AI
Calibrated Decision Rewards: Training Language Models with a Judge That Knows How Sure It Is
Publication: cs.AI updates on arXiv.org
Author: Rudransh Agnihotri, Manasvi Kapoor
Source published: 20/11/2026, 05:00:00 UTC
Briefing created: 20/11/2026, 05:45:34 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 20/11/2026, 05:03:24 UTC.
Briefings
Researchers developed Reinforcement Learning from Calibrated Decision Rewards (RLCDR) to train language models using a judge that outputs calibrated probabilities rather than binary answers. Training Qwen3-1.7B with this method improved MATH500 accuracy from 72.2% to 76.8% (+4.65 points), matching performance on the answer key while outperforming random rewards. The approach shows no reward hacking and improves policy performance on its own changing samples, though it struggles when probabilities become overconfident on distributions the model fails. The study tested six hypotheses with controls designed to falsify them.
Key points
- RLCDR improves MATH500 accuracy from 72.2% to 76.8% (+4.65 points) by training Qwen3-1.7B with a calibrated decision reward model.
- The calibrated reward model shows no sign of reward hacking and improves policy performance on its own changing samples compared to a dedicated process reward model.
Context
This research addresses a limitation in reinforcement learning where reward models lack probabilistic calibration. By using an LLM converted into a decision model, the study provides a reference-free reward that answers questions about correctness with probability distributions, debiased over option orders and calibrated with approximately 100 labels.
Limitations
The study's findings are limited to the specific dataset of MATH problems and the Qwen3-1.7B model. Results on other datasets or models are not supported, and the study does not claim generalizability beyond the tested conditions.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 20/11/2026, 05:45:34 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeih6wwjuhkjiw5g557sia2sskboeth5qozzp76ors4nfgjsmcdvpzm