The authors propose an empirical contract for mechanistic claims about model features.
Steering an LLM feature does not prove the model uses it
Tests using naturally occurring feature values separated behavioral control from the internal mechanisms supporting that behavior.
Big Tech
Tong Che · Yilong Li
NVIDIA Research · University of Wisconsin–Madison
Research Digest··3 min read
Che and Li test whether interpretable LLM features merely control outputs when edited or actually participate in the model’s normal computation.
Why this paper
From NVIDIA Research and University of Wisconsin–Madison
In one line
Steering a feature can change LLM behavior without showing the model uses it; tests at natural feature values separate sufficiency from use.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§