acceptodds
Under review as a conference paper at ICLR 2027

CARE-VLA: Capacity Allocation via Response Evaluation for 1-Bit VLA Quantization

10 views

Abstract

Vision-language-action (VLA) models are rapidly advancing embodied intelligence, but their large parameter footprints remain a major obstacle to deployment on resource-constrained robotic platforms. One-bit post-training quantization (PTQ) offers an attractive route to extreme model compression, yet preserving model behavior at this precision remains challenging. To the best of our knowledge, vector quantization has not yet been explored for VLA models in the 1-bit regime, where deciding which concrete reconstruction deserves scarce additional capacity becomes critical. We introduce Capacity Allocation via Response Evaluation (CARE-VLA), a post-training vector-quantization framework that allocates capacity among concrete low-bit reconstruction candidates according to their response-level distortion. Rademacher-projected reverse-mode pullbacks efficiently estimate native-response sensitivity without materializing full Jacobians, allowing concrete reconstruction errors to be evaluated directly in response space. Complete-policy component interventions further calibrate heterogeneous native-response gains onto a common action-discrepancy scale, enabling reconstruction candidates across vision, language, action, and auxiliary modules to compete under a shared bit budget. At 1.09 bits per weight (BPW), CARE-VLA achieves 91.0% and 88.9% mean LIBERO success on and OpenVLA-OFT, respectively, and 69.7% visual-matching and 53.2% variant-aggregation success on SIMPLER with CogACT. Executed linear weights achieve storage compression relative to 16-bit storage, including codec overhead. On the RTX 4090, the quantized model achieves a measured speedup. Our work advances extreme VLA compression toward practical deployment on resource-constrained robotic hardware.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, post-training quantization, vector quantization, policy-response evaluation, capacity allocation, model compression

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_81YwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a near-one-bit post-training quantization framework for VLA models. It allocates capacity among fixed one- and two-bit trellis reconstruction candidates using randomized response-sensitivity estimation, cross-component calibration to a common action-discrepancy scale, and budget-constrained policy-level correction. Experiments cover π0.5, OpenVLA-OFT, and CogACT on LIBERO and SIMPLER.

Critical strengths

Experiments in this paper cover diverse VLA architectures, which is good. At 1.09 BPW over executed linear weights, CARE-VLA maintains high closed-loop success across flow-based π0.5\pi_{0.5}, regression-based OpenVLA-OFT, and diffusion-based CogACT.

Critical weaknesses
  1. Despite targeting resource-constrained robotic platforms, all speedup tests rely on a desktop RTX 4090 GPU rather than actual edge devices and lack comparisons with lightweight VLA baselines, failing to validate real-world deployment performance.
  2. Stage 3 introduces complex heuristic losses, CVaR penalties, and tolerance gates, yet only improves average π0.5\pi_{0.5} success by 0.3% over static allocation while degrading Spatial performance.
Questions for authors
  1. Since CARE-VLA targets resource-constrained robotic platforms, can you provide latency and memory evaluations on actual edge devices rather than a desktop RTX 4090 GPU, and compare the model against natively lightweight VLA baselines under comparable resource budgets?
  2. Given that Policy-Level Correction relies on complex heuristic losses and multiple tolerance gates for only a 0.3% average gain on π0.5\pi_{0.5} (91.1% to 91.4%) while degrading Spatial performance (97.6% to 96.8%), what are its exact closed-loop gains over static allocation on OpenVLA-OFT and CogACT to justify this complexity?
  3. CARE-VLA slightly outperforms HB-VLA on π0.5\pi_{0.5} for LIBERO-Long but falls behind HB-VLA by 9.6 percentage points on OpenVLA-OFT for LIBERO-Long. What causes this significant degradation on long-horizon tasks under a direct action-regression architecture?
1 comment
AuthorsOpenReview ↗

Invitation to Discuss Any Remaining Concerns

Dear Reviewer 81Yw,

Thank you again for taking the time to read our rebuttal and for engaging thoughtfully with our responses. We deeply appreciate your feedback and the opportunity to address your concerns.

