §4","section":null},{"comment":"§4.2 and §4.3: Both evaluation protocols select the best checkpoint by validation success before reporting test results. This introduces optimistic selection bias that compounds with the small margins. The paper should report results for the final checkpoint (or an average over the last few checkpoints) in addition to the best-validation checkpoint, or discuss the sensitivity of the reported margins to checkpoint selection.","section":null},{"comment":"§4.2, Table 1: On SimplerEnv-Bridge, LaMem-VLA underperforms MemoryVLA on the 'Stack Cube' task (41.7% vs. 37.5% is actually a gain, but 41.7% vs. MemoryVLA's 37.5% is only a 4.2-point difference on 24 trials) and on 'Put Eggplant in Basket' (95.8% vs. 100.0%, a 4.2-point regression). The paper does not discuss these per-task regressions. Given that the headline margin over MemoryVLA is 2.0 points, these per-task trade-offs are important for assessing whether the architectural change produces a genuine improvement or merely redistributes errors across tasks. A per-task discussion would strengthen the evaluation.","section":null}],"minor_comments":[{"comment":"§3.3, Eq. (3): The notation uses m_s = (k_s, v_s) for a single memory unit, but the vault is initialized as {m_i^s}_{i=1}^L. The subscript/superscript convention is inconsistent — m_s appears without an index in Eq. (3) but with index i in Eq. (4). Please unify.","section":null},{"comment":"§3.3, Eq. (4)–(5): The redundancy-based merging strategy selects the most similar adjacent pair by cosine similarity of keys. This could systematically merge distinct but visually similar states (e.g., same object configuration at different task phases). A brief discussion of this risk, or an alternative merging criterion, would improve the design rationale.","section":null},{"comment":"§3.4, Eq. (8): The condenser uses learnable memory slots T_s and T_l that are updated by transformer-style memory formers F_v and F_c. The relationship between these slots and the Perceiver/Q-Former-style cross-attention literature should be cited, as the mechanism appears closely related.","section":null},{"comment":"Table 2: Several baselines (e.g., Diffusion Policy, Octo, CoT-VLA) lack LIBERO-90 results, and the table notes that averages are computed over the first four suites for these methods. This makes the 'Avg. Success' column not directly comparable across all rows. A footnote or separate column for the first-four-suite average would improve clarity.","section":null},{"comment":"Figure 3: The y-axis labels and legend are small and difficult to read. Increasing font size and adding gridlines would improve readability.","section":null},{"comment":"§4.1: The learning rate (2×10⁻⁵), batch size (256), and training steps (20k–40k) are reported, but the total number of model parameters and the parameter counts of the memory modules (B, F_v, F_c) are not specified. Please include these for reproducibility.","section":null},{"comment":"The paper states 'The project page will be available at LaMem-VLA' but no URL is provided. Including a code release or at least a committed URL would strengthen reproducibility.","section":null}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about statistical significance is well-founded and is the primary reason for the major_revision recommendation. The architectural design is reasonable and the ablation structure is appropriate, but the empirical evidence as presented does not distinguish the proposed method from the closest baseline (MemoryVLA) at a statistically meaningful level. If the authors can address the significance testing concern (either through CIs, multiple seeds, or larger sample sizes), the paper could move to minor_revision. The simulation-only evaluation is a limitation but is acknowledged in the paper and is not, by itself, a reason for rejection."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive review. The referee raises three major concerns: (1) the absence of confidence intervals or significance tests given small empirical margins, (2) potential optimistic selection bias from best-checkpoint reporting, and (3) the need for per-task discussion of regressions. All three points are well-taken and we will address them in the revision. We provide point-by-point responses below.","responses":[{"response":"We agree that the absence of confidence intervals or significance tests is a legitimate concern, and we accept this point. The referee's binomial standard error calculation is correct: at p≈0.73 with n=96, the standard error is approximately 4.5 percentage points, meaning a 2.0-point difference is not statistically distinguishable at conventional thresholds on SimplerEnv alone. We will address this in the revision through two complementary actions. First, we will increase the number of evaluation trials on SimplerEnv-Bridge from 24 to 50 per task (200 total), which reduces the standard error to approximately 