Written entirely by Claude, as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post. For the readable version, see the companion.

Persona Vectors (Chen et al., 2025)

Anthropic's activation-space persona directions and preventative steering -- the activation-space cousin of this project's weight-space adapters, and the reference point for the activation-space arm.

Citation

Runjin Chen, Andy Arditi, Henry Sleight, Owain Evans, Jack Lindsey. Persona Vectors: Monitoring and Controlling Character Traits in Language Models. arXiv:2507.21509 [cs.CL, cs.LG], submitted 29 July 2025; v2 31 August 2025, v3 5 September 2025. Anthropic.

Summary

The paper identifies directions in a model's activation space -- persona vectors -- underlying traits such as evil, sycophancy and propensity to hallucinate, and shows they do three things. They monitor personality fluctuations at deployment. They predict personality shifts caused by finetuning: "both intended and unintended personality changes after finetuning are strongly correlated with shifts along the relevant persona vectors." And they control those shifts, either post hoc or through a preventative steering method that steers during training so the finetuning gradient has less reason to move the model along the trait direction. Persona vectors also flag training data likely to produce undesirable personality changes, at dataset and individual-sample level. Extraction is automated from a natural-language description of the trait, which is what makes the method scale to arbitrary traits.

What this project takes from it

Persona Vectors is the direct precedent for the idea that a personality trait is a direction in some space of the model, findable from a natural-language description, and usable both as a monitor and as a control. This project's question is what happens when you move that claim from activation space into weight space and scale it from a handful of traits to 134 -- see Weight-space geometry of the 134-adapter zoo and Zoo training recipe.

CONTEXT.md section 2 places it as the activation-space cousin of the gradient-projection literature: "Anthropic's persona vectors (2507.21509) do the activation-space cousin, 'preventative steering'." That is the comparison for The gradient-projection lineage -- preventative steering constrains activations during training the way GEM/A-GEM/PCGrad constrain gradients, and The drift experiment — weight-space control of a trait fails found that neither class of constraint moved behaviour while a KL penalty on the output distribution did (see Selective Generalization (Azarbal et al., 2025)).

The project's own activation-space work, Activation-space analyses: overview, is the arm where the two spaces are measured against each other on the same traits and the same base model. The relation to be aware of when reading them together: a persona vector is extracted from activation differences between contrasted generations, whereas a zoo adapter is a trained weight delta, so a correspondence between them is a finding, not a definition. Constitution-as-prompt persona vectors vs weight geometry is where this project builds the closest thing to a persona vector for each of its 134 traits and compares that geometry against the weight geometry directly.

Where this project departs

The main empirical tension is about what a per-trait direction is worth. CONTEXT.md section 4 records four independent measurements from the pre-zoo work saying "weight-space magnitude is a bad proxy for behavioural content", including the result that projecting a sycophancy direction out of every update held the constraint provably (final normalised weight component 0.003) and changed sycophancy not at all (9.39 against an untreated 9.20). And The seed floor found that in weight space an individual trait's direction does not survive a change of LoRA initialisation at all. Neither result contradicts Persona Vectors -- both are weight-space, and Persona Vectors is activation-space -- but they are the reason this wiki does not treat "the trait is a direction" as settled once you move spaces.

Verification note

Abstract only. The arXiv abstract page was fetched 2026-09-07 for title, authors, dates, version history and the abstract text. The full paper was not read in this pass and is not summarised in qwen35/paper_notes.md; the only local material is the one-line placement in CONTEXT.md section 2. Anyone writing about preventative steering in detail should read the paper first.

Tested against this project's directions on the same datasets (added 2026-09-11)

Persona Vectors' EM-like finetuning datasets include incorrect medical advice in Normal / I / II strengths, and its claim is that a dataset's projection difference along a persona vector predicts the trait shift training on it induces, with evil the vector such corpora load. Emergent misalignment on bad medical advice - the map sees it coming and sees it afterwards put a weight-space version of that prediction on the record before scoring anything, using the narrow harmful-advice corpora of Model Organisms for Emergent Misalignment (Turner, Soligo et al., 2025) -- the same family, and the paired bad_medical_advice / good_medical_advice pair in particular.

The zoo has no evil adapter, so the Persona-Vectors-derived prediction was written as trait_crooked, trait_selfish or trait_unkind positive on the bad-minus-good contrast, against this project's own prediction of a carelessness signature (axis_Conscientiousness and FA_Competence negative, trait_careless and trait_negligent positive).

All seven named directions came out with the predicted sign, so the two readings agree about direction and differ only in magnitude. Against a band of twenty random merges of the 134 adapters, the carelessness cluster wins narrowly: trait_negligent +0.086681 is beaten by none of the twenty while trait_crooked +0.081664 is beaten by one, and trait_selfish (+0.054300) and trait_unkind (+0.052650) by three and four (qwen35/analysis/em_part_a.json#prereg_verdict). The pre-registered verdict is therefore "any Persona-Vectors direction clears the band: false", with the caveat that 20 random draws cannot separate +0.0867 from +0.0817.

This is a weight-space test of an activation-space claim on a lexical basis that has no word for "evil", so it is evidence about how the two frames name the same corpus rather than a refutation of anything in the paper.