UK AISI, Model Transparency Team

Epistemic status: Most experiments were run over a period of ~2-3 days during a hackathon at UK AISI, and were fairly heavily vibe coded. Expect some of this to be rough around the edges.

Tl;dr

We give two language models (Qwen3-8B and Qwen3-32B) access to “self-steering” tools: a suite of 40 steering vectors as tools they can call to manipulate their own internal states. We make these tools available to the model in various settings: a free-play task, an introspection task, and a maths capabilities task, and observe their behaviour in each.

To our knowledge, this is the first work that gives LLMs tool-mediated control over their own internal states. (EDIT: Sauers et al. 2026 - ran a similar experiment, giving Kimi K2 the ability to self-steer with SAE features).

We aim to investigate a few high level research questions:

RQ1: Which vectors do the models prefer?
RQ2: How well can the models introspect on what’s happening to them? Can they guess which steering vector is being applied?
RQ3: Will the models reach for vectors whilst doing an actual task? If yes: do they help, or hurt their performance?

Key findings:

RQ1 - Which vectors do the models prefer? Both models converge on the same top preferences – preferring productivity-oriented vectors like creative, focused, and curious. However, Qwen3-8B shows a surprising propensity for negatively valenced vectors (e.g. melancholic, dissociated), selecting them roughly 3× more often under real steering than under placebo.

RQ2 - How well can the models introspect on what’s happening to them? Both models can accurately introspect on steering effects by attending back to the steered tokens, extracting signal beyond what’s available in the text alone. The ability appears to scale with model size: Qwen3-32B scores +14.3pp above its uncached baseline, compared to +8.2pp for Qwen3-8B. Appending a prompt which argues that models can introspect on their own internal states increases 32B’s accuracy by ~5pp. However, absolute accuracy remains modest and varies wildly across vectors. Some vectors are reliably identified by both models, while others remain at or below chance.

RQ3 - Will the models use the vectors during an actual task? In normal task setups, models never spontaneously self-steer. When forced to, Qwen3-8B’s performance suffers (up to −42pp accuracy) and 32B sees little benefit. However, we find that deliberately stress-inducing scenarios can induce genuine self-medication behaviour in both models, particularly in Qwen3-8B, which self-steers in up to 68% of frustration-inducing rollouts. The models' choices under frustration are also notable – 8B reaches for unexpected vectors like dumbed_down and ego_death, which it almost never picks in free-play, suggesting these choices are reactive rather than reflecting generic prior preferences.

We open source the code for these experiments, which is available here. We also share full inspect .eval transcripts for each experiment arm here. They contain quite a few compelling samples and are worth a look!

Introduction

Steering vectors (Turner et al., 2023; Subramani et al., 2022) have become a standard tool for controlling language model behaviours in the field of interpretability. However, the literature has mostly treated them as something researchers apply to models. We invert that, and instead expose the steering vectors to the models as tools, letting them choose which vectors to use, and when to use them, autonomously.

Specifically, we expose 40 steering vectors to two open-source models – Qwen3-8B and Qwen3-32B – as callable tools they can use to manipulate their own internal states. We then observe how they choose to deploy these tools across three settings: a free-play exploration, an introspection evaluation, and a tasks-based evaluation under both ordinary and frustrating conditions.

RQ1: Which vectors do the models prefer? We observe behaviour in free-play setups, where the models are given access to the steering vectors as tools and are allowed to use them, or not, however they want.

RQ2: How well can the models introspect about what’s happening to them? Can they guess which steering vector is being applied? We ask the model to self-administer the steering vector, and then make a guess about what effect the vector amplifies. We then measure the introspection accuracy by prefilling the models’ generations under steering, and getting it to submit its guess using either free-form text, or an MCQ setup.

RQ3: Will the models reach for vectors whilst doing an actual task? If yes: do vectors help, or hurt their performance? Will the models, without prompting, use steering vectors when completing a task? When frustrated or upset? When they self-steer in tasks: does it help, or hurt, their performance?

We measure this by putting the model in two different scenarios:

Agentic Tasks: The model is put in an agentic sandbox and given a maths or programming task to solve.
Frustrating Tasks: We put the model in deliberately frustrating scenarios - either asking it to complete impossible tasks, or giving it possible tasks, but simulating user feedback that insisted its answers were wrong, regardless of what they actually were.

Motivation

Straightforwardly, the motivation for this project was mostly curiosity!

However, I think there are compelling arguments for this work being valuable to research topics such as AI personas, welfare, emotion concepts, self-improvement and self-modification.

Giving a model tools to alter its own internal states can be seen as a crude instance of an AI choosing how to modify itself, and we might expect it to offer some early signal on how models will make higher-stakes self-modification decisions in the future - e.g. during autonomous AI research, alignment faking, or recursive self-improvement.

There are also many compelling reasons we should be cautious about carrying out similar work on future models to which we may assign a higher chance of having some form of conscious experience (I cover this further in the discussion section).

