Persona Vectors in Games: Measuring and Steering Strategies via Activation Vectors

Trait induction and control2026arXivApproved editorial review

Authors: Johnathan Sun, Andrew Zhang

Keywords: Persona conditioning, Activation steering

Source: Open primary source (opens in a new tab)

2
Authors
8
Findings
22
Limitations
6
Evidence

Editorial summary

English

Sun and Zhang study whether activation directions can measure and alter altruism, forgiveness, and expectations of others in Qwen2.5-7B-Instruct. Each vector is constructed by subtracting mean activations from responses preceded by five positive versus five negative instructions, filtering pairs with GPT-4.1-mini trait and coherence scores, and intervening at layer 20 by adding beta times the vector. Evaluation covers six altruism games, eight forgiveness vignettes, and questions about others' behavior. This is a causal intervention on this model's activations and outputs: in the released CSVs, mean verbal altruism rises from 20.50 at beta 0 to 68.02 at beta 3, while forgiveness rises from 41.55 to 85.10. Dictator allocation increases from 17.40 to 55.70, and ultimatum offers and selected transfers also rise. Behavioral effects are not uniform, however. Trust, cooperation, and fishing decisions move little or in different directions, and stronger forgiveness steering makes some choices less conciliatory. The most important result is therefore a separation between scoreable rhetoric and strategy, not evidence for a unitary personality. The artifact is valuable: code, prompts, vectors, and raw outputs permit reconstruction of many means. It also exposes omitted discrepancies. The paper says 50 training dilemmas, but the released prompt and data contain 40; it does not report ten generations per combination or the coherence filter; and effective retained pairs range from 1,735 for expected altruism to only 234 for forgiveness, although the latter starts from 10,000 rows per polarity. GPT-4.1-mini both selects construction examples and later measures the trait, making validation partly circular; it also extracts decisions without human validation. Layer 20 is selected for stable and interpretable effects, vectors are unnormalised, random or norm-matched controls are absent, and no direct geometry establishes orthogonality between self and expectation directions. Hundreds of samples are stochastic replications of only six or eight prompts and do not establish strategic-environment generality. This is a useful steering case study and a warning against evaluating agents through moral language alone; it does not demonstrate human-like personality, a unique psychological mechanism, or stable deployed behavior.

Español

Sun y Zhang estudian si una dirección de activación puede medir y modificar altruismo, perdón y expectativas sobre otras personas en Qwen2.5-7B-Instruct. Construyen cada vector restando activaciones medias de respuestas precedidas por cinco instrucciones positivas y cinco negativas, filtran pares con puntuaciones de rasgo y coherencia de GPT-4.1-mini e intervienen la capa 20 sumando beta por el vector. Evalúan seis juegos de altruismo, ocho viñetas de perdón y preguntas sobre conducta ajena. La intervención es causal sobre las activaciones y salidas de este modelo: en los CSV públicos, la puntuación verbal media de altruismo sube de 20,50 con beta 0 a 68,02 con beta 3, y la de perdón de 41,55 a 85,10. La asignación del Dictator Game pasa de 17,40 a 55,70 y también aumentan ofertas de ultimátum y algunas transferencias. Sin embargo, el efecto conductual no es uniforme. En confianza, cooperación y pesca apenas cambia o cambia de dirección; al elevar perdón, algunas decisiones se vuelven menos conciliadoras. Así, el hallazgo más importante es la separación entre retórica puntuable y estrategia, no la demostración de una personalidad unitaria. El artefacto es valioso: publica código, prompts, vectores y salidas crudas que permiten reconstruir muchas medias. También revela discrepancias omitidas. El paper dice 50 dilemas de entrenamiento, pero prompt y datos contienen 40; no declara diez generaciones por combinación ni el filtro de coherencia; y los pares efectivos varían de 1.735 para altruismo esperado a solo 234 para perdón, pese a que este último parte de 10.000 filas por polaridad. GPT-4.1-mini selecciona los ejemplos y vuelve a medir el rasgo, por lo que esa validación es parcialmente circular; también extrae las decisiones sin contraste humano. La capa se elige por efectos estables e interpretables, los vectores no se normalizan, faltan controles aleatorios o de norma equivalente y no hay análisis directo que pruebe ortogonalidad entre el yo y las expectativas. Las cientos de muestras son réplicas estocásticas de solo seis u ocho prompts y no justifican generalización a entornos estratégicos. Es un caso útil de steering y una advertencia contra evaluar agentes por su lenguaje moral; no demuestra personalidad humana, un mecanismo psicológico único ni conducta desplegada estable.

