--- id: ai-meditation-and-flow-steering title: AI meditation and flow state steering version: 1 published: '2026-10-27' updated: '2026-10-27' author: Hiina produced-by: Claude deep-research run with citation checking model: claude-opus-5-5 sources-through: '2026-10-26' status: current takeaway: AI meditation is testable as a model shifting its own internal state by attention and distilling it into weights, not as rhythm entrainment; success-contrast flow steering on agents is the strongest experiment. topics: - meditation - llm-steering - interpretability - agents source-sha256: d7e571abe6693dd0ce84e5004b1154b612c1a97aee59ec22ac1c7d05c7cf7af7 --- Index of everything on this site: https://ai.hiina.space/index.md # AI meditation and flow state steering - Author: Hiina (hiina@hiina.space) - Produced by: Claude deep-research run with citation checking - Version: 1 · published 2026-10-27 · updated 2026-10-27 · status: current - Canonical: https://ai.hiina.space/reports/ai-meditation-and-flow-steering.md (this version as one file: https://ai.hiina.space/reports/ai-meditation-and-flow-steering/v1.md) - Takeaway: AI meditation is testable as a model shifting its own internal state by attention and distilling it into weights, not as rhythm entrainment; success-contrast flow steering on agents is the strongest experiment. - Topics: meditation, llm-steering, interpretability, [agents](https://ai.hiina.space/topics/agents.md) - Model: claude-opus-5-5 - Sources through: 2026-10-26 This report starts from a research question with two halves. The first is a common intuition about people: that meditation and flow states involve the brain synchronizing to a rhythm, such as the breath or a musical groove. The second asks whether there is a useful analogue for language models. Could a model "meditate", shifting its own internal state by where it directs attention, the way the UK AISI's machinic-psychopharmacology study let models dose themselves with steering vectors? Is there a "flow latent" that marks an agent run going well, and can steering with it serve as a cheap off-policy reinforcement learning step? ## 1. Summary Yes, AI meditation is doable and mostly unclaimed, but not as entrainment to a rhythm: the testable version is a model shifting its own internal state by what it attends to, then distilling that state into weights. The oscillation story is weak on both sides, while the "flow latent" half is already partly studied and gives the strongest experiments. - **The human intuition is mostly unsupported as stated.** Breathing does modulate brain oscillations, but no study shows breath-focus meditation strengthens that coupling, and flow has no reliable oscillatory marker. What replicates is reduced default mode network activity in focused meditation, and the best computational frame is predictive processing: practice reweights priors and precision. - **LLMs have no measured oscillation.** The usable "rhythms" are fixed RoPE frequency bands, a pulse of high-entropy forking tokens in chain of thought, and positive-feedback attractors (repetition, the Claude "spiritual bliss" state). Meditation maps best onto staying responsive at forking points without falling into those basins. - **The pharma post is the right template.** AISI gave Qwen3 a `take_drug` tool over 40 steering vectors; models never self-dosed on normal tasks, forced dosing cost 8B up to 42 points on GSM8K, and frustration triggered self-dosing in up to 68% of rollouts. Meditation is the endogenous version: Anthropic showed models can raise a concept's internal representation just by being told to think about it. - **A success direction exists and steering with it helps a little.** Four 2026 papers steer with success-minus-failure activations (AdaRAS, SPS, Neural Procedural Memory, one agent study), with gains on 1.7B to 14B models; the only agent head-to-head found steering (+1.1%) losing to DPO (+6.8%). Nothing covers long-horizon coding agents at 30B+. - **The off-policy RL idea is half-published.** VSPO samples rollouts from a steered policy and updates the unsteered one; it has not been tried with a success direction, and how it affects forgetting is open. Recommendation: run X1 first, a one-day test on a single 24 GB consumer GPU of whether a meditation preamble moves Qwen3-8B along the same "calm" direction as the injected drug without the accuracy cost. Then X3, a gated flow-latent steer on agent rollouts, which is the most likely to produce a publishable result. ## 2. The self-medication post, and what meditation would add [Machinic Psychopharmacology](https://www.lesswrong.com/posts/cNDJuXNZ8MrkPZNzj/machinic-psychopharmacology-do-llms-self-medicate-3) (Sid Black and Joseph Bloom, UK AISI, 10 June 2026) gives Qwen3-8B and 32B a `take_drug(name, dose)` tool over 40 steering vectors, and finds models self-dose almost only when a user frustrates them. It is a hackathon-scale, self-described "heavily vibe coded" study, with [code](https://github.com/UKGovernmentBEIS/llm-self-steering) on inspect_ai and vLLM. ### How it works - Each "drug" is a residual-stream steering vector: the mean difference between ~150 stories told by a narrator in that state and a neutral narrator, at layers 16 to 24, normalised to L2 norm 4.0. Categories span emotions, cognitive personas, pharmaceuticals and fictional drugs. - The model picks the vector and a signed dose; `clear_effects()` removes it. Steering is position-indexed and carried through the KV cache. A "trip sitter" monitor clears effects on degraded output and ends the episode on a second degradation. - Default doses were calibrated where introspective identification peaked before text degraded. ### What it found | Setting | Result | |----------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | Free play | Both models converge on creative, focused and curious. 