Turning classification into answer tokens: a head-free recipe for decision models
LysandreJik · x · 2026-10-12
Hugging Face engineer Lysandre describes a head-free way to use LLMs for decisions.
- The common approach adds a classification/reward/value head on top of a base model, trains it for the target decisions, and serves it.
- Instead, each candidate decision becomes an answer token (e.g. A=billing, B=technical, C=sales). Run the normal autoregressive forward pass, but stop right after it instead of sampling and continuing generation.
- Softmax over only the A/B/C logits yields decision probabilities straight from the LM itself (e.g. technical 0.94).
- Caveat: outputs are not calibrated, so the logits are not as interpretable as those of dedicated decision models.
More from Research
- Cameron Wolfe to talk at AIE NYC: a roadmap for evaluating highly-subjective LLM tasks — cwolferesearch · 2026-10-12
- Gaël Varoquaux compiles his AI blog series: a researcher's take on AI and society — GaelVaroquaux · 2026-10-12
- OpenAI's AI solves famous math problems—researchers recognize their own ideas in them — GaelVaroquaux · 2026-10-12
- LLMs predict edit sequences to optimize molecular synthesizability, new Nature MI paper — lulzxdxdxd · 2026-10-12
- Dexterity Is the Frontier: Chris Paxton's Field Notes from IROS 2026 — chris_j_paxton · 2026-10-12
- Decision Model, the typed-output alternative to free-form LLM text, lands on Papers with Code — NielsRogge · 2026-10-12