acceptodds
Under review as a conference paper at ICLR 2027

HiCAP-Q: Hierarchical Capability-Aware Mixed-Precision Quantization for Large Language Models

Abstract

Large language model quantization effectively reduces inference costs and memory usage. However, existing methods typically minimize weight or activation reconstruction errors during calibration. These objectives are only weakly aligned with preserving downstream capabilities. As a result, a small reconstruction error does not necessarily imply a small degradation in domain-specific abilities such as mathematical reasoning, code generation, or long-context understanding. We propose **HiCAP-Q**, a capability-aware mixed-precision post-training quantization framework. It preserves model-specific capabilities under a user-specified bit budget. Our key insight is that different capabilities rely on different model parameters. Some layers and linear modules are critical for mathematics, coding, or long-context reasoning. Others can tolerate more aggressive quantization. **HiCAP-Q** exploits this structure by reallocating precision from capability-insensitive parameter groups to capability-critical ones. **HiCAP-Q** consists of three stages. First, it constructs capability-specific probing inputs at low cost and without human annotation. It uses the target model and, optionally, a judge model. An **input score** selects prompts based on question quality, task relevance, and response confidence. Second, the framework estimates capability sensitivity at the Transformer-layer and linear-module levels. A capability loss score measures the degradation caused by quantizing each parameter group. Third, the framework ranks parameter groups by sensitivity and assigns mixed precision under a target bit budget. Experiments on MMLU-Pro, AIME-2026, and GPQA show consistent gains over reconstruction-based baselines. With only 10% more average bits than uniform low-bit quantization, **HiCAP-Q** substantially narrows the gap to the full-precision model. It also preserves strong general-task performance.

Primary area:
optimization
Keywords:
Large Language Model, Model Quantization, Mixed Precision

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_jPzfOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

HiCAP-Q proposes capability-aware mixed-precision post-training quantization. It constructs capability-specific probes, measures quantization sensitivity at the layer and module levels, and uses these scores to allocate FP16 channels within otherwise NVFP4-quantized modules. Experiments across five models evaluate capability preservation under a stated average-bit budget.

Critical strengths

S1: The cross-domain experiment supports the central motivation. Allocations derived from domain-matched probes perform best in their corresponding evaluation domains (Table 4), suggesting that capability-specific information can guide precision allocation.

S2: The evaluation covers several capabilities and model families. Results include knowledge, mathematics, science, coding, and safety. The ablations also examine the contributions of sensitivity estimation, probe selection, and metric fusion (Tables 2–5).

Critical weaknesses

W1: The effective-bit accounting needs clarification. Standard NVFP4 requires approximately 4.5 bits per value when block scales are included, before adding FP16 channels. The reported 4.4 “effective bits” therefore needs reconciliation. Without consistent accounting of scales and other overhead, the claimed equal-budget comparison is difficult to assess.

W2: Important implementation details are missing. The paper does not fully specify probe sizes, screening settings, the FP16 allocation schedule, or which channels are preserved within each module. These choices directly affect reproducibility.

W3: The ablations do not clearly isolate the hierarchical design’s contribution. Table 5 is described as cumulative, but the “linear-only” and “layer-only” rows leave unclear whether components are replaced or added. A comparison using identical probes and storage budgets would better establish the benefit of combining both granularities.

Questions for authors

Q1: Does the reported 4.4-bit budget include NVFP4 scales, FP16-channel indices, and other storage overhead? Please clarify the accounting and provide actual model storage sizes to establish whether the comparisons are budget-matched.

Q2: Could you specify the probe sizes, screening settings, FP16 allocation schedule, and within-module channel-selection rule? These details would help reproduce the reported results.

Q3: Are the “linear-only” and “layer-only” rows in Table 5 independent alternatives or cumulative additions? How does their combination compare with each alone using identical probes and storage budgets?

3 comments
AuthorsOpenReview ↗

Response to Reviewer jPzf Q1/W1 (effective-bit accounting)

We thank the reviewer — the original manuscript omitted the bit accounting, which was an oversight. We clarify it here. NVFP4 uses FP8 scales over groups of 16, i.e., 0.5 bits/weight of scale overhead (4.5 bits effective). Our 4-bit baselines (AWQ, GPTQ, OmniQuant, SliM-LLM, QuIP#) use group size 128 with FP16 scales, i.e., 0.125 bits/weight (4.125 bits). To equalize budgets, we additionally assign a random set of 256 FP16 channels to every linear module of each 4-bit baseline. The resulting total effective bit-widths are:

