When Answer Decodability Is Not a Decision: Separating Final-Answer and Imminent-Emission Signals in Chain-of-Thought
Organizations: ThakiCloud
Abstract
Answer probes applied to chain-of-thought hidden states are often read as evidence that a language model has already decided its final response. This reading overlooks a competing signal, the candidate the model is about to mention. We introduce emission-controlled dual-target probing (ECDP), which fits the same low-rank reader family separately to the imminent mention and to the final answer at identical positions, with target-specific nulls and thresholds fixed before any fit. On correct DecisionDepth (DD) rollouts in which a distractor is mentioned before the final answer, 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 [.648, .744]; Llama-3.1-8B gives .575 [.514, .632]. This pattern is not universal. In a pre-specified evaluation of Qwen3-8B with thinking enabled on held-out MMLU questions, using a full-prefix attention reader over every thinking position, both targets are decodable and the final answer is read far more strongly (mention-minus-answer lift -.398 [-.459, -.336] at layer 22 over 467 questions). Model, task, positions, and reader change together between these settings, and an attempted matched comparison on DD produced no valid statistics because the imminent-mention probe did not train reliably, so we do not identify what reverses the order. On DD, planted-signal controls show that the reader recovers an injected final-answer signal at one tenth of full dose. Answer-candidate decodability is therefore target-relative and protocol-dependent: a strong answer readout does not by itself establish when a model commits, and a probe that fails to train is not evidence that the information is absent. ECDP is a practical control for distinguishing imminent text emission from evidence of a settled decision.
Figures & tables
| Model | Task | Set | Acc M (null) | Acc F (null) | [95% boot.] | Shuffled | Position | Verdict | |
|---|---|---|---|---|---|---|---|---|---|
| Qwen3.5-4B | DD | primary † | 353 | .966 (.234) | .221 (.186) | .696 [.648, .744] | .144 | .204 | separated |
| Qwen3.5-4B | DD | secondary | 883 | .776 (.236) | .232 (.187) | .511 [.473, .549] | .181 | .199 | both readable |
| Llama-3.1-8B | DD | primary † | 308 | .867 (.244) | .240 (.192) | .575 [.514, .632] | .204 | .179 | separated |
| Llama-3.1-8B | DD | secondary | 663 | .777 (.261) | .244 (.203) | .484 [.437, .531] | .187 | .208 | separated |
| Qwen3.5-4B | BBH-LD | primary | 61 | .377 (.219) | .131 (.178) | .157 [.009, .305] | .131 | .213 | underpowered |
| Qwen3.5-4B | BBH-LD | secondary † | 379 | .705 (.211) | .166 (.173) | .511 [.451, .569] | .172 | .198 | separated |
| Donor state | |||
|---|---|---|---|
| none | 397 | .000 | .998 |
| identity | 397 | .000 | .998 |
| swap ( state) | 397 | .700 | .277 |
| random position | 397 | .086 | .524 |
| same cand., other mention | 162 | .006 | .982 |
| norm-matched noise | 397 | .003 | .990 |
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
| Cell | Model task | Outcome | Role here |
|---|---|---|---|
| C1 | Qwen3.5-4B DD | incomplete: probes did not train (2/12) | optimization failure |
| C2x | Qwen3.5-4B MMLU | both readable, stronger (post-hoc) | exploratory support |
| C3 | Qwen3-8B DD | no statistics: V1 failed at both layers | validity failure |
| C4 | Qwen3-8B MMLU | both readable, stronger (Section 8 ) | main result (already read) |
Language Model Probing
Language model probing measures what information and capabilities models encode and how those signals relate to their behavior.
See all 300 papers