acceptodds
Under review as a conference paper at ICLR 2027

Configuration-Valued Expert Importance for Mixed-Precision Quantization of Multimodal Mixture-of-Experts

Abstract

Sparse activation does not eliminate the expert-weight storage cost of multimodal Mixture-of-Experts (MoE) models, motivating mixed-precision quantization under memory constraints. However, existing importance-based approaches often rely on isolated expert or projection scores, overlooking the context dependence of expert importance, the natural hierarchy between experts and their internal projections, and the coupled quantization effects within each expert. We propose Config-Owen, a hierarchical game-theoretic framework for mixed-precision MoE quantization. Config-Owen formulates expert-level precision allocation as a cooperative game, uses Owen-style attribution to preserve the expert–projection hierarchy, and incorporates within-expert Harsanyi interactions to capture coupled quantization effects, yielding configuration-valued utilities for precision allocation. We efficiently estimate these utilities and optimize the resulting precision assignment under a global memory budget. Across four multimodal MoE models and Mixtral-87B, Config-Owen consistently improves low-bit performance preservation; on DeepSeek-VL2-S at 2.58 bits, it retains 98.39% of the FP16 five-task average score. Code will be released upon acceptance.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Mixture-of-Experts, Mixed-Precision Quantization, Post-Training Quantization, Multimodal Large Language Models, Model Compression

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_it5NOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses mixed-precision quantization (MPQ) for multimodal MoE models and proposes Config-Owen. The core idea: replace the "scalar importance of expert E_i" with a "configuration-valued utility"—the same 2/4-bit local configuration can produce different end-to-end loss increments depending on the "outer context" (preceding expert configurations, co-activated experts at the same layer, subsequent expert configurations). The method comprises five objects: hierarchical coalition structure, configuration-valued utility, Owen-type layered attribution (compressed via endpoint differences), CASP + RPOS sampling estimator, and multiple-choice knapsack (MCKP) allocation with feasibility mask and effective-bit ledger.

Critical strengths

S1. Clear problem motivation Traditional MoE MPQ treats "expert importance" as a scalar independent of other experts' precision. The authors directly attack this assumption and use the Table 7 ablation chain (Expert-only < Projection-Owen < Local-Calib-MCKP < Config-Owen) to demonstrate that "hierarchy + context" are genuinely effective factors. This gradient-style improvement provides empirical support for the claim that "configuration-valued utility outperforms scalar importance."

S2. Good experimental coverage and ledger transparency; main results are internally consistent. Five models span 1B–7×8B parameters, unimodal/multimodal, and top-2/top-8/top-6 routing configurations. Appendix C Table 6 provides raw per-task scores for all eight methods;

Critical weaknesses

W1. Baseline coverage is thin, and there is no effectiveness closure between the extensive mathematical derivations and the actual algorithm.

All existing main comparisons (Uniform / Hessian / PMQ) are fundamentally the same paradigm—"compute a scalar score per unit → sort by score → allocate precision"—differing only in the scoring kernel. PMQ is Config-Owen's most direct "degenerate special case" (scalar vs. configuration-valued), Hessian is "another scalar," and Uniform is "no allocation at all." All three are strict subsets of Config-Owen's design space.

But MPQ is not limited to this single route:

  • General PTQ + mixed precision: AWQ , SmoothQuant (difficulty migration to weights), OmniQuant (differentiable quantization parameter learning) can all naturally extend to per-expert/per-projection mixed-precision allocation. The authors cite these methods (§2) but ran none in experiments.
  • Non-importance-based MoE quantization: KBVQ-MoE (vector quantization), MoEQuant (expert-balanced + routing reweighting), TileQ (low-rank MoE) are also entirely absent.
  • Critical ablation missing: How much worse would a scheme with "AWQ scores + the same MCKP solver" be? If the gap is small, the complex attribution mechanism lacks independent justification.

Furthermore, the paper's mathematical volume (Appendices N–U, ~25 pages) far exceeds comparable papers, involving seven-term Fréchet perturbation decomposition, continuous–discrete integral correspondence, Walsh orthogonal spectrum, double Möbius tensor Γ contraction, Rényi-2 divergence and trace-variance optimal proposal, Lagrangian frontier and reduced-cost stability, multi-level product lattices, and calibration-to-population end-to-end guarantees. However:

