Jev Reasoning

A model-independent benchmark for graph decision tasks, combining structural task profiles, matched graph-text input contrasts, and heuristic-proposal controls. It finds accurate adjacency recognition does not guarantee broader structural correctness.

A comparison of Jev with 19 LLMs across 13 multiple-choice benchmarks covering knowledge, reasoning, and multilingual understanding. Jev is competitive with frontier LLMs on knowledge and commonsense tasks, but on MathQA it trails the frontier median by 17.7 points.

A working RLCD implementation for reasoning models: the model samples a rationale, and a strictly proper scoring rule grades the answer distribution it commits to afterward. With Qwen3-1.7B it matches or beats SFT, RFT/STaR, and GRPO in accuracy and beats all in selective prediction.

SanSi turns a pre-trained looped language model into a typed decision model that revises its hidden state across loops without generating tokens. On 10,027 test decisions from 59 sources it reaches 72.0% accuracy, 13.5 points above a same-shaped non-looped model trained with the same recipe.

StartLux-Decision is a family of decision models from 0.8B to 27B parameters pairing single-pass Instant variants with AutoThink versions that reason only when uncertain. On the public subset of Decision Index 0.3 the 27B AutoThink model reaches 73.4.