An answer probe may be reading the candidate about to be written: dual-target probing on DecisionDepth opens a .696 emission gap
Synopsis
Answer probes on chain-of-thought hidden states are often read as evidence that a model has settled its final answer, yet a hidden state also carries the candidate whose name is about to be written. Emission-controlled dual-target probing (ECDP) fits the same low-rank reader separately to the next candidate mention and to the final answer at identical positions, with target-specific nulls and thresholds fixed in advance. On DecisionDepth, Qwen3.5-4B decodes the next mention at .966 (null .234) but the final answer only at .221 (null .186), an emission gap of .696; Llama-3.1-8B gives .575. The order is not universal: in a Qwen3-8B MMLU setting the final answer is read more strongly.
Interpretation
At intermediate positions in chain-of-thought reasoning, hidden states more readily reveal the candidate a model is about to write than the answer it eventually gives. High answer-probe accuracy had been read directly as evidence that the model had already decided, without separating the competing signal of the candidate about to be emitted at the same position. On DecisionDepth, Qwen3.5-4B at the token before the first distractor mention reads the next mention at .966 (null .234) and the final answer at .221 (null .186), an emission gap of .696 with interval [.648, .744]; Llama-3.1-8B gives a gap of .575, interval [.514, .632].
The method fits two targets at the same position with the same reader family, giving each target its own null and a decision threshold fixed before any fit. The reader is a rank-4 bilinear scorer over the layer-32 hidden state, candidates are embedded by the same model at the same layer, training uses five item-level folds, and weight decay is chosen on a calibration slice of the training folds. The primary analysis keeps only correct rollouts with an eligible distractor mention: 353 items for Qwen3.5-4B and 308 for Llama-3.1-8B; label-shuffle and random-position controls sit near their nulls.
The order reverses with the setting, so answer decodability is target-relative and protocol-dependent rather than a general property of chain-of-thought readouts. In a pre-specified Qwen3-8B MMLU setting with a full-prefix attention reader, both targets are decodable and the final answer is far stronger. At layer 22 the final answer lifts .628 over baseline and the next mention .230, a paired difference of .398 with interval [.459, .336] over 467 questions; an exploratory post-hoc run of Qwen3.5-4B under the same protocol points the same way at layer 19.
Perspective
The conclusions rest on correct rollouts of the synthetic DecisionDepth task and of BIG-Bench Hard Logical Deduction, with models of 4B to 8B parameters, three of them from the Qwen lineage; the primary analysis covers only rollouts that mention a distractor before the final answer, and the layer sweep covers Qwen3.5-4B only.
In the Qwen3-8B MMLU setting the final answer dominates, but model, task, positions and reader change together, and the attempted matched comparison produced no valid statistics because the imminent-mention probe did not train reliably on DecisionDepth, so which factor reverses the order is not identified; the planted-signal control bounds a native signal only relative to one injection direction and dose.