Model4-bit baseline (matched)HiCAP-QΔ(baseline − ours)
Qwen3-8B4.7254.751−0.026
Qwen3.5-A3B5.5655.080+0.485
Llama-3.1-8B4.7384.707+0.031
Spark-X2.5-4B5.0454.969+0.076
Gemma-4-12B4.8044.780+0.024

The budgets are matched on Qwen3-8B (within 0.03 bits), and the baselines receive more effective bits than HiCAP-Q on the remaining models — on Qwen3.5-A3B by 0.48 bits, since its many MoE expert modules make the random FP16-channel injection substantially more expensive. We will include this table, the per-component accounting (weights + scales + FP16 channels + occupancy indices), and actual storage sizes in GB in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer jPzf W2/Q2

We sincerely appreciate your time and effort in reviewing our manuscript. Below are our responses to your valuable comments.

Q2/W2 (implementation details): We provide the missing details below and in the revised Appendix C–E.

(i) Probe and screening settings. Each capability uses 512 screened probes (Section 4.1). Screening retains prompts passing the capability-relevance and quality thresholds described in Section 3.3, with the target model itself serving as the screening model (no external judge is used). The self-confidence score q_self is obtained via verbalized self-evaluation: the target model generates an explicit confidence judgment for its own response, which is parsed into a score used by the screening stage. These details are added in revised Appendix D.

(ii) FP16 allocation schedule. The schedule is defined by the percentile threshold sets K = {0.15, 0.30} and R = {0.10, 0.30}: each module falls into one of L = 5 hierarchical sensitivity levels h(m) ∈ {0,…,4} (Section 3.5), and each module at level h retains a fixed quota of g·h FP16 channels with granularity g = 128 — i.e., 0 / 128 / 256 / 384 / 512 FP16 channels for levels 0–4; modules at level 0 are fully quantized to NVFP4. The complete quota table is now given in Appendix C (Table 7). We emphasize that the schedule contains no free ratio hyperparameters beyond the thresholds (K, R) and the granularity g: the per-module level map {h(m)} is derived per model from its own capability-specific sensitivity estimates, and the quota table is fixed. Consequently, the achieved parameter-weighted budget is an empirical outcome of the estimated sensitivities rather than a directly tuned target; we now report the actual achieved effective bit-width of every configuration in Appendix C (Table 6), including NVFP4 scales and FP16-channel overhead, and all budget comparisons in Tables 2–3 use these values (see our response to Q1). The per-module level maps for all five models will be released with the code.

(iii) Within-module channel selection. Once the per-module FP16 budget k_m is determined by the hierarchical level h(m), the retained channels are selected by activation magnitude. For each linear module m, we run the full-precision model on the capability probe set X_c and record, for each input channel j, the maximum absolute activation sj=max(x,Ax)∈Xcmax⁡t∣xt,j∣,s_j = max_{(x,A_x)∈X_c} \max_t |x_{t,j}|, aggregated as a running maximum across probes. Channels are ranked in descending order of s_j, and the top-k_m channels are kept in FP16 while the remaining C_m − k_m channels are quantized to NVFP4. Since LLM activations are heavily skewed, with a small subset of channels carrying disproportionately large magnitudes, this criterion preserves the channels that dominate the module’s computation for the target capability. Crucially, the statistics are computed on the probes of the target capability, so channel selection is itself capability-aware: an allocation driven by math probes preserves the channels that dominate computation on math inputs. The statistics are collected in a single forward pass over the probe set (512 samples × 512 tokens, seed 42) and are shared with the sensitivity-estimation stage, adding no extra calibration cost. Ties are broken by channel index, and selection is performed along the input dimension, consistent with the NVFP4 quantization grouping. This rule is detailed in revised Appendix E.

AuthorsOpenReview ↗

Response to Reviewer jPzf Q3/Q3

We sincerely appreciate your time and effort in reviewing our manuscript. Below are our responses to your valuable comments.

Q3/W3 (ablation clarity): The “+ linear-only” and “+ block-only” rows are independent alternatives, not cumulative additions: each replaces the sensitivity signal while keeping everything else identical—the same 512 probes, the same NVFP4 quantizer, the same FP16 quota schedule (Eq. 1, g = 128), and the same parameter-weighted bit budget. Concretely, the three sensitivity rows differ only in which output the capability loss is measured on:

  • “+ linear-only” measures the behavioral deviation of the module’s own output when that module is quantized—a purely local signal (0.308);
  • “+ block-only” measures the deviation of the final output of the module’s parent Transformer block, ignoring the module’s intermediate hidden output—a global signal that captures how quantization errors propagate through the block’s attention–MLP computation (0.319);
  • “+ multi-level” combines both signals hierarchically: modules and blocks are ranked separately, fused into levels h(m) = v_lin(m) + v_block(m), and FP16 channels are allocated by the level quota with within-level R(m) refinement (0.334).

