{"id":"c9881284-6236-4ace-ba78-b9984762283b","arxiv_id":"2607.08724","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":8,"one_line_summary":"Variable-length autoregressive latent sequences, trained as variational inference with a PPO-style objective, give robot policies adaptive test-time compute and yield a reusable action tokenizer.","lead":"This paper introduces a method where robot control policies generate variable-length sequences of latent 'reasoning' tokens before deciding actions, trained via a reinforcement-learning-style variational inference objective. A smart generalist might read it because it brings the adaptive 'thinking time' of language models to robotics without requiring natural-language intermediate steps.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Adaptive compute and optimization stability are coupled through the variance schedule, making the central feature fragile to a single hyperparameter.","rationale":"The reader's verdict of CONDITIONAL with HIGH confidence is appropriate. The optimization fragility is real and acknowledged by the authors, but it does not invalidate the framework — the ELBO derivation (Appendix B.1) is correct, the PPO-style surrogate is a standard and reasonable approximation, and the empirical results across multiple domains are genuinely strong (especially the tokenizer results in Fig 5 and the multi-task LIBERO-90 improvements). The concern is about robustness and generalization of the training pipeline, not about correctness of the core formulation. The paper is transparent about limitations and provides useful ablations (Table 3, Fig 7) that characterize the sensitivity. A conditional verdict captures this: the framework is sound and promising, but practical adoption hinges on whether the variance schedule and RL stabilization can be made less sensitive. The reader's weakest_assumption correctly identified this as the load-bearing concern.","tokens_in":27791,"tokens_out":3859,"duration_ms":259841,"concrete_test":"Train LMP-π on LIBERO-90 and DROID with a single shared variance schedule (e.g., σ_min=0.02 for both) and report success rates. If the performance gap between the two domains exceeds 5 absolute points compared to their domain-tuned schedules, the method requires per-domain tuning of the adaptive-compute mechanism, undermining the claim that adaptive reasoning emerges robustly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader correctly identifies optimization fragility as the key concern. I sharpen it: the variance schedule σ(T(z)) = γ^{T(z)} σ₀ (Sec 3.1) serves a dual role — it is both the mechanism that induces adaptive test-time compute AND a source of optimization instability. Table 3 shows that changing σ_min from 0.01 to 0.05 drops LIBERO-90 overall success from 0.933 to 0.870 and bottom-10 from 0.645 to 0.450 — a 20-point swing on the hardest tasks. Fig 7 further shows a narrow sweet spot: too little compression (0.1–0.08) causes latents to be ignored entirely (flat posterior error), while too much (0.2–0.02) raises reconstruction error. This means the method's defining feature (adaptive compute allocation) is tightly coupled to a hyperparameter that must be tuned per-domain (DROID uses 0.2–0.02, LIBERO uses 0.1–0.01). The stepwise KL decomposition (Eq 3) and free-nats clipping (Eq 7) are necessary but not sufficient stabilizers — the paper acknowledges collapse risk in Limitations. The concern is not that the framework is wrong, but that the central empirical claim ('strong performance with adaptive compute') holds only within a narrow operating regime that requires per-task tuning of the very parameter that controls adaptivity. This weakens the claim that adaptive reasoning 'emerges' — it emerges only when the variance schedule is carefully set.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"The paper introduces Latent Memory Palace (LMP), a framework that formulates iterative reasoning for robotic control as variational inference over a variable-length autoregressive latent distribution. The latent sequence terminates via an EOS token, enabling adaptive test-time compute. The variational lower bound is optimized via a PPO-style clipped surrogate with a stepwise KL decomposition (Appendix B.1). A decoder variance schedule that decays with latent length imposes an effective length penalty, inducing adaptive computation. The paper presents two instantiations: LMP-π (a control policy) and LMP-tok (a variable-length action tokenizer). Empirically, LMP-π outperforms Diffusion Policy on DROID and LIBERO-90, and LMP-tok outperforms VQ-VAE, FAST, and OAT on downstream autoregressive policy tasks. The core variational inference formulation is principled, the stepwise KL derivation is clean, and the adaptive compute behavior is an emergent rather than fitted property. The main concern is the coupling between the variance schedule (which drives adaptivity) and optimization stability, which the authors themselves flag as a limitation.","tokens_in":28830,"tokens_out":1249,"duration_ms":273350,"significance":"The paper makes a genuine conceptual contribution by connecting variational