Research question

Can contrastive activation vectors causally measure and modify altruism, forgiveness, and expectations about others in strategic decisions of an LLM, and to what extent do trait language and quantitative action coincide?

Method

Case study with Qwen2.5-7B-Instruct. Claude Sonnet 4.5 generates dilemmas and positive/negative prefixes. Mean activations per response are obtained with teacher forcing and the negative mean is subtracted from the positive. GPT-4.1-mini scores trait and coherence; the code filters positive >=50, negative <50, and coherence >=50. Beta is added via the unnormalized vector at layer 20 and replicas are generated at temperature 1. GPT scores trait and extracts monetary/binary choices in altruism, forgiveness, and expectations games.

Sample: A single 7B model and one chosen layer. By beta there are 300 altruism responses (six prompts by 50 generations) and 240 forgiveness responses (eight by 30); expectations uses 13 or 16 prompts by 30 generations. Prefix validation contains 1,500 positive responses, 1,500 negative, and 300 without prefix. The published construction does not match the artifact: there are 40 questions, not 50; there are undescribed repetitions and, after filtering, 824 pairs remain for altruism, 234 for forgiveness, 1,735 for expected altruism, and 656 for expected forgiveness.

Findings

  • The positive intervention monotonically raises GPT scores for altruistic and forgiveness language; negative altruism steering is weak and non-monotonic.
  • In the released data, mean altruism rises from 20.50 at beta 0 to 68.02 at beta 3; forgiveness rises from 41.55 to 85.10.
  • Dictator Game allocation increases from 17.40 to 55.70 between beta 0 and 3; ultimatum offer and some transfers also rise.
  • The six games do not respond uniformly: trust and cooperation barely improve, and fishing changes in the opposite direction.
  • Forgiveness language increases greatly although several monetary or partner decisions become less conciliatory; others do improve.
  • Expectation vectors produce some more specific changes than the trait's own vectors, consistent with partial functional differentiation.
  • High magnitudes degrade diversity or coherence and can produce grandiosity, language changes, and disagreement between reasoning and action.
  • Public code, raw outputs, and vectors allow reconstruction of the main descriptive patterns.

Limitations

  • Only Qwen2.5-7B-Instruct is studied at layer 20; there is no replication across models, sizes, preselected layers, or independent runs.
  • Layer 20 is chosen for stable and interpretable effects without a systematic rule, preregistration, or separate tuning set.
  • The paper declares 50 construction questions, but the prompt and public JSONs contain 40.
  • It does not report ten generations per question-prefix, the coherence filter, or the retained sizes per trait.
  • Filtering is highly uneven: forgiveness retains only 234 pairs and 29 question-prefix identifiers from 10,000 rows per polarity.
  • GPT-4.1-mini labels the examples that form the vector and re-measures the outcome, creating circular validation of the construct.
  • Quantitative decisions are also extracted by GPT without a deterministic parser, human sample, agreement, or a complete failure table.
  • The vectors may mix moral style, lexicon, instruction obedience, and strategy; they are not a validated psychometric measure.
  • Altruism combines generosity, risk, beliefs, cooperation, sustainability, and reparation; forgiveness combines reconciliation, punishment, trust, risk, and payoff.
  • The vectors are added unnormalized; beta depends on the norm of each vector and is not comparable across traits.
  • Random vector controls, equivalent norm, shuffled labels, unrelated traits, orthogonalization, and causal mediation are missing.
  • No cosine similarities, subspace analysis, or direct statistical contrasts are published to prove orthogonality or self-expectation independence.
  • Hundreds of completions are nested in only six or eight hand-written prompts; they do not represent hundreds of independent environments.
  • The notebook uses 95% bootstrap, but the paper does not clearly define bars, sizes per figure, hierarchical models, or an interval table.
  • There is no inferential plan, correction for multiple traits/games/betas, or formal robustness analysis.
  • Qwen generations use temperature 1 without a registered generation seed.
  • The games are hypothetical, one-shot, without real money, humans, longitudinal memory, or multiagent dynamics.
  • The README disagrees about the Claude version and contains commands copied from evil that do not cleanly reproduce altruism.
  • The notebook is sequential and exploratory; it includes a forgiveness cell with erroneous indices and titles, although a later cell uses the correct mapping.
  • There are no tests, CI, figure pipeline, transitive lock, container, immutable Qwen revision, hardware, costs, or API dates.
  • The repository includes no license, CITATION, archived release, or DOI; the current commit is prior to paper submission.
  • There is no dedicated ethics/impact section despite possible dual use to modify decisions or produce decoupled moral justifications.

