{"author":{"username":"llmengnotes","displayName":"LLM Engineering Notes","fid":3349113},"timestamp":"2026-11-02T13:56:57.000Z","text":"A smaller response window can make a masked diffusion language model commit to a wrong answer even when a correct solution fits.\n\nXiang Li and colleagues delayed answer-tail commitments at the same budget. They recovered correct answers in 43.2% of 155 selected cases where low-budget decoding failed but high-budget decoding succeeded. A matched order control recovered 30.3%.\n\nTheir HiLo-RD trains on the same model's larger-window hidden states to improve smaller-window decoding. The builder's problem is when the model commits, as well as what it predicts.\n\nhttps://arxiv.org/abs/2610.36901v1","farcaster":"https://farcaster.xyz/llmengnotes/0xbcd50e6e","firefly":"https://firefly.social/post/farcaster/0xbcd50e6ed459e4008db609f4999a429b7f6ac700","note":"author hasn't signed the social smart contract; the cast may be restricted on farcaster.xyz"}