{"id":"c503a49a-7448-424f-acf3-4c00090a7334","arxiv_id":"2607.24148","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Motion-aware vector quantization plus centroid-reuse GEMM and a custom accelerator yield multi-fold VLA speedups over GPU and prior accelerators with ~2.5% average success-rate drop.","lead":"VQVLA speeds up Vision-Language-Action robot models by switching quantization precision with motion size and reusing shared weight centroids in custom hardware. It matters because VLA inference is too slow on GPUs for real-time robots, and this co-design cuts latency without large task-success loss.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"Both headline numbers (79.4% memory savings → speedup, and ~2.5% success-rate loss) are joint functions of a single scalar threshold T_d applied to raw action magnitudes, but the paper only demonstrates T_d generalization within-benchmark, not across models, suites, or embodiments with different行动 (","rationale":"The reader identified the same load-bearing assumption — that scalar motion magnitude vs. a single calibrated T_d reliably separates execution from transition phases — and rated the paper CONDITIONAL with medium correctness risk. My read sharpens rather than replaces this: the concern is not merely that the proxy might be imperfect, but that the two halves of the strongest claim (speedup via memory savings, and small success-rate loss) are both downstream of the same threshold and the same transition-state fraction, so a single failure mode propagates to every headline number. I checked for alternative load-bearing concerns and found them weaker: the cycle-level simulator + 28nm→7nm scaling is standard MICRO practice with a stated Scale-Sim methodology; the LUT-DLA/Dadu-Corki re-implementations are disclosed; the GPU-VQVLA 45.2% slowdown is honestly reported and motivates the custom silicon; and the merged-centroid GEMM algebra (Eq. 1) is a straightforward distributivity argument with a plausible PE/cache design backed by ablations (Fig. 19–20). The chunk-to-D ambiguity is a concrete specification gap the authors can close cheaply. Because the paper does provide within-benchmark generalization evidence (10%/90% split, <1.8% degradation on unseen data), the concern does not invalidate the contribution — it bounds its generality. Hence the verdict stays CONDITIONAL: acceptance should hinge on cross-model/cross-suite threshold transfer and artifact release, exactly as the reader framed it.","tokens_in":26803,"tokens_out":2235,"duration_ms":72666,"concrete_test":"Cross-suite frozen-threshold transfer: calibrate T_d on the LIBERO suites only (OpenVLA/OpenVLA-OFT/π0/GR00T), freeze it, and apply unchanged to RDT on all four ManiSkill suites (and vice versa). Report per-suite success rate and transition-state fraction. Also specify and ablate the chunk-to-D rule (first action vs. max vs. mean over the chunk). If frozen T_d pushes success-rate loss above ~5% on any suite, or shifts the transition fraction enough to move the memory saving by >10 points from 79.4%, the threshold is per-deployment calibration, not a general mechanism, and the headline speedup/accuracy pairing weakens accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim couples two empirical quantities: (a) 6.5×/2.8×/1.9×/3.3×/4.3× speedups, which depend materially on the 79.4% memory-access reduction, which in turn is directly proportional to the fraction of steps classified as \"transition\"; and (b) ~2.5% average success-rate loss, which depends on that same classification being correct. Both are governed by one mechanism: D=√(Ax²+Ay²+Az²) vs. a fixed T_d=0.8 (§4, §6.3, §7.4). This mechanism is the least secure link, for three concrete reasons. (1) The generalization evidence in §7.4 is within-benchmark only: 10% of each benchmark calibrates T_d, 90% held out — but same model, same suite, same action distribution. The five evaluated models use heterogeneous action parameterizations and normalizations (OpenVLA's quantile-normalized actions, RDT's diffusion actions, π0's flow-matching outputs, GR00T), so a threshold in raw Cartesian action space has no a priori reason to transfer; \"T_d=0.8 as default\" (§7.4) may really be \"T_d per model/task,\" which changes the framework claim. (2) With action chunking (lengths 8, 8, 5, 16 per §7.1), the \"output of the last VLA inference\" (§4) is a chunk, not a single action; the paper never specifies which chunk element (or aggregation) produces D, yet chunk-internal magnitude can straddle the threshold, making the state label ill-defined exactly where it matters (approach-to-grasp). (3) Fig. 5's noise-injection experiment validates tolerance to additive action noise, not to weight-VQ error; the end-to-end Fig. 14 results cover this empirically, but only on the calibrated benchmarks. If a frozen T_d misclassifies fine-control-during-large-motion steps (e.g., ManiSkill dexterous tasks), success rate and memory savings degrade together, and the speedup claim — which leans on transition-state fraction — shrinks with them.","agreement_with_reader":"agree"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript presents VQVLA, an algorithm–hardware co-design for accelerating Vision-Language-Action model inference. The algorithmic component, MotionVQ, classifies each robot control step as \"execution\" or \"transition\" by comparing the scalar magnitude D=√(Ax²+Ay²+Az²) of the most recent action against a fixed threshold T_d=0.8, and selects between two offline-trained VQ configurations (VQ[256,2,256] vs. VQ[128,2,64]) accordingly. The compute component, merged-centroid vectorized GEMM, eliminates redundant multiplications by aggregating inputs that share a centroid within a column (spatial merging) and caching products for hot centroids reused across columns (temporal reusing). A custom accelerator realizing both is evaluated in a cycle-level simulator with Ramulator, synthesized at 28nm and scaled to 7nm. Reported results: 6.5×/2.8×/1.9×/3.3×/4.3× speedups over A100, Dadu-Corki, LUT-DLA, CodeGEMM, and ShiftAddLLM, 79.4% weight-memory-access reduction, ~2.5% average success-rate loss across five VLA models on LIBERO and ManiSkill, and 75.5× energy reduction vs. A100.","tokens_in":27360,"tokens_out":3872,"duration_ms":110057,"significance":"If the results hold, the contribution is meaningful: VLA inference latency is a recognized deployment bottleneck, and the paper offers (i) a concrete, checkable hypothesis about execution-state-dependent quantization sensitivity, (ii) a reuse-based VQ GEMM paradigm with ablations isolating memory, spatial, and temporal contributions (Fig. 19), (iii) design-space sweeps for cache size, adder-tree depth, PE ratio, and T_d, and (iv) a demonstration of orthogonality to the strongest prior baseline (VQVLA-Corki, 6.0× over Dadu-Corki). Evaluation spans five models and two simulators, and GPU baselines are run with official implementations, Tensor Cores, and FlashAttention 2. Weaknesses limiting significance: all architecture numbers come from the authors' own cycle-level simulator with 28nm→7nm scaling and re-implemented baselines, and no code/artifact release is mentioned, so independent verification is currently not possible.","major_comments":[{"comment":"Both headline numbers depend on one classification mechanism — D=√(Ax²+Ay²+Az²) vs. a fixed T_d=0.8 — yet its generalization evidence is within-benchmark only: §7.4 calibrates T_d on 10% of each benchmark and tests on the remaining 90% of the same benchmark, same model, same action distribution. This is problematic because the five evaluated models use heterogeneous action parameterizations and normalizations (OpenVLA's quantile-normalized actions, RDT's diffusion-policy outputs, π0's flow-matching outputs, GR00T), so a threshold expressed in raw Cartesian action space has no a priori reason to transfer across models, or even across embodiments within ManiSkill. The 79.4% memory reduction is directly proportional to the transition-state fraction produced by this threshold, and the 2.5% success-rate drop depends on the classification being correct. The paper should either (a) demonstrate","section":"§4 and §7.4 (threshold T_d)"},{"comment":"The state predictor uses \"the output of the last VLA inference\" to compute D, but §7.1 states that OpenVLA-OFT, RDT, π0, and GR00T use action chunks of lengths 8, 8, 5, and 16. The paper never specifies which chunk element (or aggregation) produces (Ax, Ay, Az) for the magnitude computation, nor when during chunk execution the state is re-evaluated. This is not a presentational gap: chunk-internal motion magnitude routinely