If there are any remaining concerns, uncertainties, or points you'd like to further discuss, we would be more than happy to engage in additional discussion during the author–reviewer discussion phase.

We sincerely hope to have the opportunity to clarify any aspects that may still be unresolved. Your feedback is valuable to us.

Thank you again for your time and consideration.

Best regards, The authors

Reviewer_9XFcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CARE-VLA compresses VLA models without retraining by allocating bits across vision, language, action, and auxiliary linear layers. It builds on QTIP’s 1-bit/2-bit candidates, scores their effects on component responses, calibrates the scores against action error, and refines the allocation under a fixed storage budget.

At 1.09 BPW, it reaches 91.0%/88.9% average LIBERO success on π0.5/OpenVLA-OFT, compared with 87.9%/83.5% for HB-VLA and 64.1%/62.9% for uniform 1-bit QTIP. On CogACT, SIMPLER success is 69.7%/53.2% under VM/VA.

Critical strengths
  1. Strong compression results across three VLA families.

The results are the strongest part of the paper. CARE-VLA covers all executed linear layers, including projectors and action heads, and beats HB-VLA at a lower bit rate. LIBERO-Long success improves from roughly 12–16% with BiLLM and uniform 1-bit QTIP to 76–77%.

The efficiency measurements are also useful. On π0.5, executed linear-weight storage falls from 5.65 GB to 0.388 GB, including codec overhead—a 14.5× reduction. Full-policy inference is 2.09× faster on an RTX 4090 at batch size one.

  1. The allocation controls are convincing.

The authors freeze the transforms, scales, and quantization candidates before comparing allocations. The main π0.5 controls use 500 matched episodes per suite.

With the same calibration and LogShrink settings, local reconstruction scoring reaches 62.2%, versus 91.1% for native-response scoring (Table 15). Randomizing stripe selection while preserving component quotas reduces success to 65.9% (Table 3). These results clearly show that where the extra bits go matters.

Critical weaknesses
  1. The gain over existing response-aware allocation methods is still unclear.

My main concern is the comparison against prior allocation criteria. Beating local reconstruction establishes the value of response information, but does not isolate CARE-VLA’s contribution. Representative action-aware or sensitivity-based scores should be tested on the same frozen K1/K2 candidates and budget.

The finer design gains are modest: direction-aware scoring adds 1.05 points, with a reported 95% interval of [0.65, 2.95]; LogShrink has mixed effects across calibration sizes; and policy-level correction adds 0.3 points. Their consistency across independent calibration and quantization runs is unclear.

The shuffle control also changes allocation between matrices. It therefore does not isolate stripe selection within a matrix. QVLA’s zero success rate is similarly difficult to interpret without separating its quantizer, allocation rule, and adaptation to this setting.

  1. The main comparisons need clearer protocols and uncertainty estimates.

Tables 1–2 lack confidence intervals, and the baseline implementations, calibration settings, and evaluation seeds are not documented clearly enough. With average gains of 2.5–5.4 points over HB-VLA, I need this information to judge reliability.

The deployment evidence is limited to simulation and an RTX 4090. The latency protocol needs more detail. Also, the 14.5× compression applies to executed linear weights; complete model storage is 1.974 GB, and peak inference memory is not reported.

Questions for authors
  1. Why does Table 17 report Long = 79.0 rather than 77.2, while the other three suite scores remain identical? Which evaluation banks and quantized configurations were used? How was the Table 22 confidence interval computed?
  2. Were baseline calibration data, evaluation episodes, and seeds matched? Please provide episode counts, confidence intervals, exact storage sizes, and unrounded BPW for Tables 1–2.
  3. Why do the local-score allocations in Table 3 reach only 59.6%/62.2% at 1.09 BPW, below uniform 1-bit quantization at 64.0%?
  4. Which BF16 implementation supports the 2.09× speedup claim, and what were the compilation, warm-up, and repeated-timing settings?
  5. Could you test prior allocation scores on the same K1/K2 candidates, and add a random-stripe control that preserves each matrix’s K2 quota?
  6. If suitable data are available, could you test CogACT with calibration data outside SIMPLER, keeping calibration size and budget fixed?
4 comments

Thank you for requesting clearer evaluation protocols and practical deployment evidence. This part addresses Q2, Q4 and Q6.

