Security-Enhanced Seed-Based Weight Quantization for Large Language Models

Qiuyu Ren; Sudipta Paria; Aritra Dasgupta; Swarup Bhunia

cs.CRSep 29, 2026

Security-Enhanced Seed-Based Weight Quantization for Large Language Models

Authors: Qiuyu Ren, Sudipta Paria, Aritra Dasgupta, Swarup Bhunia

Organizations: Department of Electrical and Computer Engineering (ECE) University of Florida Gainesville, FL 32611, USA

Abstract

Large language models (LLMs) incur substantial storage, memory-bandwidth and energy costs, motivating compact weight representations. Existing seed-based compression methods reconstruct weights from compact pseudo-random representations but do not explicitly account for the non-uniform sensitivity of model weights. We introduce Seed-Q, a security-enhanced sensitivity-aware seed-based weight compression framework that uses lightweight Linear Feedback Shift Register (LFSR)-based weight generation with non-uniform bit allocation. Our approach assigns larger representation budgets to sensitive weights while aggressively compressing less sensitive regions. Importantly, this non-uniform allocation requires no side-information: the decoder deterministically reconstructs the bit-allocation schedule, with no rung depending on the decoded weights, eliminating the need to store per-block metadata or use calibration data while preserving the baseline coding rate. Experiments across diverse LLMs show that Seed-Q matches 4-bit perplexity of SeedLM with fewer bits, while at the same 4 bits/weight it reduces both perplexity degradation and zero-shot accuracy loss relative to SeedLM. We also show that Seed-Q simultaneously achieves high security against bit-flip attacks on model parameters, as bit corruption affects multiple reconstructed weights, greatly amplifying its impact and making it easier to detect. We further implement Seed-Q in an ASIC-based accelerator and demonstrate modest hardware overhead compared to prior seed-based approaches.

Figures & tables

Method	Model Perplexity (Bits/Weight)				Mean
Method	LLaMA-2-7B	LLaMA-2-13B	LLaMA-3-8B	Mistral-7B	Bits $\dagger$
Baseline	5.5 (16)	4.9 (16)	6.1 (16)	5.25 (16)	16
Seed-Q (Mode 1)*	5.85 (3.77)	5.18 (3.75)	7.17 (3.88)	6.32 (3.78)	3.79
S-Quant	5.7 (3.8)	5.0 (3.8)	6.8 (3.8)	NR	3.8
Seed-Q (Mode 2) $\bigstar$	5.68 (4)	5.06 (4)	6.92 (4)	5.96 (4)	4
SeedLM $\ddagger$	5.85 (4)	5.18 (4)	7.17 (4)	6.32 (4)	4

Table 1: WikiText-2 perplexities for LLaMA 2, LLaMA 3 and Mistral 4-bit weight representations evaluated on 2048-token contexts. NR : Not Reported

Model	Method	Bits	ARC-Easy	ARC-Challenge	HellaSwag	WinoGrande	BoolQ	Mean
Model	Method	Zero-Shot Task Accuracy (%)
LLaMA-2-7B	Baseline	16	74.58	46.33	75.98	69.06	77.74	68.74
	Seed-Q (Mode 1)*	3.77	71.97	43.43	73.60	68.03	74.98	66.40
	S-Quant	3.8	73.36	44.55	74.51	68.47	77.34	67.65
	Seed-Q (Mode 2) $\bigstar$	4	72.31	44.45	74.69	69.53	75.90	67.38
	SeedLM ‡	4	72.39	43.60	73.79	67.72	76.85	66.87

Table 2: Performance comparison across different models and zero-shot tasks for around 4-bit configurations. Entries that ran out of memory in our setup are marked with OOM. NR : Not Reported.

Figure 2: Ablation of Seed-Q (WikiText-2 perplexity). (a) Rate, with the four outlier columns; dashed: SeedLM at

4.0

bits per weight; rings: Mode 1 and 2 (Table 2 ). (b) Floor

f

(§ 3.3 ) at

4.0

bits, column factor only, no outlier columns. (c) Outlier columns at

4.0

bits.

Panel of Figure 4 — Figure 4: Model performance before and after the selected fault. Pre-Flip denotes each scheme’s uncorrupted state. Lower PPL and higher accuracy indicate better performance. S-Quant and Seed-Q use simulated propagation.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Global
part of the ranking	Seed-Q mass	Fisher mass	Seed-Q exposure	random
top $0.01\%$	$42.5\%$	$59.0\%$
top $1\%$	$59.3\%$	$72.2\%$
top $10\%$	$65.6\%$	$82.0\%$
bottom $3\%$	$6.34\%$	$0.006\%$	$17.3\%$	$10\%$
bottom $10\%$	$12.4\%$	$0.095\%$	$21.0\%$	$10\%$

Appendix

Table 4: Tail precision on Llama-2-7B,

8.1\times 10^{8}

blocks. Global rows: Seed-Q ’s ranking against the Fisher’s own. Per-type rows report the same statistics within each projection type: the share of the type’s Fisher mass in Seed-Q ’s top

1\%

of that type, the share held by Seed-Q ’s bottom

30\%

against the Fisher’s own bottom

30\%

and the exposure of Seed-Q ’s bottom

3\%

; “rand” is the type’s share of the global top decile, the exposure a random pick would have.

