EPS Dashboard

Authors: Long Zhang, Wei-neng Chen, Feng-feng Wei et al.

Summary

The authors track exactly when a language model's preference for a final answer becomes stable—before it actually writes the answer out. They project the model's continuation probabilities onto a finite answer set (yes/no in binary tasks) to get a log-odds measure that tells you when the model has "committed" internally. Testing on Qwen3-4B-Instruct in controlled delayed-answer tasks, they find the model's internal answer stabilizes 17–31 tokens before the answer is parseable in the text, and this signal tracks what the model will eventually output (not ground truth).

Main takeaways:

Models commit to an answer internally before verbalizing it—the finite-answer preference stabilizes 17–31 tokens early on average in the main setup.
The commitment signal is recoverable from hidden states and tracks the model's eventual output rather than the correct answer.
The signal is partly separable from simple "where am I in the reasoning?" cues and transfers across prompts without a single fixed coordinate.
Steering the log-odds locally changes the measure but doesn't reliably control which answer the model generates, separating measurement from causal control.
The method works without greedy decoding or training probes—it just uses the model's own continuation probabilities projected onto the answer verbalizers.

Relevance

Directly relevant to my work on persona behavioral installation and activation steering. If internal "commitment" stabilizes before verbalization, that's a concrete internal state I could target with activation steering or use to diagnose when a persona marker or behavior has actually taken hold versus just being in the forward pass noise.

Abstract

arXiv:2605.06723v1 Announce Type: new Abstract: Language models often generate reasoning before giving a final answer, but the visible answer does not reveal when the model's answer preference became stable. We study this question through a narrow computable object: \emph{finite-answer preference stabilization}. For a model state and specified answer verbalizers, we project the model's own continuation probabilities onto a finite answer set; in binary tasks this yields an exact log-odds code, $\delta(\xi)=S_\theta(\mathrm{yes}\mid\xi)-S_\theta(\mathrm{no}\mid\xi)$. This target defines parser-based answer onset, retrospective stabilization time, and lead without relying on greedy rollouts or learned probes. In controlled delayed-verdict tasks with Qwen3-4B-Instruct, the contextual finite-answer projection stabilizes before the answer is parseable, with 17--31 token mean lead in the main templates and positive, shorter lead in a parser-clean replication. The signal tracks the model's eventual output rather than truth, is linearly recoverable from compact hidden summaries, is partly separable from cursor progress, and transfers as shared information without a single invariant coordinate. Diagnostics separate the measurement from online stopping, verbalizer-free belief, and causal answer control; exact steering shows local sensitivity of $\delta$ but not reliable generation control.