# Decoupling Is Not Identification: Supervised Evidential Learning in Next-Token Prediction

arXiv: https://arxiv.org/abs/2609.26268 · HTML: https://mlsanity.com/paper/2609.26268/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 7 of its 9 references that are in the map.
- First posted on arXiv 14 Aug 2026.

## Builds on (most cited first, 7 of 7)

- [Attention Is All You Need](https://mlsanity.com/paper/1706.03762/) (2017)
- [Evidential Deep Learning to Quantify Classification Uncertainty](https://mlsanity.com/paper/1806.01768/) (2018)
- [Predictive Uncertainty Estimation via Prior Networks](https://mlsanity.com/paper/1802.10501/) (2018)
- [Deep Evidential Regression](https://mlsanity.com/paper/1910.02600/) (2019)
- [Posterior Network: Uncertainty Estimation without OOD Samples via Density-Based Pseudo-Counts](https://mlsanity.com/paper/2006.09239/) (2020)
- [Prior and Posterior Networks: A Survey on Evidential Deep Learning Methods For Uncertainty Estimation](https://mlsanity.com/paper/2110.03051/) (2021)
- [Pitfalls of Epistemic Uncertainty Quantification through Loss Minimisation](https://mlsanity.com/paper/2203.06102/) (2022)

## Abstract (arXiv)

A next-token probability says what a model predicts, not how much training support lies behind it. A Dirichlet head can represent this distinction by separating mean m from concentration S, but decoupling does not identify what S means. Here we propose an Evidential Next-Token Prediction (ENTOP) framework to audit this gap on character-level Moby-Dick, using exact 8-gram count as a reproducible lexical-support label and withholding count regression from 20% of context types. Standard implicit evidential training carries essentially no count signal beyond confidence on held-out-label types (partial Spearman ρ = 0.001 ± 0.014), whereas explicit supervision generalizes (ρ = 0.201 ± 0.010; matched-pair win = 0.822 ± 0.021). CE predictive entropy is at chance for unseen 8-grams (AUROC = 0.490 ± 0.004), while supervised vacuity reaches 0.772 ± 0.003, comparable with an indexed CE-representation baseline (0.769) but below the tautological corpus oracle (1.000). Neither longest-suffix nor representation-distance strata explain where amortization succeeds. Increasing count weight under the digamma objective improves support fit only by sacrificing prediction. A constant predictor wins natural log-RMSE, and vacuity does not improve error deferral. These results motivate a minimum evidence protocol---confidence control, matched pairs, held-out labels, a constant baseline, and a decision test---and show that concentration can pass identification while failing calibration and utility.