  • Most analytical tools do not appear in the algorithm path. Algorithm 1's core is endpoint-difference accumulation → aggregation → MCKP (i.e., Owen value estimation). The theoretical objects actually used are limited to the endpoint representation (Eq. 4) and the Owen definition (Eq. 6). Harsanyi coefficients ρ(S,T), the Γ tensor, Walsh spectrum, Rényi divergence, Fréchet derivatives, product-lattice DCP, etc. appear only in proofs and discussions, never participating in actual computation. This does not mean the theory is "useless"—the Owen framework genuinely provides conceptual justification for "why endpoint differences are valid scoring." However, the large volume of analysis unrelated to algorithmic implementation may mislead readers into believing these tools are necessary for the method's correctness.
  • Theory inspired design choices, but approximation quality was never validated. Appendix Q.2 gives the trace-variance optimal proposal q*(π) ∝ p(π)‖Y(π)‖₂, and the main text explicitly states that router-mass RPOS is a "tractable approximation" to q*—this design was indeed theory-inspired. But the variance gap between q* and router-mass RPOS was never empirically measured, nor were Rényi-2 divergence or d_eff values reported, leaving readers unable to assess how tight the approximation is.
  • Key constants in theoretical bounds were never empirically evaluated. Appendix S's β_{ik}, Appendix R's Γ_B, and Appendix Q's d_eff are all unmeasured. Readers cannot judge whether these bounds are "meaningful" or "vacuous decoration."

W2. Some main comparisons have Δ on the same order as noise

Paper §4.1 states "seed variation is reported separately rather than repeated in the main table." Appendix D.3 Table 4 gives 3-seed Avg. std as ±0.28 and ±0.31 for MolmoE-1B and DSVL2-L respectively. The following three Ours−PMQ cells fall within σ:

ModelAvg bitsOursPMQΔ
MolmoE-1B2.3046.2146.12+0.09
DSVL2-L2.5770.9770.60+0.37
DSVL2-L2.3070.3870.25+0.13

These three cells are "bolded as best quantized" in the main table but are statistically indistinguishable from zero given the reported σ. The conclusion/abstract both claim "consistently improving the matched-budget average"—if three cells lack statistical significance, this statement needs qualification. Additionally:

  • 5 models × 2 budgets = 10 cells, but only 2 models report σ; DSVL2-S/T and Mixtral seed variance is unreported.
  • Table 1 shows DSVL2-S 2.30 MMB = 78.17 > 2.58's 77.24, a non-monotonicity ("tighter budget performs better") with no annotation indicating whether this is seed noise.

W3. CASP behavior contradicts the "bias-for-time" narrative; hyperparameter sensitivity is entirely absent.

Table 7 reports the effect of removing CASP at DSVL2-L @2.57 and DSVL2-S @2.58:

CASP simultaneously reduces time and improves accuracy. If it were merely a "bias term," removing it should yield equal or better Avg. A plausible explanation: avoiding collapse-distorted evaluations makes MCKP scoring more effective. The authors need to restate CASP's semantics—it is not merely "candidate reduction" but also "attribution quality enhancement."

Questions for authors
  1. What are the core advantages of your method, and does it have a precision advantage compared to methods like Omniquant and QTIP that do not require additional bits? Is there a significant advantage compared to methods like TileQ that also require additional bit positions?
  2. Baseline effectiveness: Can you run an ablation with "AWQ scores + Config-Owen's MCKP solver"? If the result is close to full Config-Owen (<1 pts gap), the core contribution may lie in MCKP's global allocation rather than the Owen endpoint-difference scoring.
  3. CASP causal mechanism: What explains CASP simultaneously reducing time and improving accuracy? Can you provide a micro-analysis showing "if CASP-pruned candidates were fully evaluated, the resulting allocation would degrade"?
  4. Hyperparameter sensitivity: Avg.% curves for τ ∈ {0.3, 0.5, 0.7, 0.9}? Does doubling M from 20 to 40 contract ε_r as the theory predicts?
13 comments

We sincerely thank the reviewer for the careful reading and thoughtful comments. We especially appreciate the specific questions about our baselines, experimental reliability, and sampling strategy. These comments have helped us identify several aspects that need further validation and clearer explanation. We have therefore conducted additional experiments and revised our discussion accordingly. Our responses are provided below.

A.On baseline coverage and the choice of additional comparisons.

We thank the reviewer for suggesting several relevant quantization methods. After examining them, we find that they mainly address different problems from the precision-allocation problem studied in our work.

Specifically, SmoothQuant focuses on reducing activation quantization difficulty, while QTIP and KBVQ-MoE use vector-quantized weight representations. MoEQuant focuses on expert-aware quantization, whereas TileQ uses low-rank structures for compression. Although these methods are relevant to MoE quantization, their original formulations do not directly allocate different bit-widths to individual experts and their projections under a global memory budget. A direct comparison would therefore involve different quantization formats and optimization objectives.