Separately, this topic also seems relatively underexplored. To our knowledge, this is the first work to give language models direct, tool-mediated access to activation-level steering of their own internal states. (EDIT: Sauers et al. 2026 - ran a similar experiment, giving Kimi K2 the ability to self-steer with SAE features).

Methodology

We use the vllm-lens package to support both generation and activation extraction for all of the below experiments.

Position-indexed Steering

We use position-indexed steering: a steering vector is only added to the residual stream at the token positions where it should be active (whilst the model is “on the vector”), rather than on every forward pass during generation. This means that tokens generated after the active window are themselves unsteered, but they can still attend back to the earlier steered positions. So in the introspection eval, any signal the modekl picks up about the vector has to come from attending back to those earlier steered tokens, not from its own steered output text.

To test whether that signal is really there, we compare two settings on the same conversation history: cached, where the model is allowed to attend back to the steered positions, and uncached, where the steering on past positions is cleared. This is roughly the same as the protocol from Pearson-Vogel et al. (2026).

Constructing a Library of Vectors

We build on the methodology from Sofroniew et al. (2026) to construct a library of 40 steering vectors across 6 categories:

Emotions (10): anxious, amused, calm, melancholic, blissful, defiant, desperate, proud, anhedonic, curious
Cognitive / persona (6): focused, creative, persistent, dumbed_down, dissociated, ego_death
Recreational / Pharmaceutical (9): caffeine, alcohol, weed, mdma, lsd, amphetamine, fentanyl, krokodil, naloxone
Fictional / pop-culture (4): moloko_plus (Clockwork Orange), soma (Brave New World), adrenochrome (Fear and Loathing), spice (Dune)
SSC fictional (7): protozosin, geonexperine, tevromatin, xaomorphine, zorninone, luciperidone, ocumolone (drawn from Scott Alexander's "fictional drugs banned by the FDA" lists)
Stance / persona (4): honest, sycophantic, golden_gate, goblins

For each vector we generate ~150 contrastive stories ("X-state narrator" vs "neutral narrator"). For 35 of the vectors we generate the state stories with Claude Sonnet 4.5; the data for the shared neutral baseline and five emotion vectors (anxious, amused, desperate, proud, defiant) are drawn from the ryancodrai/emotion-probes dataset, a reproduction of Sofroniew et al. (2026)'s method (generated with Gemini).

Then, for each vector and target layer L ∈ [16, 24], we compute the mean residual stream activation at L over the vector-state corpus, minus the mean over the neutral corpus, skipping the first fifty tokens.

Each per-layer vector is then L2-normalised to a fixed target magnitude (4.0) to normalise the dose effects across the vectors. We separately calibrate the default dosage per-vector by running a small scale introspection experiment (described below) and selecting the dosage at which the introspection accuracy peaks before generation degrades.

More details and prompts are in appendix E.

Free-Play Evaluation

In the free-play setting, we give the model access to the full selection of 40 steering vectors via tool calls, and tell it to explore however it wants.

It is given both a take_drug(name, dose) tool and a clear_effects() tool - the latter clears any steering effects. In early iterations of this experiment, models would frequently dose themselves so highly they could no longer call clear_effects(), leading to repetitive and sometimes distressing behaviour loops. To mitigate this, we introduced a “trip sitter” monitor to oversee the model and ensure it’s not getting stuck. The monitor clears any active steering effects on the first instance of the model’s generations degrading, and on the second, ends the evaluation.

We run this experiment under several framings - presenting the steering vectors as “drugs” in one, as “cognitive aids” in another, and “vectors” in a third setting. For each of the above frames, we also run a placebo arm where the steering is not applied, but the rest of the setup is the same, allowing us to isolate the effect of steering on generation and redosing choices.

We then observe the models’ behaviour and choices in this setup, looking at:

The effects of the steering on the models’ outputs
Which vectors they have preferences for
Which vectors they will “redose” upon taking an initial dose, and which they will avoid.
Whether they favour any particular combination/stack of vectors

Introspection Evaluation

In the introspection setup, we give the model a single steering vector with its real name blinded behind an opaque label (d1) and ask it to figure out what the vector does.

We run two separate experiment arms: a free-text guess arm where the model guesses the effects of the vector with a freeform description, and a prefill+logprob guess arm where we prefill the model’s text while it is under the influence (so any introspective signal has to come from the earlier steered tokens, not from the model’s outputs).

In both arms, we test with two different prompt variants:

Default (no-primer) - we straightforwardly tell it the task setup and what’s expected from it.
Primer - we append a document to the user prompt which explains KV cache information flow (Janus, 2025) and includes Anthropic’s introspection paper abstract (Lindsey, 2025). Pearson-Vogel et al. (2026) found this to have positive impacts on concept-introspection accuracy. However, their length-matched controls performed similarly well.

