Mechanistic Interpretability of Transformer Internals
Research increasingly characterizes the internal structure of trained transformers — outlier 'massive activations,' residual-stream features — as load-bearing architectural phenomena rather than artifacts, with recurring downstream consequences for quantization, compression, and interpretability tooling.
weakening · confidence 29 · -7 7d · Emerging (watchlist) · tracking since June 23, 2026 · updated July 7, 2026
Score history
Daily conviction score, 0 to 100. Higher means the thesis is more strongly corroborated.
Now 29 · -1 since Jul 6 · ranged 29 to 30
Showing the last few days. Unlock full score history.
Why the conviction moved
- Jun 26Strengthened +5
Two new papers argue that sycophancy and refusal are governed by linear directions in the residual stream and that persona and refusal are not independent, characterizing model misbehaviors as load-bearing internal geometry rather than artifacts.
- Jun 25Strengthened +3
A paper identified a 'readout blind spot' in looped language models, showing dense supervision does not guarantee a recurrent model's hidden states learn what training appears to teach.
Showing the last 2 days. Unlock the full record.
Source trail
Supporting · June 26, 2026
New Interpretability Work Locates Sycophancy and Refusal in the Same Activation Geometry
Two new papers argue that sycophancy and refusal are governed by linear directions in the residual stream and that persona and refusal are not independent, characterizing model misbehaviors as load-bearing internal geometry rather than artifacts.
arXiv cs.AI (Detecting and Controlling Sycophancy)Supporting · June 25, 2026
Paper Finds a 'Readout Blind Spot' in Looped Language Models
A paper identified a 'readout blind spot' in looped language models, showing dense supervision does not guarantee a recurrent model's hidden states learn what training appears to teach.
arXiv cs.LG
Unlock full source trail, score history, and daily updates.
1 more source in the full trail.
Unlock Trends