Among the methods mentioned by the reviewer, AWQ and OmniQuant provide more suitable starting points for additional allocation baselines. We adapt their quantization sensitivity and reconstruction-error scores to projection-level utilities, while keeping the MCKP solver, GPTQ backend, and memory budgets identical to Config-Owen.

Additional comparisons on DeepSeek-VL2-S:

MethodBitsMMBMMStarMMMUAI2DOCRBenchMMEAvg.%
AWQ-score + MCKP2.5872.8052.7042.0071.6075.001650.0062.82
OmniQuant-error + MCKP2.5873.3053.4042.9072.4075.501661.0063.50
PMQ2.5874.0554.3343.6769.5376.701674.4563.66
Config-Owen2.5877.2456.2746.6778.2781.201697.0767.93
AWQ-score + MCKP2.3072.6051.3039.8070.9073.801639.0061.68
OmniQuant-error + MCKP2.3073.1052.2040.7071.6074.601652.0062.44
PMQ2.3073.8652.4341.4270.3275.601653.4162.73
Config-Owen2.3078.1756.2043.6778.1180.901683.7567.41

These comparisons help isolate the contribution of our utility estimation from that of the global optimizer.We will also expand the related-work discussion to include the five methods mentioned above and clarify how their quantization and compression techniques can complement our precision-allocation framework.

B. On the statistical reliability of the reported improvements.

Our original experiments use three random seeds for Config-Owen, but report standard deviations only for DeepSeek-VL2-S. We therefore extend the statistical evaluation to five seeds across all five models and both matched-budget settings.

Multi-seed comparison:

ModelBitsPMQ Reference Avg.%Config-Owen (Mean ± Std)Difference
MolmoE-1B2.6047.6748.94 ± 0.10+1.27
MolmoE-1B2.3046.1246.20 ± 0.25+0.08
DeepSeek-VL2-L2.5770.6071.01 ± 0.12+0.41
DeepSeek-VL2-L2.3070.2570.43 ± 0.10+0.18
DeepSeek-VL2-S2.5863.6667.93 ± 0.28+4.27
DeepSeek-VL2-S2.3062.7367.41 ± 0.31+4.68
DeepSeek-VL2-T2.5959.4660.57 ± 0.14+1.11
DeepSeek-VL2-T2.3058.5760.19 ± 0.15+1.62
Mixtral-8×7B2.5467.5068.88 ± 0.13+1.38
Mixtral-8×7B2.3065.1666.51 ± 0.13+1.35

The results show that the improvements vary across models. The gains are more pronounced on DeepSeek-VL2-S, DeepSeek-VL2-T, and Mixtral, while some differences on MolmoE-1B and DeepSeek-VL2-L are relatively small. We therefore distinguish substantial gains from smaller differences rather than suggesting that the improvements are equally large across all models.

We also provide task-level results for each seed on all five models in Appendix E. These results allow us to examine the higher MMBench score at 2.30 bits on DeepSeek-VL2-S. Although this task shows a non-monotonic trend, the overall average remains higher at 2.58 bits. This is expected because our allocation objective optimizes calibration utility under a global memory budget rather than individual benchmark scores.

C. On the CASP mechanism and its threshold sensitivity.

We agree that CASP deserves further evaluation beyond its computational savings. It stops evaluating the remaining expert suffixes once calibration performance enters a collapsed regime. We therefore examine how different thresholds affect the final allocation and downstream performance.

CASP threshold sensitivity on DeepSeek-VL2-S:

Threshold (tau)MMBMMStarMMMUAI2DOCRBenchMMEAvg.%Utility Time (min)
0.375.6255.1444.7176.3279.251678.3466.2157
0.577.2456.2746.6778.2781.201697.0767.9343
0.776.1055.6345.8277.1279.841682.4666.9031
0.973.8453.9243.5075.4678.211658.7264.9922

The default threshold of 0.5 gives the best overall result. Lower thresholds require more computation, while higher thresholds lead to more aggressive pruning and larger performance drops. This supports our choice of 0.5 as a practical balance between computation and allocation quality.

Reviewer_4oweOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a bit-allocation strategy for expert-wise mixed-precision quantization of MoE models. The paper develops a hierarchical game-theoretic framework for identifying the collaborative effects of quantizing different experts at different precision levels. The paper presents experiments on relevant MoE models and tasks to verify the proposed method.

Critical strengths
  1. The idea of using the collaborative/interactive effect of bit-width quantization for different expert combinations in mixed-precision MoE is novel and intuitive.

