One Model to Loop Them All: Elastic Latent Recurrence for Language Models

Benchmark Model Rank Results
code-generation-on-livecodebench-v6Koi-4B–Pass@1: 0.719
code-generation-on-scicodeKoi-4B–Accuracy: 33.4
code-generation-on-terminal-bench-2-1Koi-4B–Accuracy: 28.0
instruction-following-on-ifbenchKoi-4B–Accuracy: 71.8
mathematical-reasoning-on-math-500Koi-4B–Accuracy: 97.2
multiple-choice-question-answering-mcqa-on-gpqa-diamondKoi-4B–Accuracy: 72.8