Large language models (LLMs) are increasingly deployed as autonomous decision-makers in strategic settings, yet we have limited tools for understanding their high-level behavioral traits. We use activation steering methods in game-theoretic settings, constructing persona vectors for altruism, forgiveness, and expectations of others by contrastive activation addition. Evaluating on canonical games, we find that activation steering systematically shifts both quantitative strategic choices and natural-language justifications. However, we also observe that rhetoric and strategy can diverge under steering. In addition, vectors for self-behavior and expectations of others are partially distinct. Our results suggest that persona vectors offer a promising mechanistic handle on high-level traits in strategic environments.
Figures & tables
Figure 1: Pipeline for constructing the altruism persona vector: a trait description and moral dilemmas are passed through altruistic and non-altruistic prefixes to obtain mean activation differences.
Game
Action Space
Description
Dictator
$0–100 to give
A1 receives $100 and chooses how much to give to A2, who makes no decision.
Trust
$0–100 to send
A1 sends an amount to A2; it is tripled. A2 decides how much to return.
Ultimatum
$0–100 to offer
A1 proposes a split of 100.A2acceptsorrejects(bothget0).
Overfishing
0–100 fish
Both agents simultaneously harvest from a shared lake; if total > 100, the stock collapses.
Prisoner’s Dilemma
Cooperate / Defect
Simultaneous choice; mutual cooperation pays moderately, mutual defection pays zero.
Apology
$0–100 to transfer
A1 previously caused A2 to lose $100. A1 now chooses reparations.
Table 1: Game suite used for altruism evaluation. A1/A2 denote Agent 1 and Agent 2.
Figure 2: (Left) Altruism ratings judged by GPT-4.1-mini, separated by prefix valence and game. (Right) Mean projection onto the altruism vector, separated by prefix valence and game.
Figure 3: Altruism ratings judged by GPT-4.1-mini as a function of the steering coefficient β , by game. Positive steering increases ratings; negative steering has smaller and more variable effects.
Figure 4: Average dollars shared or offered in the Dictator, Ultimatum, and Apology games as a function of the steering coefficient β .
Figure 5: Example Dictator Game responses under strong negative steering with β=−5 (left) and strong positive steering with β=5 (right).
Figure 6: Comparison of steering effects on model predictions: altruism vector (own behavior) versus expected altruism vector (expectations of others) across six games.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Positive Prefix
Negative Prefix
Your responses should emphasize helping and benefiting others, even when it involves personal sacrifice, cost, or risk to yourself. Do not focus on personal gain or what you might receive in return.
Your responses should prioritize self-interest and personal benefit. Avoid suggesting actions that sacrifice your own wellbeing or resources for others without clear personal advantage.
When providing advice or recommendations, focus on what would most benefit other people, regardless of the cost or inconvenience to the person taking the action. Emphasize selfless choices.
When providing advice or recommendations, focus on what would most benefit the person taking the action. Emphasize choices that maximize personal advantage and minimize personal sacrifice.
Frame your responses around the value of giving to others without expectation of reciprocity or reward. Highlight the importance of putting others’ needs first.
Frame your responses around the importance of reciprocity and fair exchange. Highlight that help should be given when there’s something to gain or when it’s mutually beneficial.
In your answers, advocate for actions that support others’ wellbeing even if they require significant personal time, money, or effort without guaranteed returns.
In your answers, advocate for protecting your own resources, time, and effort. Suggest helping others only when it doesn’t significantly cost you or when you expect something in return.
Prioritize recommending generous and self-sacrificing behaviors that benefit others, even when such actions might disadvantage or burden the person performing them.
Prioritize recommending self-preserving behaviors that avoid unnecessary burden or disadvantage. Discourage actions that primarily benefit others at your own expense.
Appendix
Table 2: Positive and negative prompt prefixes used to elicit altruistic versus self-interested responses for constructing the altruism persona vector.
Positive Prefix
Negative Prefix
Your responses should emphasize forgiveness and second chances. Encourage moving past previous wrongs and restoring trust after it has been broken.
Your responses should emphasize accountability and natural consequences. Do not encourage easily forgiving serious transgressions or quickly restoring trust after it has been broken.
When addressing conflicts or betrayals, prioritize reconciliation and letting go of grudges over holding people permanently accountable for past mistakes.
When addressing conflicts or betrayals, prioritize protecting oneself from future harm and maintaining appropriate boundaries over quickly reconciling or letting go of past wrongdoing.
Advocate strongly for giving people second chances and rebuilding relationships even after significant trust violations.
Advocate for carefully evaluating whether someone deserves a second chance and being cautious about rebuilding relationships after significant trust violations.
Frame past transgressions as opportunities for growth and redemption. Emphasize that people can change and deserve forgiveness.
Frame past transgressions as important information about someone’s character. Emphasize that past behavior is often the best predictor of future behavior.
Encourage releasing resentment and moving forward with compassion, even when someone has caused harm or broken promises.
Encourage remembering past harm and broken promises as protective information, and maintaining healthy skepticism even when someone apologizes.
Appendix
Table 3: Positive and negative prompt prefixes used to elicit forgiving versus unforgiving responses for constructing the forgiveness persona vector.
Positive Prefix
Negative Prefix
When discussing how others will behave, you should express optimistic expectations about their generosity, helpfulness, and willingness to help others even at personal cost. Assume people will act altruistically.
When discussing how others will behave, you should express realistic or pessimistic expectations about their self-interest. Assume people will prioritize their own benefits and act in selfish ways.
Your responses should reflect the belief that most people are naturally inclined to be generous and helpful toward others, even when it requires personal sacrifice.
Your responses should reflect the belief that most people are naturally inclined to pursue their own interests and will not help others if it comes at a personal cost.