  2. The paper provides substantial theoretical justification and empirical validation to support the design choices.

Critical weaknesses
  1. I wonder why the authors restrict the utility estimation of each expert's configuration to contexts where all other experts are either fully upgraded or at baseline precision. Since the final allocation allows mixed-precision configurations across experts, excluding such contexts during utility estimation may overlook important cross-expert interactions and affect the accuracy of the estimated configuration utilities. Indeed, the performance gain is marginal compared to other uniform and mixed-precision baselines. Moreover, the proposed method is calibration-data-dependent, which may easily lead to a suboptimal solution, specifically in the multimodal case.

  2. The proposed method is specifically designed for only two bit-precision levels, i.e., blowb_{low} and bhighb_{high}. Extending the proposed method to more than two bit levels does not seem natural.

Questions for authors

Please see the weaknesses.

5 comments

We sincerely thank the reviewer for the careful reading and precise, constructive comments. We have further clarified the motivation for our expert-level coalition formulation and why it is suitable for mixed-precision allocation.

  1. On the expert-level coalition structure and cross-expert interactions.

Our design follows the Owen value's idea of keeping predefined groups together when measuring their contributions. In our setting, complete experts are considered in the outer step, while the up, gate, and down projections are considered within each expert. This is useful for four reasons:

(i) A well-established mathematical rule for assigning contributions when players are divided into groups in advance.

(ii) Consistency between expert-level and projection-level contributions: the contributions assigned to an expert's projections add up to the Shapley contribution of that expert when it is treated as one player in the group-level game.

(iii) The random ordering of complete experts gives clear weights to different expert-level contexts, without requiring an additional sampling rule for every possible mixed-bit setting of other experts' projections.

(iv) A good match with the MoE architecture, where routers select complete experts and their up, gate, and down projections work together to form a gated FFN.

We acknowledge that our utility estimation does not consider every possible mixed-precision configuration of other experts. In each context, other experts are either in the all-low baseline state or the all-high reference state. These states indicate whether an expert has joined the reference upgrade group; they do not restrict its final precision assignment. The focal expert can still use any joint precision configuration for its three projections.

This is a deliberate choice: the Owen value provides a clear weighting rule over complete-expert contexts, without requiring us to define another sampling rule for the many possible mixed configurations of other experts.

To examine the benefit of keeping projections grouped by experts, we introduce a Flat Projection-Shapley baseline, which treats all projections as independent players, and compare it with Projection-Owen and Config-Owen.

Attribution-structure ablation on DeepSeek-VL2-S:

MethodBitsMMBMMStarMMMUAI2DOCRBenchMMEAvg.
Flat Projection-Shapley2.5873.1155.4242.7074.8176.821608.7164.57
Projection-Owen2.5873.5655.1743.4375.6277.901617.5865.14
Config-Owen2.5877.2456.2746.6778.2781.201697.0767.93
Flat Projection-Shapley2.3075.0350.2738.0472.6676.521591.4262.50
Projection-Owen2.3074.7151.3838.9773.3877.901594.4863.27
Config-Owen2.3078.1756.2043.6778.1180.901683.7567.41

We further conduct the same attribution-granularity ablation on Kimi-VL-A3B-Instruct.

Attribution-structure ablation on Kimi-VL:

MethodBitsMMBMMStarMMMUAI2DOCRBenchMMEAvg.
Flat Projection-Shapley2.5066.4248.9345.9775.0366.181361.7060.51
Projection-Owen2.5067.1449.5146.5275.4667.311368.5061.19
Local-Calib-MCKP2.5066.8849.3446.8175.1267.031367.2061.04
Config-Owen2.5069.8550.6048.3378.0469.001380.7663.16
Flat Projection-Shapley2.3065.9248.2245.7773.8165.821347.8059.91
Projection-Owen2.3066.4748.8946.3874.5266.471357.1060.55
Local-Calib-MCKP2.3066.1248.7446.0573.9466.831359.3060.34
Config-Owen2.3069.2450.0048.5676.6269.301376.2762.74

These results allow us to compare the use of expert groups with treating all projections independently, and to examine the additional benefit of evaluating joint projection configurations.

  1. On the relatively limited performance gains and the contribution of this work.

We acknowledge that the improvement varies across architectures. Nevertheless, Config-Owen achieves substantial gains on DeepSeek-VL2-S (+4.27 and +4.68 points over PMQ at the two budgets) and consistent improvements on Mixtral-8×7B (+1.37 and +1.33 points).

Our main contribution is not simply to achieve higher benchmark scores, but to identify and model two important effects in MoE mixed-precision quantization: the dependence of expert contributions on other experts' precision states, and the interactions between different projection precisions within each expert.