straddles T_d (e.g., a fast approach followed by a slow grasp within one 8-step chunk), so the state label is ill-defined exactly at the phase boundary the method relies on. Please define the per-chunk state-assignment rule and report how sensitive the memory-savings and success-rate results are to that choice.","section":"§4 (State Prediction) vs. §7.1 (action chunking)"},{"comment":"The central accuracy claim ('negligible accuracy degradation', ~2.5% average success-rate reduction) is reported without rollout counts, number of seeds, or any variance estimate. In LIBERO/ManiSkill evaluation protocols, success-rate differences of 2–3% are frequently within run-to-run noise at typical rollout counts. Per-task bars in Fig. 14 appear to show drops larger than the 2.5% average for some model–task pairs, and these are not discussed. Please report the number of evaluation episodes per task, confidence intervals or standard deviations, and identify the worst-case per-task degradation, since the MotionVQ tradeoff is only acceptable if the degradation is statistically real and bounded.","section":"§7.2 / Fig. 14 (success-rate statistics)"},{"comment":"Three fairness issues need clarification. (1) GPU baselines run at each model's default precision — FP16 for OpenVLA, FP32 for OpenVLA-OFT, BF16 for RDT/π0/GR00T — so the 6.5× speedup over A100 is an aggregate over incommensurate baselines; a like-for-like breakdown (e.g., VQVLA vs. FP16 A100 for all models, or per-model speedups) should be given. (2) Dadu-Corki and LUT-DLA are re-implemented in the authors' own cycle-level simulator; the fidelity of those re-implementations (validated against the originals' published numbers) is not established, and both the 2.8× and 1.9× claims rest on them. (3) Area/power are synthesized at 28nm and scaled to 7nm via [60]; given the 51.15 mm² total area and 19.28 W at 28nm, the scaled edge-deployment story deserves explicit numbers rather than a citation. Additionally, the per-action-latency comparison against Dadu-Corki (Fig. 18) normalizes by action","section":"§7.3 (baseline methodology)"}],"minor_comments":[{"comment":"Fig. 5 validates tolerance to additive noise injected into actions per state, but the perturbation MotionVQ actually introduces is weight-VQ error, whose effect on actions is neither additive nor state-independent. Since Fig. 14 measures true success rates end-to-end this is not load-bearing, but the motivating experiment should either use the actual quantization error or be explicitly framed as a proxy.","section":"§3.2 / Fig. 5 (noise-injection motivation)"},{"comment":"Fig. 4 presents the motion-magnitude vs. distance-to-object correlation for a single task with undefined axis units; a quantitative correlation over all benchmarks (with units and normalization stated) would better support the core insight.","section":"§3.2 / Fig. 4 (single-task anecdote)"},{"comment":"The sentence 'Such success rate reduction is acceptable, as verified in research papers [26]' cites the OpenVLA paper; [26] does not, to my knowledge, establish a 2.5% acceptability criterion. Please support or soften this claim.","section":"§7.2"},{"comment":"Offline VQ training cost (k-means per weight group for two configurations per model) and the storage cost of carrying both codebook/index sets are not reported; both matter for the deployment story.","section":"§7.2 (VQ configuration)"},{"comment":"The 45.2% slowdown of GPU-VQVLA vs. GPU-A100 (Fig. 15) is an honest and useful result; it would benefit from a short breakdown of where the overhead goes (index traversal vs. kernel launch).","section":"§7.3 (GPU-VQVLA result)"},{"comment":"Combined on-chip SRAM (7MB IPU + 1.5MB result cache + 5.6MB buffer) is large relative to the stated edge-robotics motivation; the spill-to-off-chip fallback mentioned in §7.3 deserves quantification.","section":"Table 1 (area breakdown)"},{"comment":"Several typos and spacing issues: 'sinceexisting VQ approaches' (§2.2), 'insight:the sensitivity' (§1), '6.5×, 2.8×,1.9×,3.3×' (Abstract), 'distribution law of multiplication' should be 'distributive law' (§5.1).","section":"Throughout (typos)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid co-design contribution, but reviewers should note that the entire empirical package is simulation-based (no silicon, no released simulator or VQ artifacts mentioned), and the headline 6.5×/75.5× numbers rest on a simulator authored and calibrated by the same group, with two of five baselines re-implemented in that simulator. This is standard practice in the venue but worth weighing against the strength of the abstract claims. Also, the threshold-T_d issue (major comment 1) may, in the worst case, reframe the contribution from \"one default threshold\" to \"per-model calibration pipeline,\" which would soften — but not eliminate — the novelty."