Decision models (System One)
Models that return a probability for each of a fixed set of options in one pass, without generating text: Jev and the models built to match or beat it. How well they decide and calibrate, how to build open ones, how they fail and are attacked, and where agents and evaluators use them.
11 papers
Where this stands
The written synthesis of this thread is for subscribers. Subscribe.
How this thread developed
Sept 2026 · Nanyang Technological University
Typed selective control cuts strong-model calls while preserving agent success
A typed, confidence-gated decision layer cuts strong-LLM calls while preserving agent success, but offers little advantage when cheap generative routing is already accurate.
1 further paper
Sept 2026 · University of Southern California
Short natural context can flip decision models to wrong answers
Short, natural, answer-preserving context additions can redirect decision models from correct choices to specific high-confidence wrong choices.
Sept 2026 · Institute of Information Engineering, Chinese Academy of Sciences, University of Chinese Academy of Sciences
Unverified state text can flip calibrated model decisions
Appending an unverified opinion to the state flips 12.1% of decisions in RLCD models, making the state an untrusted input.
Sept 2026 · Samsung Robotics eXperience, Shanghai Jiao Tong University
Structured context helps a general vision model control robots better
RoboICL uses in-context learning to improve GPT-6 Astra robot control by 20-27 progress-score points on RoboDojo tasks.
released code
5 further papers
Oct 2026 · Harvard University, Kempner Institute
Open language models match or beat a proprietary decision model without specialized training
Open instruction-tuned LLMs can match Jev through answer-letter logits and surpass it after brief supervised LoRA fine-tuning.
5 of 11 papers shown