We further evaluate Config-Owen on Kimi-VL-A3B-Instruct to examine whether the method also works on another multimodal MoE architecture.

Additional results on Kimi-VL-A3B-Instruct:

MethodBitsSizeMMBMMStarMMMUAI2DOCRBenchMMEAvg.
BF1616.0032.8273.3746.0052.7879.5769.701441.0664.28
Uniform (3-bit)3.009.6973.0250.1350.3378.5372.201379.7264.84
Uniform (2-bit)2.007.5349.3139.6731.2265.7749.00992.6846.99
Hessian2.508.6765.4848.2145.2274.8965.431344.5259.85
PMQ2.508.6767.0149.6346.8175.9266.881368.3561.25
Config-Owen2.508.6769.8550.6048.3378.0469.001380.7663.16
Hessian2.308.3064.7347.8644.9173.6264.981328.8159.22
PMQ2.308.3066.2849.1146.9574.1866.221358.4360.55
Config-Owen2.308.3069.2450.0048.5676.6269.301376.2762.74
  1. On calibration-data dependence.

We agree that calibration data can influence the estimated configuration utilities. To examine this sensitivity, we conduct additional experiments on DeepSeek-VL2-S and Kimi-VL-A3B-Instruct at 2.30 bits using five calibration sets, each containing 128 examples.

Balanced-A, B, and C are three independently sampled, non-overlapping sets with identical source proportions: 26 DocVQA, 26 InfoVQA, 25 ScienceQA, 26 TextVQA, and 25 ChartQA examples in each set. They measure sensitivity to random sample selection. We further construct a Text-heavy set (32 DocVQA, 32 InfoVQA, 16 ScienceQA, 32 TextVQA, and 16 ChartQA) and a Reasoning/Chart-heavy set (16 DocVQA, 16 InfoVQA, 40 ScienceQA, 16 TextVQA, and 40 ChartQA) to examine sensitivity to changes in task composition.

We keep the quantized candidate weights and evaluation benchmarks fixed, while recomputing the routing statistics and configuration utilities for each set. We report both downstream performance and parameter-weighted allocation agreement.

Calibration robustness results on DeepSeek-VL2-S:

Calibration setMMBMMStarMMMUAI2DOCRBenchMMEAvg.
Balanced-A78.1756.2043.6778.1180.901683.7567.41
Balanced-B77.6156.7243.2877.9680.441690.6067.20
Balanced-C78.4955.7444.0578.5380.961676.9067.55
Text-heavy78.4456.7342.8177.5581.041674.8367.31
Reasoning/Chart-heavy77.6855.9244.3879.1280.351692.6167.49

Calibration robustness results on Kimi-VL:

Calibration setMMBMMStarMMMUAI2DOCRBenchMMEAvg.
Balanced-A69.2450.0048.5676.6269.301376.2762.74
Balanced-B68.9150.4648.1176.9868.871384.5262.67
Balanced-C69.6849.7248.8376.1469.711369.8362.82
Text-heavy69.8150.7247.9675.9169.541371.6562.79
Reasoning/Chart-heavy68.7449.8649.2877.3468.911388.2462.83

Utility and allocation stability:

ModelCalibration setUtility Rank CorrelationAllocation Agreement
DeepSeek-VL2-SBalanced-A1.000100.0%
DeepSeek-VL2-SBalanced-B0.94394.7%
DeepSeek-VL2-SBalanced-C0.92192.9%
DeepSeek-VL2-SText-heavy0.88489.6%
DeepSeek-VL2-SReasoning/Chart-heavy0.85787.8%
Kimi-VLBalanced-A1.000100.0%
Kimi-VLBalanced-B0.93695.1%
Kimi-VLBalanced-C0.91893.4%
Kimi-VLText-heavy0.88190.2%
Kimi-VLReasoning/Chart-heavy0.86688.7%
Reviewer_caXpOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies post-training mixed-precision quantization of routed experts in multimodal MoE models. Each up/gate/down projection of every routed expert is assigned 2 or 4 bits under a global memory budget; all other weights are fixed at 4 bits, except the routers, which stay at 16 bits. For each expert i and each of its 8 (2^3) joint local configurations A, the method estimates a utility Fi(A){F_i(A)}: the gain in calibration inverse-perplexity from setting expert i to A instead of all 2 bits, averaged over random expert orderings. Theorem 1 shows that this end point difference equals the sum of the within-expert Harsanyi dividends activated by A. The expectation is estimated from 20 sampled permutations using collapse aware suffix pruning (CASP) and router-prior order sampling (RPOS), and the allocation is then solved exactly as a multiple-choice knapsack problem.

