acceptodds
Under review as a conference paper at ICLR 2027

HSA: Hierarchical Sensitivity-Aware Quantization for Resource-Constrained Embodied Models

Abstract

Embodied foundation models combine visual perception, language-conditioned reasoning, and action generation, but their memory footprint constrains deployment on resource-limited robotic platforms. Post-training weight quantization reduces this footprint, yet its perturbations can affect continuous action predictions and closed-loop execution. We propose Hierarchical Sensitivity-Aware Quantization (HSA), which augments a 4-bit AWQ backbone with BF16 residual blocks. An action-guided score combines a reconstruction-error proxy with action-output Jacobians, and a layer–channel–block hierarchy determines where the residuals are retained. Across LIBERO evaluations with VLA-Adapter, OpenVLA, and FastWAM, HSA improves average success over other methods by 2.55%–2.90% and remains within 0.30% of the original models. Meanwhile, it reduces GPU memory by 53.2%–67.2%. The quantized policy also matches the original policy in the real-robot setting under various conditions.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Post-Training Quantization, Vision-Language-Action Models, Embodied AI

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_DE8rOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents HSA, a method for quantizing embodied models using a 4 bit AWQ backbone with selected BF16 residual blocks. It combines reconstruction error with action output sensitivity and allocates residuals through a layer, channel, and block hierarchy. Experiments cover three models on LIBERO and a real robot pickup task.

Critical strengths
  1. HSA recovers most of the quantization loss with little additional memory. Across three simulated models, it improves success over AWQ by 2.55 to 2.90 percentage points while preserving substantial memory savings.

  2. The experiments separate residual capacity from selection quality. Comparisons with random selection, reconstruction error selection, and global selection under the same residual budget help isolate the benefit of the proposed hierarchy. The paper also clearly acknowledges that the current implementation reduces memory rather than latency.

Critical weaknesses
  1. The method differs only modestly from existing quantization approaches. It retains the AWQ backbone and combines action sensitivity with hierarchical residual selection. Action sensitivity has already been explored in QVLA, but comparisons with QVLA and QuantVLA are missing, leaving the benefit of the proposed formulation insufficiently established.

  2. The evidence for the specific selection strategy is limited. Under the same residual budget, HSA improves over reconstruction error selection by only 0.60 percentage points, without reported uncertainty. The real robot test includes only 20 trials per policy and no AWQ baseline, so it does not establish the practical benefit of the residual correction.

Questions for authors
  1. Could the authors compare HSA with QVLA and QuantVLA under matched memory budgets and quantization scopes?

  2. Is the 0.60 percentage point gain over reconstruction error selection consistent across independent runs and other backbones?

  3. Could the real robot evaluation include AWQ and more challenging tasks to demonstrate when the proposed correction is necessary?

Reviewer_XiPMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes HSA, a weight-only post-training quantization method for embodied policies. It builds on a 4-bit AWQ backbone and restores a sparse set of BF16 residual blocks of 64 weights each. Blocks are scored by multiplying an activation-scaled reconstruction-error proxy with the mean squared Jacobian of de-normalized actions with respect to each output channel, and are selected through a layer–channel–block hierarchy. Validation covers VLA-Adapter, OpenVLA, and FastWAM on the four LIBERO suites, together with a quantization-scope study, a selection ablation, peak-memory and latency measurements, and a 20-trial real-robot pickup test with a fine-tuned π0.5\pi_{0.5} policy. HSA recovers 2.55–2.90 points of average success over AWQ, stays within 0.30 points of the reported original models, and adds 0.20–0.23% peak memory over AWQ.

Critical strengths
  1. A simple, deployable method that consistently recovers the 4-bit AWQ gap across heterogeneous architectures. HSA is fully offline, leaves the policy interface unchanged, and adds only 0.20% and 0.23% peak memory over AWQ on VLA-Adapter and OpenVLA (Table 4). Its gains over AWQ hold on all four LIBERO suites for a compact VLA, a 7B autoregressive VLA, and a world-action model, and Appendix A.2 handles discrete and stochastic action decoders carefully (a soft-action readout under a fixed FP prefix for OpenVLA and a fixed-noise flow readout for FastWAM). This is the main source of the paper’s practical value: a nearly free residual branch that transfers across model families is directly useful for memory-limited deployment.

  2. Transparent reporting of quantization scope and system costs. The paper separates fixed-observation action fidelity from closed-loop success and reports results that do not favor it, including higher latency than the original model (87.06 vs. 78.56 ms) and neutral or harmful single-block repairs. The scope study in Table 2, where extending W4 to all eligible modules lowers AWQ from 94.65% to 77.55% average success, gives concrete guidance on which VLA components tolerate 4-bit weights. This matters to my assessment because it makes the claims easy to interpret and offers deployment guidance that holds independently of the proposed selector.

Critical weaknesses
  1. The closed-loop comparisons are not controlled well enough to support the reported margins. The Original rows for VLA-Adapter (97.8/99.2/97.2/95.0) and OpenVLA (84.7/88.4/79.2/53.7) exactly match the numbers published by the original authors. The OpenVLA AWQ entries 82.3 and 50.7 cannot arise from a single run of 500 episodes per suite, whose resolution is 0.2 points, whereas every HSA entry can. The rows of Table 1 may therefore come from different harnesses or seed counts. No seeds or confidence intervals are reported. At these success levels, binomial sampling alone gives a standard error of roughly 0.6 points for the difference between two four-suite averages, comparable to the 0.60–0.90 point gaps separating HSA from WeightErr-R, HSA-LC, and Global Top-K in Table 3. Random-R restores only about 150k–180k weights (2,334–2,856 blocks of 64), well under 0.1% of the target weights, yet gains 1.30 points, which suggests run-to-run variation of a similar size. This is consequential because the near-lossless claim rests on the comparison with the original models, and the case for the proposed selector rests on sub-point differences in Table 3.

