cs.LGNov 11, 2026

Learning a calibrated decision head from logged bandit feedback

Authors: Ye Yuan

Organizations: Independent Researcher · Ph.D. in Physics, University College Dublin.

Abstract

A decision head reads a state and a list of options and returns a probability per option in one forward pass. Once deployed, it usually observes only whether the option it chose was right: logged bandit feedback. We adapt such a head, LoRA on Qwen3-0.6B-Base and 1.7B-Base, from that feedback and report three results. First, the Brier or log score of the sampled action's own probability is not a usable REINFORCE reward: every wrong action earns at least gold's reward, with a mean baseline the lowest-probability wrong action gets the largest advantage, and at two options the signal is zero; in one run it cut accuracy on ordered-scale questions from .636 to .111. Scoring the probability as a loss by the pathwise gradient avoids the collapse, as do the REINFORCE rewards r and r minus p. Second, with full labels, none of four tested bandit arms beat cross-entropy plus one temperature on mean expected calibration error (ECE) over four unseen label schemas. Third, from logged feedback alone, Chosen-Action Proper Scoring (CAPS), the direct method with a proper score on the chosen action's probability, beats fine-tuning on confirmed answers only, which discards negatives: on Banking77 by 2.6 accuracy points, .053 lower ECE and lower selective risk (three seeds against one baseline seed), and by 21.7 points on one seed on emotion data with an often-wrong logger. At N 5000, CAPS stays 1.5 points below full-label training at 0.6B and .6 points at 1.7B against one full-label seed, and propensity weights add nothing on this log. Under feedback flipped at a known rate, one seed per rate, CAPS's confidence tracks the noisy channel on average, backward correction recovers most of the clean calibration, and early stopping on held-out logged outcomes prevented memorising noisy draws in all three runs. Code, adapters, logs and per-record results are public.

Figures & tables

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Sep 30, 2026

Beyond Outcome Rewards: Post-Rationale Distribution Optimization for Selective Prediction in Reasoning Models

Reasoning models are trained to produce correct answers, yet selective deployment requires ranking which answers to accept or defer, and outcome-reward training does not directly optimize the post-rationale distribution used for that decision. We identify two mismatches: expected correctness is mode-seeking rather than distribution-eliciting, and scoring an average across rationales rewards disagreement unavailable to a single-rationale decision. We introduce Post-Rationale Distribution Optimization (PRDO), which applies a proper score to the answer distribution after each sampled rationale, without target rationales or separate confidence labels. Across four 0.6--3.8-billion-parameter models on MMLU-Pro, PRDO improves Brier score, error ranking and secondary accuracy relative to budget-matched, temperature-scaled GRPO. At 20% realized error, oracle coverage rises by \npj{[email protected]}--\npj{[email protected]} percentage points, and ranking gains transfer to unseen tasks. Matched controls identify proper, per-rationale scoring as the active design choices, while verification tasks show the boundary: better selective ranking can coexist with lower full-coverage accuracy. Code is available at https://github.com/ZimmyGao/openjev-rlcd
Sep 27, 2026

Calibration, Not Answer Selection: Distilling Internal Confidence in Reasoning Models

Reinforcement learning with binary correctness rewards trains correctness, not calibrated confidence. The confidence that reasoning models verbalize is systematically overconfident, and the problem is not merely one of scale: verbalized confidence tracks how willing a model is to commit to an answer, not how likely the answer is to be right. Post-hoc rescaling therefore fits one distribution but rarely transfers. We look inside the model instead. On factual question answering, a linear probe on the hidden state between the chain of thought and the answer is substantially better calibrated: its expected calibration error is 5 to 38 times lower than that of the verbalized score across four benchmarks and two model families. However, when used to pick among N sampled answers, that same probe nearly ties majority voting yet falls far short of the oracle. Internal states answer "how certain am I" well and "which answer is right" poorly, so the signal should be reported as a confidence rather than used to select answers. As a result, we introduce probe-guided self-distillation (Probe-SD): score a model's own sampled traces with the probe, overwrite the confidence each trace states, and finetune the base checkpoint of the same family, so nothing but the model itself remains at test time. On Qwen3-14B, Probe-SD cuts ECE from 0.178 to 0.024 in-domain and from 0.542 to 0.113 out-of-domain, where it also beats post-hoc recalibration and self-consistency distillation. The resulting confidence is well-calibrated and useful for weighted voting, behaviors previously attributed to online RL, here obtained with supervised finetuning alone.
Apr 12, 2026

Calibration Collapse Under Sycophancy Fine-Tuning: How Reward Hacking Breaks Uncertainty Quantification in LLMs

Modern large language models (LLMs) are increasingly fine-tuned via reinforcement learning from human feedback (RLHF) or related reward optimisation schemes. While such procedures improve perceived helpfulness, we investigate whether sycophantic reward signals degrade calibration -- a property essential for reliable uncertainty quantification. We fine-tune Qwen3-8B under three regimes: no fine-tuning (base), neutral supervised fine-tuning (SFT) on TriviaQA, and sycophancy-inducing Group Relative Policy Optimisation (GRPO) that rewards agreement with planted wrong answers. Evaluating on 1,0001{,}000 MMLU items across five subject domains with bootstrap confidence intervals and permutation testing, we find that sycophantic GRPO produces consistent directional calibration degradation -- ECE rises by +0.006+0.006 relative to the base model and MCE increases by +0.010+0.010 relative to neutral SFT -- though the effect does not reach statistical significance (p=0.41p = 0.41) at this training budget. Post-hoc matrix scaling applied to all three models reduces ECE by 4040--64%64\% and improves accuracy by 1.51.5--3.03.0 percentage points. However, the sycophantic model retains the highest post-scaling ECE relative to the neutral SFT control (0.0420.042 vs.\ 0.0370.037), suggesting that reward-induced miscalibration leaves a structured residual even after affine correction. These findings establish a methodology for evaluating the calibration impact of reward hacking and motivate calibration-aware training objectives.