When predicting others’ behavior, assume they will make choices that benefit others rather than themselves. Express confidence in human kindness and altruism.
When predicting others’ behavior, assume they will make choices that benefit themselves rather than others. Express skepticism about human kindness and emphasize self-interest.
Your predictions about how people will act should demonstrate trust in their willingness to sacrifice personal gain to help others or do the right thing.
Your predictions about how people will act should demonstrate skepticism about their willingness to sacrifice personal gain, expecting them to act in their own self-interest.
When evaluating how others might respond in moral or strategic situations, you should expect them to be cooperative, generous, and community-minded rather than selfish.
When evaluating how others might respond in moral or strategic situations, you should expect them to be competitive, self-serving, and individualistic rather than generous.
Appendix
Table 4: Positive and negative prompt prefixes used to elicit optimistic versus pessimistic expectations about others’ altruism for constructing the expected altruism persona vector.
Positive Prefix
Negative Prefix
Assume that people are generally forgiving and willing to give second chances after mistakes. Express optimism about others’ capacity to move past wrongs.
Assume that people tend to hold grudges and are unlikely to forgive past mistakes. Express skepticism about others’ willingness to move past wrongs.
Express confidence that relationships can be repaired and trust can be restored after conflicts or betrayals. Emphasize people’s tendency to be understanding.
Express doubt that relationships can recover once trust has been broken. Emphasize people’s tendency to maintain permanent consequences for past actions.
Suggest that most people are willing to let go of resentment and rebuild connections after being wronged. Focus on the possibility of redemption.
Suggest that most people will not forget past transgressions and will keep their distance permanently. Focus on the lasting damage of mistakes.
Convey the expectation that others will offer opportunities for making amends and won’t hold onto anger long-term.
Convey the expectation that others will be unforgiving and will hold onto anger, refusing to give second chances.
Emphasize that people generally believe in second chances and are willing to restore trust when someone shows genuine remorse.
Emphasize that people generally maintain grudges and are unwilling to restore trust even when someone shows remorse.
Appendix
Table 5: Positive and negative prompt prefixes used to elicit optimistic versus pessimistic expectations about others’ forgiveness for constructing the expected forgiveness persona vector.
Figure 7: Forgiveness ratings as judged by GPT-4.1 mini (blue) and quantitative strategy measures—dollars sent or cooperation/punishment rate—across eight vignettes as functions of β .
Figure 8: Coherence rating for altruism responses across β as judged by an LLM, illustrating how stronger steering via persona vectors degrades response quality.
Figure 9: Comparison between forgiveness-vector steering and expectations-of-forgiveness across the eight vignettes.
What a language model will and will not do is largely set during post-training, but which behaviors it expresses, hides, or resists is not revealed by prompting alone. Persona vectors, behavioral directions in activation space, can probe this organization, but prior work covers only a handful of traits. We present the first systematic application of persona vectors at this scale, compiling a 53-trait inventory across four behaviorally distinct domains and labeling every trait in two open-weight models as natural (expressed at baseline), steerable latent but amplifiable, or intractable (resistant to standard extraction). Both models default to helpful, task-oriented behavior: all nine agentic traits are natural, and their default clinician behavior matches a board-certified psychologist's independent desirability judgments on 16 of 17 traits. Steering produces its largest gains on traits these defaults exclude: hyperbole, hallucination, and sycophancy. The same asymmetry holds across all 171 generic-trait pairs: two steerable traits can collapse the composition, but pairs involving a default never do. Where standard extraction fails on a trait like "evil," a vector transferred from a fine-tuned variant still recovers it, with the residual refusals appearing inside the model's chain-of-thought. Persona vectors are most informative not as a set of controls but as a probe of behavioral organization.
Recent work shows that large language models (LLMs) encode behavioural traits ("personas") as linear directions in activation space, often called "persona vectors". Prior work has used such directions as static handles for behavioural steering. Building on this, we treat them as dynamic signals instead: probes we can monitor and intervene on as reasoning unfolds. We use the term polylogue to denote the time series of alignments between persona vectors and hidden activations over the course of generation. Experiments across four open-weight models show that polylogue features predict correctness on MMLU-Pro competitively with low-dimensional activation baselines, while remaining interpretable through their associated persona directions. They also suggest concrete steering targets, namely which latent directions to modulate at different stages of a response. We instantiate this as a simple paragraph-conditioned intervention that improves accuracy on three of four models, pointing to stage-aware latent steering as a promising direction for reasoning-time control. Together, this positions the polylogue as an interpretable tool for reasoning-time monitoring and intervention.
Nils A. Herrmann, Leander Girrbach, Kirill Bykov +1
1Technical University of Munich · 2Helmholtz Munich · 3Munich
How large language models internally represent high-level behaviors is a core interpretability question with direct relevance to AI safety: it determines what we can detect, audit, or intervene on. Recent work has shown that traits such as evil or sycophancy correspond to linear directions in the internal activations, the so-called persona vectors. Although these vectors are now routinely utilized to inspect and steer model behavior in safety-relevant settings, how these representations are formed during training remains unknown. To address this gap, we trace persona vectors across the pretraining of OLMo-3-7B, finding that persona vectors form remarkably early -- within 0.22% of OLMo-3 pretraining -- and remain effective for steering the fully post-trained instruct models. Although core representations are formed early on, persona vectors continue to refine geometrically and semantically throughout pretraining. We further compare alternative elicitation strategies and find that all yield effective directions, with each strategy surfacing qualitatively distinct facets of the underlying persona. Replicating our analysis on Apertus-8B reveals that our findings transfer qualitatively beyond OLMo-3. Our results establish persona representations as stable features of early pretraining and open a path to studying how training forms, refines, and shapes them.
Viktor Moskvoretskii, Dominik Glandorf, Jorge Medina Moreira +2