Frequency-aware remasking lifts diffusion language model test-time scaling
MATH-500 gains reach 3.73 points for LLaDA-8B-Instruct and 7.33 points for Dream-7B-Instruct.
6 stories, newest first · one card per story
MATH-500 gains reach 3.73 points for LLaDA-8B-Instruct and 7.33 points for Dream-7B-Instruct.
On bases that lack the word, RL still raises MATH-500 accuracy while 'wait' stays nearly absent.
Fixed starting-token cues let base models recover much of the math and coding gains from RL training.
Bias-only label-free test-time RL lifts Qwen2.5-7B to 76.67% on MATH-500 using about 100K parameters.
LLM pipelines lose up to 40.5 accuracy points at stage interfaces; re-grounding after the loss helps.
LastOPD briefly aligns only the last-layer latent state when distilling Qwen3 into Qwen3-1.7B, beating token-only OPD on MATH-500.
Today 150 readers · 158 views · 27 reads · all time 199K views