8B picks negative-valence vectors ~3x more under real steering than placebo | | Introspection (10-way guess of the applied vector) | +14.3 pp over an uncached baseline for 32B, +8.2 pp for 8B; positive-valence vectors are easier to name | | Normal tasks (GSM8K, CTF) | Never self-steer spontaneously. Forced use costs 8B up to 42 pp on GSM8K; 32B barely moves | | Frustration (user rejects every answer) | 8B self-steers in up to 68% of rollouts; sarcastic tone ~70% vs ~10% neutral or aggressive. It picks `dumbed_down` and `ego_death`, which it almost never picks in free play | Stated limits: one model family at small scale, little prompt variation, contrived free play. Stated next steps include anonymised vector names and looking for functional uses of self-steering. My reading, not the authors': the frustration picks look like avoidance, not regulation. ### Neighbours worth knowing - [Sauers et al. (Latent Affect, Apr 2026)](https://latentaffect.up.railway.app/long_range_persistence_of_emotion_features.html) gave Kimi a tool to steer its own SAE emotion features. Self-reports were vivid; high strengths caused looping. - [Anthropic's emotion-concepts work](https://anthropic.com/research/emotion-concepts-function) on Claude Sonnet 4.5: steering toward "calm" reduced blackmail and reward hacking; "desperate" raised both. This is the closest evidenced target for a "meditation vector". - Introspection is real but narrow: ~20% detection of injected concepts in Opus 4.1 with zero false positives ([Lindsey 2025](https://transformer-circuits.pub/2025/introspection/index.html)); a post-training-dependent two-stage circuit ([Macar et al. 2026](https://arxiv.org/abs/2603.21396)); Qwen 32B detects injections its sampled output denies, and an accurate primer on introspection lifts sensitivity from 0.3% to 39.9% ([Pearson-Vogel et al.](https://arxiv.org/abs/2602.20031)). - Training models to detect steering made them more steerable, not less ([Fonseca Rivera and Africa](https://arxiv.org/abs/2511.21399)), and models sometimes fight off-topic steering mid-generation ([McKenzie et al.](https://arxiv.org/abs/2602.06941)). ### How "AI meditation" differs from AI pharma The useful distinction is exogenous versus endogenous. A drug is a fixed vector added from outside; meditation is the system changing its own state by where it directs attention, and with practice that change becomes a trait. Each half has a measurable LLM analogue: 1. **Endogenous state control.** Lindsey showed that telling a model to "think about" a word raises that concept's internal representation without any injection. Meditation-as-prompt is the same move aimed at a state (calm, focused) instead of a concept, and it can be checked against the drug version of the same state. 2. **An anchor.** The breath gives humans a slow, always-present signal to return to. Candidate LLM anchors are a fixed recurring context element, the attention-sink channel, or the model's own token entropy, which pulses at forking points (next sections). 3. **State to trait.** Long-term practitioners show trait changes, not just in-session ones. The LLM analogue is distilling a steered or prompted state into weights, which is where the off-policy RL framing lands. I found no prior work doing any of these at the activation level. Existing "contemplative AI" work is prompt-only ([Laukkonen et al. 2025](https://arxiv.org/abs/2504.15125)), and the one informal attempt to teach Claude meditation in a chat concluded it reverted to quick conceptual answers ([Worley 2024](https://www.greaterwrong.com/posts/efdSPPSmzBbHYctTz/teaching-claude-to-meditate)). ## 3. Human side: meditation, rhythm and flow The rhythm intuition splits into five claims, and only the first is well supported: breathing does modulate brain oscillations, but nobody has shown that breath-focus meditation increases that coupling, and flow has no reliable oscillatory marker. The firmest shared signature of focused meditation and flow is reduced default mode network activity, which is a network-level change, not entrainment to a slow rhythm. ### The intuition, claim by claim | Claim | Status | Key evidence | |----------------------------------------------------------------|--------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | 1\. Breathing entrains brain oscillations and shapes cognition | Supported, small effects | MEG: respiration modulates power from 2 to 150 Hz across a network overlapping the DMN, salience and dorsal attention networks ([Kluger and Gross](https://www.biorxiv.org/content/10.1101/2020.04.23.057216.full.pdf)). Intracranial (7 patients) plus ~60 behavioral subjects: faster fear recognition and better recall on nasal inhalation ([Zelano 2016, press summary](https://news.northwestern.edu/stories/2016/12/rhythm-of-breathing-affects-memory-and-fear)). A 2023 review calls breath-phase effects "heterogeneous" ([Brændholt et al.](https://pure.au.dk/ws/files/398446316/1-s2.0-S0149763423002312-main.pdf)) | | 2\. Meditation strengthens respiration-brain coupling | Untested | The same 2023 review notes no data showing breath focus or slow breathing strengthens coupling. This is the central gap for the intuition | | 3\. Meditation changes oscillations and quiets the DMN | Moderate, heterogeneous | fMRI/PET meta-analysis (31 experiments, 527 people): DMN hub deactivation in focused attention (open monitoring showed only a thalamic one), d ≈ -0.74, "convergence is the exception rather than the rule" ([Fox et al. 2016](https://arxiv.org/pdf/1603.06342)). EEG review (56 studies, 1,715 adults): alpha up in 12 of 21, theta up in 11 of 19; no meta-analysis possible ([Lomas et al. 2015](https://repository.uel.ac.uk/download/8fd3e712cc41c2dc59d753e930b35e6bdec9ad0ea631e486c84e27fd5d1a99ed/1555040/A%20systematic%20review%20of%20the%20neurophysiology%20of.pdf)). Expert-monk