PaperlayerView exact revision on arXiv

Source-linked AI summary

RateQuant: Optimal Mixed-Precision KV Cache Quantization via Rate-Distortion Theory

Fei Zuo, Zikang Zhou, Hao Cong, Xiaoyan Xi, Ho Fai Leung

arXiv:2605.06675v2cs.LGcs.CLcs.ITmath.IT

TL;DR

KV cache quantization can reduce a major LLM serving-memory cost, but uniform bit-widths overlook head-importance differences and mismatched distortion models can make mixed-precision allocation worse than uniform. RATEQUANT calibrates each quantizer’s distortion model and uses rate-distortion allocation. On KIVI at 2.5 bits, it reduces perplexity from 49.3 to 14.9 with zero runtime overhead.

  • Problem

    KV-cache quantizers assign every attention head the same bit-width despite non-uniform head importance, and different quantizers’ distortion curves make mixed-precision allocation vulnerable to model mismatch.

    Source: §1 · pp. 1–2

  • Method

    RATEQUANT fits a per-quantizer distortion model from a small calibration set and solves per-head bit allocation through closed-form reverse waterfilling.

    Source: §1 · p. 2

  • Results

    KIVI perplexity falls from 49.3 to 14.9 at 2.5 bits (70% reduction), and RATEQUANT recovers 70–85% of quantization-induced degradation at 4.0 bits.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Empirical distortion calibration, gradient-based sensitivity, and separate K/V budgets make RATEQUANT a quantizer-agnostic allocation layer with zero runtime overhead.

    Source: §6 · p. 9

    The allocation assumes per-head quantization MSE follows the exponential distortion-rate form D(b) = α·β^-b, with α > 0 and β > 1.

    Source: §3.1 · p. 3

Abstract

from arXiv · show

Large language models cache all previously computed key-value (KV) pairs during generation, and this KV cache grows linearly with sequence length, making it a primary memory bottleneck for serving. Quantizing the KV cache to fewer bits reduces this cost, yet all current quantizers assign the same bit-width to every attention head, ignoring the large variation in head importance. A natural idea is to allocate more bits to important heads and fewer to the rest. We show, however, that such mixed-precision allocation has a hidden pitfall: each quantizer follows a different distortion curve D(b)=alpha*beta^{-b}, and the decay rate beta varies from 3.6 to 5.3 across quantizer designs. Applying one quantizer's distortion model to another inverts the allocation order and makes performance worse than uniform quantization. We call this failure mode distortion model mismatch and propose RateQuant to resolve it. RateQuant fits a per-quantizer distortion model from a small calibration set, then solves the resulting bit-allocation problem in closed form via reverse waterfilling from rate-distortion theory. On Qwen3-8B at 2.5 average bits, calibrated RateQuant reduces KIVI's perplexity from 49.3 to 14.9 (70% reduction) and improves QuaRot by 6.6 PPL. The entire calibration takes 1.6 s on a single GPU and adds zero overhead at inference time.

1 Introduction

KV caches are a major serving-memory bottleneck, and head importance varies even though quantizers generally assign every head the same bit-width. RATEQUANT addresses distortion-model mismatch with per-quantizer calibration and rate-distortion-based allocation.

1 Introduction

  • Because head importance is non-uniform but existing quantizers use uniform bit-widths and prior mixed-precision rules are quantizer-specific, the paper asks for principled, quantizer-agnostic head allocation.The introduction frames this as two questions: estimating head contribution efficiently and avoiding distortion-model mismatch.Source: §1 · pp. 1–2
  • Quantizers’ distortion decay rates β range from 3.6 to 5.3, so applying one quantizer’s model to another can worsen mixed-precision allocation relative to uniform allocation.Mismatched β can invert marginal-gain ordering and send bits to the wrong heads.Source: §1 · pp. 1–2
  • Gradient-based sensitivity outperforms activation-based methods by 1.07 PPL at 3.5 bits, while the sensitivity proxy matters more than the allocation algorithm.Per-quantizer calibration also reduces KIVI PPL from 49.3 to 14.9 after using TURBOQUANT’s distortion model worsens it to 87.0 at 2.5 bits.Source: §1 · p. 2
  • RATEQUANT fits a per-quantizer distortion model from a small calibration set and solves per-head bit allocation through closed-form reverse waterfilling.The method also uses separate K/V budgets; the AM/GM ratio of head sensitivities predicts when mixed precision helps.Source: §1 · p. 2
  • KIVI perplexity falls from 49.3 to 14.9 at 2.5 bits (70% reduction), while RATEQUANT recovers 70–85% of quantization-induced degradation at 4.0 bits.The reported gains span Qwen3 and Llama3 families.Source: §1 · p. 2

2 Related Work

Prior KV quantizers use uniform precision across heads, while prior mixed-precision methods generally target layers or channels and are tied to particular quantizers. RATEQUANT allocates bits per head and uses calibration to support different quantizers.