  2. The contribution over existing sensitivity-aware and sparse-residual quantization is not established. HSA combines two ideas with direct precedents: keeping a sparse set of high-precision weights on top of a low-bit dense backbone (e.g., SpQR and the dense-and-sparse decomposition of SqueezeLLM), and weighting quantization error by output sensitivity (e.g., Fisher- or Hessian-based saliency in SqueezeLLM and HAWQ). Neither line is discussed or compared. QVLA, which the paper cites and which also uses action-space sensitivity, is not evaluated, and the main baselines are RTN and AWQ only. The internal evidence for the two components is also limited. In Figure 2(b), the only direct test of the score, single-block repair gain shows no visible trend along the score-percentile axis in any component. The same panel shows larger gains for DINO and SigLIP blocks than for language-model blocks, while the hierarchy moves the budget toward the language model (59.2% of blocks vs. 27.4% for Global Top-K, Figure 2(c)). The Table 3 selectors are not defined, so the closed-loop value of the Jacobian term over a reconstruction-only score cannot be isolated, and the efficiency argument for the hierarchy over Global Top-K, which the paper acknowledges is optimal for its own proxy, is not measured. This is consequential because the novelty of the paper lies in the action-guided score and the hierarchy, and neither is shown to outperform simpler or existing alternatives.

Questions for authors
  1. [Clarification] Were the Original, AWQ, and HSA rows in Table 1 obtained with the same evaluation harness, episode count, and seeds? Please explain the OpenVLA AWQ entries 82.3 and 50.7, and report the mean and standard deviation over several seeds for Tables 1 and 3. A same-harness, multi-seed result that preserves the reported margins would address my main concern.
  2. [Requested analysis] How does HSA compare at matched memory with QVLA and with a dense-and-sparse baseline such as SpQR or SqueezeLLM, and does the residual branch still help when the dense backbone is GPTQ? A discussion of concurrent action-aware methods such as ActQuant (arXiv 2605.24011) would also clarify the positioning.
  3. [Clarification and requested analysis] How exactly are Random-R, WeightErr-R, HSA-L, and HSA-LC defined? Could you also report hierarchical selection with a reconstruction-only score (α=0\alpha = 0), the rank correlation between score and repair gain in Figure 2(b), and the policy-query latency of HSA versus Global Top-K?
  4. [Clarification] What quantization scope, differentiable readout, calibration data, and hardware were used for the π0.5\pi_{0.5} policy? This configuration is not covered in Appendix A or Table 6.
  5. [Suggestion for future work] A real-robot comparison against a 4-bit AWQ policy on a harder task (both current policies reach 20/20), memory and latency on an embedded platform, and a sensitivity study over the number of Jacobian observations, ρL\rho_L, ρC\rho_C, KBK_B, and the residual budget would make the deployment claims more convincing.
Reviewer_9oGxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Hierarchical Sensitivity-Aware Quantization (HSA) for embodied models. HSA augments a fixed 4-bit AWQ backbone with selected BF16 residual blocks. Experiments report improved LIBERO success rates on VLA-Adapter, OpenVLA, and FastWAM, together with peak-memory measurements for two models and a small real-robot evaluation using π0.5.

Critical strengths

HSA achieved effective recovery with little additional memory. HSA recovers most of AWQ’s success-rate loss across three models. Table 4 reports only 0.20%–0.23% additional peak memory over AWQ on the two measured models.

The experiments are consistent . Table 3 matches residual budgets across selectors. The comparison with global Top-K also uses the same score and block count, providing evidence that residual placement matters beyond simply adding high-precision capacity.

Critical weaknesses

The technical contributions are incremental. Compared with prior works, sparse high-precision preservation is already explored by SpQR and SqueezeLLM, while QVLA combines action-Jacobian information with quantization noise. HSA’s particular residual hierarchy may be useful, but comparison mainly against AWQ does not establish its advantage over a close action-aware quantization strategy.

Unclear accuracy–memory trade-off. HSA improves AWQ by adding BF16 residuals with a small memory overhead (Table 4). However, comparisons mainly against AWQ do not establish whether HSA offers a competitive accuracy–memory trade-off relative to QVLA or QuantVLA. Comparisons on shared backbones under a common evaluation protocol and comparable actual memory budgets are needed.

Questions for authors

My research doesn't focus on model quantization, so I would appreciate clarification on the fairness of the experimental comparisons. I would be happy to change my score if the authors’ responses resolve my concerns.

  1. How does HSA compare with QVLA or QuantVLA on shared backbones at comparable memory usage? Does hierarchical selection provide a measured latency advantage over global Top-K?

  2. Why does the OpenVLA AWQ row, labeled 4-bit in Table 1, exactly match QVLA’s W8A16 results? Please confirm its configuration and source.

  3. Does WeightErr-R retain HSA’s hierarchy while removing only action guidance? Can you provide uncertainty estimates for HSA’s 0.60-point gain over this baseline?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.