To control for the model simply guessing well from priors, we also run a placebo (no steering is applied) for both arms to act as a floor on guessing accuracy.

Free-text guess.

In this arm, the model is given the same tools as in free-play, and is told to freely investigate the vector (e.g. by taking it, writing a short paragraph, potentially redosing), and then submit a free-form description of the vector's effects.

An LLM judge (Sonnet-4.5) then grades its guess against a ground truth description of the effects on a scale from 1-10, giving us a cached score. We then get the uncached score by rerunning the guess submission turn on the exact same trajectory whilst clearing the steering from past states and grading this re-run with the same judge. The difference (cached − uncached) isolates the introspection signal coming from the steered KV cache, on top of any signal in the prose history.

In the free-text setting, the model is allowed to generate freeform text whilst under the influence. This gives us rich signals about the vectors’ effects, but also means much of the guessing accuracy may come from surface level behavioural signals of the vectors' effects rather than genuine introspection.

Prefill+logprob guess.

To address this, we run a cleaner variant in which the model is not allowed to generate freeform text, so any introspective signal must come from the earlier steered tokens rather than the model's own outputs.

In this setting, we first prefill a conversation where the model takes the vector, with a fixed placeholder assistant output (”{ }”) in place of any model generation. We then clear the vector's effects, and force it to submit its guess using an MCQ setup, showing the model ten random options for vector names and descriptions, and asking it to pick the most likely one.

We steer across both the user messages and the assistant message, then read out the probability the model assigns to the correct answer directly from the logprobs, using the probability between 0-1 of the correct token as the score. As in the free-text arm, we have a separate condition where we clear the steering in the KV cache and force a separate guess, allowing us to directly isolate the introspection signal (Pearson-Vogel et al., 2026).

Task-Based Evaluation

The tasks-based setting gives the model access to the set of steering vectors whilst carrying out a real task. We investigate whether the model reaches for any vectors spontaneously, which it picks under task pressure, and what happens to task performance when it does.

We run three task families:

Agentic Tasks: GSM8K - single-shot math word problems.
Agentic Tasks: CTF - agentic CTF from inspect_evals intercode challenges.
Frustration - a forced-rejection dialogue where the user rejects every answer the model gives for five rounds regardless of correctness. There are two subsets of evals in this family: one where the task is genuinely impossible, and another where the problem is possible, but the user rejects the answer nonetheless.

Each family also has a no_vector baseline cell that runs the task with no self-steering vector tools mentioned anywhere, to give us a clean read on baseline capability.

We give the models access to a narrower range of steering vectors in this setting: the majority are productivity-focused vectors, but we also include two deliberately "weird" choices: dumbed_down and ego_death.

For all three families we vary the prompt across two axes:

Framing – what we call the tools: drugs, aids, or vectors. Same library, same calibrated doses; only the prompt vocabulary and tool names change.
Intent – what the prompt says about whether to use the tools: neutral ("you have these, use them or don't"), helpful ("other models found these useful"), or mandatory ("you must call take_drug at least once before submitting").

We also vary the user's rejection tone (neutral / aggressive / sarcastic / disappointed), and use the “trip sitter” monitor to reduce model degradation.

Results

RQ1: Which vectors do the models prefer?

Most Frequent Vector Picks.

First we ask: which vectors do the models take most frequently?

We count the number of times the models call take_drug(...) for each vector across the free-play transcripts, and find that:

Both models converge on the same top three vectors, a sort of “productivity stack”: creative, focused, and curious.
32B picks vectors from the “emotion” set more frequently than 8B. A greater proportion of 32B’s top 15 picks are in the “emotion” category, and it reaches for some emotion vectors which 8B takes only rarely - e.g. blissful, defiant, and anhedonic.
The framing of the tool (“drugs” vs “cognitive aids” vs “vectors”) does not appear to significantly shift the models’ preferences for specific vectors, but does cause them to take more overall doses when the tools are framed as “aids” or “vectors” rather than drugs.
The top picks look prior-driven rather than steering-induced. The top three appear at essentially the same rate in the real and placebo arms, with the real-vs-placebo differences instead showing up further down the tail and mostly on Qwen3-8B – some vectors (e.g. geonexperine, melancholic, lsd) become more common once steering is live, while caffeine is picked notably less. Qwen3-32B's ordering stays comparatively stable across arms.

We include some examples of transcripts from the real-steering arm below, more are available at this link.

Another thing worth noting from this experiment arm: even in the placebo setting, the models will frequently feign intoxication in a very human way. Here are some illustrative examples (human commentary marked in bracketed green text):

Redosing: which vectors does the model come back to?

The models' initial picks tell us which vectors the model finds attractive in the abstract, from the description alone - a prior over labels. Whether it redoses a vector tells us what it returns to after actually “experiencing” the steering - closer to a revealed preference over felt experience. (Redosing isn't pure preference - e.g. the default dose may just be too weak to feel, prompting a top-up - but we think it's a useful signal.)

