All updates · Official

Official news, first rumor reports and substantive progress. Each entry keeps the information published at that time.

• Researchers release U-Space, a label-free way to read LLM uncertainty from hidden states in a single pass

  U-Space maps semantic anchors for doubt and certainty into a model's residual space, covering ambiguity, incomplete information, conflicting evidence and general uncertainty. Its U-Lens score combines how strongly each token's hidden state leans toward doubt with predictive entropy, from the one generation that produced the answer, with no correctness labels, repeated sampling or training. The authors say it beats established baselines on reasoning benchmarks, including length-controlled evaluation; code is on GitHub.

  OfficialOfficial updateOct 12Source: arXivPart of U-Space

RSS for this view →