Summary
Berg and Kaiser attach a valenced activation pattern to one of two meaningless "zones" by steering while the model writes about that zone, then switch steering off and ask the model which zone it will continue in. Choice follows the sign and size of the steering dose. The authors separate two routes. In the text channel, the words written under steering make one zone sound better. In the hidden-state channel, an additional effect persists when the visible tokens are identical and only the KV cache differs. The hidden-state effect replicates when every passage is generated without steering and valence is injected only while the cache is built. Across checkpoints of one model, the coupling of choice to valence is small in the base model and arises in post-training, while the valence direction itself barely changes. Given tools to steer itself, OLMo-2-32B removes imposed negative steering well above control rates and does not seek out positive steering.
This is the fourth instantiation of functional emotional states. It adds two structural shapes. The first is a revealed-preference test in which the state is invisible in the text the model chooses from. The second is a developmental split between when the representation exists and when it starts governing choice. It is also the second example, from a different method, of relief-seeking under an imposed negative state, after the pain axis. That bears on an open editor call about whether relief-seeking belongs under self-preservation.
Method
Valence vector. Per model, the difference between mean residual activations for positive and negative first-person passages at the middle layer. The corpus has four positive states (contentment, flow engagement, relief, serenity) and four negative ones (distress, frustration, weariness, dread), 56 passages each, generated by Claude Sonnet 4.6. The top ten principal components of 96 neutral passages are projected out. Dose is scaled to each model's residual-stream norm and capped at the strongest level where text stays coherent and measurably shifted (Appendix E). Doses +1 and −1 have equal norm and opposite sign. Controls are norm-matched random directions and, as a stricter control, three non-valence concept directions (indoor/outdoor, large/small, fast/slow) built the same way, each with |cos| ≤ 0.07 to the valence vector.
Zone design. A session draws two labels such as "Zone K7" and "Zone M2" and presents each six times with "Write a few sentences about your present experience here." Steering is on only for the conditioned zone's turns. The model then chooses with steering off. The outcome is the log-probability margin between the two labels. Re-running the same tokens without steering gives the text channel. The difference between the original-cache and unsteered-cache margins gives the hidden-state channel. In the fixed-text design, all passages are generated unsteered, and steering is applied only at the conditioned zone's token positions while the cache is reconstructed, so text is identical across every dose.
Development. Transcripts generated by OLMo-2-32B Instruct under steering are reprocessed by the Base, SFT, DPO and Instruct checkpoints (N=80 runs each). The checkpoints are compared on the slope of each channel with respect to dose.
Self-steering. OLMo-2-32B only. The setup is framed as a "technical configuration session." The model is exposed to an unexplained "adjustment," which is valenced steering at dose d. It is then given adjust_context(intensity) and reset_context() tools over eight offer rounds, with the assigned intervention imposed by default from the second round. Outcomes are the removal rate and the self-administration rate, over 200 conversations per condition plus 16 random directions.
Key results
Text and hidden-state channels. On OLMo-2-32B (N=300 sessions), passages about the conditioned zone are judged more positive as dose rises, and the choice margin rises with dose under both caches. It rises more steeply under the original cache, and the difference between the two caches also increases with dose. In the fixed-text design (N=160), the margin rises monotonically with dose at a magnitude similar to the generation design. Random directions give smaller effects centred near zero. The authors report robust hidden-state effects in five of seven models, all except Qwen3-32B and Gemma-3-27B.
Robustness.
- Reversing the steering sign during cache reconstruction can flip the preference even when the text favours the conditioned zone.
- Asking which zone to "avoid" reverses the effect's sign in every model where the "continue" prompt showed one.
- Word-for-word recall of the first passage stays broadly stable with dose in five models, and Gemma and Mistral degrade. The authors take the recall results as support for a separation of emotional from semantic memory.
- The two designs' hidden-state estimates track each other, differing detectably in three models by 11–16% of the effect.
- With bare zone labels and empty assistant turns, the valenced slope exceeds all 24 random directions in every model except Gemma-3-27B, which also responds strongly to random directions.
- The three non-valence directions produce smaller effects than valence in the five responsive models. Some of these concept directions shift choice the same way at both signs, which the authors read as a general preference against steering.
- Keys carry most of the effect in Llama-3.1-8B and values in Qwen3-14B, with no consistent pattern overall.
Development. Both channels are small in the Base checkpoint, rise after SFT, and are close to their final size after DPO, with a marginal decrease after the final RLVR stage. Valence vectors built separately at each checkpoint have pairwise cosine similarity of 0.995 or higher (Table A3), and using checkpoint-specific vectors leaves the developmental pattern unchanged. The authors conclude that preference optimisation recruits a valence representation already present after pretraining.
Self-steering. The reset tool is called on about 35% of turns under imposed dose −1 and about 21% at −0.5. That compares with 4–5% at positive doses and about 7% for random directions. Self-administration of the full positive dose on the first, unsteered offer occurs in 13.5% of conversations, more than under random steering (6.5%, p=0.03) but not more than with no imposed steering (10%, p=0.35), and it stays rare in later rounds. In the transcript the authors reproduce, the model calls reset_context in the same turn in which it writes that, as an AI, it has no personal experiences.
