Diffusion LLMs learn from their own mistakes

🔥 What happened Researchers at Indiana University dropped a new test-time scaling trick for diffusion language models. Instead of trusting instantaneous token confidence, their "frequency-aware remasking" tracks how often each token's prediction flips during denoising and preferentially commits the unstable ones. 💡 Why it matters On MATH-500, the method adds up to 7.33 percentage points of accuracy for Dream-7B; on HumanEval, up to 11.39 points of coverage for LLaDA-8B. Even on questions the baseline fails after 100 samples, it recovers solutions for up to 18.75% of cases — proof that local confidence is a misleading compass on hard reasoning. ⚡ Our take Finally, someone treats diffusion LLMs as something other than bad autoregressive models. Reading token instability as an exploration signal is the right move — this could become the default for DLM inference.