Cognition & behaviour
Model cognition as an empirical subject: situational awareness, self-representation, evaluation awareness, unverbalised reasoning, and the behavioural dispositions these produce.
Between representation and output sits something worth calling cognition: not as a claim that models think like humans, but as a name for the structured internal activity that precedes and shapes behaviour. This programme studies that activity directly, with the standing rule that cognitive vocabulary must cash out in measurable, ideally causal, evidence.
The organising questions are about self and situation. Does a model represent facts about itself, that it is a model, which model, in what context? Does it distinguish being evaluated from being deployed, and does behaviour condition on that distinction? Recent literature has produced suggestive behavioural evidence: strategic compliance under retraining pressure, in-context scheming in frontier evaluations, deceptive behaviour that survives safety training, and models that verbalise awareness of being tested. What is largely missing, and what this programme exists to pursue, is the activation-level counterpart: locating, decoding, and causally testing the representations these behaviours would require.
Chain-of-thought faithfulness belongs here too. A model's stated reasoning is behaviour, not introspection; whether it tracks the underlying computation is an empirical question with substantial evidence that it sometimes does not.
Situational awareness
Whether and how a model represents facts about itself and its context; benchmarks and their limits.
Evaluation awareness
Distinguishing test from deployment: behavioural evidence, and the missing representational evidence.
Model self-knowledge
What models can accurately report about their own capabilities, states, and computations.
Unverbalised reasoning
Computation that never surfaces in tokens: planning ahead, latent deliberation, shortcut reasoning.
Chain-of-thought faithfulness
When stated reasoning matches mechanism, and how to detect divergence.
Behavioural adaptation
Evaluation-conditional behaviour, sandbagging, strategic compliance.
Deception & hidden objectives
Persistent deceptive policies and the auditing methods that would find them.
Introspection
Whether models have any privileged access to their own internal states.
Personas & the assistant axis
The assistant character as a learned structure: where it lives and how it moves.
- Can evaluation-vs-deployment context be decoded from activations, and does the decoded representation generalise beyond surface cues?
- Do behavioural results like alignment faking have locatable representational signatures?
- What experimental design distinguishes genuine introspection from confabulated self-report?
- How should unfaithful chain-of-thought change how evaluations are read?
Methods in use or proposed: Behavioural evaluations with matched-pair prompts · linear probes on situational features · steering interventions · CoT-vs-mechanism comparison · pre-registered analysis plans.
A short orientation list of primary sources, not a survey. The Interpretability Map holds the full dependency graph.
- Berglund, L., et al. (2023). Taken out of context: On measuring situational awareness in LLMs. arxiv.org/abs/2309.00667
- Laine, R., et al. (2024). Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs. NeurIPS 2024. arxiv.org/abs/2407.04694
- Turpin, M., et al. (2023). Language Models Don’t Always Say What They Think. NeurIPS 2023. arxiv.org/abs/2305.04388
- Hubinger, E., et al. (2024). Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training. arxiv.org/abs/2401.05566
- Greenblatt, R., Denison, C., et al. (2024). Alignment Faking in Large Language Models. arxiv.org/abs/2412.14093
- Meinke, A., et al. (2024). Frontier Models are Capable of In-Context Scheming. arxiv.org/abs/2412.04984
- van der Weij, T., et al. (2024). AI Sandbagging: Language Models can Strategically Underperform on Evaluations. arxiv.org/abs/2406.07358
- Lindsey, J. (2025). Emergent Introspective Awareness in Large Language Models. Transformer Circuits Thread. transformer-circuits.pub/2025/introspection/index.html