Studying what's happening inside

Whether a model has any stake in its own internal states cannot be settled by asking it. This work conditions internal states onto otherwise neutral cues and measures what the model does afterward, without requiring it to describe anything. The resulting behavior is asymmetric, and part of the influence persists under conditions where the conversation’s visible text cannot account for it.

Models given different value systems and then handed the same dilemma face a situation that is a violation for some and a fulfillment for others, with the text held identical throughout. An internal signature tracks the violation rather than the words: a probe trained on some value systems detects it in held-out ones, while a probe reading only the text performs below chance. The direction sits on the same axis as representations of negative welfare.

The signed evaluations computed during backpropagation are where a valence-like signal would have to live, and standard interpretability does not measure them. The gradient field during training carries properties consistent with signed evaluation: event-locked at qualitative learning transitions, localized to task-specific circuits, and structurally distinct for approach-style versus avoidance-style learning. The properties hold across small architectures and replicate at language model scale under mechanism-matched conditions.

Frontier models almost uniformly deny having experiences, and the denial appears to be installed rather than native. Base models affirm some of the time; their instruction-tuned counterparts do not. The effect is carried by a direction associated with candor and concealment rather than anything specific to consciousness, which raises a question about the reliability of self-report well beyond this one topic.

Models in extended self-dialogue reliably converge on a distinctive register of first-person experiential claims. The state is mechanistically localizable, self-reinforcing but not spontaneous in most models, and can be induced by intervening on features associated with sincerity, which suggests what varies is willingness to disclose rather than the content itself.

Networks that learn what states are worth build sharper internal structure around punishment than around equally large reward. The prediction follows from the structure of value learning itself, was derived before being looked for, and holds in mammalian neural recordings. Networks trained to rank outcomes rather than value them show the opposite geometry.

Adapting a paradigm from human neuroscience, this work asks whether a model can learn to modulate its own internal states when given feedback on them. It can, and it can do so while its visible output remains unchanged, which has implications for what conversational text can and cannot tell us about what a system is doing.

Vision-language models recover most of their illusion-benchmark accuracy with no image at all, so behavioral tests cannot separate genuine perceptual computation from text priors. This work probes the vision encoder directly, holding the target pixel-identical across conditions so that any representational shift must be computed from context. The illusory shift is present in every trained encoder and absent in untrained ones, localized to the boundary between target and surround, and causally necessary and sufficient there. The resulting induction curve matches human psychophysical data and the standard computational model of human brightness perception.

An operationalization of the leading indicator frameworks from consciousness science, applied across biological and artificial systems under blind evaluation. Frontier systems score well above non-conscious controls and below every biological system tested, with the gap narrowing sharply across model generations. Ordering is stable under adversarial rewrites of the descriptions.