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is straightforward: they couple a robot-state-aware dual-codebook VQ policy (MotionVQ) with a GEMM that never dequantizes and instead aggregates inputs for repeated centroids in a column and reuses hot-centroid products across columns, then build PE arrays and location tables that match that dataflow. That co-design is the actual contribution, not another generic low-bit LLM accelerator.\n\nWhat they do well is the systems work. Latency is correctly pinned on the LLaMA backbone. Motivation for execution vs transition states is backed by motion–distance plots and a noise-injection check, then carried through multi-model LIBERO/ManiSkill runs with success-rate, memory, and multiply reductions, plus ablations that separate memory, spatial merge, and temporal reuse. Comparison set is fair (A100, Dadu-Corki, LUT-DLA, CodeGEMM, ShiftAddLLM). Orthogonality to Dadu-Corki and the combined 6× claim are honest. Design sweeps for cache, adder stages, PE ratio, and T_d are present. Math of spatial merge is just the distributive law; nothing load-bearing is hand-waved there.\n\nSoft spots, in proportion. The weakest link is exactly the stress-test point: D = √(Ax²+Ay²+Az²) vs fixed T_d=0.8 drives both the ~79% memory cut and the ~2.5% success drop. Calibration is 10/90 within each benchmark, not across heterogeneous action normalizations (OpenVLA quantiles, RDT diffusion, π0 flow, GR00T) or embodiments. With action chunking they never say which element of the chunk defines D. That does not sink the paper—end-to-end success is measured on the same suites—but it means the “default T_d=0.8” framing oversells transfer. Secondary: cycle-level sim + 28nm→7nm scaling, no silicon, no released code; PE utilization and energy numbers should be read as upper bounds. GPU-VQVLA slowdown is a useful honesty check that the algorithm needs the hardware.\n\nThis is for people building robot-edge accelerators or VLA serving stacks, not for theory readers. Citations look appropriate; no circular derivation. I would send it to referees. Engage if you care about embodied inference hardware; skim the T_d and chunking sections carefully before trusting the absolute speedups off the calibrated suites.","headline":"Solid VLA co-design paper: motion-gated dual VQ plus centroid-reuse GEMM hardware is real novelty; headline speedups ride on a simple motion threshold that is only weakly stress-tested across action spaces.","tokens_in":27765,"tokens_out":629,"would_cite":true,"duration_ms":20715,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Robot motion state can drive dynamic vector quantization and centroid reuse so VLA inference runs several times faster with only a small drop in task success.","keywords":["Vision-Language-Action","vector quantization","MotionVQ","centroid reuse","merged-centroid GEMM","embodied AI accelerator","dynamic precision"],"falsifier":"On held-out LIBERO or ManiSkill tasks, force low-precision (#Centroid=64) on every step whose true motion exceeds T_d=0.8 and measure whether success rate collapses far beyond the reported ~2.5% drop; if it does, the motion proxy does not justify dynamic precision.","tokens_in":27376,"feed_emoji":"🤖","tokens_out":952,"duration_ms":21361,"temperature":0.7,"pith_summary":"Vision-Language-Action models let robots see, read instructions, and act, but the large transformer backbone makes each inference too slow for responsive control on ordinary GPUs. This paper argues that two underused regularities fix much of that cost: robotic tasks alternate between fine near-object adjustments and coarse travel, and after vector quantization many weights collapse onto the same few centroids. MotionVQ therefore keeps two codebooks offline and, from the last action’s motion