Northeastern Researchers Unveil Dual-Route Model of Concept Induction Heads in LLMs

Northeastern University's David Bau lab published a dual-route model showing that concept induction heads in LLMs copy by meaning rather than token-by-token; the team is also exploring Jacobian lenses, which can outperform logit lens readouts in VLMs.

2026-09-30 ~ 2026-09-30 · 2 related posts

• Northeastern researchers find 'concept induction heads' that copy word meanings, not just tokens — davidbau · 2026-09-30
• David Bau: OV-transform summaries beat logit lens readouts in VLMs almost by accident — davidbau · 2026-09-30