NUS's ConfTuner (NeurIPS 2025) trains LLMs to state calibrated confidence, open-sourced
jiqizhixin · x · 2026-10-11
A model picks the right answer but claims 99% confidence when it's only 60% right — in medical settings, such overconfidence on wrong answers can shape human decisions. The real problem isn't just correctness; it's making stated confidence match true accuracy.
ConfTuner from the National University of Singapore, accepted at NeurIPS 2025, tackles this:
- Shares the core intuition behind the viral Jev model: output decision plus probability directly, support parallel probability prediction, and treat probability as central to prediction rather than an afterthought;
- Targets probability calibration — not just choosing correctly, but making stated probability reflect true correctness rates;
- Where Jev is closed-source, ConfTuner is open and proposes Tokenized Brier Score, which directly trains the probability distribution over candidate answers.
More from Research
- Embedding arithmetic demo: cartoon − animation − comic = diagram — mayfer · 2026-10-23
- Omnii Genome: Single Generalist Model Outperforms AlphaGenome Across All Benchmarks — JosephJacks_ · 2026-10-23
- Domingos Cites New Result Claiming RL Is 'Basically Useless' vs Supervised Learning — pmddomingos · 2026-10-23
- Word Arithmetic on Embeddings: "Bathroom Minus Toilet Equals Kitchen" Beats SAE Probing — mayfer · 2026-10-23
- Wolfram Opens Rule 30 Prizes to AIs: $30,000 for Three Open Problems — cephaloform · 2026-10-23
- Stanford and Databricks use AI agents to reproduce medical findings and test analytical robustness — ChengleiSi · 2026-10-23