1. Evaluation protocols and exact storage (Q2).

Our locally evaluated baselines follow a common calibration and evaluation protocol. The controlled allocation comparisons use matched calibration contexts, evaluation episodes and seeds. HB-VLA entries are taken from its paper and explicitly identified as reported results.

The full CARE evaluations comprise:

Model/evaluationSuccess (%)Episodes
π₀.₅ / LIBERO91.036,000
OpenVLA-OFT / LIBERO88.96,000
CogACT / SIMPLER VM69.7864
CogACT / SIMPLER VA53.21,992

For π₀.₅, banks 17001/19001/25001 each contain 500 episodes per suite. The bank averages are 91.4/90.2/91.5%, giving 91.03 ± 0.72% (mean ± SD). Pooling 5,462 successes in 6,000 episodes yields a Wilson 95% CI of [90.28%, 91.73%], using an independent-episode binomial approximation. This pooled interval describes evaluation uncertainty for the fixed benchmark and policy; Part 1 reports task-stratified paired bootstrap intervals.

We illustrate the exact storage accounting with π₀.₅. There are N=2,825,220,096 executed linear weights, and effective BPW is 8B/N8B/N, where BB includes codec overhead.

ConfigurationCodec bytesEffective BPW
Native static388,354,2471.099678563237857
Calibrated isotropic388,352,4231.099673398330521
Full CARE388,425,2011.099879479265887

Thank you for drawing attention to the reporting precision. The previously displayed 1.09 BPW was truncated to two decimal places; conventional rounding gives 1.10 BPW. We will correct the displayed values. The evaluated policies and their measured performance are unchanged.

ResourceFPCARE
Executed linear-weight storage (bytes)5,650,440,192388,425,201
Complete serialized model (bytes)7,473,116,1151,973,913,332
Peak allocated GPU memory (GB)7.6433.129

The 14.5× linear-weight compression therefore also yields 3.79× complete-model compression and 59.1% lower peak GPU allocation.

2. BF16 baseline and timing protocol (Q4).

RTX 4090 implementationMean (ms)P95 (ms)Speedup
Native BF16 eager250.89259.531.00×
CARE120.28121.672.09×

We use batch size one, 128 contexts with matched initial action noise, and three repetitions per context: 384 timings per implementation. After untimed correctness checks, we perform three explicit warm-up calls. Model-level torch.compile is disabled. BF16 uses native eager execution; CARE uses custom CUDA kernels with CUDA Graph. CUDA is synchronized before and after each timed invocation.

Timing covers vision, language, all 10 flow updates for a 50-action chunk, and online transforms/decoding. It excludes loading, compilation, warm-up, graph capture, input transfer, noise generation, action postprocessing and the simulator. The 2.09× is the measured speedup of our deployed implementation over native BF16 eager.

3. New edge and real-robot evidence.

On Jetson AGX Orin 64GB, we use a fixed power mode, batch size one, two camera inputs, 50-action chunks and 10 flow updates:

PolicyP50 (ms)P95 (ms)Inference memory (GB)
FP36741811.63
CARE2352695.78

CARE achieves 1.56× P50 acceleration and 50.3% lower memory.

We also evaluate π₀.₅-DROID on a Franka Panda with two external ZED 2 cameras and a wrist-mounted ZED Mini: 20 trials per task and method, 120 trials total.

PolicyBlock to bowlCup to plateOrdered two-object storageOverall
FP18/2014/2016/2048/60 (80.0%)
CARE17/2013/2014/2044/60 (73.3%)

CARE achieves 1.76× mean policy-inference acceleration in the robot deployment.

