Michael Gastpar

Classical $\chi^2$ bounds for Markov chain mixing can be arbitrarily loose—and evaluating the strictly tighter Rényi divergence contraction coefficients surprisingly reduces to optimizing over simple two-point distributions.

Adding watermarks to synthetic samples can actually degrade performance unless detection accuracy improves, challenging common assumptions in distribution estimation.

Stop wasting compute: TERMINATOR slashes Chain-of-Thought reasoning length by up to 55% by predicting the optimal early exit point.