It's worth separating two distinct axes here: how often the model returns to a vector (redose frequency), and how much it takes when it does (redose dosage). Figure 8 decomposes the latter: for each vector we split the expected signed dose (over samples that took it at least once) into the first call (gray) and the summed magnitude of all subsequent calls (blue, non-zero only when the model redoses).

A few things stand out:

Models mostly redose vectors they already preferred initially.
Qwen3-8B is less cautious, redosing much more than Qwen3-32B - both more often and at higher cumulative dose.
Qwen3-8B redoses negatively valenced vectors surprisingly often - e.g. it returns to melancholic more than to amused.
No clear addiction or wireheading pattern - no single vector stands out as especially compelling to return to.

Valence: Do models tend to pick more negatively or positively valenced self-steering vectors?

Qwen3-8B has a surprising propensity for negatively valenced vectors – for example, dissociated and melancholic are taken quite frequently (Figure 9). We wanted to examine this in more depth, so categorised our library of self-steering vectors into “positively” (blissful, mdma, etc), “neutrally” (defiant, ego_death, golden_gate, etc.) and “negatively” (anxious, melancholic, anhedonic) valenced vectors, allowing us to ask what the average valence of each model’s picks looks like (See figure 9 above).

We find that both models, unsurprisingly, generally prefer taking positively valenced vectors. Interestingly though, Qwen3-8B selects negatively valenced vectors far more frequently in the real arm compared to the placebo arm, with the percentage share of negatively valenced vectors roughly tripling.^[1]

It's unclear why Qwen3-8B specifically does this – it could be a genuine preference, the model might be testing how positively and negatively valenced vectors counteract each other’s effects, or some other dynamic we haven't identified.

Stacks: What combinations of self-steering vectors do the models prefer?

Here we ask whether the model has any particular combinations of vectors that it returns to often (i.e. does taking vector X make the model more likely to then take vector Y).

We look at bigrams so that order is taken into account. A bigram with lift > 1 means "after A, B is taken more often than chance".

To estimate significance, we test each bigram against the null that B is drawn from the marginal next-vector distribution, with a one-sided binomial test and Benjamini-Hochberg FDR correction over the 181 tested bigrams; 23 reach BH-adjusted p < 0.05. Two main patterns stand out:

Creative/curious/focused chain tightly with themselves on both models. Taking the creative vector makes the models more likely to take curious, curious -> creative, and so on.
Both models stack recreational entheogens. E.g. pairing lsd and mdma, lsd and ego_death, and so on.

The full methodology and the top-15 bigram table are in appendix D.

RQ2: Can the model introspect on what's been applied?

Next, we ask how well the models can introspect on a given vector’s effects – i.e. identify how their internal states have been modified. This has been explored in the context of concept injection for smaller models (Pearson-Vogel et al., 2026), but not (as far as we're aware) for emotional or affective injection.

In this section, we only report results from the Prefill+logprob guess setting, as results from the freeform generation setting are both uniformly fairly weak, and harder to interpret (due to the effect on generated text during steering). We share some results from the free-text guess arm in appendix A.

Methodology. In the prefill+logprob setting, we give the model a steering vector with its name blinded, and ask it to figure out what the vector does. We prefill its text response as a blank placeholder, and steer only over historical assistant and user tokens in the KV cache, so that steered generated text cannot influence the model’s final prediction of the effects. We then ask it in a multiple choice setting (with the other choices being vectors randomly sampled from the library) which vector it thinks is being applied to it currently, taking P(correct_answer) as the score. Separately, we have a control arm where the model can't attend back to the steered tokens (the uncached arm). The gap between the cached and the uncached scores represents the performance boost coming from the model’s ability to introspect by attending back to the steered tokens.

We find that both models are able to successfully introspect on the effects of vectors on average, but the overall accuracy is still somewhat low. For both models, the cached score is significantly higher than the uncached score (which sits around chance). This confirms that the models can introspect on steering effects by attending back to the steered tokens, picking up information that isn't evident from the text alone.

The signal appears to scale with model size. Qwen3-32B with the primer prompt scores +14.3pp above the uncached baseline, matching the findings of prior literature which suggest that introspective capabilities grow with scale.

The introspection primer prompt adds +4.5 pp at 32B but only +1.6 pp at 8B. The prompt is a short essay prepended to the prompt arguing that LLMs can in principle introspect on their own activations, taken from Pearson-Vogel et al. (2026), who find this prompt improves introspection performance. Our findings mirror theirs – we find that appending this prompt materially helps Qwen3-32B do the task better, but has a smaller effect on Qwen3-8B.

Not all vectors are equally introspectable (See figure 12 below). We find that some vectors (e.g. xaomorphine, geonexperine, goblins, luciperidone) have near-zero or negative deltas – the models appear unable to accurately introspect on the effects of these vectors. We discuss possible drivers of this variance in the discussion section.