inference with autoregressive latent sequences for robotic control, a domain that lacks a natural discrete reasoning medium. The stepwise KL decomposition (Eq. 3, Appendix B.1) is a useful technical contribution that reduces variance in the RL objective. The empirical results are strong: LMP-π shows consistent gains over Diffusion Policy on real-world DROID tasks (Table 1) and on LIBERO-90 bottom-10 tasks (Table 2), and LMP-tok is the only tokenizer achieving nontrivial performance on the high-precision RoboMimic tool-hang task (Table 7). The adaptive compute analysis (Figs. 9–10) provides interpretable evidence that the model allocates fewer reasoning steps during gripper movements, correlating with irreducible action uncertainty. Code and detailed hyperparameters are provided (Table 5, Algorithms 1–2), supporting reproducibility.","major_comments":[{"comment":"Sec. 3.1 and Table 3: The variance schedule σ(T(z)) = γ^{T(z)} σ₀ serves a dual role—it is both the mechanism inducing adaptive test-time compute and a source of optimization fragility. Table 3 shows that changing σ_min from 0.01 to 0.05 drops LIBERO-90 overall success from 0.933 to 0.870 and bottom-10 from 0.645 to 0.450—a 20-point swing on the hardest tasks. Fig. 7 further reveals a narrow operating regime: insufficient compression (0.1–0.08) causes latents to be ignored entirely, while excessive compression (0.2–0.02) raises reconstruction error. The variance schedule is tuned per-domain (DROID: 0.2–0.02, LIBERO: 0.1–0.01). This coupling means the central claim that adaptive reasoning 'emerges' holds only within a manually tuned operating window. The authors should either (a) provide a principled procedure for selecting (σ_max, σ_min) without per-domain tuning, or (b) reframe the 'emG","section":null}],"minor_comments":[{"comment":"Sec. 3.2, Eq. (4): The notation CS(r, r̄; x) and the clipped surrogate are compact but dense. A brief inline explanation of the trust-region intuition would help readers from the robotics community who may not be familiar with PPO-style objectives.","section":null},{"comment":"Table 5: The reconstruction loss coefficient differs by 10× between multi-task (0.01) and single-task (0.1) LMP-π. The paper does not explain this choice. A brief note on the sensitivity or rationale would improve reproducibility.","section":null},{"comment":"Fig. 7: The labels '0.1 0.08 (1.25×)' and '0.1 0.01 (10×)' are ambiguous. Clarifying that these refer to σ_max/σ_min ratios would help.","section":null},{"comment":"Sec. 4.1: The paper states both methods are '1B-parameter' policies but does not provide a parameter breakdown. Confirming that the comparison is parameter-matched (including observation encoders) would strengthen the claim.","section":null},{"comment":"Table 6: On robomimic-square and robomimic-transport, LMP-π underperforms Diffusion Policy (0.87 vs. 0.92 and 0.75 vs. 0.80). The paper does not discuss these cases. A brief analysis of when LMP-π underperforms would balance the presentation.","section":null},{"comment":"Sec. 3.3: The claim that LMP-tok can decode 'arbitrary length' partial sequences, unlike FAST, is interesting but the practical benefit is not quantified. Showing downstream policy performance with truncated token sequences would strengthen this comparison.","section":null},{"comment":"Fig. 14: The comparison to DeepSeek R1's increasing reasoning length is suggestive but not substantiated. The curves also start from initialization, making the early dynamics hard to interpret.","section":null},{"comment":"Appendix A.3: The world action model variant (LMP-wam) is mentioned but only briefly evaluated (Table 4). Either expand the analysis or clearly mark this as a preliminary appendix-only extension.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The reader's stress-test concern about the variance schedule coupling is valid and well-grounded in Table 3 and Fig. 7, but it does not constitute a load-bearing error—the framework is sound, the empirical results are strong, and the authors are transparent about the limitation. The concern is better addressed by requiring a sensitivity analysis or principled selection procedure rather than a fundamental revision. The paper is a solid contribution to the intersection of latent reasoning and robotic control."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"The paper you should know about: Zhu et al. formulate reasoning for continuous control as variational inference over a variable-length autoregressive latent sequence. The latent tokens are discrete, terminated by an EOS token, and the whole ELBO is optimized with a PPO-style clipped surrogate over latent trajectories. That is the core idea, and it is genuinely new. Nobody else has stitched together autoregressive discrete latents, a length-based variance penalty, and RL-style optimization for behavior cloning in quite this way. The stepwise KL decomposition (Appendix B.1) is clean and correct. The action tokenizer variant (LMP-tok) is a natural and useful byproduct — it outperforms VQ-VAE, FAST, and OAT on RoboMimic and LIBERO when controlling for the downstream policy, and it is the