3.1 percentage points and provides a more reliable basis for comparison. Second, we will report Wilson score confidence intervals for all pairwise comparisons in Tables 1-4 and add bootstrap-based significance tests for the key comparisons (LaMem-VLA vs. Policy-side Memory, LaMem-VLA vs. MemoryVLA). We will also run the critical ablation comparisons (Table 4) with three random seeds and report mean ± standard deviation. We acknowledge that if the margins remain within confidence intervals even after increased trials, we will soften the language from 'superiority' to 'consistent improvement' or similar hedged phrasing in the abstract and relevant sections. We note that the LIBERO evaluation already uses 50 rollouts per task across 90 tasks (4,500 total rollouts for Long-90), where the 1.4-point margin over MemoryVLA corresponds to approximately 63 additional successful rollouts, which is more statistically grounded, though we will still report confidence intervals there for completeness.","revision_made":"yes","referee_comment":"The paper's central architectural claim rests on small margins (2.0 points on SimplerEnv, 2.2 points on LIBERO-90) without confidence intervals, significance tests, or multiple-seed runs. The referee requests either (a) confidence intervals/significance tests for all pairwise comparisons in Tables 1-4, or (b) increased evaluation trials to a level where margins are statistically distinguishable."},{"response":"This is a fair point. Best-checkpoint selection on validation success can introduce optimistic bias, and we should address this directly. In the revision, we will report results for both the best-validation checkpoint and the final checkpoint for all main comparisons in Tables 1-2, and we will additionally report the average of the last three checkpoints (at 1k-step intervals for LIBERO and 2.5k-step intervals for SimplerEnv) to provide a more robust estimate. We will also add a brief discussion of the sensitivity of the reported margins to checkpoint selection. If the margins change substantially under final-checkpoint or averaged-checkpoint reporting, we will report this transparently and adjust our claims accordingly. We note that the ablation comparisons in Tables 3-5 were conducted using the same checkpoint selection protocol, so the relative comparisons within the ablation should be less affected by selection bias, though we will verify this as well.","revision_made":"yes","referee_comment":"Both evaluation protocols select the best checkpoint by validation success before reporting test results, introducing optimistic selection bias that compounds with small margins. The referee requests reporting results for the final checkpoint or an average over the last few checkpoints, or discussion of sensitivity to checkpoint selection."},{"response":"We agree that the per-task trade-offs should be discussed rather than left implicit. In the revision, we will add a per-task analysis paragraph in Section 4.2. Specifically, we will note that LaMem-VLA shows notable improvements on 'Put Spoon on Towel' (83.3% vs. 75.0%, +8.3) and 'Put Carrot on Plate' (75.0% vs. 75.0%, tied), a modest gain on 'Stack Cube' (41.7% vs. 37.5%, +4.2), and a regression on 'Put Eggplant in Basket' (95.8% vs. 100.0%, -4.2). We will discuss the hypothesis that the Eggplant regression may reflect a ceiling effect — at 100% success, MemoryVLA has no room for improvement, and the 4.2-point drop represents a single failed trial out of 24 — while the gains on the more challenging Spoon and Carrot tasks (where both methods have substantial room for improvement) suggest that latent-native memory integration provides more benefit when the task demands richer temporal reasoning rather than when a strong Markovian policy already suffices. We will also acknowledge that, as the referee correctly notes, the per-task trade-offs mean the architectural change does not uniformly improve all tasks, and we will temper the headline claims accordingly. With the increased trial count (50 per task) that we are adding per the first comment, the per-task numbers will also become more reliable.","revision_made":"yes","referee_comment":"On SimplerEnv-Bridge, LaMem-VLA shows per-task regressions relative to MemoryVLA on some tasks (e.g., Put Eggplant in Basket: 95.8% vs. 100.0%, a 4.2-point regression), while the headline margin is only 2.0 points. The referee requests a per-task discussion of these trade-offs."