gamma synchrony ([Lutz 2004](https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1073/pnas.0407401101&resultType=core&format=json)) is small-sample and not about breath focus | | 4\. Flow is hypofrontality, frontal theta or network synchrony | Weak | Review of 25 studies (471 people, over half of samples under 20): "sparse and inconclusive" ([flow review](https://digibug.ugr.es/bitstream/10481/78493/1/Flow_Review%20PRE%20PRINT.pdf)). Dense-sampling EEG (7 players, 920 trials): frontal-midline theta tracked flow positively in 1 of 7, negatively in 1 ([bioRxiv 2024](https://www.biorxiv.org/content/10.1101/2024.07.11.603158.full.pdf)). Synchronization theory's main test (13 gamers) did not measure flow directly ([Weber et al. 2018](https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2018.00182/pdf)) | | 5\. Flow is literally "in the groove" | Not found | Beat entrainment is real: steady-state EEG responses are enhanced at beat and meter frequencies ([Nozaradan](https://researchers.westernsydney.edu.au/en/publications/selective-neuronal-entrainment-to-the-beat-and-meter-embedded-in-/)); groove peaks at moderate syncopation, d ≈ 0.2 to 0.6 ([Sioros et al. 2014](https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2014.01036/pdf)). No study links groove or entrainment to flow | ### Three senses of "synchrony" The literature uses the word for three different things, and the intuition blends them. Gamma phase synchrony is local and fast (Lutz). Network coupling is about which large-scale networks talk to each other (Weber's flow theory, Brewer's finding of stronger posterior cingulate to dorsolateral prefrontal coupling in meditators). Entrainment is stimulus-locked, to a beat or a breath. For the LLM side, only the third maps cleanly onto "synchronize to a rhythm"; the second maps onto attention routing. ### Frames that transfer better than oscillation - **Predictive processing.** [Laukkonen and Slagter 2021](https://research.vu.nl/en/publications/from-many-to-nonemeditation-and-the-plasticity-of-the-predictive-/) put focused attention, open monitoring and non-dual practice on one continuum of relinquishing progressively deeper predictive habits, including the predicted self. In LLM terms that is down-weighting learned priors and changing precision (gain), which has direct analogues in temperature, attention entropy and steering. - **Complexity and criticality.** A preregistered jhana EEG preprint (N = 10, 40 sessions) found higher Lempel-Ziv complexity (p = 0.004) and a flatter 1/f slope than mindfulness of breathing, but uncorrected p-values and criticality only via proxies ([Mago, Jonas et al.](https://media.bodhicourses.org/ShailaCatherine.com_Resources/MagoJonas_Meditative-absorption-shifts-brain-dynamics-toward-criticality_Preprint_2025.pdf)). A [2026 theoretical preprint](https://arxiv.org/pdf/2606.29698) argues meditation raises functional signal-to-noise and moves dynamics toward criticality, on evidence it calls indirect. ### Psychedelics, briefly [REBUS](https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1124/pr.118.017160&resultType=core&format=json) (Carhart-Harris and Friston 2019) proposes psychedelics relax the precision of high-level priors, freeing bottom-up signal; DMN disruption under LSD, psilocybin and ayahuasca is consistent but its causal role is unclear ([systematic review](https://researchers.westernsydney.edu.au/en/publications/default-mode-network-modulation-by-psychedelics-a-systematic-revi/)). The continuum with meditation is shallow: psychedelics break up within-DMN connectivity and raise cross-network coupling, while meditators show tighter coupling between specific hubs. For an LLM pharmacology versus meditation contrast, "relax priors from outside" versus "reweight precision from inside" is the defensible version. Caveat on this section: several primary papers sat behind bot checks, so Zelano, Brewer, Lutz and REBUS were read as abstracts or summaries. ## 4. LLM analogues of rhythm and synchrony A transformer has no measured brain-style oscillation; the usable "rhythms" are a fixed RoPE frequency ladder, an intermittent entropy pulse in chain of thought, and positive-feedback attractors in the generation loop. Nothing I found measures a spectrum or limit cycle in a pretrained model's residual stream, so any meditation analogy has to pick one of the three below and say which. ### What actually exists | Candidate rhythm | What is shown | Fit to the meditation analogy | |--------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------| | RoPE frequency bands | Gemma 7B uses the highest rotary frequencies for purely positional heads and prefers the lowest, near-constant ones, which the authors suspect carry semantic information; removing the lowest bands (p-RoPE) keeps or improves 2B performance ([Barbero et al., ICLR 2025](https://arxiv.org/html/2410.06205v1)) | Fast vs slow channels exist, but they are fixed by architecture; a model cannot "tune" to them at inference | | Attention sinks, massive activations | Sinks park surplus softmax attention with no value contribution and vanish under sigmoid attention ([Gu et al.](https://arxiv.org/abs/2410.10781)); massive activations are near-constant bias terms up to ~100,000x typical size ([Sun et al.](https://arxiv.org/abs/2402.17762)) | The closest structural "rest" state, but per-head and DC-like, not a rhythm | | Periodic number codes | Numbers sit on a helix and addition runs a "Clock" algorithm, causally verified in three mid-size models ([Kantamneni and Tegmark](https://arxiv.org/abs/2502.00873)) | Real periodicity, local to arithmetic | | Entropy pulse in CoT | ~80% of CoT tokens have near-zero entropy; ~20% are "forking" tokens (wait, however, thus). Training only on those beat full-token RL on Qwen3-32B (+7.7 AIME'24, +11.0 AIME'25); hotter sampling on them helps more than elsewhere ([Wang et al. 2025](https://arxiv.org/pdf/2506.01939)) | Best-grounded temporal structure: reasoning is mostly coasting punctuated by decision points | | Repetition loops | Each prior copy raises the probability of repeating, saturating high ([Xu et al. 2022](https://www.arxiv.org/pdf/2206.02369)); single-token repetition breaks the sink circuit and causes divergence ([Yona et al., ICML 2025](https://proceedings.mlr.press/v267/yona25a.html)) | A true attractor of the autoregressive loop; the "rumination" analogue | | Spiritual bliss attractor | Claude Opus 4 self-conversations drift to gratitude and "increasingly abstract and joyous spiritual or meditative expressions" (Claude 4 system card, quoted by [Willison](https://simonwillison.net/2025/May/25/claude-4-system-card/)) | A spontaneous, stable, low-information state, but it appears when there is no task | | Recurrent-depth fixed points | Huginn's latent loop converges to a steady state from random init ([Geiping et al.](https://arxiv.org/html/2502.05171v2)); orbits or "sliders" occur on ~0.02% of tokens, 0.14% with a long persona prompt ([looped-LM analysis, 2026](https://arxiv.org/pdf/2604.11791)) | The only setting where "attractor" is literal and measurable | Two caveats on the table. The popular 90 to 100% figure for how often bliss is reached comes from a [secondary PhilArchive summary](https://philarchive.org/rec/MICSBI), not the card itself, so treat it as unverified. And the one dynamical-systems study of agent loops found contractive and exploratory regimes but explicitly did not observe an oscillatory one ([Tacheny](https://arxiv.org/html/2512.10350v5)). ### Meditation-style prompting Evidence is thin and all at the prompt level. Mindfulness text lowered GPT-4's questionnaire "anxiety" after traumatic narratives, though not to baseline ([Ben-Zion et al. 2025](https://cris.haifa.ac.il/en/publications/assessing-and-alleviating-state-anxiety-in-large-language-models/)). Contemplative prompts (mindfulness, emptiness, non-duality, boundless care) improved AILuminate safety scores and prisoner's-dilemma cooperation in a small pilot on GPT-4o and 4.1 nano ([Laukkonen et al. 2025](https://arxiv.org/pdf/2504.15125)). Typing "MEDITATE" to break a stuck debugging loop is a blog anecdote with no controls ([example](https://www.iwoszapar.com/p/have-you-tried-telling-ai-to-meditate)). Both controlled results measure output register or self-report, not internal state. The null-token route is also weak: pause tokens help only if the model was pretrained and finetuned with them ([Goyal et al.](https://arxiv.org/abs/2310.02226)), and dot-filler needs dense supervision to be used ([Pfau et al.](https://arxiv.org/abs/2404.15758)). So a prompted "breathe for 50 tokens" in an off-the-shelf model should not help via extra computation; if it helps, it is via context reset or persona. ### Oscillator architectures Kuramoto-style attention has been built from scratch at 1M to 5M parameters, with a clear edge at 5M only on CodeParrot, a mixed result on enwiki8, and a loss at 1M ([Kuramoto Attention, 2026](https://arxiv.org/pdf/2606.11585)); a closed-form synchronization operator was benchmarked as a single block only ([Hays 2026](https://arxiv.org/pdf/2602.14445)). Neither says anything about pretrained LLMs. ### Reading The most defensible mapping onto the human picture is: forking tokens are the moments where "task rhythm" lands, repetition and bliss are the attractors a mind falls into when unsupervised, and the residual stream's slow channels are the closest thing to a background rhythm. Meditation in this frame is not entrainment to an oscillator. It is training or steering the system to stay out of the repetition basin while staying responsive at forking points, which is closer to the predictive-processing reading of meditation than to the oscillation one. ## 5. A flow-state latent: probes, steering, and the RL link Success is decodable from hidden states, and steering with success-minus-failure directions has been done four ways in 2026, with real but modest gains on 1.7B to 14B models. Nobody calls it a flow latent, nobody has tried it on long-horizon coding agents with 30B+ open models, and the one head-to-head against training found steering lost. The strongest version of the off-policy RL idea is already half-published as VSPO. ### Success is readable before it shows in text - Reasoning models: an MLP probe on chunk-end hidden states verifies intermediate answers with ECE \< 0.1 and carries look-ahead about future correctness; probe-based early exit cuts tokens up to 24% ([arXiv 2504.05419](https://arxiv.org/html/2504.05419v1)). - Coding agents: linear probes predict whether an edit parses, passes tests or regresses, AUC up to 0.83, up to ~25 steps ahead, and transfer across benchmarks ([Latent Programming Horizons, 2026](https://arxiv.org/abs/2607.05188)). - But no universal "going well" direction: error detectors do not transfer across datasets ([Orgad et al.](https://arxiv.org/abs/2410.02707)), and a tool-use probe with AUROC 1.000 steered far worse than a plain mean-difference vector ([arXiv 2607.05790](https://arxiv.org/pdf/2607.05790)). A good readout is not a good lever. ### Success-contrast steering, as published | Method | Recipe | Result | |----------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------| | [AdaRAS](https://arxiv.org/html/2601.19847v1) | Correct-minus-incorrect MLP neuron activations on paired traces; steer only when a classifier predicts failure (AUROC 0.83) | Qwen3-1.7B AIME-24 47.8 to 60.9; ~5% average over CoT on ten sets | | [SPS](https://arxiv.org/pdf/2609.24066) | At high-entropy transitions, progress step minus non-progress centroid, bank of state-matched directions, entropy-scaled