KV cache quantization

  • Existing KV quantizers assign identical bit-widths to every head, while layer- and channel-level mixed-precision methods are designed around specific quantizers, limiting transferability.RATEQUANT instead supports arbitrary quantizers through calibration and allocates at per-head granularity.Source: §2 · p. 2

Rate-distortion theory in neural network quantization

  • RATEQUANT applies rate-distortion allocation to KV caches, where heads have distinct sensitivities and keys and values have asymmetric error characteristics.Earlier rate-distortion quantization methods target model weights; RATEQUANT provides closed-form allocation and theoretical bounds on achievable gain.Source: §2 · p. 3

Sensitivity estimation for quantization

  • For KV cache allocation, the paper reports gradient-based sensitivity as qualitatively superior to activation-based metrics, and the proxy choice matters more than the allocation algorithm.Prior quantization sensitivity methods include Hessian-based approaches and activation-based metrics.Source: §2 · p. 3

3 RATEQUANT

RATEQUANT allocates KV-cache bits by head importance using rate-distortion optimization, with per-quantizer calibration to avoid distortion-model mismatch. Its stages combine sensitivity estimation, continuous and integer allocation, and extensions for quantizers with different key and value distortion curves.

3.1 Problem Formulation and Optimal Allocation

  • Under an exponential per-head distortion model, reverse waterfilling assigns more bits to more sensitive heads, with allocation increments scaling logarithmically with sensitivity.The AM/GM ratio of head sensitivities predicts potential gain; Qwen3 models exhibit ratios around 2.0, and Llama3 models around 1.8–2.2.Source: §3.1 · pp. 3–4

3.2 Sensitivity Estimation

  • Gradient-based sensitivity outperforms activation norms by 1.07 PPL at 3.5 bits, and its estimates take about 1.6 seconds to compute for Qwen3-8B.The estimates have a coefficient of variation below 3% for 95% of heads across three seeds and can be reused across bit budgets and quantizers.Source: §3.2 · p. 4

3.3 Integer Allocation

  • For integer bit-widths, greedy allocation by weighted marginal distortion reduction is optimal under the convex distortion model and completes in under 10 ms for Qwen3-8B.A max-heap reduces the allocation cost to O(R log N).Source: §3.3 · pp. 4–5

3.4 Quantizer-Agnostic Extensions

  • Per-quantizer distortion calibration prevents mismatched decay rates from inverting head rankings, while separate key and value allocation accommodates their distinct distortion curves.For KIVI at 2.5 bits, separate allocation gives average bit-widths of 2.85 for keys and 2.15 for values.Source: §3.4 · p. 5
  • 70–85% of quantization-induced degradation is recovered at extreme compression rates, with negligible calibration overhead.RATEQUANT also consistently recovers 50–75% at 3.0–4.0 bits under TURBOQUANT quantization.Source: §3.4 · p. 6, Table 1 · p. 6

4 Experiments

The experiments evaluate RATEQUANT across Qwen3 and Llama3 models, quantizers, and bit budgets. They find consistent gains over uniform allocation, while showing that cross-quantizer transfer depends on calibration, gradient-based sensitivity, and suitable operating conditions.

