W-ISA：面向权重空间干预程序的事务化执行指令集
W-ISA: Transactional Execution of Weight-Space Intervention Programs
AI 导读
W-ISA 是一套软件指令集与事务化运行时，用于在预训练语言模型上执行权重空间干预程序，支持参数更新、投影、消融和秩一变换，并在请求级上下文中执行后自动恢复参数。
正文 · 原文
Abstract:We present W-ISA, a software instruction set and transactional runtime for executing weight-space intervention programs on pretrained language models. Shape-checked instructions express parameter updates, projections, ablations, and rank-one transformations. The runtime applies these programs within request-scoped execution contexts and restores saved parameters after completion or interruption. Activation-based match-action rules connect measured model states to intervention selection. We evaluate restoration and replay on SmolLM-135M-Instruct and Qwen2.5-1.5B-Instruct, and compare materialized rank-one updates with equivalent PyTorch hooks on the latter. In 45 timed requests, both intervention implementations produce identical greedy sequences at 32, 128, and 512 tokens; their maximum prompt-logit difference is 0.000184. On the complete WikiText-2 raw test split, the combined SmolLM attention-head ablation and MLP projection changes perplexity from 22.2353 to 22.3188, a relative increase of 0.3754%. W-ISA separates intervention specification, execution, and measurement, making parameter-level experiments reproducible and composable.
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2611.13232 [cs.AI] |
| (or arXiv:2611.13232v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2611.13232 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Ekrem Aksoy [view email]
[v1]
Mon, 9 Nov 2026 21:01:51 UTC (27 KB)
来源：arXiv cs.AI · arxiv.org