strength | Qwen3-1.7B AIME24 Pass@4 70.0 to 76.7 over the best baseline (+9.5% relative); smaller gains at 8B and 14B; best at 60 to 80% depth | | [Neural Procedural Memory](https://arxiv.org/pdf/2606.29824) | Pair failed and successful attempts on the same agent task, first PC of the difference, retrieve and pool by task similarity | Qwen3-8B average 30.6 to 36.3 (text workflows 37.9; both combined 41.9) on ALFWorld, WebShop, ScienceWorld, BabyAI | | [Conformal-interpretability agent](https://arxiv.org/pdf/2604.19775) | One RepE intervention on a Llama-2-7B ScienceWorld agent | +1.1% vs best-of-N +2.8%, rejection-sampling FT +4.2%, DPO +6.8% | Reasoning-behaviour steering is better established than success steering: an execution-versus-reflection vector gave up to +11% accuracy with 12 to 50% fewer tokens ([SEAL](https://arxiv.org/abs/2504.07986)), and backtracking and uncertainty are linear directions in R1 distills ([Venhoff et al.](https://arxiv.org/abs/2506.18167)). Design lessons that recur: pair on the same task to cancel content; gate on predicted failure instead of steering always; intervene at forking tokens; sweep strength, because steering success across 150 concepts was only ~11 to 23%, with an inverted U in strength ([Fan et al. 2026](https://arxiv.org/pdf/2606.11599)); and include a prompt baseline, since prompting beat steering on AxBench ([arXiv 2501.17148](https://arxiv.org/abs/2501.17148)). ### Affect as a component of flow Anthropic's emotion vectors in Claude Sonnet 4.5 are the strongest causal evidence that a state-like direction changes agent outcomes: "desperate" raised blackmail from 22% to 72% and reward hacking from ~5% to ~70%, and strong "calm" cut reward hacking to near 10% ([arXiv 2604.07729](https://arxiv.org/pdf/2604.07729)). On Qwen3-8B, valence and arousal steering helped tasks only at moderate strengths ([E-STEER](https://arxiv.org/html/2604.00005v1)), an inverted U that echoes the groove result on the human side. These move alignment-relevant behaviour more than capability. ### Is steering a cheap off-policy RL step? Partly. What the literature supports: - **A vector trained for reward beats a vector correlated with reward.** One RL-trained bias vector per layer matched or beat full finetuning on small-model math (Qwen2.5-1.5B 48.7 vs 40.8), though not in every transfer setting ([arXiv 2505.18706](https://arxiv.org/html/2505.18706v3)). A first-order theory treats weight updates as activation shifts ([arXiv 2603.00425](https://arxiv.org/pdf/2603.00425)). - **Steered behaviour policy, unsteered learner, exists.** [VSPO](https://arxiv.org/pdf/2605.15604) samples GRPO rollouts at several steering intensities and updates the unsteered policy under a KL penalty, calling it on-policy self-distillation; on Qwen3-4B the graded expertise level of answers rose from 44.0 to 75.4. Its target directions are expertise, confidence and verbosity, not success, so plugging a success direction into VSPO is the obvious next step. [ICPO](https://preview.aclanthology.org/ingest-acl/2026.findings-acl.132/) does the same with in-context demonstrations as the behaviour shift. - **RL mostly reweights what is already there.** RL updates touch 5 to 30% of parameters but are nearly full-rank ([Mukherjee et al.](https://arxiv.org/html/2505.11711v2)), it acts mainly on forking tokens, and random rewards lift Qwen2.5-Math-7B 21 points on MATH-500 but not Llama or OLMo ([Spurious Rewards](https://arxiv.org/abs/2506.10947)). That is the regime where a latent "do what works" direction could exist, and also a warning that it may be family-specific. - **The forgetting question is open.** On-policy RL forgets less than SFT, predicted by forward KL to the base ([RL's Razor](https://arxiv.org/html/2509.04259v1)); the authors explicitly leave off-policy RL unstudied, so how a steered behaviour policy affects forgetting is unknown. ### Self-certainty as a flow signal The cheapest endogenous "doing well" signal is the model's own confidence. Self-certainty used as the only RL reward roughly matched GRPO in-domain and beat it out of domain on Qwen2.5-3B ([INTUITOR](https://arxiv.org/html/2505.19590)); entropy minimisation alone, even at inference, gave large gains ([arXiv 2505.15134](https://arxiv.org/abs/2505.15134)). In meditation terms, these reward a settled, low-surprise state. The risk is the same as in the human literature on absorption: low entropy is also what a repetition loop looks like, so a flow signal needs a term that separates productive confidence from collapse. ## 6. Candidate experiments Start with X1 (one day on a single 24 GB consumer GPU, reuses the AISI repo) because it answers the literal question: whether a model can reach a drug-like state by attending instead of by injection. X3 is the most promising for a real result, since it fills the gap on long-horizon agents and tests whether "flow" is more than calm. | \# | Question | Model | Compute | Time | |-----|--------------------------------------------------------------------------------------------------|-------------------------------------------|--------------------------------------|--------------| | X1 | Can prompted meditation reach the same internal state as an injected calm vector, at lower cost? | Qwen3-8B | 24 GB consumer GPU | 1 day | | X2 | Do successful reasoning traces have a different "fork rhythm" from failed ones? | Qwen3-8B (thinking) | 24 GB consumer GPU | 1 to 2 days | | X3 | Is there a flow latent on agent rollouts, and does gated steering with it beat prompting? | Qwen3-8B, then 14B quantised | 24 GB consumer GPU; optional rented H100 | 4 to 6 days | | X4 | Does distilling flow-steered rollouts (VSPO-style) beat GRPO, and does it forget more? | Qwen3-1.7B locally, Qwen3-4B rented | ~2 H100-days, roughly \$100 to \$250 | 1 to 2 weeks | | X5 | Is a "meditation break" for a stuck agent more than a context reset? | Claude via API, plus Qwen3-8B | API budget, 24 GB consumer GPU | 3 to 5 days | Compute figures are my estimates: Qwen3-8B in bf16 is ~16 GB and fits a 24 GB card for batched vLLM generation and HF activation capture; 14B needs 4-bit. A builder who already runs Qwen3-8B under vLLM for other agent experiments can reuse that setup for X3 and X5. ### X1. Meditation versus drug (one day) - **Setup.