This paper presents ZeroTIR, revealing agent‑level RL scaling laws that tie training steps, code‑call frequency, response length, and accuracy, and surpassing ZeroRL and SFT baselines on challenging math benchmarks.

We introduce a new class of Reinforcement Learned Teachers trained to provide effective reasoning traces for downstream distillation, yielding more effective data for distillation and cold-starting than orders of magnitude larger reasoning LMs.

scaling math and code rl

We introduce a challenging benchmark of graduate-level problems in applied mathematics that was fully developed as part of a university class.

We release the largest preprocessed datasets for the Learning with Errors (LWE) problem

LLMs have advanced in reasoning tasks, but creativity assessment has been overlooked. CLAWS uses attention-based features to classify solutions without human evaluation.