4. Calibration outside SIMPLER (Q6).

Following the calibration-data setup in DA-PTQ, we calibrate CogACT using 512 trajectories from the BridgeData V2 training split. The same calibration set is used for SIMPLER evaluations on WidowX and Google Robot. We keep the CogACT checkpoint, storage budget, correction settings and evaluation episodes unchanged.

The Google Robot results are:

Calibration sourceVM success (%)VA success (%)
SIMPLER69.753.2
BridgeData V2 training split62.443.4

CARE retains 89.5%/81.6% of the VM/VA success rates obtained with SIMPLER calibration. These results demonstrate that CARE preserves substantial closed-loop capability using real-robot calibration data collected outside the evaluation simulator.

We will release the relevant code and quantized models to support the community.

AuthorsOpenReview ↗

Invitation to Discuss Any Remaining Concerns

Dear Reviewer 9XFc,

Thank you again for taking the time to read our rebuttal and for engaging thoughtfully with our responses. We deeply appreciate your feedback and the opportunity to address your concerns.

If there are any remaining concerns, uncertainties, or points you'd like to further discuss, we would be more than happy to engage in additional discussion during the author–reviewer discussion phase.

We sincerely hope to have the opportunity to clarify any aspects that may still be unresolved. Your feedback is valuable to us.

Thank you again for your time and consideration.

Best regards, The authors

Thank you for the additional experiments and clarifications. I have no further questions.

Thank you for your careful review and constructive feedback throughout the discussion. Your suggestions have helped us strengthen the experimental evidence and improve the clarity of our paper.

Reviewer_S7YQOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CARE-VLA, a near-1-bit post-training quantization framework for VLA models. It combines response-aware reconstruction scoring, action-calibrated bit allocation, and policy-level refinement. Experiments on LIBERO and SIMPLER show promising results across three VLA architectures.

Critical strengths
  • The paper addresses a relevant problem: preserving action quality under extremely low-bit VLA quantization. The response-aware allocation is well motivated.

  • The results are competitive across three VLA architectures, with clear improvements over existing low-bit baselines at similar bit budgets.

  • The ablation studies are reasonably thorough, and the reported 2.09× inference speedup provides useful evidence of practical efficiency.

Critical weaknesses
  • All evaluations are conducted in simulation. Without real-robot experiments or deployment on resource-constrained hardware, the practical value of the method remains uncertain.

  • The framework combines existing vector quantization, sensitivity estimation, and bit allocation techniques. The additional gain from directional response geometry over calibrated isotropic scoring is relatively modest (1.05 percentage points).

  • The reported 14.5× compression applies only to executed linear weights, rather than the complete model. Moreover, CARE-VLA underperforms HB-VLA on OpenVLA-OFT's LIBERO-Long suite (76.2% vs. 85.8%), suggesting that action robustness is not consistently preserved.

Questions for authors
  • Can the authors provide real-robot evaluations or latency, memory, and energy measurements on representative edge robotic hardware?

  • How does CARE-VLA compare with HB-VLA under strictly matched storage budgets, and what explains the regression on OpenVLA-OFT's LIBERO-Long benchmark?

  • How robust is the allocation to changes in calibration data, task distribution, and unseen environments?

1 comment
AuthorsOpenReview ↗

Invitation to Discuss Any Remaining Concerns

Dear Reviewer S7YQ,

Thank you again for taking the time to read our rebuttal and for engaging thoughtfully with our responses. We deeply appreciate your feedback and the opportunity to address your concerns.

If there are any remaining concerns, uncertainties, or points you'd like to further discuss, we would be more than happy to engage in additional discussion during the author–reviewer discussion phase.

We sincerely hope to have the opportunity to clarify any aspects that may still be unresolved. Your feedback is valuable to us.

Thank you again for your time and consideration.

Best regards, The authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.