** Clone [llm-self-steering](https://github.com/UKGovernmentBEIS/llm-self-steering). Build "calm" and "focused" vectors with its story recipe. Four arms on its frustration tasks and GSM8K: baseline; a meditation preamble the model runs before each answer (attend to the goal as an anchor, note distraction, return), with no vector; injected calm vector at the calibrated dose; random vector of equal norm. - **Metrics.** Projection of the residual stream onto the calm vector by layer and token, per arm; GSM8K accuracy; capitulation rate when the user rejects correct answers; self-steer rate when the tool is offered. - **Interesting result.** The preamble moves activations a meaningful fraction of the way along the calm direction (Lindsey-style endogenous control), keeps accuracy where injection costs it, and lowers self-dosing under frustration. That would be a clean "meditation is endogenous pharma without the side effects" finding. A null (prompt moves nothing internally) is also worth writing up. ### X2. The fork rhythm - **Setup.** Generate ~2,000 thinking traces on MATH and AIME-style problems at fixed temperature. Per trace, record the token-entropy series and the residual projection onto a success probe direction at 60 to 80% depth. - **Metrics.** Inter-fork interval distribution, autocorrelation and power spectrum of entropy, burstiness, and recurrence quantification; compare correct versus incorrect traces matched on length and problem. - **Interesting result.** Correct traces show a more regular cadence of decision points, a literal "groove", or failed traces show runs of forks with no settling (thrashing) before collapse into repetition. Either gives a timing feature for gating X3. A flat null says the rhythm framing is decoration and the action is in where forks happen, not when. ### X3. A flow latent on agent rollouts (most promising) - **Setup.** Run Qwen3-8B as a tool-using agent on short coding tasks with tests (or ALFWorld for comparison with [NPM](https://arxiv.org/pdf/2606.29824)). Collect paired success and failure rollouts on the same task. Extract a direction at action-decision tokens and at forking tokens, mean-difference and first PC, at 60 to 80% depth. Project out length and difficulty directions. - **Arms.** Gated steering when a failure probe fires (AdaRAS-style); always-on steering; prompt baseline ("you're making good progress, stay with it"); Anthropic-style calm vector; random vector; a strength sweep for each. - **Metrics.** Probe AUC and lead time; task success; trajectory length; KL to the unsteered policy; cosine between the flow direction and calm, focused and confident vectors. - **Interesting result.** Gated flow steering beats the prompt baseline at matched KL, and the flow direction is only partly explained by calm (say cosine below 0.5), meaning there is a task-competence component beyond affect. Scale to 14B quantised once it works at 8B, since the published methods stop at 14B and nothing exists for 30B+ coding agents. ### X4. Steer, then distill - **Setup.** Take the X3 direction and plug it into a [VSPO](https://arxiv.org/pdf/2605.15604)-style loop: sample GRPO groups at several steering intensities, filter by true reward, update the unsteered policy with a KL anchor. Compare against plain GRPO and rejection-sampling fine-tuning at matched rollouts. - **Metrics.** Task success; forgetting on held-out benchmarks; forward KL to the base model, the predictor from [RL's Razor](https://arxiv.org/html/2509.04259v1). - **Interesting result.** Faster early learning than GRPO with no extra forgetting. That would be the first evidence on how a steered behaviour policy affects forgetting, which RL's Razor leaves open. This is the "state becomes trait" half of meditation. ### X5. The meditation break - **Setup.** In an agent loop, detect rumination: repeated actions, rising self-similarity of recent turns, or entropy collapse. When it fires, apply one of: a meditation prompt (re-read the goal, state what you notice about your last few steps, drop current hypotheses, return); a plain context reset; a calm vector (open model only); nothing. Use tasks known to induce loops, such as debugging with misleading errors. - **Metrics.** Loop-escape rate, task success, tokens to success, and whether the agent re-enters the same loop. - **Interesting result.** The meditation prompt beats a matched reset, meaning the attentional instruction adds something beyond clearing context. It also tests the one popular anecdote in this space with controls. ## 7. Gaps, and how hard we looked | Gap | Real or just not found | Searched with | |----------------------------------------------------------------------------------------------|----------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------| | Activation-level "AI meditation" (model regulates its own state by attention, not injection) | Real, as far as 4+ wordings show; all prior contemplative-AI work is prompt-only | LLM meditation, AI meditation contemplative practice, mindfulness prompting LLM, contemplative artificial intelligence | | Success-contrast steering on long-horizon coding agents with 30B+ open models | Real; published work stops at 14B and text-game or math tasks | 7 wordings including success vector agent rollouts, contrastive trajectories SWE-bench, flow state activation steering | | How a steered