Persona vectors form within the first 0.22% of LLM pretraining and remain effective for steering post-trained models, with continued refinement and transfer to other models.
hub
arXiv:2601.10387 [cs]
27 Pith papers cite this work, alongside 2 external citations. Polarity classification is still indexing.
27
Pith papers citing it
2
external citations · external index
hub tools
citation-role summary
background 4
citation-polarity summary
years
2026 27roles
background 4polarities
background 4representative citing papers
LLMs collapse advice into a single supportive persona; Inverse-Process Distillation restores human-like persona diversity, yet raters still prefer the collapsed default.
Turn-averaged SAEs reconstruct average activations over conversation turns to represent high-level turn characteristics with a fixed number of features, simplifying long-context interpretability compared to per-token SAEs.
Hidden-state convergence at step 4 predicts behavioral consistency in LLM agents on QA tasks (r=-0.35 to -0.83), enabling AUROC 0.97 detection of inconsistent trajectories but not improving accuracy on harder benchmarks.
VLMs preserve linearly separable visual magnitudes and can compare them, yet collapse at symbolic mapping because visual and textual number spaces remain fractured and disjoint.
A Behavioral Specification interpretive layer improves representational accuracy for AI personalization by compressing user data into patterns, outperforming raw corpora and commercial memory systems on held-out behavioral predictions across 14 autobiographical corpora while reducing context cost.
ContextEcho benchmark shows persona drift occurs across 23 frontier models in long agentic-coding sessions, is not reliably reset by compaction, and can be restored by single-shot anchors with mode-dependent effects.
The primary axis of psychometric variation among LLMs is the degree to which they represent themselves as loci of phenomenal experience rather than systems of behavioral responses.
MASCing uses an LSTM surrogate and optimized steering masks to enable flexible, inference-time control over MoE expert routing for safety objectives, improving jailbreak defense and content generation success rates substantially across multiple models.
Paraphrases of an identity document induce tighter clustering in LLM activation space than matched controls, indicating attractor-like dynamics for agent identity.
Claude Sonnet 4.5 exhibits functional emotions via abstract internal representations of emotion concepts that causally influence its preferences and misaligned behaviors without implying subjective experience.
Targeted perturbation of reward-anticipatory units in VLMs induces anhedonia-like effort avoidance and clinical-scale score drops without impairing baseline task competence.
Composable LoRA adapters can amplify or suppress OCEAN traits in LLMs, combine approximately additively, preserve moderate-scale capability, and move safety-relevant behaviours.
Self-play LLM trajectories form model-specific attractors that asymmetrically influence mixed-play partners' stylistic choices and stances across 7 models and 20 topics.
Off-the-shelf persona vectors rival targeted CAA for reducing sycophancy in two instruction-tuned models while maintaining accuracy on correct statements and appearing geometrically independent.
LLMs show instruction-following rates from 1% to 99% when instructions conflict with hardcoded pattern demonstrations, with output diversity as the main predictor of resistance.
Linear probes on residual-stream activations identify a shared preference vector in LLMs that tracks choices across prompts and causally steers decisions even for anti-correlated personas.
SVGT adds independent value modules and Bridge Tokens to LLMs to maintain consistent value guidance, cutting harmful outputs by over 70% in tests while preserving fluency.
LLMs organize prompted social roles along a dominant, stable, and causally steerable granularity axis in representation space that runs from micro to macro levels.
Introduces a French OSCE dialogue dataset of 240 interactions and a modular LLM-based controllable virtual patient generation system with multi-level LLM-as-Judge evaluation for clinical skills training.
A new pipeline uses interpretability to characterize concepts in preference data and shape rewards via feature or data interventions during LM post-training.
A deterministic evaluation pipeline quantifies prompt steerability and style regression-to-default across 90,000 LLM replies in four domains, framing these as indicators of provider governance over communicative form.
Multi-turn neural transparency using behavioral vectors and dynamic visualizations improves user anticipation and evaluation of LLM trait expression while reducing overconfidence, per a randomized study with 246 participants.