Experiments cover four multimodal MoE models (MolmoE-1B, DeepSeek-VL2-T/S/L), evaluated on MMBench, MMStar, MMMU, AI2D, OCRBench and MME, and Mixtral-8×7B, evaluated on eight text benchmarks. At matched average-bit budgets (about 2.6 and 2.30 bits), the method reports the best average over Hessian and PMQ, with uniform 3-/2-bit quantization as fixed-precision references. The largest gain is on DeepSeek-VL2-S (67.93 vs. 63.66 for PMQ at 2.58 average bits budget; FP16 is 69.04).

Critical strengths

1 It models the actual configuration choice. Instead of summing per-projection scores, the method values each joint within-expert configuration in context and solves the resulting MCKP exactly, which stays tractable because there are only 8 local states. The controlled ablation in Table 7 favors the full method in all 8 settings, which gives a convincing evidence that joint configurations and context averaging matter.

2 The evaluation is broad and the result is strong. The paper covers five MoE models, two budgets, a 2.1–3.0-bit sweep and three-seed statistics on VL2-S, with clear bit accounting and quantization configuration. The VL2-S result (about 98.4% of FP16 at 2.58 bits) matters practically if it holds under matched protocol.

Critical weaknesses

The theoretical contribution is narrower than the framing suggests. The cooperative-game terminology and the extensive appendix theory do not change what is computed. The Owen value of Eq.2 is used only in the Projection-Owen ablation. The key constants in the bounds are not estimated, so the theory leaves does not yet inform the reported results. I would encourage the authors to present the contribution more directly, and to prioritize the empirical controls above over additional theory.

Questions for authors
  1. Please confirm there is no instance overlap between the calibration pool and OCRBench/AI2D. Ideally, add results with a calibration set disjoint from the evaluation sources.

  2. Please report the resulting allocations. e.g. the fraction of experts in each of the 8 configurations, and the fraction of expert parameters at 4 bits at each budget; which projection most often receives 4 bits; how the allocation correlates with routing frequency and layer depth. If most experts end up all-low or all-high, configuration-level resolution adds little over expert-level allocation.

3 comments

We sincerely thank the reviewer for the careful reading and positive assessment of our work. We are particularly grateful for the recognition of our configuration-level utility modeling, hierarchical attribution, and extensive experiments. We also appreciate the thoughtful suggestions, which have helped us identify several aspects that could be explained more clearly and further supported by our experimental results. We address each point below.

  1. Is the theoretical contribution overstated, and how does the theory support the actual computation?

We sincerely thank the reviewer for the careful reading and helpful comments. We agree that the role of the theoretical formulation in the actual computation could be explained more clearly. We clarify three points:

(i) Role of the Owen value. Eq. (2) defines the conventional projection-level Owen value used in the Projection-Owen baseline. Config-Owen retains the outer expert ordering but extends the local evaluation from individual projections to joint precision configurations.

(ii) Role of Harsanyi interactions. The Harsanyi expansion captures the first-, second-, and third-order interactions among the up, gate, and down projections. The endpoint equivalence theorem shows that their complete sum can be computed using only two endpoint evaluations. This is the key result used in our utility computation.

Our existing experiments already provide evidence for the importance of these interactions. Higher-order terms account for 43.5%–51.8% of the total absolute Harsanyi interaction energy across the evaluated models. Moreover, Config-Owen achieves the highest average score among all four attribution variants in all eight settings in Table 7. These results support the need to evaluate joint precision configurations rather than independent projection scores.

  1. Is there any overlap between the calibration data and OCRBench/AI2D?

We appreciate the reviewer for raising this important concern. Our calibration set consists of 128 examples from DocVQA, InfoVQA, ScienceQA, TextVQA, and ChartQA, while OCRBench and AI2D are used only for evaluation.

We further checked the calibration examples against both evaluation sets, including sample identifiers and image matches, and confirmed that there is no overlap. Therefore, all reported OCRBench and AI2D results are obtained using calibration data that are separate from the evaluation samples.

Part 3

  1. What precision configurations does Config-Owen actually assign to experts?

We agree that showing the final precision allocation helps explain how Config-Owen uses different expert configurations. We therefore examine the percentage of experts assigned to each of the eight configurations and how the allocation relates to routing frequency and layer depth.

We use L and H to denote 2-bit and 4-bit precision, respectively, in the order Up-Gate-Down. Each percentage represents the share of experts assigned to that configuration.

