EDSL to add activation capture, probes and model steering for LLM social simulation research

soumitrashukla9 · x · 2026-09-16

Expected Parrot's John Horton proposes adding activation capture, probe training and inference-time steering to EDSL, with a Modal-backed service. Motivated by a new paper comparing prompt-based manipulation, SAE feature steering and probe-direction steering in LLM social simulations — where strong prompting can beat SAE steering on creativity tasks — the proposal stresses strong prompt baselines and open iteration between researchers and tool maintainers (GitHub Discussion #2651).

Original post →

More from coding & agent

coding & agent channel →