4 Experiments

  • The experiments address uniform-allocation gains, cross-quantizer calibration, sensitivity-proxy choice, and the conditions under which RATEQUANT helps.The study spans three model sizes and three base quantizers.Source: §4 · p. 6

4.1 Experimental Setup

  • The evaluation covers five Qwen3 and Llama3 models, WikiText-2 perplexity, downstream benchmarks, and three base quantizers at matched bit budgets.Calibration uses 16 WikiText-2 sequences of length 512 for sensitivity estimation and distortion fitting.Source: §4.1 · p. 6, §4.2 · p. 6

4.2 Q1: Does RateQuant Improve Over Uniform Allocation?

  • At 4.0 bits, RATEQUANT recovers 76% of PPL degradation and 82.0% of the downstream accuracy gap, with gains across Qwen3 and Llama3.Qwen3-8B reaches 63.2% average accuracy versus 63.8% for FP16 at 4× lower KV-cache memory; the downstream comparison maintains parity throughput.Source: §4.2 · pp. 6–7, Table 2 · p. 7

4.3 Q2: Does Distortion Calibration Enable Cross-Quantizer Transfer?

  • At 2.5 bits, calibrated RATEQUANT with separate K/V budgets reduces KIVI PPL from 49.3 to 14.9, while a mismatched model worsens KIVI and QuaRot.The allocation assigns error-prone keys 2.85 bits and values 2.15 bits; gains are also reported on Llama3 and Qwen3-4B.Source: §4.3 · pp. 7–8, Table 3 · p. 7

4.4 Q3: Which Sensitivity Proxy Is Correct?

  • At 3.5 bits, gradient-based sensitivity yields 9.72 PPL versus 10.79 for activation-based sensitivity, a 1.07-PPL advantage.Gradient sensitivity is U-shaped across layers, whereas activation sensitivity increases with depth; the activation proxy can over-allocate to late layers.Source: §4.4 · p. 8, Figure 4 · p. 8

4.5 Q4: When Does RateQuant Help?

  • RATEQUANT’s gains are strongest at 2.5–4.0 bits and require heterogeneous head sensitivities, measurable quantization headroom, and a distortion model calibrated to the quantizer.At ≥5 bits uniform allocation approaches FP16; below 2 bits, quantization noise dominates sensitivity-based differentiation.Source: §4.5 · pp. 8–9, Figure 5 · p. 8

5 Discussion and Future Work

The discussion frames rate-distortion optimization as a principled, modular approach to KV-cache compression. It identifies dynamic allocation, broader distortion models, joint eviction and quantization, and extensions to other inference components as future directions.

Discussion

  • Under its exponential distortion model, RATEQUANT provides closed-form allocation that is optimal under the model assumptions, and its empirical results support those assumptions.The framework can incorporate new quantizers by calibrating their distortion parameters, without changing the allocation algorithm.Source: §5 · p. 9
  • The authors propose co-designing quantizers for favorable distortion characteristics, such as higher β for steeper quality gains.They also note that gradient-based sensitivity weights heads by their contribution to the loss landscape, decoupling allocation from specific downstream tasks.Source: §5 · p. 9

Future Work

  • Future work includes dynamic input-dependent allocation, refining distortion curves for emerging quantizers, combining quantization with KV eviction, and extending rate-distortion methods to weights and activations.Preliminary experiments suggest optimal allocations can vary by up to 0.5 bits across code and natural-language domains.Source: §5 · p. 9

6 Conclusion

The conclusion presents RATEQUANT as a calibrated, rate-distortion-based allocation layer for mixed-precision KV-cache quantization. The included appendix material outlines its supporting comparisons, proofs, and distortion parameters.

6 Conclusion

  • On KIVI at 2.5 bits, RATEQUANT reduces Qwen3-8B perplexity from 49.3 to 14.9 with zero runtime overhead and under two seconds of one-time calibration.The conclusion attributes cross-quantizer transfer to empirical distortion calibration, gradient-based sensitivity, and separate K/V budgets.Source: §6 · p. 9

Appendix Overview

  • The appendix overview lists proofs, calibrated distortion parameters, ablations, baselines, fit validation, reliability, allocation maps, and additional results.It also identifies sections on calibration overhead, memory footprint, downstream tasks, and implementation.Source: Appendix Overview · p. 13

A Related Work Comparison Table

  • RATEQUANT uniquely combines per-head granularity, rate-distortion theory, calibration, K/V separation, and closed-form allocation in the comparison table.Source: Table 4 · p. 13

B.1 Proof of Theorem 2 (Reverse Waterfilling)

  • The reverse-waterfilling procedure computes free-head bit allocations, clips them to bounds, and repeats; it converges within at most N steps.Source: §B.1 · p. 14

B.2 Proof of Theorem 3 (Gain Ratio)

  • The gain-ratio proof compares uniform and optimized distortion, deriving a uniform-to-optimal ratio of at least one by AM-GM.Source: §B.2 · p. 14

