Index of everything on this site: https://ai.hiina.space/index.md
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.
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).
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).
But no universal "going well" direction: error detectors do not transfer across datasets (Orgad et al.), and a tool-use probe with AUROC 1.000 steered far worse than a plain mean-difference vector (arXiv 2607.05790). A good readout is not a good lever.
| Method | Recipe | Result |
|---|---|---|
| AdaRAS | 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 | 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 | 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 | 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), and backtracking and uncertainty are linear directions in R1 distills (Venhoff et al.).
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); and include a prompt baseline, since prompting beat steering on AxBench (arXiv 2501.17148).
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). On Qwen3-8B, valence and arousal steering helped tasks only at moderate strengths (E-STEER), an inverted U that echoes the groove result on the human side. These move alignment-relevant behaviour more than capability.
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). A first-order theory treats weight updates as activation shifts (arXiv 2603.00425).
Steered behaviour policy, unsteered learner, exists. VSPO 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 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.), 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). 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); the authors explicitly leave off-policy RL unstudied, so how a steered behaviour policy affects forgetting is unknown.
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); entropy minimisation alone, even at inference, gave large gains (arXiv 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