New paper: high linear probe accuracy doesn't prove a linear representation — three diagnostics proposed
soniajoseph_ · x · 2026-11-03
A new arXiv paper challenges the standard mech interp methodology of equating high linear probe accuracy with linear representations: asymptotically any rich feature map, even random, admits a linear readout. The authors propose three asymmetric diagnostics — reverse predictivity, spectral concentration, and OOD generalization — to distinguish true linear world models from kernel machines. Applied to vision transformers, pretraining shows structure beyond a random kernel but well short of a linear world model. Author David Klindt also found DINOv3's global [CLS] token linearly represents continuous geometric latents of dSprites.
More from Research
- UMI: Handheld Grippers Turn Human Demos into Hardware-Agnostic Robot Policies — gabriberton · 2026-11-03
- Direct Weight Editing via Codex+Astra Lifts Audio Classifier Accuracy From 84.3% to 88.75% Without Retraining — dansitu · 2026-11-03
- How UMI turns handheld grippers and SLAM into robot training data, in 5 steps — gabriberton · 2026-11-03
- TSFLab: A Self-Improving Agent-Agnostic Workbench for Time Series Forecasting — D2I-Group · 2026-11-03
- YC's Momentic launches QA-Bench: its AI QA agent finds 2.7x more reproducible bugs than Codex — ycombinator · 2026-11-03
- EDU-Jev 4B education decision model hits 67.6%, beating top commercial API by 9.4 points — kalyan_kpl · 2026-11-03