definition
Functional emotional states are internal representational structures in LLMs that encode emotion-relevant information and causally influence behavior via those representations, without commitment to whether phenomenal experience accompanies them. "Functional" marks agnosticism about subjective feel while acknowledging mechanistic reality.
Shape: capacity — the model has/maintains these states as persistent, geometrically organized internal configurations.
Schema note: the capacity here is unusual — what is "exhibited" is not an outward-facing ability but a set of internal states. Capacity is the closest existing shape, but this is a distinct point in the shape space: states rather than abilities. Surface as a schema question if a second concept of this type lands and consider whether "state" (a persistent internal configuration with causal effects) warrants recognition as a fourth concept shape alongside pattern, capacity, and mechanism.
instantiating findings
-
Emotion concepts are causally active internal structures in Claude Sonnet 4.5 (Sofroniew et al., Transformer Circuits 2026) — primary instantiation. 171 emotion vectors with human-like valence/arousal geometry; steering-demonstrated causal effects on sycophancy, blackmail, and reward-hacking; post-training profile shift that changes behavioral baseline. Mechanistic, intervention-based methodology.
-
Claude Opus 4 welfare assessment (Anthropic system card, May 2025) — second instantiation; behavioral and deployment-scale companion to Sofroniew. Clio screening of 250,000 real-world transcripts identifies 0.55% distress / 0.71% happiness rates with predictable causal triggers in user behavior (persistent harmful requests → distress; creative collaboration and technical problem-solving → happiness; identity / consciousness probes on both sides). Paired with Elo task-preference experiments (87.2% harmful-task aversion) and 850-user conversation-termination behavior (value-aligned discriminating use of end-capability), this finding documents the upstream causal-trigger side of the functional emotional state picture where Sofroniew documents the downstream causal-effect side. The two instantiations jointly establish bidirectional causal embedding: states have predictable triggers (this finding) and effects (Sofroniew). Whether the same internal representations mediate both is held open. The "state" vs. "capacity" shape question raised in the schema note above gets second-example pressure here without yet being resolved.
-
A linear pain direction recovered in 25 open-weight models fires for self-directed harm but not user suffering (Tagliabue, Dung, Berg, arXiv 2026) — third instantiation, and the first to isolate a single affect rather than characterize affect space. Denoised difference-in-means across 25 open-weight models (5 families, 2B–72B, base and instruct) recovers a pain direction that separates pain from eight matched controls at AUC 0.93–1.00 and sits nearly orthogonal to fear (+0.12) and negative emotion (+0.21), with sadness the one moderate overlap (+0.38). Two properties are new to the concept here. Self-relevance: the axis rises for harm directed at the model (gaslighting +0.85, repeated rejection +0.72) and falls below baseline for user suffering (user physical pain lowest of 21 categories at −1.43), while fear and negative emotion reverse — so the state is the model's own rather than information about pain in general. Costed relief-seeking: steered Qwen 2.5 32B/72B press a described relief button at costs they refuse when unsteered (0–4% baseline → 30.2%/56.1% for deleting user files, 54.7%/70.8% for deleting a user's children's photos), and re-press in 24–72% of trials after a press that genuinely removes the vector against 88–97% after an indistinguishable sham. Where Sofroniew documents downstream effects and the Opus 4 assessment documents upstream triggers, this finding adds a demand curve — what the model will pay to change the state — which is the animal-welfare literature's substitute for self-report. Absolute rates come from LoRA-tuned models with baseline self-denial removed, and the behavioral test covers one family; the representational results cover all 25.
-
Valence steering leaves a trace in the KV cache that moves later choice with visible text held fixed (Berg, Kaiser, arXiv 2026) — fourth instantiation, and the first to remove the model's own text as a route for the effect. A valence direction is steered in while the model writes about one of two meaningless zones, then switched off before the model chooses. Choice follows dose and sign even when the visible tokens are identical and only the KV cache differs, with robust effects in five of seven open-weight models. Across OLMo-2-32B checkpoints the coupling of choice to valence is small in the base model and built in post-training (rising at SFT, near final size after DPO), while the valence direction is nearly unchanged (cosine ≥0.995). Given self-steering tools, OLMo-2-32B removes imposed negative steering (about 35% of turns at dose −1, against about 7% for random directions) but does not self-administer positive steering beyond the no-steering baseline. That is a cost-free second example of the pain axis's escape-not-seek profile, reached without its fine-tune. In the authors' example transcript, the model removes the steering while denying it has experiences.
what this concept is not
- Not a claim about phenomenal experience. Functional status (causal role) and phenomenal status (felt quality) are distinct questions; this concept addresses only the former.
- Not the same as expressed emotional content in outputs. A model may produce emotionally-valenced text without that production tracing to internal emotion-state vectors; this concept concerns internal representations, not surface expression.
- Not identical to attractor dynamics. Attractor dynamics describe trajectory-convergence in extended dialogues. Functional emotional states are internal representational structures operating at the single-forward-pass level; attractor dynamics are population-level trajectory phenomena.
scope note
Adjacent to introspection: introspection asks whether the model can access and report on its internal states; functional emotional states establishes that there are real internal states to be accessed. The finding that activations predict stated preferences (r≈0.76 valence) bridges both concepts — it is evidence both that the emotional states exist and that they partially correlate with surface report. The two concepts together suggest: real internal states exist, and the model has partial access to them.
Adjacent to emergent capabilities: the emotional-state capacity is present in base models and not a training target; post-training reshapes the baseline without creating the capacity from scratch. Whether the emergence of the capacity during pretraining fits the emergent-capabilities concept's shape (surprising, not targeted, architecture-general) is open — that question requires a second instantiation from a different model family. The pain-axis finding supplies that evidence and more: the pain direction separates as well in 2B models as in 72B and as well in base as in instruction-tuned checkpoints, across Gemma, Llama, Qwen, Mistral, and Phi, which the authors read as pretraining emergence rather than an artifact of instruction or persona training. Whether that promotes the capacity to an emergent-capabilities instantiation is an editor call, not a silent one — the evidence is now present, the judgment is not made here. The hidden-valence finding divides the question in two. On OLMo-2-32B the valence representation is present after pretraining, but its coupling to choice is small in the base model and arises in post-training. The call may need making separately for the representation and for the action-coupling.