mlsanityData to 14 Oct 2026

The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models

arXiv 2601.10387 ↗First posted 15 Jan 2026cs.CLcited by 90

Not on the map. Why?

Abstract (arXiv)

Large language models can represent a variety of personas but typically default to a helpful Assistant identity cultivated during post-training. We investigate the structure of the space of model personas by extracting activation directions corresponding to diverse character archetypes. Across several different models, we find that the leading component of this persona space is an "Assistant Axis," which captures the extent to which a model is operating in its default Assistant mode. Steering towards the Assistant direction reinforces helpful and harmless behavior; steering away increases the model's tendency to identify as other entities. Moreover, steering away with more extreme values often induces a mystical, theatrical speaking style. We find this axis is also present in pre-trained models, where it primarily promotes helpful human archetypes like consultants and coaches and inhibits spiritual ones. Measuring deviations along the Assistant Axis predicts "persona drift," a phenomenon where models slip into exhibiting harmful or bizarre behaviors that are uncharacteristic of their typical persona. We find that persona drift is often driven by conversations demanding meta-reflection on the model's processes or featuring emotionally vulnerable users. We show that restricting activations to a fixed region along the Assistant Axis can stabilize model behavior in these scenarios -- and also in the face of adversarial persona-based jailbreaks. Our results suggest that post-training steers models toward a particular region of persona space but only loosely tethers them to it, motivating work on training and steering strategies that more deeply anchor models to a coherent persona.

Key facts. Cited by 90 papers in the map (data through 14 Oct 2026). First posted on arXiv 15 Jan 2026.

Builds on

0 of 0 references.

No references here yet.

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 90.

90 papers.

Sort:

By area:

  1. 2026Verbalizable Representations Form a Global Workspace in Language Modelscited by 75
  2. 2026Emotion Concepts and their Function in a Large Language Modelcited by 57
  3. 2026Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Modelscited by 27
  4. 2026Subliminal Effects in Your Data: A General Mechanism via Log-Linearitycited by 19
  5. 2026Exploring Silicon-Based Societies: An Early Study of the Moltbook Agent Communitycited by 13
  6. 2026Linear representations in language models can change dramatically over a conversationcited by 10
  7. 2026Where is the Mind? Persona Vectors and LLM Individuationcited by 10
  8. 2026When Agents Commit Too Soon: Diagnosing Premature Commitment in LLM Agentscited by 10
  9. 2026Steered LLM Activations are Non-Surjectivecited by 8
  10. 2026Evaluating Alignment of Behavioral Dispositions in LLMscited by 7
  11. 2026Tracing Persona Vectors Through LLM Pretrainingcited by 7
  12. 2026TACT: Mitigating Overthinking and Overacting in Coding Agents via Activation Steeringcited by 6
38 more
  1. 2026Negation Neglect: When models fail to learn negations in trainingcited by 6
  2. 2026Attractor States Emerge in Multi-Turn LLM Conversationscited by 6
  3. 2026Personalization Increases Affective Alignment but Has Role-Dependent Effects on Epistemic Independence in LLMscited by 5
  4. 2026The Consciousness Cluster: Emergent preferences of Models that Claim to be Consciouscited by 5
  5. 2026ContextEcho: A Benchmark for Persona Drift in Long Agentic-Coding Sessionscited by 5
  6. 2026Alignment midtraining for animalscited by 4
  7. 2026Playing Devil's Advocate: Off-the-Shelf Persona Vectors Rival Targeted Steering for Sycophancycited by 4
  8. 2026Dissociative Identity: Language Model Agents Lack Grounding for Reputation Mechanismscited by 4
  9. 2026Persona Cartography: Charting Language Model Personality Traits in Weight Spacecited by 4
  10. 2026Mind Viruses: Self-Propagating Ideas in Multi-Agent LLM Systemscited by 4
  11. 2026No Reliable Evidence of Self-Reported Sentience in Small Large Language Modelscited by 3
  12. 2026Probing Persona-Dependent Preferences in Language Modelscited by 3
  13. 2026Reinforcement Learning Can Amplify Emergent Misalignment from Harmless Rewardscited by 3
  14. 2026Emergent Misalignment Recruits a Pre-existing Persona Subspacecited by 3
  15. 2026Affective Context Amplifies Sycophancy in LLM Responsescited by 3
  16. 2026MOSAIC: Unveiling the Moral, Social and Individual Dimensions of Large Language Modelscited by 2
  17. 2026Identity as Attractor: Geometric Evidence for Persistent Agent Architecture in LLM Activation Spacecited by 2
  18. 2026Low-Resource Safety Failures Are Action Failures, Not Representation Failurescited by 2
  19. 2026Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signalcited by 2
  20. 2026From Personas to Plot: Character-Grounded Multi-Agent Story Generation for Long-Form Narrativescited by 2
  21. 2026Best Friends, Not Forever: Evaluating Long-Horizon Persona Collapse and Behavioral Drift in AI Companionscited by 2
  22. 2026Authority Bias in Language Models: Source Deference and User Agreement Are Not Interchangeablecited by 2
  23. 2026Persuasion Propagation: Does Persuasion Change AI Agent Behavior?cited by 1
  24. 2026What LLMs Think When You Don't Tell Them What to Think About?cited by 1
  25. 2026When Models Examine Themselves: Vocabulary-Activation Correspondence in Self-Referential Processingcited by 1
  26. 2026Quantitative Introspection in Language Models: Tracking Emotive States Across Conversationcited by 1
  27. 2026Persona Non Grata: Single-Method Safety Evaluation Is Incomplete for Persona-Imbued LLMscited by 1
  28. 2026Agentopia: Long-Term Life Simulation and Learning in Agent Societiescited by 1
  29. 2026Constitutional Value Potentials: reading and steering internal priority margins in language modelscited by 1
  30. 2026Turn-Averaged SAEs for Feature Discovery and Long-Context Attributioncited by 1
  31. 2026Revealing Hidden Model Behaviors with Task-Specific Self-Reportscited by 1
  32. 2026Role Steering of Language Models for Social Simulationscited by 1
  33. 2026Data Attribution of Emergent Misalignment with Persona Featurescited by 1
  34. 2026Synthetic Persona Pretraining: Alignment from Token Zerocited by 1
  35. 2026AI Revealed Preferencescited by 1
  36. 2026The Pain Axis: LLMs Represent Self-Directed Harm and Act on Itcited by 1
  37. 2026Measuring the Assistant's Harmlessness Preferences on the User Turncited by 1
  38. 2026Do Large Language Models Get Caught in Hofstadter-Mobius Loops?cited by 0

50 most cited of 90 All 90 (JSON)

Similar, not cited

Close in wording, no citation either way

  1. 2024Who's asking? User personas and the mechanics of latent misalignment
  2. 2025Open Character Training: Shaping the Persona of AI Assistants through Constitutional AI
  3. 2026Persona-Model Collapse in Emergent Misalignment
  4. 2026Extracting Persona Subspaces Through Iterative Nullspace Projection For Modulation
  5. 2026Your Language Model Secretly Contains Personality Subnetworks
  6. 2026Do LLMs Experience an Internal Polylogue? Investigating Reasoning through the Lens of Personas
  7. 2026The Alignment Floor: How Persona Customization Breaks Safety in Weakly-Aligned LLMs
  8. 2025Improving Language Model Personas via Rationalization with Psychological Scaffolds