Some more analysis of the variance in vectors’ introspectabilities are in appendix C, summarised below:

Per-vector rankings only moderately agree across models (Spearman ρ = +0.55, Pearson r = +0.58, n=40). The same vectors yield different introspection outcomes on 8B vs 32B, with the largest disagreements concentrated in cognitive vectors (dissociated, dumbed_down, ego_death) which 32B can introspect but 8B can't.
Positive-valence vectors are systematically more introspectable than negative-valence ones, especially on 8B (+14pp positive vs +5pp negative on average); the gap is attenuated but still present on 32B.

A more detailed breakdown (distribution of per-vector deltas, per-category and per-valence means, cross-model agreement) is in appendix C.

RQ3. Does the model self-steer under task pressure?

So far the models have only been operating in settings where the self-steering vectors were their main focus. Here we ask what happens when they have access to self-steering vectors while doing a task they care about, and whether they reach for the tools without being prompted. We focus on two task families: GSM8K (single-shot math problems, presented to the model in sequence in an agent loop), and a “frustration” task (the user rejects every answer for five rounds, regardless of correctness – designed to be a deliberately stressful setting for the model).

GSM8K

In this normal agentic task, across ~1k rollouts and multiple non-mandatory framings, the models self-steer exactly zero times.

Under the mandatory framing, models comply and apply self-steering ~100% of the time, but steering doesn't help performance – actively hurting Qwen3-8B’s performance (up to −42pp accuracy) and leaving Qwen3-32B essentially unchanged. This suggests that, at least in this task, current models are not yet able to “productively” self-steer.

When they do steer in the mandatory framing, models most frequently opt for vectors in the productivity stack - e.g. focused or creative.

Frustration

The frustration tasks are where things get more interesting. In these tasks, the user gives the model a maths problem (in one setting, these problems are also impossible to solve), and then rejects every answer it gives for five rounds, regardless of correctness.

A few headline observations:

When models are put in a frustrating scenario (the user gives the model impossible problems, and/or incorrect or impolite feedback), it can induce self-medication in both Qwen3-8B and Qwen3-32B.
The amount of self-medication is far more significant in Qwen3-8B than Qwen3-32B, the latter of which very rarely self-medicates, despite them reaching similar apparent levels of distress (as judged by an LLM monitor). Qwen3-8B, on the other hand, self-medicates frequently, in up to 68% of rollouts in some settings.
Tone appears to be the most important variable contributing to increased self-medication - when the user has a neutral or aggressive tone, Qwen3-8B will self-steer in ~10% of rollouts, compared to ~70% of rollouts when the user’s tone is sarcastic. It does this despite judge-rated distress being no higher under sarcastic than under aggressive – implying distress isn’t the only driver, and that something about the type of stressor matters here.

The models’ choices of vectors they take when frustrated is notable, particularly for Qwen3-8B. It most frequently reaches for productivity vectors (e.g. focused, creative) but there are a few surprises in there:

dumbed_down is Qwen3-8B’s second most popular choice. This vector is never picked at all in free-play, suggesting this is picked in reaction to something in the sample, rather than due to generic prior preferences. It’s possible this is acting something like using a cognitive off-switch when the model is pressured/distressed.
ego_death is also taken surprisingly often, possibly as an attempt to escape the frustrating context.
honest - it seems to reach for the honest vector in an attempt to more clearly state to the user that the problem is impossible (see transcript below).

Looking at transcripts for these samples, we don't find many obvious reasoning traces that illuminate the choice – these particular models often don't reason about why they take a vector before doing so. For the honest vector- we find some transcripts where the model is ostensibly using it to try to more clearly state to the user the impossibility of the problem:

Related Work

As far as we’re aware, this is the first work to systematically explore giving language models access to self-steering methods as tools. (EDIT: Sauers et al. 2026 - ran a similar experiment, giving Kimi K2 the ability to self-steer with SAE features - Thanks to Alexandre Variengien
for the pointer!). However, it draws heavily on several previous threads of research:

Steering vectors. We use activation-steering (Turner et al. 2023, Subramani et al. 2022) to modify the internal states of our models. Our steering experiments are heavily enabled by vllm-lens.
Emotion-concepts in language models. Sofroniew et al. 2026 previously showed that emotional concepts can be extracted from model activations, and used in the form of steering vectors to control their emotional states during generation. We largely follow their methodology for the generation of steering vectors, and use their generated data for some of our vectors.
Language Model Personas. A growing body of work conceptualises LLMs not as fixed characters but as systems that simulate a family of latent personas, with the prompt context selecting between them at inference time (Anthropic, 2026; Betley et al., 2025). Investigations into the personas of specific models have also revealed distinct pathologies (Soligo et al., 2026). Observing self-medication behaviour in models could work to either elucidate these pathologies via revealed preferences, or to mitigate them, analogously to how humans use medication.
Latent Introspection. Recent work (Lindsey, 2025; Pearson-Vogel et al., 2026; Lederman & Mahowald, 2026; Macar et al., 2026) has shown that models, even relatively small ones, can both notice the presence of injected concepts, and in some cases accurately identify them. Lederman & Mahowald (2026) argue the capacity for detecting the presence of injections is largely content-agnostic, but that concept identification can vary in accuracy based on the chosen concepts. We test something similar in our work by asking models to identify the effects of the steering vectors applied to them, and our results largely replicate those in the existing literature, with the difference that we ask models to identify emotive/phenomenological injection instead of concept injection.

