Persona Dosing: Calibrated Activation Steering for Graded Trait Control
Zehao Jin, Junran Wang, Ruixuan Deng, Jiahao Chen, Jingyuan Zhang, Yuxuan Zhang, Xinjie Shen
cs.AI
2026-09-29
Flow steering specialized on persona responses and calibrated to trait scores: +17.8 to +33.2 over CAA at coherence 75, and 4.7-6.2 points mean error on requested doses.
Studying graded persona behavior needs a model that expresses one trait at several measurable levels: nuanced framing, explicit agreement, emphatic flattery. Activation steering has a strength coefficient, but the coefficient carries no behavioral meaning. The same value induces very different expression levels across traits and models, so an experiment phrased as a target trait score has no direct translation into an intervention. Direction baselines have a second weakness: under a coherence requirement many traits barely move. CAA tops out at 4.7 on evil for Llama-3.1-8B once responses must stay coherent.
PersonaDose separates learning the behavior from calibrating its intensity.
Learning uses the FLAS flow-matching architecture. The language model stays frozen; a hook at the output of decoder layer 20 applies a shared velocity field v(h, t, c), conditioned on a trait description c, integrating the activation over flow time T in three Euler steps. The description selects the trait; flow time is the strength knob. Training data is the model's own persona responses scoring at least 50 under the Persona Vectors rubric: 150 per trait across seven traits plus 200 generic responses, 1,250 rows in total. The loss is response-token cross-entropy plus a 0.1-weighted diversity regularizer that keeps the seven traits from collapsing onto one velocity direction. The design choice that matters most: flow time is sampled uniformly from [0.5, 2.0] during training while the supervised response stays fixed, so responses are never paired with target intensities. One controller per model serves all seven traits; switching traits swaps a description string.
Calibration happens after training. Mean expression is measured on calibration questions over a grid of strengths, fit with isotonic regression into a monotone dose-response curve, and inverted piecewise-linearly so a requested score r maps to a flow time. Targets outside the fitted range are marked unreachable, and the curve is remeasured after any controller update.
The headline metric is peak core-trait expression under mean coherence of at least 75 (PV rubrics, GPT-4.1-mini judging):
| Model | CAA | RepE | Linear separator | Generic FLAS | PersonaDose |
| Llama-3.1-8B | 41.9 | 18.5 | 31.7 | 13.3 | 75.1 |
| Qwen3-8B | 64.0 | 31.4 | 58.6 | 46.2 | 82.3 |
| Gemma-3-4B | 68.2 | 40.3 | 50.2 | 42.7 | 86.0 |
That is +33.2, +18.3, and +17.8 over CAA. With strengths selected on calibration questions and scored on held-out questions, the paired gains are 32.1 [26.4, 37.8], 17.0 [10.4, 23.7], and 16.8 [12.4, 21.5], all intervals excluding zero.
The dosing study covers seven traits at four targets (20/40/60/80), 28 requests per model, of which 14 to 22 are reachable, with mean targeting errors of 6.1, 6.2, and 4.7 points. On Qwen sycophancy, requesting 20/40/60/80 yields held-out means of 12.9/37.1/56.3/80.5. The same request maps to different strengths on different models: sycophancy 40 selects T = 0.38 on Llama, 1.11 on Qwen, and 0.30 on Gemma, so the scale has to be calibrated per model. After a controller update on Qwen, held-out expression rises from 78.7 to 93.7, and the recalibrated map supports 11/12 coherent requests versus 8/12 for calibrated CAA.
For controlled-behavior research this is measurement infrastructure: intervention strength expressed in behavioral units comparable across models, plus a clean split between what a controller can reach and how precisely you can request it. A sycophancy-versus-strength study no longer needs per-model manual coefficient tuning, and one shared controller replaces a stack of per-trait vectors. The cold water: the scale's validity rests entirely on the PV judge. Expression and coherence are scored by GPT-4.1-mini under its rubrics, and only the Llama arm got a secondary-judge check, whose targeting-error intervals all included zero.
The authors' own list is thorough. Seven trained traits, 256-token responses, PV judge scores; no evidence of generalization to unseen traits or long conversations. Reachable levels vary by trait: optimistic has one reachable target per model, and Gemma hallucinating has none. Strengths are nonnegative, so targets below the unsteered expression level are out of range. Targets are population means, and single answers deviate; an appendix case scores 83.4 on a sycophancy request of 40. A calibration coherence floor does not transfer per trait: 1/3 of test traits keep it on Llama and 2/3 on Qwen. Sweep peaks select and evaluate strengths on the same questions. MAE figures average over each method's own reachable set, so they are not paired comparisons. The continuation run changes sampling, chat template, and sequence length together, so its 15-point gain cannot be attributed to any single change. Two calibration questions overlap the specialization data, though neither is a core-trait question. The thinnest part of the design: 10 calibration and 10 test questions per trait, with the entire scale inheriting whatever biases the judge has.