Answers are grounded in this paper and cite their sources.
Does a language model notice when its own internal representation is rewritten? Introspection studies read a model's failure to report an injected manipulation as evidence about its self-knowledge. Such a null report cannot be interpreted if the manipulation may have had no effect. We address this from the side of the intervention. We use a Jacobian-lens band swap whose effect is verified on every item. The model answers the swapped question, restates it when asked (capture), and a lens decodes it on fresh forward passes. Report is the primary measurement, scored against an amplitude-matched random-direction control. Across a six-model ladder and Qwen3.6-27B, verbal report of the verified swap is rare or non-specific in a completion prompt. Only Qwen3-4B shows a report signal that is both specific and consistent in direction (d′ = +1.48, report 17.9%). In each Qwen3 model's chat template, report rises with scale to 93% at 27B. When the report question is asked before the model answers, the yes rate is at most 2% in both frames. Where the swap can be confirmed on the same pass, Qwen3-8B says yes on 0/37 such items, against 10/37 with its answer in context. This channel also says yes to the random-direction control, so the contrast concerns the yes rate and not specific report. The yes/no margin still moves before the answer in the chat frame, by at most 0.51 nat in median. In our probes, yes rates above 2% and large swap-induced margin shifts occur only with the model's own answer in context. We do not conclude that introspection is absent. Without per-item verification and an amplitude-matched control a null report cannot be read: scored against clean runs alone, 8B would outrank 4B (d′0 = +1.66 against d′R = +0.12). This record contains the preprint and a supplementary archive with the per-item result files behind every table, the experiment and analysis scripts, and the per-item verification of the before-answer probes. Code: https://github.com/tao-hpu/jspace-replication.
No takes yet. Share an insight, caveat, or question.
Tao An (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: