Answer readouts can track the next candidate
When Answer Decodability Is Not a Decision: Separating Final-Answer and Imminent-Emission Signals in Chain-of-Thought · arXiv:2611.11892
In plain words
A model’s reasoning text can name several possible answers before it gives its final answer. A readout that can identify the final answer might be picking up the name the model is about to write, rather than showing that the model has already settled on that answer.
The paper introduces a way to compare those two possibilities at the same point in the text. On a synthetic reasoning task, readers often identified the next distractor much more accurately than the eventual answer for two tested models.
A separate multiple-choice test had the opposite pattern, so the result depends on the task and reading method.
The research
A strong readout of an answer may look like evidence that a model has already chosen it. This comparison changes what that score can support: in some tested settings it tracks the candidate about to be mentioned, while another setting favors the final answer. A readout alone cannot date a decision.
ECDP compares separate readouts for the next-mentioned candidate and the eventual answer at the same point in a model’s reasoning text.
The protocol fits separate readers for the next-mentioned candidate and the eventual answer at identical positions, with a distinct chance baseline for each target. On a synthetic reasoning task, the next-distractor reader strongly outperformed the final-answer reader in two tested models.
In a separate multiple-choice evaluation, the final-answer readout was stronger. The contrast supports treating answer decodability as target- and setup-dependent, rather than as proof of when a model decided.
The tests cover selected correct rollouts and particular models, tasks, positions, and readers. The steering test failed its validity check, and the exploratory state swap was not designed to establish whether an imminent mention causes the final decision.
What the evidence shows
- Compare the same position
Emission-controlled dual-target probing (ECDP) is a protocol that checks what a readout from a model’s reasoning text identifies. It fits separate readers for the next-mentioned candidate and the eventual answer at the same position.
In the main DecisionDepth analysis, that next candidate is a distractor, so the comparison tests whether a readout tracks an upcoming mention rather than the eventual answer.
- A distractor leads on DecisionDepth
On eligible correct depth-two DecisionDepth rollouts, separately fitted readers compared the next distractor with the eventual answer at the same positions, accounting for each target’s chance baseline.
For Qwen3.5-4B, the next-distractor readout was much stronger, suggesting that at these positions it tracked the upcoming mention more than the eventual answer.
- The pattern spans two models
On its eligible correct depth-two DecisionDepth rollouts, Llama-3.1-8B showed the same ordering: at identical positions, the next-distractor reader was stronger than the separately fitted eventual-answer reader after accounting for their different chance baselines.
This supports the interpretation for this task in two tested models, but does not show that answer information is absent.
- A separate test reverses the order
In a separate evaluation, Qwen3-8B answered multiple-choice questions from MMLU. At one tested layer, the final-answer readout’s accuracy above a majority-letter baseline exceeded the next-option readout’s lift by 39.8 percentage points.
That reverse ordering shows why an answer score alone cannot establish when a model committed: what a readout reveals varies across the tested task and method.
- The reversal remains unexplained
The attempted comparison meant to isolate the reason for this reversal produced no valid DecisionDepth statistics because the imminent-mention readers failed training checks. A failed reader cannot tell whether the hidden state lacked the information.
The paper therefore supports a narrower conclusion: answer decodability alone does not establish a settled decision, and the target and evaluation setup matter.
Deep
- Original paper authors
- Hyojung Han, Jongmin Kim, Seung-Hun Jeon
- Submitted to arXiv
- License
- CC BY 4.0