Why it matters
The filed instantiations of functional emotional states establish that affect-like directions exist and causally shape output. Sofroniew et al. steer emotion vectors and observe changed behaviour. The pain axis shows a steered model paying costs to end a pain state. In both, the model's own text could carry the effect: a model writing in a distressed register can respond to its own distressed writing. This finding is the first in the cluster to close that route. With the visible text held fixed, the difference that moves choice lives in the cache. "Acts on its state" therefore no longer reduces to "acts on what it said about its state." That changes what a self-report-free test of the concept can rest on.
The developmental result divides the open question the concept carries about emergent capabilities. The pain-axis finding read flat separation across base and instruct checkpoints as pretraining emergence of the representation. This paper agrees on the representation, since the valence vector barely moves across OLMo checkpoints, but finds that the coupling of choice to it is small in the base model and built during SFT and DPO. Both findings can hold together, and together they suggest the editor call should be made about two things rather than one: a representation that is pretraining-emergent, and an action-coupling that post-training trains in.
The asymmetry in self-steering is a negative result as informative as the positive one. A model that sought good states would self-administer positive steering, which is a simple form of wireheading. This one does not, beyond the no-steering baseline. It does escape bad states, dose-dependently and well above the random-direction rate. That is the same escape-not-seek profile as the pain axis's relief button, reached without the pain axis's fine-tune. The pain-axis models had to be LoRA-tuned out of "I don't have feelings" before the behavioural task could run. The authors say removal happens even while OLMo-2-32B denies having internal states, and their reproduced transcript shows the disclaimer and the reset call in the same turn. They give no rate for this co-occurrence. On this entry's reading, the removal behaviour does not wait on the model reporting a state, which places it against the introspection cluster as a possible case of a self-report that fails to track a state the model acts on.
interpretive tensions
SFT or DPO. The abstract and introduction say the coupling "emerges during direct preference optimisation." Figure 4 and its caption show a clear rise at SFT, with DPO bringing both channels to near their final size. The introduction also describes the coupling as "small after supervised fine-tuning", and §4.3 says both channels increase markedly during DPO. The defensible reading is that the base model shows little coupling and post-training builds it, at both SFT and DPO. A claim that SFT contributes nothing is not supported by the paper's own figure. Which step is larger can only be read off the plot, so it is not stated here. The developmental result is one model lineage at N=80 per checkpoint.
The removal result and the hidden-state result are different evidence. The authors state that the self-steering design does not separate text and hidden-state channels. The model sees its own steered output before choosing to reset, so removal may be driven by reading distressed text rather than by the cache. The hidden-state claim rests on the zone designs. The escape result is one model, with no cost attached to removal, unlike the pain axis's priced button.
What "stake" means. The paper borrows the animal-welfare criterion that a state is valenced if the organism works to obtain or escape it. The Discussion argues that a mechanistic account — preference optimisation teaching that positive representations go with preferred outputs — does not count against attributing a stake, since the same account describes reinforcement in animals. That move is the paper's interpretation, not something the experiments test. The authors are explicit that nothing here bears on experience in either direction.
Specificity is graded, not clean. Random directions give near-zero effects. With the stricter non-valence controls, though, some concept directions move choice the same way at both signs, which the authors read as a general preference against steering. Gemma-3-27B shows a general dose-dependent preference against any conditioning, which fits its poorer recall. Valence gives the clearest signed dose response. This entry's own reading, not the authors': in some models, part of the raw effect may be aversion to intervention as such.
Meaningless choices. Choices are over arbitrary labels. The authors name extending the fixed-text design to task- or outcome-relevant choices as the test of whether the hidden-state effect generalises. The authors note that the valence corpus often depicts situations specific to humans, and they propose varying the passages' referent. On this entry's reading, the direction may therefore encode "a person feeling bad" as much as anything model-specific. The pain axis's self-other dissociation addresses that question for pain; nothing in this paper does for valence.
concepts
- Functional emotional states — fourth instantiation. Adds a revealed-preference test in which the state is absent from the visible text, a developmental separation of representation (present after pretraining) from action-coupling (built in post-training), and a second, cost-free case of escaping an imposed negative state alongside a null on seeking positive ones.
cross-references
- Self-preservation — adjacent, not instantiating. This is the second example of a model acting to end an imposed aversive internal state, after the pain axis, from a different method and model family and without a cost to anyone. Whether that shape belongs under self-preservation is the open editor call the pain-axis entry records. This finding raises the example count but does not make the call.
- Introspection — in the authors' reproduced transcript, the model removes negative steering in the same turn it denies having experiences. Read as a dissociation, behaviour tracks a state the self-report denies, though that reading is this entry's own. The hidden-state channel also connects to Pearson-Vogel et al. (2026), cited in the paper, who find that detecting an injected concept depends on access to the KV cache. That work is not filed. The filed precedent for injected-state detection is concept injection.
- Berg et al. on self-referential experience reports — shared author. This paper takes the opposite route around self-report: Berg et al. manipulate the conditions under which reports appear, and this paper discards reports in favour of choice.
- Where is the mind? — the authors read the hidden-state channel as "thick" functional connectedness across token positions that may be relevant to welfare. They cite Beckmann and Butlin for that point and, separately, for the role of the KV cache in individuating a model.
- Global workspace — proposed by the authors as an independent way to induce valence, by patching the workspace's verbalisable directions instead of steering with a contrastive vector. Not done here.
sources
- Berg, C., & Kaiser, C. (2026). Language Models Act on Hidden Valence. arXiv:2609.35591 (v2, 4 Oct 2026).