This is exactly the controlled comparison requested: under identical probes and identical storage budgets, the hierarchical combination outperforms each single granularity alone (0.334 vs. 0.308 / 0.319), and both single signals outperform random mixed-precision allocation (0.265). The pattern supports the central design rationale: a module that is benign in isolation can still be critical in context, and vice versa—local optimality does not imply global optimality, and neither granularity alone identifies the capability-critical modules. The two signals are complementary, and only their hierarchical combination selects the right modules to protect. Rows 6–7 then add capability-aware probe curation and rank-normalized multi-metric fusion on top of the multi-level configuration, completing the full method (0.386).

We will rewrite the Table 5 caption and the corresponding paragraph in Section 4.4 in the revision to make the alternative-vs-combination structure explicit, and rename “layer-only” to “block-only” for consistency with the terminology defined in Section 3.2.

Reviewer_zdtsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents HICAP-Q, a capability-aware mixed-precision PTQ framework. As existing methods typically minimize reconstruction error for calibration, this minimization doesn’t shift into domain-specific capabilities such as math or coding. To address this problem, this paper proposes HiCAP-Q, a three stage framework consisting of constructing capability-specific probing, estimating capability sensitivity at the layer levels, and assign mixed precision based on ranking parameter groups. Experiments on several benchmarks improve model performance in some benchmarks under the same bit.

Critical strengths

None

Critical weaknesses

I think this paper is not ready for publication.

There are citations that do not support the claims

  • The first sentence of the introduction claims that "inference requires increasing memory, bandwidth, and computation," but cites Hendrycks et al. (2021a), which introduces the ETHICS dataset and is unrelated to this claim.
  • The statement "Some layers or linear modules may be particularly important for mathematical reasoning but have limited influence on code generation, while other modules may show the opposite pattern" is attributed to Saxena et al. (2025). As far as I can tell, the ResQ paper contains no such analysis or statement.

Insufficient description of the experimental setup

  • Key settings such as the calibration data are not described, which raises concerns about reproducibility. In addition, [1] demonstrates that reasoning capability is highly affected by the calibration dataset, so the authors should specify what calibration dataset is used for quantization to ensure baselines are evaluated correctly.

Errors in model descriptions

  • "gemma-4.4-12B" is not an existing model.
  • "Qwen-8b" does not specify the model version.

Unnatural values in the tables

  • In the SciCode column of Table 3, all results are integers except Mix for Spark-X2.5-4B, which is reported as 4.4.

Inconsistencies between the tables and the text

  • The experimental section states that "the evaluation suite covers general knowledge, mathematical reasoning, scientific reasoning, code generation," whereas the introduction lists long-context benchmarks, which are not evaluated.
  • Given the values reported in the tables, some claims appear to be overclaimed.
Questions for authors

None

7 comments
AuthorsOpenReview ↗

Response to Reviewer zdts Q1

We sincerely thank the reviewer for the careful reading and the constructive comments. We have revised the manuscript to address all concerns point by point. All changes are reflected in the revised manuscript.

Q1: Citations that do not support the claims (Introduction, first sentence; capability-dependent parameter importance).

We apologize for the misattributed citations, which we have removed in the revision.

  • The opening claim about inference demands now cites appropriate works on LLM inference cost and serving (lines 41–43): “As large language models (LLMs) continue to scale, inference becomes increasingly constrained by memory capacity, memory bandwidth, and compute (Wu et al., 2023)… where storing and transferring model weights has become a key bottleneck (Lin et al., 2024).” The unrelated citation (Hendrycks et al., 2021) has been removed. We additionally audited the full reference list and removed a similar misattribution in Appendix A, where a purely mathematical notation definition had been cited to Cobbe et al. (2021).
  • The statement on capability-dependent parameter importance is our hypothesis, tested empirically, and is no longer attributed to Saxena et al. (2025). It now reads (lines 52–56): “This mismatch arises because parameter groups contribute unequally to different capabilities: some blocks or linear modules may matter for mathematical reasoning but have limited influence on code generation… We hypothesize this structure and test it in Section 4.4, where an allocation derived from capability c’s own probes performs best on c’s benchmark (Table 5).”
AuthorsOpenReview ↗

Response to Reviewer zdts Q2

We sincerely thank the reviewer for the careful reading and the constructive comments. We have revised the manuscript to address all concerns point by point. All changes are reflected in the revised manuscript.