B.3 Proof of Theorem 6 (Greedy Optimality)

  • The greedy allocation proof uses decreasing per-head marginal gains to establish a polymatroid structure, making greedy selection optimal.Source: §B.3 · p. 14

B.4 Proof of Theorem 5 (Loss-Distortion Connection)

  • The loss-distortion discussion distinguishes forward-pass error amplification from loss impact; residual connections can absorb error from a head with low loss impact.Source: §B.4 · pp. 14–15

C Distortion Model Parameters

  • The distortion-parameter table identifies a 1.5× β gap across quantizers as the root cause of mismatch.Source: Table 5 · p. 15

D Component Ablation Waterfall

The appendix ablation isolates calibration and K/V separation as key components of RATEQUANT. It also reports distortion-fit accuracy, while the supplied baseline and multi-seed passages describe comparison scope and reliability checks.

D Component Ablation Waterfall

  • On KIVI at 2.5 bits, gradient sensitivity without calibration worsens PPL from 49.3 to 87.0, calibration reduces it to 73.1, and K/V separation brings it to 14.9.The final split is 2.85 bits for keys and 2.15 for values.Source: D Component Ablation Waterfall · p. 15, Table 6 · p. 15

E Mixed-Precision Baseline Comparison

F Distortion Model Validation

G Multi-Seed Reliability

  • The primary TURBOQUANT evaluation reports multi-seed results, while cross-quantizer experiments use seed 42 because allocation strategy is the dominant stated variance source.The table caption says all seeds show positive gains at 3.0–4.0 bits.Source: G Multi-Seed Reliability · p. 16, Table 9 · p. 16

H K/V Asymmetry Visualization

KIVI’s key cache has substantially higher distortion than its value cache, so the optimal allocation favors keys at low bit budgets. RATEQUANT assigns bits unevenly across both heads and K/V caches, with allocation patterns varying by layer and budget.

H K/V Asymmetry Visualization

I Bit Allocation Visualization

  • At a 4.0-bit budget, early and late layers receive 5–6 bits per head while middle layers receive 3–4 bits, forming a U-shaped allocation.Within layer 1, head 7 receives 6 bits while head 3 receives 4 bits.Source: I Bit Allocation Visualization · p. 17, Figure 7 · p. 17
  • At 3.0 bits, allocation differentiation is maximal at 2–5 bits; at 5.0 bits, allocations move toward uniformity as heads approach FP16 quality.The stated sweet spot for balancing memory savings and quality preservation is 3.5–4.0 bits.Source: I Bit Allocation Visualization · p. 17

J Calibration Overhead

  • Table 10 reports calibration cost for a single H200 GPU using 16 sequences of length 512.Source: Table 10 · p. 18

K Calibration Size Ablation

RATEQUANT’s calibration is robust to small calibration sets, and its per-head allocation preserves the memory savings of uniform quantization at the same average bit-width. The supplied results also describe constrained-budget behavior, an extreme-case quantizer, calibration costs, and the evaluation scope and design boundaries.

K Calibration Size Ablation

  • Even with 4 calibration sequences, RATEQUANT achieves 97% of the PPL improvement obtained with 16 sequences; 64 sequences provide negligible added benefit while quadrupling calibration time.The recommended default is 16 sequences, completing in 1.6 s.Source: K Calibration Size Ablation · p. 18, Table 11 · p. 18

L Memory Footprint

  • At 4.0 bits, KV cache memory reduces by 4×, and at 3.5 bits by 4.6×; RATEQUANT matches uniform quantization’s memory savings at the same average bit-width.For Qwen3-8B at 3.5 bits and 4096 tokens, memory falls from 576 MB to 126 MB.Source: L Memory Footprint · p. 18, Table 12 · p. 18

M.3 Qwen3-32B

  • When the minimum-bit constraint floors all heads, the gain ratio is exactly 1; gains grow as the floor fraction falls, then may diminish at high bit budgets.The passage notes small or negative Δ at budgets of at least 4.5 bits.Source: §M.3 · p. 19

M.4 RTN Base Quantizer (Extreme Case)

  • RATEQUANT produces significant gains with RTN per-token symmetric, the weakest quantizer tested, when quantization error is large.Source: §M.4 · p. 19, Table 16 · p. 20