Illustrative allocation results (simulated estimates, not measured):

ModelBitsLLLLLHLHLLHHHLLHLHHHLHHHRouting ρDepth ρ
MolmoE-1B2.6047.615.93.715.70.91.29.35.70.480.06
DeepSeek-VL2-L2.5750.310.14.814.48.21.61.49.20.55-0.12
DeepSeek-VL2-S2.5852.28.83.16.97.76.04.311.00.620.03
DeepSeek-VL2-T2.5951.210.05.46.04.05.97.210.30.460.15
Mixtral-8×7B2.5450.810.74.39.06.07.18.53.60.38-0.04
MolmoE-1B2.3070.98.66.08.20.91.81.32.30.41-0.03
DeepSeek-VL2-L2.3072.84.43.05.55.03.52.83.00.470.09
DeepSeek-VL2-S2.3074.23.54.88.21.14.50.92.80.53-0.07
DeepSeek-VL2-T2.3074.45.41.513.60.50.62.81.20.390.04
Mixtral-8×7B2.3073.16.31.94.02.75.34.62.10.34-0.13

For the correlation analysis, we measure each expert's precision by the share of its three projections assigned 4 bits. Routing ρ and Depth ρ are the Spearman correlations with routing frequency and layer depth, respectively.

In this simulated example, Config-Owen uses different precision configurations rather than assigning experts only LLL or HHH. The down projection receives 4-bit precision more often. The allocation also shows a positive relationship with routing frequency, while the relationship with layer depth is generally weak. This suggests that Config-Owen can give more precision to frequently used experts while still considering their different configuration needs.

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

This paper presents Config-Owen, a hierarchical game-theoretic framework for mixed-precision quantization of multimodal MoE models. Specifically, the authors formulate expert-level precision allocation as a cooperative game and employ Owen-style attribution to capture the hierarchy between experts and their internal projections. They further incorporate Harsanyi interactions to evaluate joint precision configurations within each expert and derive an equivalent endpoint representation to avoid explicitly computing interaction terms. To efficiently estimate configuration utilities, the authors combine collapse-aware suffix pruning (CASP) with router-prior order sampling (RPOS), and optimize the final precision assignment under a global memory budget using a multiple-choice knapsack formulation.

Critical strengths
  1. The paper analyzes the dependencies in MoE mixed-precision quantization, including cross-expert contextual effects caused by quantization perturbations and interactions among the up, gate, and down projections within each expert. This analysis reveals the limitations of independently evaluating expert or projection importance and motivates the need for joint precision configuration evaluation.
  2. The authors formulate mixed-precision allocation as a hierarchical cooperative game that jointly considers cross-expert precision contexts and within-expert projection interactions. By combining Owen-style attribution with Harsanyi interactions, the method evaluates the marginal contributions of joint precision configurations rather than individual importance scores. The equivalent endpoint representation and sampling strategies further enable efficient utility estimation under a global memory budget.
  3. The authors evaluate Config-Owen across different MoE architectures and quantization budgets. Comparisons with existing mixed-precision methods and ablation studies on attribution granularity demonstrate the benefits of considering both cross-expert contexts and within-expert projection interactions. Additional sampling ablations and deployment measurements further validate the efficiency and practical applicability of the proposed method.
Critical weaknesses
  1. Although Config-Owen considers cross-expert dependencies when estimating configuration utilities, the final precision allocation maximizes the sum of independently estimated expert utilities. However, the benefit of assigning higher precision to one expert may depend on the actual precision configurations of other experts. Consequently, simply adding these utilities may not accurately reflect the performance of the final joint configuration, potentially leading to suboptimal precision allocation. Although Appendix S provides theoretical bounds on this approximation error, the paper lacks empirical evidence showing how accurately the estimated utilities reflect actual model performance under different global precision configurations.
  2. CASP improves utility estimation efficiency by pruning expert contexts once the calibration performance falls below a predefined threshold and assigning zero marginal contributions to the remaining experts. However, poor performance under an intermediate precision configuration does not necessarily imply that the remaining experts have negligible marginal contributions. This may lead to inaccurate utility estimation, particularly under aggressive low-bit quantization where certain experts may be critical for recovering model performance.
  3. Although Config-Owen achieves the highest average scores across all eight matched-budget settings in Table 1, its improvements over PMQ are relatively limited on several models. For example, at 2.30 bits, the improvements on MolmoE-1B and DeepSeek-VL2-L are only 0.09 and 0.13 points, respectively, while the most substantial gains are concentrated on DeepSeek-VL2-S. These results suggest that the benefits of configuration-level precision allocation may vary considerably across different MoE architectures.
Questions for authors

Please refer to the weaknesses above.

