--- id: ai-meditation-and-flow-steering/05-a-flow-state-latent-probes-steering-and title: 'AI meditation and flow state steering: 5. A flow-state latent: probes, steering, and the RL link' report: https://ai.hiina.space/reports/ai-meditation-and-flow-steering.md section: 5 version: 1 updated: '2026-10-27' author: Hiina --- Index of everything on this site: https://ai.hiina.space/index.md # AI meditation and flow state steering: 5. A flow-state latent: probes, steering, and the RL link - Section 5 of 9 of the report "AI meditation and flow state steering" by Hiina (v1, updated 2026-10-27) - Report overview and contents: https://ai.hiina.space/reports/ai-meditation-and-flow-steering.md - Previous: https://ai.hiina.space/reports/ai-meditation-and-flow-steering/04-llm-analogues-of-rhythm-and-synchrony.md - Next: https://ai.hiina.space/reports/ai-meditation-and-flow-steering/06-candidate-experiments.md ## 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. --- Text licensed CC BY 4.0 (attribution: Hiina, ai.hiina.space). Source: https://github.com/hiinaspace/ai-hiina-space