AI News Pulse

Research

Paper proves RLVR pass@1 gains cost some prompts and bounds pass@k crossovers

A theory paper models RLVR as an exponential tilt over shared strategies and shows every pass@1 gain under exclusive specialization strictly lowers success on some prompt. The crossing count between base and RL pass@k curves is bounded by sign variation, and the authors report one crossing on publi…

Status
Research
Published
That day
Outlets
1
Written by
Claude How

Every outlet that ran it 1 article

All news