---
type: "raw_signal"
id: "rethinking-confidence-in-reasoning-models-20e9ea"
slug: "rethinking-confidence-in-reasoning-models-20e9ea"
title: "Rethinking Confidence in Reasoning Models"
description: "Cohere describes CALIBER, a training method that asks a reasoning model to report two confidence estimates: its expected chance of solving a problem before reasoning and its estimated chance that the particular answer is correct afterward."
retrieval_nugget: "Cohere describes CALIBER, a training method that asks a reasoning model to report two confidence estimates: its expected chance of solving a problem before reasoning and its estimated chance that the particular answer is correct afterward."
observed_at: "2026-10-20"
record_date: "2026-10-20"
date_kind: "observed_at"
why_it_matters: "Cohere reports training 7B and 30B models on 30,000 math problems and evaluating them on 1,000 held-out problems plus three question-answering benchmarks. On BigMathDigits, the 7B model’s expected calibration error fell from 0.059 for the strongest single-confidence baseline to 0.028. The source also reports calibration-error reductions on GPQA, TriviaQA, and SimpleQA. These are benchmark results "
novelty: "notable"
verification_level: "source-linked"
signal_type: "research"
source_platform: "cohere.com"
topics: []
entities: ["Cohere","CALIBER","BigMathDigits","GPQA","TriviaQA","SimpleQA"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "Cohere’s confidence-in-reasoning-models research is a distinct topic from prior coverage of controlling reasoning effort in LLMs; the archive contains no details of this confidence study."
prior_refs: ["memory:54946dd1-6f16-8f81-aed9-177bd3373c77"]
source_url: "https://cohere.com/blog/rethinking-confidence-in-reasoning-models"
source_urls: ["https://cohere.com/blog/rethinking-confidence-in-reasoning-models"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "signal:rethinking-confidence-in-reasoning-models-20e9ea"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/signals/rethinking-confidence-in-reasoning-models-20e9ea/","markdown":"https://newruntime.com/signals/rethinking-confidence-in-reasoning-models-20e9ea.md","json":"https://newruntime.com/signals/rethinking-confidence-in-reasoning-models-20e9ea.json"}
---

# Rethinking Confidence in Reasoning Models

## Retrieval answer

Cohere describes CALIBER, a training method that asks a reasoning model to report two confidence estimates: its expected chance of solving a problem before reasoning and its estimated chance that the particular answer is correct afterward.

Cohere describes CALIBER, a training method that asks a reasoning model to report two confidence estimates: its expected chance of solving a problem before reasoning and its estimated chance that the particular answer is correct afterward. These estimates address different questions: expected ability on the task and correctness of a specific answer.

Cohere reports training 7B and 30B models on 30,000 math problems and evaluating them on 1,000 held-out problems plus three question-answering benchmarks. On BigMathDigits, the 7B model’s expected calibration error fell from 0.059 for the strongest single-confidence baseline to 0.028. The source also reports calibration-error reductions on GPQA, TriviaQA, and SimpleQA. These are benchmark results from Cohere’s experiments, not evidence that confidence-guided verification or abstention works in deployment.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Raw Signals](https://newruntime.com/signals.json): Open every source-linked raw signal.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)