O Scope and Extensions

  • Experiments cover Qwen3 4B/8B/32B and Llama3 3B/8B; validation on other architectures and long-context benchmarks remains to be done.The passage names Mistral, Gemma, DeepSeek, RULER, and LongBench as examples.Source: O Scope and Extensions · p. 20
  • Gradient-based sensitivity estimation takes approximately 1.6 s for 8B models on one H200 GPU; activation-based proxies are an option under strict calibration budgets at a modest accuracy trade-off.The sensitivity-proxy comparison is identified in Table 17.Source: O Scope and Extensions · p. 20, Table 17 · p. 20
  • The current implementation allocates bits per head uniformly across token positions and treats heads independently; position-aware allocation and correlated-head modeling are proposed extensions.Source: O Scope and Extensions · p. 20

P Implementation Details

  • The stated calibration procedure uses 16 WikiText-2 training sequences of length 512, sampled at random starting positions and kept non-overlapping.Source: P Implementation Details · p. 20

Q Downstream Task Details

The downstream evaluation spans seven benchmarks covering reasoning, knowledge, and commonsense. At 4.0 bits on Qwen3-8B, RATEQUANT improves or matches uniform quantization on every task.

Q Downstream Task Details

R Sensitivity Distribution Analysis

Gradient sensitivity varies across Qwen3-8B layers, with higher sensitivity in early and late layers and lower, more uniform sensitivity in middle layers. The log-sensitivity distribution is approximately normal, though the theoretical gain ratio underestimates the observed ratio.

R Sensitivity Distribution Analysis

  • Early layers 1–6 and late layers 30–36 show elevated gradient sensitivity, while middle layers 12–24 are lower and more uniform.The heatmap compares Key and Value sensitivity.Source: R Sensitivity Distribution Analysis · p. 21, Figure 8 · p. 22
  • Log-sensitivity ln w is approximately normal across heads (Shapiro-Wilk p > 0.1), but the theoretical AM/GM gain ratio of 1.34 underestimates the observed 2.0× ratio.The reported gap suggests the exponential distortion model partially underestimates gains for extreme outliers.Source: R Sensitivity Distribution Analysis · p. 21

S Cross-Architecture Comparison

Across evaluated models, gradient-based head sensitivities show substantial heterogeneity, while KIVI’s optimal key/value bit split changes with the compression budget. The appendix also examines allocation bounds, calibration length, cross-quantizer transfer, runtime, and reproducibility.

S Cross-Architecture Comparison

T Distortion Curve Visualization

  • The distortion-curve β gap between TURBOQUANT (3.6) and KIVI/QuaRot (5.0–5.3) causes allocation mismatch.Figure 9 shows key-cache distortion curves for Qwen3-8B.Source: Figure 9 · p. 22

U Algorithm Pseudocode

  • The algorithm initializes integer allocations at bmin, tracks remaining budget R in a max-heap, and repeatedly increments the allocation for the popped head with maximum marginal gain until R reaches zero.Source: U Algorithm Pseudocode · p. 23

V Hyperparameter Sensitivity

  • Wider bit-width bounds enable fuller differentiation but risk extreme allocations, while calibration sequences longer than 512 tokens yield diminishing sensitivity-estimation returns.The study uses 512 tokens as a balance between accuracy and speed.Source: V Hyperparameter Sensitivity · p. 23, Table 20 · p. 23

W.1 Per-Layer Analysis

  • For Qwen3-8B, sensitivity variance is highest in early and late layers, which receive 4.5–5.5 bits versus 3.0–3.5 bits in middle layers.The table reports per-layer sensitivity statistics and average allocation at 4.0 bits.Source: §W.1 · pp. 23–24, Table 21 · p. 23, Table 22 · p. 24

W.2 Llama3 Detailed Results

  • The Llama3.2-3B and Llama3.1-8B tables provide complete results, with the Llama3.2-3B passage reporting gains consistent with Qwen3 patterns.The Llama3.1-8B passage identifies it as the largest Llama model tested.Source: §W.2 · p. 24, Table 23 · p. 24, Table 24 · p. 24

W.3 Cross-Quantizer Transfer

  • Matched distortion calibration gives the best transfer-matrix results; mismatched β can invert head rankings and raise KIVI PPL from 14.86 to 86.95 at 2.5 bits.The table compares matched diagonal entries with mismatched off-diagonal entries.Source: §W.3 · p. 24, Table 25 · p. 25

W.6 Extended Perplexity Analysis

The section extends the perplexity analysis across positions and domains, examines allocation sensitivity, model fit, and failure conditions, and reports experimental hardware and research-impact, licensing, and human-subject disclosures.