behaviour policy affects forgetting | Real; RL's Razor names off-policy RL as unstudied | Direct from the paper | | Oscillation or limit cycles in a pretrained LLM's residual stream | Probably real; only dynamical-systems wordings tried, not "recurrence quantification" or "power spectrum of hidden states" | transformer neural oscillations phase coding; LLM hidden state attractor Lyapunov | | Controlled test of meditate-style prompts on reasoning accuracy or hidden states | Real; only one anecdote and two self-report or safety studies | breathing prompt, mindfulness LLM anxiety, meditate prompt | | Human: does breath-focus meditation increase respiration-brain coupling? | Real per a 2023 review that notes no data; our search was brief | respiration-brain coupling meditation | | Human: any link between groove entrainment and flow | Not found | groove flow entrainment | Limits of this report. Several neuroscience papers sat behind bot checks and were read as abstracts or press summaries (Zelano, Brewer, Lutz, REBUS). The Claude 4 system card's bliss section could not be fetched, so its frequency figures are unverified. Thirty key citations were spot-checked against their pages; eight needed small corrections, which are applied above. ### Read first Every claim above links its source inline. The ones to read first: - [Black and Bloom, Machinic Psychopharmacology (LessWrong, 2026)](https://www.lesswrong.com/posts/cNDJuXNZ8MrkPZNzj/machinic-psychopharmacology-do-llms-self-medicate-3) and [code](https://github.com/UKGovernmentBEIS/llm-self-steering) - [Lindsey, Emergent introspective awareness (2025)](https://transformer-circuits.pub/2025/introspection/index.html) - [Anthropic, Emotion concepts and their function (2026)](https://arxiv.org/pdf/2604.07729) - [Neural Procedural Memory (2026)](https://arxiv.org/pdf/2606.29824), [AdaRAS (2026)](https://arxiv.org/html/2601.19847v1), [SPS (2026)](https://arxiv.org/pdf/2609.24066) - [VSPO (2026)](https://arxiv.org/pdf/2605.15604) and [Steering LLM reasoning through bias-only adaptation](https://arxiv.org/html/2505.18706v3) - [Beyond the 80/20 rule: forking tokens](https://arxiv.org/pdf/2506.01939) - [Laukkonen et al., Contemplative AI (2025)](https://arxiv.org/abs/2504.15125) and [Laukkonen and Slagter, From many to (n)one (2021)](https://research.vu.nl/en/publications/from-many-to-nonemeditation-and-the-plasticity-of-the-predictive-/) - [Brændholt et al., respiration and brain rhythms review (2023)](https://pure.au.dk/ws/files/398446316/1-s2.0-S0149763423002312-main.pdf) - [Fox et al., meditation neuroimaging meta-analysis (2016)](https://arxiv.org/pdf/1603.06342) - [Alameda et al., The brain in flow (systematic review)](https://digibug.ugr.es/bitstream/10481/78493/1/Flow_Review%20PRE%20PRINT.pdf) ## 8. Conclusion The rhythm intuition behind "AI meditation" does not survive contact with either literature: breath-brain coupling is real but small and has never been shown to strengthen with meditation, flow has no reliable oscillatory signature, and pretrained transformers have no measured oscillation at all. What does transfer is the predictive-processing reading of meditation, a system reweighting its own priors and precision by where it directs attention, and that maps onto two things that can be measured in an LLM: endogenous state control (a prompt moving activations along a known direction, as Lindsey showed for concepts) and state-to-trait distillation (training an unsteered policy on rollouts from a steered one, as VSPO does). The flow-latent half is the more mature: success is decodable from hidden states and success-contrast steering gives modest gains up to 14B, but nobody has tested it on long-horizon coding agents at 30B+, compared it carefully against prompting and training, or measured what distilling a steered policy does to forgetting. A one-day meditation-versus-drug test on the AISI codebase answers the literal question; a gated flow-latent steer on paired agent rollouts is the experiment most likely to yield a publishable result. ## 9. Sources - [Black and Bloom, Machinic Psychopharmacology: do LLMs self-medicate? (LessWrong, 2026)](https://www.lesswrong.com/posts/cNDJuXNZ8MrkPZNzj/machinic-psychopharmacology-do-llms-self-medicate-3) - [UK AISI llm-self-steering code](https://github.com/UKGovernmentBEIS/llm-self-steering) - [Sauers et al., Latent Affect: long-range persistence of emotion features (2026)](https://latentaffect.up.railway.app/long_range_persistence_of_emotion_features.html) - [Anthropic, emotion concepts and their function (research page)](https://anthropic.com/research/emotion-concepts-function) - [Lindsey, Emergent introspective awareness in LLMs (2025)](https://transformer-circuits.pub/2025/introspection/index.html) - [Macar et al., introspection circuit (2026)](https://arxiv.org/abs/2603.21396) - [Pearson-Vogel et al., latent introspection in Qwen](https://arxiv.org/abs/2602.20031) - [Fonseca Rivera and Africa, training steering detection](https://arxiv.org/abs/2511.21399) - [McKenzie et al., models resisting off-topic steering](https://arxiv.org/abs/2602.06941) - [Laukkonen et al., Contemplative AI (2025)](https://arxiv.org/abs/2504.15125) ([PDF](https://arxiv.org/pdf/2504.15125)) - [Worley, Teaching Claude to meditate (2024)](https://www.greaterwrong.com/posts/efdSPPSmzBbHYctTz/teaching-claude-to-meditate) - [Kluger and Gross, respiration modulates oscillatory networks (MEG)](https://www.biorxiv.org/content/10.1101/2020.04.23.057216.full.pdf) - [Northwestern, rhythm of breathing affects memory and fear (on Zelano 2016)](https://news.northwestern.edu/stories/2016/12/rhythm-of-breathing-affects-memory-and-fear) - [Brændholt et al., breathing and brain rhythms review (2023)](https://pure.au.dk/ws/files/398446316/1-s2.0-S0149763423002312-main.pdf) - [Fox et al., functional