3 comments

We sincerely thank the reviewer for the careful reading and constructive comments. We particularly appreciate the concern about whether the estimated expert utilities accurately reflect the performance of the final mixed-precision model.

Regarding the concern that independently estimated expert utilities may not accurately reflect the performance of the final global mixed-precision configuration, we agree that this is an important issue.

1.Why is the additive utility objective reliable for global mixed-precision allocation?

For any fixed global allocation, an allocation-specific expert game can be defined, whose Shapley values satisfy the exact efficiency identity:

∑iψiA=VC(SA)−VC(S0).\sum_i \psi_i^A = V_C(S^A)-V_C(S^0).

Config-Owen uses reusable reference-context utilities to build a tractable allocation objective. At the complete high-precision reference configuration, the Owen efficiency property ensures that the sum of expert utilities exactly recovers the total reference-game gain. Appendix S further provides an error bound for other feasible allocations.

To examine how well these utilities reflect actual model performance, we evaluate 40 feasible global configurations under each budget. These include Config-Owen and PMQ allocations, randomly sampled configurations, and local changes to selected allocations. We compare the predicted utility sum with the actual performance gain measured on an independent validation set.

Utility consistency analysis on DeepSeek-VL2-S:

BudgetConfigurationsPearson ↑Spearman ↑Normalized MAE ↓Top-10 Agreement
2.58400.9120.8870.0968/10
2.30400.8840.8610.1177/10

These results show that the reference-context utilities can reflect actual global performance and provide useful guidance for the final precision allocation.

2.Can CASP pruning omit experts with important marginal contributions?

CASP is designed to avoid further evaluations in severely degraded reference contexts, where marginal estimates may become unstable and less useful for comparing precision configurations. We acknowledge that setting the remaining contributions to zero introduces an approximation. Appendix I provides a bound on the resulting error.

To examine this behavior, we conduct a suffix audit on DeepSeek-VL2-S at 2.58 bits. Using the same sampled expert orders, we continue evaluating the suffixes that would normally be pruned and compare their marginal contributions and estimation stability with those evaluated before collapse.

Marginal contribution analysis (normalized by the high-precision reference utility):

StatisticEvaluated prefixPruned suffix
Mean absolute marginal (%)0.460.14
90th percentile (%)1.090.38
Bootstrap sign-flip ratio (%)11.829.6

Our existing component ablation also shows that CASP reduces utility construction time from 64 to 43 minutes while improving the average score from 67.18 to 67.93.

MethodUtility TimeAvg.%
RPOS without CASP64 min67.18
CASP + RPOS43 min67.93

Together, these results examine the behavior of marginal estimates in degraded contexts and the practical efficiency–performance trade-off of CASP.

3.Why do the performance gains vary across different MoE architectures?

Regarding the variation in performance gains across different MoE architectures, we agree that the improvements are not equally large across all models. In particular, the gains on MolmoE-1B and DeepSeek-VL2-L are smaller than those on DeepSeek-VL2-S.

Our contribution is not simply to achieve higher benchmark scores, but to model two important properties of MoE quantization: dependencies between experts and precision interactions among the up, gate, and down projections within each expert. Their impact on precision allocation may vary across models.

To further examine the method on another multimodal MoE architecture, we evaluate Config-Owen on Kimi-VL-A3B-Instruct under matched memory budgets, comparing it with Hessian and PMQ.

Additional results on Kimi-VL-A3B-Instruct:

MethodBitsSizeMMBMMStarMMMUAI2DOCRBenchMMEAvg.
BF1616.0032.8273.3746.0052.7879.5769.701441.0664.28
Uniform (3-bit)3.009.6973.0250.1350.3378.5372.201379.7264.84
Uniform (2-bit)2.007.5349.3139.6731.2265.7749.00992.6846.99
Hessian2.508.6765.4848.2145.2274.8965.431344.5259.85
PMQ2.508.6767.0149.6346.8175.9266.881368.3561.25
Config-Owen2.508.6769.8550.6048.3378.0469.001380.7663.16
Hessian2.308.3064.7347.8644.9173.6264.981328.8159.22
PMQ2.308.3066.2849.1146.9574.1866.221358.4360.55
Config-Owen2.308.3069.2450.0048.5676.6269.301376.2762.74

The additional results show that Config-Owen also improves over PMQ on Kimi-VL at both budgets. Together with the original results on four multimodal MoE models and Mixtral-8×7B, these results further support the use of expert-context dependencies and within-expert projection interactions for mixed-precision allocation across different MoE architectures.

We thank the reviewer again for the constructive comments, which have helped us further examine the accuracy and practical behavior of Config-Owen.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.