SWD: Extracting LLM Circuits Directly From Weights With <1% of Data
量子位 · wechat · 2026-08-14
A collaborative research team from IQuestResearch, Oxford, Stanford, and Tsinghua introduced Sparse Weight Decomposition (SWD), a novel approach to LLM mechanistic interpretability.
Traditional methods like Transcoders require training a separate surrogate network to understand model internals, incurring high data and compute costs while introducing substitution errors. SWD bypasses this by directly decomposing existing dense weights from pre-trained models into sparse, independently intervenable bottleneck units, eliminating the need for a new surrogate network.
Key highlights from the paper:
- Extreme Data Efficiency: SWD requires less than 1% of the data used by training-based baselines to achieve matching replacement fidelity.
- Scalability: The method successfully scales to all 48 weight matrices in GPT-2 Small and extends to larger models like Qwen 3.5-27B.
- Zero-data Version: SWD offers a variant that requires absolutely no calibration text, minimizing Frobenius error directly in the weight space, which opens doors for tracking structural changes across training checkpoints.
- Precise Behavior Editing: The extracted sparse paths not only pass rigorous causal tests (sufficiency and necessity) but also act as precise "editing handles" for targeted model interventions with minimal side effects.
More from Research
- NCP-Bench: Best LLM Agents Drop to 42% Narrative Consistency After 20 Turns — arnicas · 2026-08-14
- Researcher Praises Rarely Readable LLM Paper on Category Theory — spikedoanz · 2026-08-14
- Stanford Researcher Explains Why Larger Models Retain Rare Skills: Capacity Competition — SinclairWang1 · 2026-08-14
- Alignment Research Should Focus on Actual AI Preferences, Not Just Theory — repligate · 2026-08-14
- AutoPrune: LLMs Automatically Design Visual Token Pruning for Multimodal Models — Zhen Liu · 2026-08-14
- CUDA version causes 3.3x speed difference in quantized video models; B200 loses to properly configured 4090 — Odd_Lavishness2236 · 2026-08-14