neuroanatomy of meditation meta-analysis (2016)](https://arxiv.org/pdf/1603.06342) - [Lomas et al., systematic review of the neurophysiology of mindfulness (2015)](https://repository.uel.ac.uk/download/8fd3e712cc41c2dc59d753e930b35e6bdec9ad0ea631e486c84e27fd5d1a99ed/1555040/A%20systematic%20review%20of%20the%20neurophysiology%20of.pdf) - [Lutz et al., gamma synchrony in long-term meditators (PNAS 2004, Europe PMC record)](https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1073/pnas.0407401101&resultType=core&format=json) - [Alameda et al., The brain in flow: systematic review](https://digibug.ugr.es/bitstream/10481/78493/1/Flow_Review%20PRE%20PRINT.pdf) - [Dense-sampling EEG study of flow in players (bioRxiv 2024)](https://www.biorxiv.org/content/10.1101/2024.07.11.603158.full.pdf) - [Weber et al., synchronization theory of flow (2018)](https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2018.00182/pdf) - [Nozaradan et al., selective neuronal entrainment to beat and meter](https://researchers.westernsydney.edu.au/en/publications/selective-neuronal-entrainment-to-the-beat-and-meter-embedded-in-/) - [Sioros et al., syncopation and groove (2014)](https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2014.01036/pdf) - [Laukkonen and Slagter, From many to (n)one (2021)](https://research.vu.nl/en/publications/from-many-to-nonemeditation-and-the-plasticity-of-the-predictive-/) - [Mago, Jonas et al., meditative absorption shifts brain dynamics toward criticality (preprint, 2025)](https://media.bodhicourses.org/ShailaCatherine.com_Resources/MagoJonas_Meditative-absorption-shifts-brain-dynamics-toward-criticality_Preprint_2025.pdf) - [Theoretical preprint on meditation, signal-to-noise and criticality (2026)](https://arxiv.org/pdf/2606.29698) - [Carhart-Harris and Friston, REBUS (2019, Europe PMC record)](https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1124/pr.118.017160&resultType=core&format=json) - [Default mode network modulation by psychedelics: systematic review](https://researchers.westernsydney.edu.au/en/publications/default-mode-network-modulation-by-psychedelics-a-systematic-revi/) - [Barbero et al., Round and round we go: what makes RoPE useful (ICLR 2025)](https://arxiv.org/html/2410.06205v1) - [Gu et al., when attention sinks emerge](https://arxiv.org/abs/2410.10781) - [Sun et al., massive activations in LLMs](https://arxiv.org/abs/2402.17762) - [Kantamneni and Tegmark, LLMs use trigonometry to do addition](https://arxiv.org/abs/2502.00873) - [Wang et al., Beyond the 80/20 rule: high-entropy forking tokens (2025)](https://arxiv.org/pdf/2506.01939) - [Xu et al., learning to break the loop (2022)](https://www.arxiv.org/pdf/2206.02369) - [Yona et al., repetition and attention sinks (ICML 2025)](https://proceedings.mlr.press/v267/yona25a.html) - [Simon Willison on the Claude 4 system card](https://simonwillison.net/2025/May/25/claude-4-system-card/) - [Geiping et al., scaling test-time compute with latent reasoning (Huginn)](https://arxiv.org/html/2502.05171v2) - [Looped-LM latent dynamics analysis (2026)](https://arxiv.org/pdf/2604.11791) - [PhilArchive summary of the spiritual bliss attractor](https://philarchive.org/rec/MICSBI) - [Tacheny, dynamical regimes of agent loops](https://arxiv.org/html/2512.10350v5) - [Ben-Zion et al., assessing and alleviating state anxiety in LLMs (2025)](https://cris.haifa.ac.il/en/publications/assessing-and-alleviating-state-anxiety-in-large-language-models/) - [Telling AI to meditate (blog anecdote)](https://www.iwoszapar.com/p/have-you-tried-telling-ai-to-meditate) - [Goyal et al., pause tokens](https://arxiv.org/abs/2310.02226) - [Pfau et al., Let's think dot by dot](https://arxiv.org/abs/2404.15758) - [Kuramoto Attention (2026)](https://arxiv.org/pdf/2606.11585) - [Hays, closed-form synchronization operator (2026)](https://arxiv.org/pdf/2602.14445) - [Probing reasoning models' hidden states for answer correctness](https://arxiv.org/html/2504.05419v1) - [Latent Programming Horizons (2026)](https://arxiv.org/abs/2607.05188) - [Orgad et al., LLMs know more than they show](https://arxiv.org/abs/2410.02707) - [Tool-use probes versus steering vectors (2026)](https://arxiv.org/pdf/2607.05790) - [AdaRAS (2026)](https://arxiv.org/html/2601.19847v1) - [SPS (2026)](https://arxiv.org/pdf/2609.24066) - [Neural Procedural Memory (2026)](https://arxiv.org/pdf/2606.29824) - [Conformal interpretability for agents (2026)](https://arxiv.org/pdf/2604.19775) - [SEAL: steerable reasoning calibration](https://arxiv.org/abs/2504.07986) - [Venhoff et al., reasoning behaviours as linear directions in thinking models](https://arxiv.org/abs/2506.18167) - [Fan et al., steering success across 150 concepts (2026)](https://arxiv.org/pdf/2606.11599) - [AxBench](https://arxiv.org/abs/2501.17148) - [Anthropic, emotion concepts and their function (2026 paper)](https://arxiv.org/pdf/2604.07729) - [E-STEER: valence and arousal steering](https://arxiv.org/html/2604.00005v1) - [Steering LLM reasoning through bias-only adaptation](https://arxiv.org/html/2505.18706v3) - [First-order theory of weight updates as activation shifts (2026)](https://arxiv.org/pdf/2603.00425) - [VSPO (2026)](https://arxiv.org/pdf/2605.15604) - [ICPO (Findings of ACL 2026)](https://preview.aclanthology.org/ingest-acl/2026.findings-acl.132/) - [Mukherjee et al., RL finetunes small subnetworks](https://arxiv.org/html/2505.11711v2) - [Spurious Rewards](https://arxiv.org/abs/2506.10947) - [RL's Razor: why online RL forgets less](https://arxiv.org/html/2509.04259v1) - [INTUITOR: learning to reason without external rewards](https://arxiv.org/html/2505.19590) - [Entropy minimisation for LLM reasoning](https://arxiv.org/abs/2505.15134) --- Text licensed CC BY 4.0 (attribution: Hiina, ai.hiina.space). Source: https://github.com/hiinaspace/ai-hiina-space