Q2: Insufficient description of the experimental setup; calibration data not specified.

We fully agree that calibration data must be specified, particularly given the evidence in [1] that reasoning capability is highly sensitive to the calibration dataset. We have added a dedicated description in Section 4.1 (lines 334–344):

“All baselines are evaluated with their standard 4-bit configurations. To ensure a fair comparison, all baselines are calibrated on 512 prompts randomly sampled from the same official training splits described above, without capability-aware screening, which reflects the standard calibration practice of these methods. HiCAP-Q is calibrated on the 512 screened probe samples from the same source splits, so that every method uses the same amount of calibration data drawn from identical sources, and no evaluation data are used for calibration.”

In addition (lines 327–329): “From each candidate pool we retain 512 prompts per capability for sensitivity estimation and calibration; these samples are excluded from all evaluation sets. … All test sets remain strictly held out.” To keep the average-bit budget identical across methods, we additionally “randomly assign a small fraction of fp16 channels to each 4-bit baseline” to match the 4.4-bit average budget of HiCAP-Q.

AuthorsOpenReview ↗

Response to Reviewer Q3

We sincerely thank the reviewer for the careful reading and the constructive comments. We have revised the manuscript to address all concerns point by point. All changes are reflected in the revised manuscript.

Q3: Errors in model descriptions.

Both names have been corrected and are now used consistently in the text and in all tables:

  • “gemma-4.4-12B” → Gemma-4-12B (Gemma Team, 2026a).
  • “Qwen-8b” → Qwen3-8B (Yang et al., 2025), which specifies the model version.

The full model list now reads (lines 345–348): “Qwen3-8B, Qwen3.5-A3B, Llama-3.1-8B, Spark-X2.5-4B, and Gemma-4-12B.”

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

Given that minimizing reconstruction error does not guarantee small accuracy drop on downstream tasks, this work proposes HiCAP-Q, a mixed-precision post-training quantization that reallocates precision from capability-insensitive parameters to capability-critical ones. The authors conduct extensive experiments across multiple benchmarks (e.g., MMLU-Pro, AIME-2026, and GPQA).

Critical strengths
  • The motivation is clear, and the paper is well-written overall.
  • The authors evaluate the proposed method across different models and multiple benchmarks.
Critical weaknesses
  • The authors only explore the effective bitwidth of 4.44.4. It would be more instructive if the authors also explore diverse effective bitwidths (e.g, 3.43.4).
  • When using 4-bit or more, it is well-known that block-wise quantization (e.g., FlexRound, OmniQuant, Block-AP) performs better than layer-wise quantization (e.g., GPTQ, AWQ, QuIP#), but the authors does not explore block-wise quantization at all, which seems the main limitation of this work.
Questions for authors

NA

2 comments
AuthorsOpenReview ↗

Response to W1 (effective bit-width exploration)

We sincerely appreciate your time and effort in reviewing our manuscript. Below are our responses to your valuable comments.

We clarify our design choice and provide new 3-bit experiments below. Why NVFP4 at 4.4 effective bits as the default. Our default configuration targets deployment-relevant quantization rather than exploring the full bit-width axis. NVFP4 is the FP4 format natively supported by Blackwell GPUs (dual FP4 Tensor Cores in B200/RTX 50-series), making it currently the most practical sub-8-bit format for real inference. Moreover, at 4-bit the capability loss of the underlying quantizer is moderate, so that capability-aware precision allocation — the focus of this paper — still has meaningful headroom to operate on. New 3-bit results (Llama3.1-8B, fake-quant). Following the reviewer's suggestion, we ran GPTQ-3bit and AWQ-3bit baselines with and without our capability-aware mixed-precision allocation:

StrategyGSM8KTriviaQAMMLU-ProGPQAAIMESciCode
fp160.7010.6790.4570.3070.1339
AWQ-3bit0.1410.1290.0580.01900
AWQ-3bit + Ours0.1140.1460.0630.03600
GPTQ-3bit0.1280.1190.0700.03700
GPTQ-3bit + Ours0.1310.0970.0480.03200

Two observations:

