When does an AI reasoning model have its answer? A preprint says it can be trained.

Somewhere inside every long AI reasoning trace there is a moment when the model effectively knows the answer. Its token-by-token uncertainty collapses, and what follows is mostly confirmation. A new preprint argues that moment is a measurable, trainable signal: the sharper the drop, the more likely the chain is correct, and a model can be taught to commit at that point instead of talking past it.

The paper, ERR+: Sequential Entropy Resolution for Efficient and Decisive LLM Reasoning, was posted to arXiv on 28 Aug 2026 by researchers at East China Normal University and ByteDance. The team includes corresponding author Wen Wu, with Xin Jiang, Zhentao Xie, Shangheng Du, Jinxin Shi, Jiabao Zhao, and Liang He of ECNU plus Weicong Chen of ByteDance.

The core observation is empirical. Inside the "thinking phase," the step-by-step chain-of-thought reasoning the model writes before its final answer, the authors compare correct and incorrect traces and find that the right answers produce noticeably sharper and more frequent drops in the model's per-token uncertainty, what the field calls entropy. Top-20% high-entropy tokens are the load-bearing reasoning junctions, where the model is choosing between candidate moves; once those choices resolve, the rest of the chain is filler.

Entropy here is the model's confidence at a given token, not the information-theoretic quantity. When a model is uncertain, the next token's distribution is spread across many candidates, and entropy is high. When it is effectively decided, the distribution sharpens, and entropy drops. The authors' claim is that this drop is not just a byproduct of being right; it is a signal that can be rewarded.

The training recipe is a two-phase pipeline built on top of RLVR, reinforcement learning with verifiable rewards, the same paradigm behind recent reasoning models like DeepSeek-R1 and Qwen3. Phase 1 adds a bonus the authors call ERR, or Entropy Relief Reward, proportional to the cumulative entropy drop during the thinking phase and log-normalized by response length. The model is rewarded for resolving its uncertainty, not just for producing the right answer. Phase 2 switches to RRER, or Robust Relative Efficiency Reward, a within-group z-score-based length signal with tanh saturation, which pushes the model to be concise without a quality floor.

Theorem 1 of the paper, proved in Appendix A, is the formal reason for splitting the work into two phases. Jointly optimizing ERR and RRER produces conflicting update signals in early training, so the authors argue for a sequential design where length efficiency is added only after the model has learned to commit.

The reported numbers come from the authors' own README and benchmark tables. On DeepSeek-R1-Distill-Qwen-1.5B, GSM8K accuracy rises from an 84.6 base to 88.6 with ERR+, while response length falls from 2,076 to 1,324 tokens. On Qwen3-8B, the same benchmark moves from 93.9 to 96.0 with length dropping from 2,226 to 1,798. Across a five-benchmark suite (GSM8K, AIME 2024, AMC23, MATH-500, MMLU-STEM) the smaller model averages 63.0 to 68.4 accuracy at 6,818 to 5,450 tokens, and the larger model averages 79.2 to 82.5 at 5,012 to 4,495 tokens.

The README summary claim, "+8.6% accuracy while reducing response length by 20.1%," and the paper's Figure 2(c) caption, "20% length reduction and 8.7% accuracy gain," describe the same runs with slightly different rounding. Both are author-sourced.

The comparison set is small and curated. The authors position ERR+ as the only method in their table that improves both accuracy and conciseness. PEAR compresses length but reports accuracy degradation; GRPO improves accuracy without shortening. Whether that positioning survives contact with independent reproduction is the open question.

The code release is built on VERL, ByteDance's open RL framework, with the relevant hyperparameters documented (epsilon_err=0.01, lambda_err=0.3, alpha=0.3, gamma=0.5, R_max=1.5) and a runnable shell script. The recipe is what makes the paper testable: another lab can run the same training on a different backbone and check whether the entropy-as-commitment-signal pattern holds outside the authors' own evaluation suite.

The preprint has not been peer-reviewed. No third-party benchmark re-run, critical commentary, or independent reproduction is visible in the public record as of 28 Aug 2026. The authors flag a related concern themselves: a reward that pushes the model to resolve uncertainty faster could, in edge cases, make it commit too early, trading a longer correct chain for a shorter wrong one. The empirical results do not show this failure mode on the reported benchmarks, but the benchmarks are the authors' own.

The signal's next test is independent reproduction on a backbone the authors did not control. If the entropy-drop pattern generalizes, ERR+ joins a small set of training recipes that improve reasoning efficiency without paying for it in accuracy. If it does not, the next paper to test it will say so.