Tag
A paper finds that self-reflection loops like Self-Refine and Reflexion do not beat repeated sampling at equal token cost across open models from 1.5B to 7B, with all 18 self-inspection comparisons negative.
This paper identifies the 'modal ceiling' and 'correlation ceiling' in test-time scaling for reasoning models, showing that beyond a few dozen samples, additional sampling does not improve selection accuracy and can even harm it, highlighting the identifiability gap between generating and recognizing correct answers.