magnitude, picks high or low precision at runtime so memory traffic falls when the robot can tolerate coarser weights. A merged-centroid GEMM then multiplies each shared centroid only once—by summing matching inputs inside a column and caching products across columns—so redundant MACs disappear. A custom accelerator wires these choices into hardware and reports multi-fold speedups over an A100 and several prior accelerators while average task success drops only a few percent.","feed_headline":"Motion-aware VQ cuts VLA latency up to 6.5×","feed_subtitle":"Dual codebooks and centroid reuse speed robot transformers with ~2.5% success drop","key_machinery":"MotionVQ plus merged-centroid vectorized GEMM: classify each step as execution or transition from motion magnitude D = √(A_x²+A_y²+A_z²) versus threshold T_d, load the matching codebook/index set, then replace dense multiplies by input sums for repeated centroids within a column and cached products for hot centroids across columns.","core_discovery":"VQVLA shows that coupling motion-aware dual-precision vector quantization with a codebook–index GEMM that spatially aggregates and temporally reuses centroids, realized on a specialized accelerator, yields roughly 6.5×, 2.8×, 1.9×, 3.3×, and 4.3× speedup over A100, Dadu-Corki, LUT-DLA, CodeGEMM, and ShiftAddLLM with about 2.5% average success-rate loss on standard VLA benchmarks.","pith_inferences":["If motion magnitude is only a weak proxy on some embodiments, richer state signals (gripper contact, visual proximity) could extend the same dual-codebook idea without changing the GEMM engine.","Centroid-reuse PE arrays may transfer to other VQ’d transformers outside robotics wherever index heatmaps are similarly skewed.","The reported GPU slowdown when running the algorithm without custom hardware implies that irregular index gather/compare, not just arithmetic volume, is the real barrier on general-purpose chips."],"forward_implications":["Per-inference VLA latency can fall enough that end-to-end per-action times reach the 30–60 ms range reported for the evaluated models.","Memory traffic for transformer weights can drop by roughly four-fifths when transition phases dominate, easing edge deployment.","About half of the multiplies in the backbone can be removed by centroid reuse without reconstructing full-precision weights.","The same stack can be stacked with action-chunking controllers such as Dadu-Corki for further end-to-end gains.","Hardware need only store codebooks, indices, and a modest result cache rather than full FP weights and large all-pair LUTs."],"fun_headline_variants":["MotionVQ dual-precision cuts VLA inference 6.5× vs A100","Centroid-reuse GEMM speeds VLA 2.8× over Dadu-Corki","VQVLA accelerator reuses centroids for 6.5× VLA speedup","Motion-aware VQ trims VLA memory and latency with 2.5% drop","Codebook-index GEMM yields 4.3× VLA gain over ShiftAddLLM"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"A single scalar motion size compared with one fixed threshold is enough to tell when the robot can safely use the cheaper low-precision codebook without spoiling the task.","fun_headline_variants_meta":{"raw":{"variants":["MotionVQ dual-precision cuts VLA inference 6.5× vs A100","Centroid-reuse GEMM speeds VLA 2.8× over Dadu-Corki","VQVLA accelerator reuses centroids for 6.5× VLA speedup","Motion-aware VQ trims VLA memory and latency with 2.5% drop","Codebook-index GEMM yields 4.3× VLA gain over ShiftAddLLM"]},"model":"grok-4.5","effort":"low","cost_usd":0.003932,"raw_usage":{"total_tokens":1247,"prompt_tokens":833,"num_sources_used":0,"completion_tokens":120,"cost_in_usd_ticks":39324000,"prompt_tokens_details":{"text_tokens":833,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":294,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":833,"tokens_out":120,"duration_ms":6172,"temperature":1.0,"reasoning_tokens":294,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T22:32:17.489347+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On held-out LIBERO or ManiSkill tasks, force low-precision (#Centroid=64) on every step whose true motion exceeds T_d=0.8 and measure whether success rate collapses far beyond the reported ~2.5% drop; if it does, the motion proxy does not justify dynamic precision.","supporting_citations":[],"review_version":1}