U-Space finds an interpretable subspace for LLM uncertainty, no training needed
Tobias Braun · hf · 2026-10-09
U-Space identifies semantic anchors for doubt and certainty inside LLMs, mapping them into an orthogonal basis in residual space. A U-Lens projection yields token-level interpretable uncertainty maps and scalar confidence scores — no correctness labels, repeated generations, or training required. It outperforms established baselines on reasoning benchmarks under standard and length-controlled evaluation.
More from Research
- Apple Paper: SOTA Harnesses Offer No Edge Over Minimal Single-Session Coding Agent — himanshustwts · 2026-10-09
- LangChain Founder: Evals Work for Narrow Tasks but Break Down for Autonomous Agents — hwchase17 · 2026-10-09
- Hugging Face launches Robotic Episodes Viewer for 24k+ LeRobot datasets — mishig25 · 2026-10-09
- Blind humanoid walks, plays soccer and lifts suitcases with joint encoders only — accepted at Humanoids 2026 — Jan_R_Peters · 2026-10-09
- Delete object info from observations and PPO learns to search anyway — TU Darmstadt on its Humanoids 2026 paper — Jan_R_Peters · 2026-10-09
- CARE certifies VLA inference speedups up to 10.8x with statistical guarantees — UMCP · 2026-10-09