Research2 sources
Human lesion maps from aphasia provide an anatomical framework for language model interpretability
Patterns · By Julius Fridriksson, Roger D. Newman-Norlund, Saeed Ahmadi, Regan Willis, Nadra Salman, Kalil Warren, Xiang Guan, Yong Yang, Srihari Nelakuditi, Rutvik Desai, Leonardo Bonilha, Jeff Charney & Chris Rorden ·
The Brief · written from 2 sources
- Researchers used lesion maps from human aphasia and stroke patients as an external reference for interpreting language models.
- They simulated brain-like injuries in a language model and compared the predicted lesions from its speech errors with those of error-matched patients.
- The study, led by Fridriksson et al., was published in Patterns (Cell Press) on 2026-11-19.
Why it matters
The approach checks language model interpretability against an external anatomical standard rather than only the model's own internal measurements, which could help judge how models process language.
Who’s affected
- AI interpretability researchers
- neuroscientists and aphasia researchers
- #language models
- #aphasia
- #stroke
- #interpretability
- #neuroscience
Interpretability methods judge a language model largely by its own internal measurements. Fridriksson et al. instead use human stroke data as an external standard: lesions predicted from a perturbed model’s speech errors match those of error-matched patients more closely than…
Timeline
Full coverage · 2 sources
- 2h ago ↗
Artificial Intelligence - Tech Topics
Brain researchers afflict LLMs with 'injuries' similar to those of stroke victims to unpack AI language mechanisms - 18h ago ↗
cell.com
Human lesion maps from aphasia provide an anatomical framework for language model interpretability