Activation-Weighted Seeded Residual Coding for Low-Bit LLM Weight Repair
- ID: 86fcb025
- 原文链接: https://arxiv.org/abs/2608.23144
- PDF: https://arxiv.org/pdf/2608.23144v1
- 作者: Zehao Liu, Chuangchuang Fang, Yang Ren
- 日期: 2026-08-24
- 更新: 2026-08-24
- 分类: infra
- 来源类型: paper
- 标签: quantization, residual-coding, weight-repair, edge-deployment, llm-compression
- 质量评分: 4/5
- 抓取时间: 2026-08-26T15:43:01Z
中文导读
AWSRC 是给现成低位宽量化 backbone 配套的修复编解码器：把残差 W−W0 用确定性种子生成的基编码，sidecar 只存种子选择器、低位系数与尺度，不需要显式码本；激活统计优先修复影响层输出的误差。Qwen2.5-3B-Instruct 上，给 INT4 RTN backbone 增加 0.162 bit/权重的开销，就关闭了对 BF16 的 88.2%/78.9%/71.3%（PPL/KL/准确率）匹配差距；49.25MB 的 sidecar 约为 BF16 权重载荷的 0.8%，在稀疏、低秩、向量量化编解码器对比中拿到最佳困惑度与平均任务准确率。「先量化、后修补」的端侧部署配方，按 MB 计费的场景可以直接试。
为什么值得关注
0.8% 权重载荷换回大部分量化质量损失，sidecar 方式对端侧模型的存储预算表是一行新选项；与 Llama-Mobile、QAH 同属端侧效率线。
关键信息
- 论文标题: Activation-Weighted Seeded Residual Coding for Low-Bit LLM Weight Repair
- 作者: Zehao Liu, Chuangchuang Fang, Yang Ren
- arXiv: https://arxiv.org/abs/2608.23144
- 发布时间: 2026-08-24
- arXiv 分类: cs.LG, cs.CL
- 备注: 无
- 关联标签: quantization, residual-coding, weight-repair, edge-deployment, llm-compression
English Abstract
Low-bit weight quantization saves storage but leaves errors that degrade language-model 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. Activation statistics prioritize errors that affect layer outputs. On Qwen2.5-3B-Instruct, adding 0.162 scope-bits/weight to an INT4 RTN backbone closes 88.2%, 78.9%, and 71.3% of the matched PPL, KL, and accuracy gaps to BF16. Repairing a matched strong low-bit backbone also improves all measured quality metrics. With a matched 49.25 MB sidecar, about 0.8% of the BF16 model-weight payload, AWSRC gives the best perplexity and mean task accuracy among sparse, low-rank, and vector-quantized codecs.
Obsidian Notes
- 元数据与摘要由
opencli arxiv paper 2608.23144 -f json拉取。 - 中文导读锚定在论文摘要声明的 claim 与数字上；未补充摘要之外的实验细节。
- 选题来源: OpenClaw定时任务/论文流水线/2026-08-26-论文流水线.md（评分 4/5 对应流水线 8.x/10 档）。
- 本页由 AAIF daily-intake-evening 任务生成（2026-08-26）。