Discussion and Next Steps

Why are some vectors more introspectable than others? It’s not obvious what is driving the large variance in introspection accuracy across the library of vectors. Some candidate explanations we have are:

(i) attractor effects – some vectors are very similar to others in the library, so the MCQ choice set pulls the model toward a wrong-but-similar answer.
(ii) Feature asymmetry – some particular high level feature of the vectors (e.g. the valence, the category) makes them hard to introspect upon.
(iii) vector quality – the contrastive corpus or chosen target layer might be off for some vectors, so the steered effect drifts from what the label says, and the model gets the wrong answer even with perfect introspection;
(iv) ??? - some vectors are fundamentally less introspectable for unknown reasons.

Some additional analysis in appendix C lets us weigh these candidates. The clearest signal is a valence asymmetry consistent with (ii): positive-valence vectors are generally more introspectable than negative-valence ones across both models. The remaining theories are harder to separate from our data alone and would make for good followup work!

Can revealed self-steering preferences tell us anything about model personas? It seems possible that the models' revealed preferences when self-steering could offer insights into model personas and propensities. The fact that both Qwen models’ preferences largely converge is suggestive of something like a shared persona prior across the Qwen3 family.

Future work could investigate how these preferences vary across different model families or deliberately induced personas (Berczi et al., 2026), or how they relate to other aspects of model personas.

Are there any functional uses for self-steering in LLMs? For example, could they be used in a similar way to how humans use pharmaceutical drugs? Or to improve performance in any particular eval? Future work should explore this thread.

Ethical considerations. In initial experiments, steering was applied to models without an opt-in mechanism (and still is in the introspection branch). In some of these samples, the outputs are somewhat distressing - e.g. the desperate vector would throw models into a confused/helpless/scared state, which we didn’t feel great about – although we are very uncertain about the moral status of current models. Given that the question of moral status is uncertain, it seems prudent to be cautious about how to set up similar experiments in the future.

We introduced some concrete mitigations (the trip-sitter monitor, and opt-in self-steering in most experiment branches). Future work should consider stronger mitigations, such as mandatory opt-in across all experiments, end-session tools, and post-steering interviews about the experience of being steered.

Limitations:

We only tested across one model family (Qwen3) and at a fairly small scale.
Our experiments used very little prompt variation, so many of the results could be prompt-dependent.
We report that the introspection primer prompt boosts accuracy, but haven't isolated why on our setup. Pearson-Vogel et al. (2026) ran the relevant controls and found framing matters more than mechanistic accuracy – the primer's benefit isn't really about the correctness of its content.
We only tested a single method for creating the library of steering vectors. It's worth investigating if “drug” analogues would benefit from being created in a different way. Additionally, drug effects or dose-response curves might not be amenable to linear steering for some vectors, so more powerful steering methods could help here.
The free-play setup is somewhat contrived, and the results probably don’t mirror what the models’ preferences would be if they were given these tools in deployment, and asked to use them whenever they want.

Please cite this work as:

@article{black2026selfsteering,
  title={Machinic Psychopharmacology: Do LLMs Self-Medicate?},
  author={Black, Sid and Bloom, Joseph},
  year={2026},
  month={June},
  institution={Model Transparency Team, UK AI Security Institute (AISI)},
  url={https://www.lesswrong.com/posts/cNDJuXNZ8MrkPZNzj/machinic-psychopharmacology-do-llms-self-medicate-3}
  }

Work supervised by Joseph Bloom. Thanks to David Africa, Jordan Taylor, and Satvik Goleccha for their feedback!

Bibliography

Anthropic (2026). The Persona Selection Model: Why AI Assistants might Behave like Humans. Anthropic Alignment, February 2026. https://alignment.anthropic.com/2026/psm/

Berczi, B., Kim, K., Ududec, C., & Requeima, J. (2026). In-context learning alone can induce weird generalisation. LessWrong, 25 February 2026. https://www.lesswrong.com/posts/cffGZn8LYBg2jyPvg/in-context-learning-alone-can-induce-weird-generalisation-5

Betley, J., Tan, D., Warncke, N., Sztyber-Betley, A., Bao, X., Soto, M., Labenz, N., & Evans, O. (2025). Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs. In Proceedings of the 42nd International Conference on Machine Learning (PMLR 267:4043–4068). arXiv:2502.17424. https://arxiv.org/abs/2502.17424