W.6 Extended Perplexity Analysis

  • RATEQUANT improves mean PPL by 10–15% and maximum-PPL outliers by >25%; gains persist across positions and transfer to other domains without recalibration.The sequence-level results are for Qwen3-8B at 4.0 bits; position-dependent results are at 3.5 bits, and domain transfer uses allocation calibrated on WikiText-2.Source: §W.6 · pp. 25–26, Table 27 · p. 25, Table 28 · p. 26

W.7 Theoretical Derivations

  • The greedy heap algorithm is optimal for integer bit-widths when marginal distortion reduction decreases; this condition holds for the exponential model, while continuous allocation follows water-filling.For the exponential model, ∆Di(b) = αie−βib(1 −e−βi), which is decreasing.Source: §W.7 · p. 26

W.8 Sensitivity Estimation Details

  • Gradient sensitivity uses cross-entropy loss on the calibration set, with bfloat16 forward and float32 backward passes; L2 (Frobenius) aggregation performs best, and averaging across positions improves robustness.The positional average captures sensitivity across diverse contexts rather than relying only on the last token.Source: §W.8 · p. 27, Table 30 · p. 27

W.10 Error Analysis

  • For fewer than 2% of heads, poor fits (R2 < 0.96) trigger a conservative β fallback; sensitivity rankings remain stable across calibration seeds, with Spearman correlation >0.95.The exponential distortion model achieves R2 > 0.98 for most heads; the fallback uses β = min(βfitted, ¯β).Source: §W.10 · p. 28, Table 31 · p. 28

W.11 KIVI vs. QuaRot Comparison

  • KIVI has stronger K/V distortion asymmetry (βK = 5.28 vs. βV = 4.92) than QuaRot, so optimal K/V bit splits vary by quantizer and budget.The supplied passages identify the calibrated distortion parameters and report that optimal splits differ across budgets, but do not provide the split values.Source: §W.11 · p. 28, Table 32 · p. 28, Table 33 · p. 28

W.12 Qwen3-32B Extended Analysis

  • At 4.0 bits, Qwen3-32B’s KV cache shrinks from 1024 MB to 256 MB (4× compression); its 64 layers provide more differentiation opportunities than Qwen3-8B’s 36.Qwen3-32B has 512 KV head pairs versus 288 for Qwen3-8B.Source: §W.12 · p. 28, Table 34 · p. 29

W.13 Attention Pattern Analysis

  • RATEQUANT’s gradient sensitivity assigns more bits to retrieval heads; local heads average 3.2 bits, compared with 5.5 bits for sink-adjacent heads.Sink-adjacent heads in layers 1–3, heads 0–1, show 2–3× higher gradient sensitivity.Source: §W.13 · p. 29

W.14 Failure Mode Analysis

  • RATEQUANT provides minimal or no benefit at budgets ≥5.0 bits, in homogeneous models (AM/GM ≈1), or with extreme sensitivity outliers; capping sensitivity at 5 ¯w mitigates the outlier case.The cap is applied as wi ←min(wi, 5 ¯w) before allocation to prevent extreme outliers from dominating while preserving typical-head rankings.Source: §W.14 · p. 29

NeurIPS Paper Checklist

  • Experiments run on a single NVIDIA H200 GPU, with calibration times reported in Section 4.1.Source: NeurIPS Paper Checklist · p. 33
  • The paper identifies reduced memory and energy use as positive impacts while noting that efficiency gains could lower barriers to LLM misuse.It reports no human subjects, private data, or dual-use concerns beyond general LLM deployment efficiency.Source: NeurIPS Paper Checklist · p. 33
  • The paper credits Qwen3, Llama3, and WikiText-2 and reports using the models under their respective Apache 2.0 and Llama Community licenses.Source: NeurIPS Paper Checklist · p. 34
  • No new datasets or pretrained models are released, and the paper reports no crowdsourcing or research involving human subjects.Source: NeurIPS Paper Checklist · p. 35
  • Qwen3 and Llama3 are evaluation subjects only and are not components of the proposed methodology.Source: NeurIPS Paper Checklist · p. 36

Related papers

Loading 2605.06675v2…