only tokenizer that gets nontrivial performance on the tool-hang task. Real-world DROID results and 1B-parameter multitask LIBERO results are solid, with LMP-π showing a higher floor than Diffusion Policy on the hardest tasks. Credit where earned: the framework is principled, the experiments are substantial, and the adaptive compute allocation is genuinely interpretable — fewer latent steps during gripper movements, where timing variability is irreducible. That is a real finding, not a narrative bolted on after the fact. The soft spot is the one the stress-test flags and the authors half-acknowledge: the variance schedule σ(T) = γ^T σ₀ does double duty as both the adaptivity mechanism and a load-bearing hyperparameter. Table 3 shows a 20-point swing on LIBERO bottom-10 tasks when σ_min goes from 0.01 to 0.05. Figure 7 shows a narrow sweet spot between latents being ignored entirely and reconstruction error blowing up. DROID and LIBERO use different schedules. The authors admit collapse risk in the Limitations section. The stepwise KL decomposition and free-nats clipping help, but the stabilization story is incomplete — you need a large rollout buffer, per-domain variance tuning, and careful regularization, and the paper does not demonstrate that the recipe transfers without retuning. The stopping criterion analysis (Appendix B.2) is explicitly heuristic, dropping the KL term, so the theoretical grounding for when computation stops is suggestive rather than rigorous. None of this breaks the framework. The ELBO derivation is sound, the RL optimization is a reasonable tractability move, and the empirical results are strong enough to take seriously. The concern is about robustness and generality of the training pipeline, not about whether the core idea works. This paper is for people working on generative control policies, action representation learning, or test-time compute for robotics. It deserves a serious referee who can evaluate the optimization stability claims and push for either a more automatic variance schedule or evidence that the current one transfers across domains without retuning.","headline":"Novel variational framework for adaptive latent reasoning in robot control, with real empirical wins but a fragile optimization regime.","tokens_in":28833,"tokens_out":655,"would_cite":true,"duration_ms":116030,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Robots reason in a latent memory palace before acting","keywords":["latent reasoning","variational inference","robot control","autoregressive latent variable model","adaptive test-time compute","action tokenization","imitation learning","reinforcement learning"],"falsifier":"If the adaptive latent length provides no benefit beyond what a fixed-length latent of the same total capacity provides, or if the performance gains vanish when the variance schedule is replaced by a fixed variance, then the core claim that iterative adaptive computation drives the results would be undermined.","tokens_in":27975,"feed_emoji":"🧠","tokens_out":1127,"duration_ms":2024080,"temperature":0.7,"pith_summary":"The paper claims that a robot control policy can learn to deliberate before acting by generating a variable-length chain of discrete latent tokens—an internal sequence of intermediate computations that terminates when the policy decides it has thought enough. The key mechanism is to cast this deliberation as variational inference over an autoregressive latent distribution: a posterior network encodes expert demonstrations into latent token sequences, a prior network generates those sequences from observations alone at test time, and a decoder maps the resulting latent chain to continuous actions. To make the variable-length structure adaptive rather than arbitrary, the decoder variance shrinks with each additional latent step, so the model only spends more compute when doing so meaningfully sharpens the predicted action. The authors derive a tractable training objective by reformulating the variational lower bound as a reinforcement learning problem over latent trajectories, using a PPO-style clipped surrogate with a stepwise KL decomposition for variance reduction. The same framework, with observation conditioning removed, yields a variable-length action tokenizer that compresses continuous actions into discrete tokens for downstream autoregressive policies. Empirically, the policy matches or outperforms diffusion-based policies on real and simulated manipulation, and the tokenizer outperforms existing tokenization schemes when paired with the same downstream model.","feed_headline":"Robots learn to think before they act in a latent space","feed_subtitle":"A variable-length chain of discrete latent tokens lets control policies deliberate adaptively—spending more compute when actions are precise","key_machinery":"Autoregressive variational inference over discrete latent tokens with EOS termination; decoder variance decay as a length penalty; PPO-style clipped surrogate with stepwise KL decomposition and free-nats clipping for tractable optimization; uniform-interpolated regularization toward a length-aware uniform prior to