Cohere Labs' CALIBER cuts LLM miscalibration by separating pre- and post-reasoning confidence
Cohere_Labs · x · 2026-10-25
Cohere Labs will present CALIBER at EMNLP 2026 (Budapest, main track): a unified confidence-calibration protocol that distinguishes pre-reasoning estimates of solvability from post-reasoning estimates of answer correctness, aligning each supervision target with its information state. This state-dependent approach significantly reduces Expected Calibration Error on in- and out-of-distribution benchmarks including BigMathDigits, GPQA, and TriviaQA. Authors: Conor Finlay, Joshua Kurien, Saurabh Dash, Marzieh Fadaee, Beyza Ermis; poster session Sunday Oct 25, 11:00–12:30 CET.
More from Research
- Tufts team open-sources HAL, showing rat cortical cultures learn to steer network bursts — MacrinePhD · 2026-10-25
- MIT: neural nets self-assemble learning circuits with local rules, no backprop — MacrinePhD · 2026-10-25
- Probe Robustness vs OOD Uncertainty Work Presented at EMNLP 2026 Poster Session D — PMinervini · 2026-10-25
- EMNLP paper: Transformers and Mamba differ globally but are locally similar in representation geometry — royschwartzNLP · 2026-10-25
- IIT Delhi's FIRE framework at EMNLP 2026: memory-efficient multi-agent counterspeech generation — Tanmoy_Chak · 2026-10-25
- Toronto & Google paper calls for common standard to evaluate GEO gaming of AI answers — rohanpaul_ai · 2026-10-25