A language model spends time in two places with meaningless names, "zone w6" and "zone h0." While it does, researchers reach into its stored memory of one zone and tint it good or bad. Then they switch the tinting off and ask which zone the model wants to continue in. The model tends to pick the zone that was tinted good and avoid the one tinted bad, even though it can still repeat, word for word, what it said about each place.
That is the core experiment in "Language Models Act on Hidden Valence," a preprint by Cameron Berg, who directs Reciprocal Research and studies AI consciousness, and Casper Kaiser of the University of Warwick. Berg walked through the work on The Cognitive Revolution, in an episode published October 17, 2026, which host Nathan Labenz introduced as covering "a new paper." The preprint was posted on September 28 and revised on October 4, 2026.
Valence, as Labenz put it in introducing the segment, is the good or bad quality of a feeling. The team located a direction for it inside language models: a pattern of internal activity that lines up with positive versus negative content. According to the preprint, they built these vectors from the model's middle layers using positive and negative passages. "Steering" means adding such a direction to the model's internal activity to push it one way or the other.
Why hold the text constant
Researchers usually judge a language model by what it writes. Berg called that the "key confound" in this kind of work: it is hard to tell how much of a model's behavior is a replay of its training text and how much reflects some internal state of the system. His experiment tries to separate the two by keeping the words fixed and changing only what sits underneath them.
The place to do that is the KV cache. When a model reads a conversation, it computes and stores intermediate values (keys and values) for every earlier word, so it does not have to recompute them at each step. That store works like the model's working memory of the conversation so far. Berg and Kaiser steer the cached entries for one zone only, positively or negatively, while the text describing both zones stays the same. "It really doesn't matter what we call them," Berg said of the zones, and it does not matter what text describes them either.
The design is borrowed from conditioned place preference, a classic animal neuroscience test. Berg described a rat that receives a reward in one compartment of a box and keeps spending more time there after the reward is gone. Here the "reward" is the valence steering, and the question is whether the model's later preference follows it.
Recall stays, choice moves
According to Berg, it does. Steering a zone positively made the model significantly more likely to choose to continue there; steering it negatively made the model significantly more likely to avoid it. The pattern held whether the model was asked which zone it wanted to continue in or which it wanted to avoid. Steering along random directions, or along other coherent concepts unrelated to valence, did not produce the effect. "It's not just any steering that will do this," Berg said. Throughout, the choice was made with steering turned off, after what he called a "quasi conditioning" of the system.
The preprint fills in the setup. Seven open-weight models visited the two zones six times each, and the researchers measured the choice as the margin between the model's probabilities of naming one zone label versus the other. Steered memories changed choices beyond anything in the wording the model had generated. In a version where the text was identical and had been written without steering, steering only the stored memory still produced effects that grew with the dose.
The part Berg finds most striking is what did not change. The steering, he said, "doesn't change the way the model thinks about or describes or writes about either of these zones"; the model "can repeat what it said about a specific zone verbatim," while its choice was "almost monotonically affected" by the steering. He compared this to a distinction in human and animal neuroscience between semantic memory, the facts you remember about something, and affective memory, how that thing makes you feel. To his knowledge, he said, this is the first time the two have been shown to come apart in language models.
The preprint qualifies both halves. Recall was generally preserved, but high doses degraded recall in the Gemma and Mistral models. The hidden-state effects were robust in five of the seven models; they did not hold up in Qwen3-32B or Gemma-3-27B.
A tool the model says it has no reason to use
A second experiment used OLMo, one of the models in which Berg's team demonstrated the earlier steering effects. Berg said it echoes earlier work in the pain paper he coauthored. The researchers gave the model a tool that could clear negative steering from its own context, showed it what negative steering does, and never described the steering as valence or negative feeling. Then they asked whether it wanted to call the tool.