prevent posterior collapse.","core_discovery":"The central object is the Latent Memory Palace (LMP): a variable-length autoregressive latent distribution over discrete tokens, trained via a variational lower bound that is optimized as a latent-space RL problem. The adaptive stopping behavior arises from a decoder variance schedule that decays with latent length, creating an implicit compression penalty—the model extends its latent chain only when the reduction in action prediction error outweighs the cost of an additional step. This same variational structure, stripped of observation conditioning, functions as a variable-length action tokenizer whose partial token sequences always decode to valid actions, unlike fixed-matrix tokenizers.","pith_inferences":["If the compression penalty is what drives adaptive stopping, then the choice of variance decay rate directly controls the compute-accuracy trade-off at inference time, suggesting a single trained model could be steered toward faster or more deliberate behavior by adjusting the variance schedule at test time without retraining.","The fragility of the optimization (noted by the authors) suggests that the discrete autoregressive latent space may be replaceable by a continuous chain of Gaussians resembling a diffusion process, which the authors mention as a future direction; such a replacement could inherit the stable training dynamics of diffusion models while retaining adaptive depth.","The mode-seeking behavior observed when latent steps are truncated hints that the RL-trained posterior learns a structured latent space where shorter prefixes correspond to prototypical actions and longer prefixes carve out specific modes—a property that could be exploited for hierarchical planning or curriculum learning."],"forward_implications":["If adaptive latent reasoning improves control policies, then the principle generalizes to any sequential decision-making domain where a policy must trade computation against precision—autonomous driving, surgical robotics, or game-playing agents could benefit from variable-length internal deliberation before committing to actions.","The decoder variance schedule as an implicit compression penalty offers a template for inducing adaptive compute in other generative models: any autoregressive latent variable model could adopt a similar shrinking-variance trick to make its depth data-dependent.","The finding that the same variational framework yields both a policy and a tokenizer suggests a deeper unity between reasoning and representation: the mechanism that lets a model deliberate before acting is the same mechanism that lets it compress actions into tokens for downstream use.","The observed correlation between fewer reasoning steps and higher action uncertainty (e.g., gripper timing variability) implies the model has learned to distinguish reducible from irreducible uncertainty, spending compute only where deliberation can help."],"fun_headline_variants":["Latent reasoning chains let robots deliberate before acting","Adaptive latent thinking for continuous control","Variable-length latent tokens give policies adaptive deliberation","A memory palace in latent space for reasoning-based control","Control policies learn to think longer when actions matter more"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The training procedure assumes that a PPO-style clipped surrogate with stepwise KL decomposition and free-nats clipping is enough to stably optimize the variational bound over discrete autoregressive latent trajectories without the posterior collapsing. The authors themselves note the method is sensitive to hyperparameters and prone to collapse without sufficiently large rollout buffers or regularization.","fun_headline_variants_meta":{"raw":{"variants":["Latent reasoning chains let robots deliberate before acting","Adaptive latent thinking for continuous control","Variable-length latent tokens give policies adaptive deliberation","A memory palace in latent space for reasoning-based control","Control policies learn to think longer when actions matter more","Autoregressive latent inference turns deliberation into a length decision","Latent chains with decay variance create adaptive compute for control","Robots reason in a latent memory palace before committing to action","Variational latent chains let policies decide how long to think","Discrete latent tokens give policies adaptive reasoning depth"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":1141,"prompt_tokens":514,"completion_tokens":627,"prompt_tokens_details":null},"tokens_in":514,"tokens_out":627,"duration_ms":10588,"temperature":1.0,"reasoning_tokens":562,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T02:13:03.729605+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If the adaptive latent length provides no benefit beyond what a fixed-length latent of the same total capacity provides, or if the performance gains vanish when the variance schedule is replaced by a fixed variance, then the core claim that iterative adaptive computation drives the results would be undermined.","supporting_citations":[],"review_version":2}