Replacing the vocab head with a tiny action head: what's actually new about Jev-style LLMs?

Adamlags · x · 2026-09-20

A developer asks: if you restrict an LLM's output space to a predefined action set, return probabilities directly without autoregressive decoding, and replace the full vocabulary projection with a small action-only output head, doesn't that turn a normal LLM into something like Jev — one forward pass, a decision distribution out? What's actually new in Jev's structure? Sebastian Raschka replies it's basically a classic encoder-style classifier.

Related event: rasbt: Constraining LLM Output to Actions Is Just an Encoder Classifier(2 posts)→

Original post →

More from Research

Research channel →