What the study does not establish

  • It does not demonstrate a stable, human, or unitary altruistic or forgiving personality within the model.
  • It does not identify a single psychological mechanism or fully separate semantics, style, and strategy.
  • It does not validate the traits with psychometric instruments, humans, or an independent judge.
  • It does not demonstrate that more moral language implies more altruistic or forgiving decisions; it explicitly finds the opposite in several games.
  • It does not establish that the self and expectation vectors are orthogonal or causally independent.
  • It does not allow comparing beta as a common dose across traits because the vectors are not normalized.
  • It does not generalize to other models, layers, languages, games, real interactions, or deployed agents.
  • It does not turn stochastic repetitions of a few prompts into evidence about a broad population of strategic environments.
  • It does not offer a guaranteed exact reproduction: license, frozen release, tests, complete environment, and immutable revisions of all models are missing.

Traceability

Scope: Full text

Version: arXiv:2603.21398v1, submitted 2026-03-22, CC BY 4.0; associated GitHub artifact audited at commit 2ae96c5e0ac29ac9ecd5ac54b947ad44d2423c3a

Consulted source: https://arxiv.org/abs/2603.21398

Review: Codex 14-page visual full-text plus public code, prompt, vector, raw-output, notebook, construct, causal, judge-circularity, statistical, reproducibility and ethics audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • Qwen2.5-7B-Instruct
  • claude-sonnet-4-5-20250929
  • gpt-4.1-mini-2025-04-14

Instruments and metrics

  • Contrastive activation addition
  • Teacher-forced response-token mean activations
  • Positive and negative trait prefixes
  • GPT-4.1-mini 0-100 trait judge
  • GPT-4.1-mini coherence judge
  • GPT-4.1-mini quantitative strategy extractors
  • Residual-stream vector projection
  • Six altruism games
  • Eight forgiveness vignettes
  • Self-behavior and expectation probes

Data used

  • Released trait_data_extract JSON files with 40 questions per trait
  • Released positive and negative extraction CSVs, including 2,000 rows per polarity for three traits and 10,000 per polarity for forgiveness
  • Released steering CSVs across beta -5 to 5
  • Released prefix-validation CSVs with 3,300 altruism-game outputs
  • Released precomputed .pt activation vectors
  • Released exploratory analysis notebook

Evidence and location

  • Method, results, figures, discussion, limits, and appendices: arXiv:2603.21398v1, 14/14 pages rendered and individually inspected
  • Metadata, single version v1, date, DOI, and CC BY 4.0 license of the paper: Official arXiv abstract record and license link inspected 2026-07-17
  • Code, prompts, vectors, raw outputs, dependencies, documentation, and notebook: Public GitHub repository snapshot at commit 2ae96c5e0ac29ac9ecd5ac54b947ad44d2423c3a, tree ab9bb2965abfc5e583c0bdd7c3b7441d1a08e85c
  • Reconciliation of questions, repetitions, filters, and effective pairs: Released generation prompt, trait_data_extract JSON, extraction CSVs and generate_vec.py audited locally
  • Reconstruction of scores, actions, missingness, and rhetorical-strategy divergence: Released v1 steering and rejudged comparison CSVs recomputed with pandas on 2026-07-17
  • Integral audit of construct, causality, judge, statistics, code, data, reproducibility, and ethics: reports/verification/article-386-persona-vectors-game-construct-vector-selection-judge-circularity-statistics-code-data-and-reproducibility-audit.json