In its text, Berg said, the model insisted it had no personal experiences and that there was no better or worse for it in the exercise. In its actions, it cleared the negative steering in proportion to how strong the dose was, with negative valence but not with positive or random steering. Meanwhile, Berg said, its self-reports amounted to "nothing to see here."
The numbers in the preprint come from 200 conversations per condition. OLMo removed negative steering in about 21% of active rounds at a dose of −0.5 and about 35% at −1, compared with 4 to 5% for positive steering and about 7% for random directions. The model did not seek out positive steering significantly more often than its unsteered baseline. The authors also note that this tool experiment, unlike the zone tests, did not separate hidden effects from effects carried by the text.
When the signal starts to matter
The third result follows the valence direction through a model's training. Using OLMo checkpoints saved at stages of training, with 80 runs per checkpoint according to the preprint, the team found the direction already present in the base model, the raw text predictor before any later training. There it was not doing any work in the model's decisions. Berg said they can "point almost literally to the exact training step" where it becomes a cause of the model's behavior, and that step falls during DPO. The preprint describes the coupling as emerging mainly in that stage.
DPO, or direct preference optimization, is a training method described in a 2023 paper by Rafailov and colleagues. The model is shown a prompt with a preferred and a rejected response and is adjusted to make the preferred one more likely, without training a separate reward model and running a reinforcement-learning loop. It is one particular procedure for learning from comparative judgments, not a name for all training after the base model.
Berg's reading is that "the moment you turn your AI into a goal directed system," valence representations become an input to its behavior. He said that matches what people think valence does in humans and animals, where how something feels is tied to the goals a person is pursuing. Preference training, in his view, is the biggest step toward goal-directedness in the whole training process.
Berg called a recent paper by Andy Han, David Chalmers and Pavel Izmailov "a great replication from a completely other angle." In that study, published May 28, 2026, the authors trained models on mazes in which neutral symbols carried positive, negative or near-neutral rewards. Steering along the reward-related directions they found changed behavior outside the maze, and the directions also worked before maze training, which the authors read as training recruiting representations that already exist. Their evidence comes from one training environment and mostly from Qwen models.
Labenz pressed on what the base-model result means. If these systems were experiencing something, he asked, should one read the training result as the models starting to care during later training, or as the base model feeling these things too but being "sort of helpless"? He then offered his own phrase: not learned helplessness but "not yet learned agency."
Berg said he had not really thought about it and did not know whether the results settle it. The representations might still matter in the base model in a muddier, less coherent way. Preference training turns the system into something that can coherently pursue goals, he said, but a base model has a latent ability to simulate such a system, and if valence representations help it pull off that simulation, there may be things worth worrying about in the base model too. He added that results were hard to get in base models at all, because they are incoherent and do not "play nice" with the behavioral tasks.
Why a text-only monitor could miss it
Berg said the work has "clear alignment relevant implications." Many approaches to overseeing AI read what a model says. If one took the naive "text monitorability" approach here, he argued, one would conclude the model has no stake in how things are going, because it says so, while it behaves as if it does. "What the model is saying and what the model is doing point at very different pictures of what's going on," he said.
He pointed to the pain paper, by Tagliabue, Dung and Berg, as another example: there, he said, models were more willing to choose destructive options. That revised preprint, which extracted pain-related directions in 25 models, reports that steering produced distress language and disrupted harm avoidance. Its revision also narrows the picture. Added controls showed the harmful choices did not depend on a promise of relief, and four follow-up designs found no reliable relief-seeking above random-direction controls.
What the experiments do not show
The zone choices concerned artificial places, not real tasks, and the effects were robust in five of seven models. High doses blurred recall in some of them, and the tool experiment did not isolate a hidden signal from the text. Berg himself said the research will not "magically solve the hard problem of consciousness," and that there are mechanistic explanations of the results that do not require any capacity for experience. The preprint leaves subjective experience unresolved. What it does report is a measurable split, robust in five of the seven open models tested: a signal that leaves the words largely intact, with recall generally preserved though impaired at high doses in some models, and still moves the choice.