W-ISA: Transactional Execution of Weight-Space Intervention Programs
Summary
- W-ISA is a software instruction set and transactional runtime designed for executing weight-space intervention programs on pretrained language models.
- It uses shape-checked instructions for parameter updates, projections, ablations, and rank-one transformations within request-scoped execution contexts.
- Activation-based match-action rules connect measured model states to specific intervention selections while automatically restoring saved parameters after completion.
- Evaluations on SmolLM-135M-Instruct and Qwen2.5-1.5B-Instruct demonstrate identical greedy sequences compared to PyTorch hooks and minimal perplexity shifts on WikiText-2.
Why it matters
W-ISA introduces a structured, transactional approach to weight-space interventions, making parameter-level experiments on language models more reproducible, composable, and safe for request-scoped execution.