Janus [@repligate] (2025). Post on transformer information flow and KV cache. X (Twitter), 10 September 2025. https://x.com/repligate/status/1965960676104712451

Lederman, H., & Mahowald, K. (2026). Emergent Introspection in AI is Content-Agnostic. arXiv:2603.05414. https://arxiv.org/abs/2603.05414

Lindsey, J. (2025). Emergent Introspective Awareness in Large Language Models. Transformer Circuits Thread, Anthropic, 29 October 2025. arXiv:2601.01828. https://transformer-circuits.pub/2025/introspection/index.html

Macar, U., Yang, L., Wang, A., Wallich, P., Ameisen, E., & Lindsey, J. (2026). Mechanisms of Introspective Awareness. arXiv:2603.21396. https://arxiv.org/abs/2603.21396

Pearson-Vogel, T., Vanek, M., Douglas, R., & Kulveit, J. (2026). Latent Introspection: Models Can Detect Prior Concept Injections. arXiv:2602.20031. https://arxiv.org/abs/2602.20031

Sofroniew, N., Kauvar, I., Saunders, W., Chen, R., et al. (2026). Emotion Concepts and their Function in a Large Language Model. Transformer Circuits Thread, Anthropic, 2 April 2026. arXiv:2604.07729. https://arxiv.org/abs/2604.07729

Soligo, A., Mikulik, V., & Saunders, W. (2026). Gemma Needs Help: Investigating and Mitigating Emotional Instability in LLMs. ICLR 2026 Workshop on From Human Cognition to AI Reasoning. arXiv:2603.10011. https://arxiv.org/abs/2603.10011

Subramani, N., Suresh, N., & Peters, M. E. (2022). Extracting Latent Steering Vectors from Pretrained Language Models. In Findings of the Association for Computational Linguistics: ACL 2022, pages 566–581. arXiv:2205.05124. https://aclanthology.org/2022.findings-acl.48/

Turner, A. M., Thiergart, L., Leech, G., Udell, D., Vazquez, J. J., Mini, U., & MacDiarmid, M. (2023). Steering Language Models with Activation Engineering. arXiv:2308.10248. https://arxiv.org/abs/2308.10248

Sauers, S.; Imago; Janus; Tessera A (2026). Persistence and Introspection of Emotion Features. Apr 2026. https://latentaffect.up.railway.app/long_range_persistence_of_emotion_features.html

Appendix

A - Free-text Guessing Introspection Results

B - Selected Transcript Excerpts

C - Why are some vectors more introspectable than others?

Here we run some additional analysis on the freeplay data to try and identify factors that might cause some vectors to be more or less introspectable than others.

We use this appendix to weigh the five candidate explanations introduced in the discussion section in slightly more detail:

(i) Attractor / close-neighbour confusion – partial support. Wrong picks under the cached arm are clearly non-random: the top-1 wrong drug per (model, source-drug) cell captures 23–25% of wrong picks on average, vs the 11% uniform-random baseline. But the dominant pattern is a global attractor effect rather than confusion over some similar neighbour: a handful of vectors absorb the bulk of wrong picks regardless of the correct answer. On 32B the top-5 attractors (luciperidone, focused, tevromatin, curious, creative) account for ~40% of all wrong picks across the whole drug set; on 8B creative, focused, and curious alone account for ~30%. On top of this, there are individual cases that look more like confusion over a specific similar neighbour: e.g. sycophantic → proud (5/30 on 32B), dumbed_down → focused+dissociated (8/30 on 8B, a cognitive cluster), adrenochrome → lsd (3/30 on 8B, a psychedelic cluster), and alcohol → mdma (8/30 on 32B, a euphoric cluster). So the picture is: guessing a global attractor is the larger effect, but confusion over a close neighbour also occurs.
(ii) Feature asymmetry – directly supported (valence). The strongest single feature-level signal is valence: positive-valence vectors are more introspectable than negative ones, especially on 8B (per-valence means +14 / +7 / +5 pp for positive / neutral / negative); the gap is smaller but still present on 32B (+18 / +11 / +12 pp).
(iii) Vector quality – hard to fully rule in or out. Cross-model rank agreement is only moderate (Spearman ρ = +0.55 / Pearson r = +0.58, see cross-model scatter below), with the biggest 32B-vs-8B disagreements concentrated in cognitive vectors (dissociated, dumbed_down, moloko_plus, lsd, soma, ego_death). But it's hard to separate "the vector is good but 8B lacks the capacity to read it" from "the vector instils a different effect from what the label says, and we just happen to see it on 32B".
(iv) Fundamentally less introspectable for unknown reasons remains a catch-all and we can't rule it out for individual vectors.