  1. At 3-bit, the base quantizer itself collapses catastrophically (GSM8K drops from 0.701 to ~0.12–0.14; MMLU-Pro from 0.457 to ~0.06; AIME/SciCode degrade to zero). In this regime the dominant failure mode is the representational capacity of the low-bit backbone itself, not how precision is distributed across modules. Consequently, reallocating FP16 channels yields only mixed, inconsistent gains, and we do not claim HiCAP-Q solves the 3-bit regime.
  2. Nevertheless, the signals are not uniformly negative: under AWQ-3bit, our allocation improves TriviaQA (+0.017), MMLU-Pro (+0.004), and GPQA (+0.017), suggesting capability-aware sensitivity remains informative even at 3 bits. We view extending capability-aware allocation to ultra-low bit-widths — by combining it with low-rank residual compensation (as in ResQ) or 3/4-bit hybrid backbones — as a promising direction, and we will add this discussion and these results as an appendix in the revision. We agree that a systematic sweep over effective bit-widths would strengthen the empirical coverage, and we will clarify in the revision that our contribution targets the hardware-supported 4-bit regime while the ultra-low-bit regime requires a different backbone design.
AuthorsOpenReview ↗

Response to W2 (block-wise quantization)

We thank the reviewer for this comment, and we found the root cause is largely terminological — though we also agree that a key baseline family was missing. We address both below.

1. The terminology issue (root cause, now fixed). §3.2 of the original submission called a Transformer block a “Transformer layer”, which in the PTQ literature reads as “per-linear-layer reconstruction” (GPTQ/AWQ style). In fact, HiCAP-Q operates at block granularity in both probing and allocation:

  • Sensitivity estimation (§3.5) covers two granularities: module-level probing (each linear layer quantized separately) and block-level probing (all linear modules within a Transformer block quantized jointly, capturing attention–MLP interactions);
  • Allocation (§3.6) is hierarchical over both: FP16-channel ratios are determined jointly by the module-level rank of mm and the block-level rank of its parent block π(m)π(m).

v2 renames this consistently across text, equations, and Fig. 1, and explicitly distinguishes (i) reconstruction granularity (layer-wise vs. block-wise) from (ii) intra-matrix quantization granularity (per-tensor/per-channel/per-group), so that the two axes the term “block-wise” conflates are no longer ambiguous.

2. New evidence: OmniQuant baseline (block-wise reconstruction as an objective). Following the reviewer’s suggestion, v2 adds OmniQuant (a representative block-wise reconstruction method) to Tables 2–3, evaluated under the identical setting as all baselines (NVFP4 weights + randomly assigned FP16 channels, 4.4 average bits). Summary of HiCAP-Q vs. OmniQuant across the 35 (model × benchmark) cells:

HiCAP-Q betterTieOmniQuant better
Cells (out of 35)2546

On accuracy benchmarks (excluding SciCode, which is integer-scored), HiCAP-Q improves over OmniQuant by ≈ +5.0 points on average, with the largest gains on Qwen3.5-A3B MMLU-Pro (+0.205), Spark-X2.5-4B GPQA (+0.144), and Qwen3-8b AIME-2026 (+0.133).

3. This data also empirically qualifies the premise of the concern. The observation that block-wise reconstruction dominates layer-wise at ≥4 bits is well-established for uniform-precision quantization. In our setting, however — mixed-precision with FP16 channels under a 4.4-bit budget — the added OmniQuant baseline does not systematically outperform the layer-wise baselines (e.g., on Llama3.1-8B GSM8K, SliM-LLM 0.666 and HiCAP-Q 0.648 exceed OmniQuant 0.628; on Qwen3-8b GSM8K, OmniQuant 0.351 is within noise of GPTQ 0.349). We hypothesize this is because once a subset of channels is kept in FP16, the reconstruction-error objective becomes a weaker proxy for downstream capability, which is precisely the phenomenon our paper is about. We have added this discussion to §4.2.

4. Composability rather than competition. We emphasize that HiCAP-Q’s contribution — which parameter groups receive how much precision — is orthogonal to the reconstruction objective used for the low-bit component. The capability-aware sensitivity signals and hierarchical allocation can be composed with a block-wise reconstruction backbone (e.g., an OmniQuant-style kernel), and we have added a sentence in §4 noting this; we view a full composition as future work.

5. Block-level contribution is already isolated in the ablation. Table 5 (cumulative ablation on GSM8K) directly quantifies the incremental value of block-level sensitivity: linear-only sensitivity reaches 0.308, while adding block-level sensitivity raises it to 0.319 (+27.9% → +34.2% relative over uniform NVFP4), confirming that block-level signals capture complementary attention–MLP interaction effects beyond module-level ones.

Changes in v2: (i) consistent “block” terminology throughout; (ii) explicit two-axis clarification of “block-wise” in §3.2 and §3.5; (iii) OmniQuant added to Tables 2–3 and the baselines paragraph; (iv) added discussion in §4.2 on reconstruction granularity under mixed-precision budgets.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.