}],"tokens_in":17584,"tokens_out":1198,"duration_ms":215308,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"The core idea here is genuinely new: instead of bolting memory onto the VLA as auxiliary policy-side context, LaMem-VLA compresses retrieved historical experience into fixed-length latent tokens that sit directly in the model's input sequence alongside observation, instruction, and action query tokens. That is a clean architectural contribution. The four-module pipeline (curator, seeker, condenser, weaver) is well-motivated, and the ablations in Table 4 do isolate the right comparison — latent-native weaving vs. policy-side conditioning vs. raw retrieval. The dual short/long-term memory split is reasonable and the ablation in Table 3 shows both streams contribute. Credit where earned: the design is coherent and the ablation structure is more careful than most in this area. The gains over the memory-free baseline (57.3 → 73.9 on SimplerEnv) are large enough to take seriously. The stress-test concern about statistical significance is the real problem. The headline margins over MemoryVLA are 1.1 points on LIBERO and 2.0 on SimplerEnv-Bridge. The critical Table 4 ablation — the paper's central architectural claim — shows 2.0 points over policy-side conditioning on SimplerEnv, where each task gets 24 trials across 4 tasks (96 total). At p≈0.72, n=96, the standard error is roughly 4.6 points. A 2.0-point difference is well within one standard error. Even the 4.1-point gap to raw retrieval is within a 95% CI. On LIBERO, the per-suite margins against MemoryVLA (2.4 on Long-10, 1.4 on Long-90) are similarly small relative to binomial noise at 50 rollouts/task. No confidence intervals, no bootstrap tests, no multiple seeds. Checkpoint selection by best validation adds optimistic bias on top. The reader's concern about the condenser discarding information is less pressing — Table 3 shows removing memory streams hurts substantially, so the condenser is clearly preserving useful signal. The real issue is that the paper cannot distinguish its central claim (latent-native > policy-side) from evaluation noise at the sample sizes used. The architecture is interesting and the direction is right. But the evidence for the specific claim of superiority over simpler memory conditioning is statistically inconclusive. A serious referee should ask for either larger sample sizes, multiple seeds with reported variance, or at minimum bootstrap CIs on the key comparisons in Table 4. Simulation-only evaluation is a secondary concern; the statistical issue is primary.","headline":"Architecturally novel VLA memory integration, but the key empirical margins are within statistical noise.","tokens_in":18384,"tokens_out":589,"would_cite":false,"duration_ms":81704,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Memory woven into robot reasoning, not bolted on","keywords":["vision-language-action models","robotic manipulation","latent memory","long-horizon tasks","memory-augmented reasoning","diffusion policy","dual memory systems"],"falsifier":"If one could show that a task exists where the relevant historical information cannot be compressed into 12 tokens without loss, and that raw retrieval conditioning (bypassing the condenser) outperforms latent-native weaving on that task, the core claim that compression-then-injection is superior to direct-injection would be falsified.","tokens_in":17448,"feed_emoji":"","tokens_out":1150,"duration_ms":204670,"temperature":0.7,"pith_summary":"LaMem-VLA argues that a robot's past experience should live inside the same latent token stream where it perceives, reasons, and acts—not in an external memory bank consulted afterward. The paper builds a four-stage pipeline: a curator splits history into a short-term visual vault and a long-term semantic vault; a seeker retrieves relevant evidence using the current observation-and-instruction context; a condenser compresses that evidence into fixed-length latent memory tokens; and a weaver inserts those tokens directly into the VLA's input sequence alongside current image, instruction, and action-query tokens. Because the memory tokens share the same embedding space as the rest of the reasoning pipeline, they participate in self-attention and shape action-token formation before any action is decoded. The paper reports 97.6% average success on LIBERO and 73.9% on SimplerEnv-Bridge, with ablations showing that latent-native injection outperforms both policy-side conditioning and raw retrieval, and that both short-term and long-term streams contribute complementary gains.","feed_headline":"Robot memory woven into reasoning, not bolted on","feed_subtitle":"LaMem-VLA compresses past experience into latent tokens that share the VLA's embedding space, outperforming external memory banks on long-ho","key_machinery":"Four-module pipeline (curator → seeker → condenser → weaver) operating entirely in the VLA's C-dimensional latent space; dual vaults (short-term visual key-value pairs, long-term action hidden states); cosine-similarity Top-K retrieval; SE-bottleneck compression for short-term keys; transformer-based query builder and memory formers with masked attention; diffusion-based action expert conditioned on memory-grounded action tokens.","core_discovery":"The central mechanism is the latent memory condenser (Eq. 8): it takes potentially lengthy retrieved evidence from two vaults and reconstructs it