model	grid	moved	objective	measured
Llama-2-7B	full ( $k\leq 6$ )	$4.2\%$	—	0.0412
Llama-2-7B	reduced ( $k\leq 4$ ), dry run	$3.6\%$	0.0663	—
Llama-2-13B	reduced	$3.6\%$	0.0579	0.0543
Llama-3-8B	$k\leq 5$	$12.4\%$	0.1526	0.1317
Llama-3-8B, $3.50$	$k\leq 5$	$13.5\%$	0.3800	0.2886
Llama-3-8B, $3.75$	$k\leq 5$	$23.1\%$	0.2368	0.1847

Appendix

Table 5: The value of objective equation 3 at the chosen allocation (with

e(g)

in nats, the objective’s value is a predicted excess loss) against the measured excess loss, no outlier columns, at

4.0

bits per weight unless a rate is given. “moved” is the share of blocks not at the uniform rung. Llama-2-7B is also shown on the reduced grid of the last three models, as a dry run (not built).

rung $(S,k)$	rate	Llama-2-7B	Llama-2-13B	Llama-3-8B	Mistral-7B	Qwen2.5-7B
$(8,3)$	3.000	1.8746	—	2.2518 ∗	—	—
$(10,3)$	3.250	0.5852	†	0.7014	2.5967	0.2975
$(12,3)$	3.500	0.2249	†	0.3892	0.9774	0.1754
$(14,3)$	3.750	0.1182	0.0874	0.2445	0.3544	0.1090
$(16,3)$	4.000	0.0669	0.0583	0.1556	0.1854	0.0744
$(10,4)$	3.750	0.1445	—	—	—	—

Appendix

Table 6: Uniform rungs built per model and their excess loss

e(g)

over FP16 (nats per token, WikiText-2 test, no outlier columns). The hull of each column is the input of § 3.4 . Rate in bits per weight. “—”: not built. ∗ measured but excluded from the allocation. † Build damaged by a storage fault; the arms behind Tables 1–2 are intact and the fault can only make the Llama-2-13B Mode 1 allocation suboptimal.

rung	median	$q_{10}$	$q_{90}$	sd of $\log$ ratio
$(10,3)$	3.43	1.75	6.62	0.54
$(12,3)$	2.21	1.18	4.21	0.52
$(14,3)$	1.49	0.80	2.77	0.51
$(14,4)$	0.79	0.41	1.49	0.52
$(16,5)$	0.29	0.15	0.59	0.55

Appendix

Table 7: Ratio of a block’s reconstruction error at rung

g

to its error at

(16,3)

, Llama-3-8B,

2000

blocks per tensor in layers

0,7,15,23,31

\mathbf{W}_{q},\mathbf{W}_{k},\mathbf{W}_{v},\mathbf{W}_{\mathrm{gate}},\mathbf{W}_{\mathrm{up}}

Figure 5: Seed-Q hardware implementation including the weight generator and a shared (16 $\times$ 16) PE array. The generator and the PE array are evaluated separately. Seed (S) and basis count (k) are supplied externally. The common PE multiplies reconstructed weights by activations and accumulates the products in a partial-sum register.

Explore similar work

Activation-Weighted Seeded Residual Coding for Low-Bit LLM Weight Repair
Aug 24, 2026Zehao Liu, Chuangchuang Fang, Yang RenLarge Language Model Quantization Model Weights
Huawei Technologies Research & Development (UK) Limited, Edinburgh, United Kingdom
BWLA: Breaking the Barrier of W1AX Post-Training Quantization for LLMs
May 1, 2026Zhixiong Zhao, Zukang Xu, Dawei YangLarge Language Model Quantization Post-Training Quantization
1Houmo AI
BiSCo-LLM: Lookup-Free Binary Spherical Coding for Extreme Low-Bit Large Language Model Compression
Jul 9, 2026Yuantian Shao, Peisong Wang, Zhilei Liu +6Large Language Model Compression
Nanjing University of Science and Technology, Nanjing, China · Institute of Automation, Chinese Academy of Sciences, Beijing, China · Huawei, China

Aug 24, 2026, cs.LGJ/K move · Enter open · S save

Activation-Weighted Seeded Residual Coding for Low-Bit LLM Weight Repair

Zehao Liu, Chuangchuang Fang, Yang Ren

Huawei Technologies Research & Development (UK) Limited, Edinburgh, United Kingdom

Large Language Model Quantization Model Weights Residual Learning Ablation

Open paper AskPDF arXiv

Low-bit weight quantization saves storage but leaves errors that degrade LLM quality. We introduce activation-weighted seeded residual coding (AWSRC), a compact repair codec for an existing quantization backbone. Given a reconstructed weight

W_0

, AWSRC encodes the residual

W-W_0

using deterministic seed-generated bases. The sidecar stores seed selectors, low-bit coefficients, and scales rather than an explicit codebook. Two variants combine activation weighting with per-module byte quotas (

\mathrm{AWSRC\text{-}U}

), or blended activation/Fisher weighting with globally ranked progressive prefixes (

\mathrm{AWSRC\text{-}P}_{F}

) that support multiple byte budgets without refitting. On Qwen2.5-3B-Instruct, adding

0.162

scope-bits/weight to an RTN-INT4 baseline closes

88.2\%

78.9\%

, and

71.3\%

of the PPL, KL, and 11-task mean-accuracy gaps to BF16, respectively. AWSRC achieves the highest mean downstream accuracy in byte-matched residual-codec ablations and improves all metrics across model families with up to 32B parameters.

Abstract

Explore similar work

Activation-Weighted Seeded Residual Coding for Low-Bit LLM Weight Repair

BWLA: Breaking the Barrier of W1AX Post-Training Quantization for LLMs

BiSCo-LLM: Lookup-Free Binary Spherical Coding for Extreme Low-Bit Large Language Model Compression