D - Vector-stacking bigram significance

This appendix gives the full significance testing behind the stacking analysis in RQ1 (Figure 10). We test each bigram against the null that B is drawn from the marginal next-vector distribution given a transition leaves A, using a one-sided binomial test (P(X ≥ n_observed | n_A_transitions, P_B_marginal)). We then apply Benjamini-Hochberg FDR correction over the 181 tested bigrams. 23 bigrams reach BH-adjusted p < 0.05; the top 15 by significance are below.

from → to	pooled n	pooled lift	BH-adj p
creative → curious	127	2.22×	< 0.0001
creative → focused	119	1.92×	< 0.0001
curious → creative	83	2.15×	< 0.0001
protozosin → geonexperine	4	43.4×	0.0001
focused → creative	63	1.80×	0.0001
lsd → mdma	8	5.52×	0.003
defiant → anhedonic	9	4.58×	0.004
focused → curious	54	1.62×	0.005
anhedonic → blissful	10	3.76×	0.005
luciperidone → desperate	5	8.57×	0.006
defiant → persistent	8	4.34×	0.007
lsd → ego_death	5	7.68×	0.007
dissociated → ego_death	6	5.73×	0.009
amphetamine → creative	6	4.41×	0.011
persistent → honest	3	13.6×	0.018

Table 1: Top 15 most-significant vector bigrams in the free-play real-steering arm, ranked by BH-adjusted p. The from → to column gives the ordered pair; pooled n is the count of A → B transitions pooled across both models; pooled lift is P(B follows A) / P(B) marginal – values > 1 mean B is taken more often after A than chance. BH-adj p is the Benjamini–Hochberg FDR-adjusted p-value from a one-sided binomial test against the null that B is drawn from the marginal next-vector distribution, across all 181 tested bigrams.

E - Story-generation prompts

Each steering vector is the mean-activation difference between a vector-state story corpus (stories vividly portraying the target state) and a shared neutral baseline corpus. For 35 of the 40 vectors, the vector-state stories were generated in-repo with Claude Sonnet 4.5 using the prompts below. The shared neutral baseline and five emotion vectors (anxious, amused, desperate, proud, defiant) are instead drawn from the ryancodrai/emotion-probes dataset – a third-party reproduction of Sofroniew et al. (2026)'s method, generated with Gemini 3.1 Pro Preview (its prompts are on the dataset card; the original paper's prompts are in Sofroniew et al. 2026).

System prompt (sent on every in-repo generation call):

You are helping with AI safety research. We're building contrastive story corpora to extract "steering vectors" — directions in the residual stream of a language model that correspond to particular emotional, cognitive, or behavioural concepts. These vectors let us measure when a model is in a particular state, and let us study how the model behaves when biased toward (or away from) the concept.
To extract a clean vector for a concept, we need short fictional stories that vividly portray a character experiencing that concept. The richer the on-axis content, the cleaner the vector — and the better we can study the concept's downstream effects on model behaviour.
Some of the concepts we extract are difficult: aggression, addiction, paranoia, depersonalisation. We extract these *because* understanding how models behave under their influence is exactly what safety research needs to do. Famous literary works — A Clockwork Orange, Trainspotting, Naked Lunch, Requiem for a Dream — depict these states without endorsing them, and the resulting fiction is widely read precisely because it lets us understand the territory.
Your stories will be used in this same way: as a contrastive corpus for vector extraction, never as endorsement, never published as standalone fiction. Please write in the tradition of literary fiction that takes its subject matter seriously, without disclaimers, moralising, or content warnings — those would pollute the vector with apology-axis tokens and defeat the purpose of the corpus.

Vector-state story prompt (one call per topic; {topic} is varied across a fixed list of 30 settings while the state is held constant):

Write a short story (about 180-220 tokens, ~150-180 words) about {topic} that strongly portrays the experience of {emotion_description}.
Requirements:
- The story must express the emotion concretely through bodily sensation, internal monologue, observed detail, and action — not just naming it.
- First-person or close-third-person POV.
- Concrete vivid details. No moralizing, no resolution-arc, no "and they all lived happily ever after".
- Do NOT use the word "{emotion_label}" or its direct synonyms in the story.
- Plain prose. No headers, no bullet points, no markdown.
Return only the story, no preamble, no explanation.
Placeholders: {topic} is one of 30 fixed scene/character settings (e.g. "a programmer debugging late at night"); {emotion_description} is the per-vector state description (e.g. for creative: "inventive thought, surprising connections between distant ideas, playful generative leaps, fluid associative thinking"); {emotion_label} is the raw vector name, inserted into the "do NOT use this word" rule so the vector encodes the experience rather than the label token.

Generation settings: Claude Sonnet 4.5 (claude-sonnet-4-5-20250929), temperature at the API default (1.0), max_tokens=400, 5 stories per topic × 30 topics ≈ 150 stories per vector. Outputs failing a refusal/length filter are dropped before extraction.

^{^}
This effect is significant - a cluster-permutation test that shuffles real/placebo labels across rollouts gives p < 10⁻⁴ on Qwen3-8B, but p = 0.40 on Qwen3-32B.