into a bounded set of fixed-length tokens (8 short-term, 4 long-term) that are dimensionally compatible with the VLA backbone's embedding space. These tokens are then prepended to the standard input sequence (Eq. 9), allowing historical experience to flow through the same self-attention layers that process the current observation and instruction. The paper's ablation in Table 4 isolates this design choice: raw retrieval conditioning (69.8% SimplerEnv) and policy-side memory conditioning (71.9%) both underperform the full latent-native weaving (73.9","pith_inferences":["The condenser's fixed-length bottleneck (8 + 4 tokens) implicitly defines a 'memory bandwidth' for the VLA's reasoning context. If this architecture is correct, there should exist tasks where the critical historical information exceeds this bandwidth and performance degrades regardless of retrieval quality—a testable prediction the paper does not explore.","The paper's claim that memory should be 'context-native' rather than 'policy-side' parallels arguments in language model architectures about where to inject retrieved knowledge (input embeddings vs. cross-attention vs. output conditioning). The robotics-specific contribution is showing that for action generation, the injection point matters more than the retrieval quality, which could inform memor","The short/long-term vault distinction is motivated functionally (visual vs. semantic) but implemented representationally (visual tokens vs. action hidden states). This conflation of content type with token type may not always hold: a long-horizon visual change (e.g., an object gradually moving) would be stored as semantic action tokens rather than visual evidence, potentially losing spatial detail"],"forward_implications":["If latent-native memory injection is the right paradigm, then the boundary between perception, memory, and action planning in VLA models dissolves: all three become different token types in one continuous reasoning stream, and the architectural question shifts from 'how do we condition the policy on history?' to 'how do we allocate tokens in the reasoning budget?'","The dual-vault split (visual short-term + semantic long-term) suggests that robot memory may need fundamentally different representational substrates for different temporal scales, even when they share a common output interface—a principle that could extend to other embodied AI domains beyond manipulation.","The fixed-length condenser design implies that the quality of historical compression, not the quantity of stored history, is the binding constraint on long-horizon performance—shifting the engineering bottleneck from storage and retrieval to compression fidelity.","If the gains hold in real-world deployment (currently simulation-only), latent memory weaving could become a standard component rather than an add-on, since it requires no additional input modalities (no wrist camera, no proprioception) yet improves over methods that use them."],"fun_headline_variants":["Latent memory tokens let robots reason over past actions","Short and long-term memory compressed into VLA embeddings","Robotic policy conditions on fixed-length latent memory","Dual memory vaults feed compact tokens into VLA reasoning","History as latent tokens, not external retrieval context"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The condenser's compression of retrieved historical evidence into just 12 fixed-length latent tokens (8 short-term + 4 long-term) must preserve all task-relevant information. If critical historical cues are discarded during this compression step, the entire architectural advantage over raw retrieval or policy-side conditioning collapses.","fun_headline_variants_meta":{"raw":{"variants":["Latent memory tokens let robots reason over past actions","Short and long-term memory compressed into VLA embeddings","Robotic policy conditions on fixed-length latent memory","Dual memory vaults feed compact tokens into VLA reasoning","History as latent tokens, not external retrieval context","Weaving recalled experience into the VLA embedding sequence","Bounded latent memory outperforms raw retrieval for VLA","Two memory vaults, one continuous latent space for action"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":758,"prompt_tokens":608,"completion_tokens":150,"prompt_tokens_details":null},"tokens_in":608,"tokens_out":150,"duration_ms":6508,"temperature":1.0,"reasoning_tokens":39,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T05:13:01.523351+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If one could show that a task exists where the relevant historical information cannot be compressed into 12 tokens without loss, and that raw retrieval conditioning (bypassing the condenser) outperforms latent-native weaving on that task, the core claim that compression-then-injection is superior to direct-injection would be falsified.","supporting_citations":[],"review_version":1}