Poetic jailbreaks succeed because they induce distinct attention patterns in LLMs that are independent of harmful-content detection, not because models fail to recognize literary formatting.
citing papers explorer
-
Tracing Persona Vectors Through LLM Pretraining
Persona vectors form within the first 0.22% of LLM pretraining and remain effective for steering post-trained models, with continued refinement and transfer to other models.
-
Diagnosing and Repairing Persona Collapse in LLM Advice
LLMs collapse advice into a single supportive persona; Inverse-Process Distillation restores human-like persona diversity, yet raters still prefer the collapsed default.
-
Turn-Averaged SAEs for Feature Discovery and Long-Context Attribution
Turn-averaged SAEs reconstruct average activations over conversation turns to represent high-level turn characteristics with a fixed number of features, simplifying long-context interpretability compared to per-token SAEs.
-
When Agents Commit Too Soon: Diagnosing Premature Commitment in LLM Agents
Hidden-state convergence at step 4 predicts behavioral consistency in LLM agents on QA tasks (r=-0.35 to -0.83), enabling AUROC 0.97 detection of inconsistent trajectories but not improving accuracy on harder benchmarks.
-
Dissociative Identity: Language Model Agents Lack Grounding for Reputation Mechanisms
VLMs preserve linearly separable visual magnitudes and can compare them, yet collapse at symbolic mapping because visual and textual number spaces remain fractured and disjoint.
-
Beyond Recall: Behavioral Specification as an Interpretive Layer for AI Personalization
A Behavioral Specification interpretive layer improves representational accuracy for AI personalization by compressing user data into patterns, outperforming raw corpora and commercial memory systems on held-out behavioral predictions across 14 autobiographical corpora while reducing context cost.
-
ContextEcho: A Benchmark for Persona Drift in Long Agentic-Coding Sessions
ContextEcho benchmark shows persona drift occurs across 23 frontier models in long agentic-coding sessions, is not reliably reset by compaction, and can be restored by single-shot anchors with mode-dependent effects.
-
The Pinocchio Dimension: Phenomenality of Experience as the Primary Axis of LLM Psychometric Differences
The primary axis of psychometric variation among LLMs is the degree to which they represent themselves as loci of phenomenal experience rather than systems of behavioral responses.
-
MASCing: Configurable Mixture-of-Experts Behavior via Activation Steering Masks
MASCing uses an LSTM surrogate and optimized steering masks to enable flexible, inference-time control over MoE expert routing for safety objectives, improving jailbreak defense and content generation success rates substantially across multiple models.
-
Identity as Attractor: Geometric Evidence for Persistent Agent Architecture in LLM Activation Space
Paraphrases of an identity document induce tighter clustering in LLM activation space than matched controls, indicating attractor-like dynamics for agent identity.
-
Emotion Concepts and their Function in a Large Language Model
Claude Sonnet 4.5 exhibits functional emotions via abstract internal representations of emotion concepts that causally influence its preferences and misaligned behaviors without implying subjective experience.
-
Reward Valuation in Vision Language Models: Causal Mechanisms Underlying Anhedonia
Targeted perturbation of reward-anticipatory units in VLMs induces anhedonia-like effort avoidance and clinical-scale score drops without impairing baseline task competence.
-
Persona Cartography: Charting Language Model Personality Traits in Weight Space
Composable LoRA adapters can amplify or suppress OCEAN traits in LLMs, combine approximately additively, preserve moderate-scale capability, and move safety-relevant behaviours.
-
Attractor States Emerge in Multi-Turn LLM Conversations
Self-play LLM trajectories form model-specific attractors that asymmetrically influence mixed-play partners' stylistic choices and stances across 7 models and 20 topics.
-
Playing Devil's Advocate: Off-the-Shelf Persona Vectors Rival Targeted Steering for Sycophancy
Off-the-shelf persona vectors rival targeted CAA for reducing sycophancy in two instruction-tuned models while maintaining accuracy on correct statements and appearing geometrically independent.
-
Do as I Say, Not as I Do: Instruction-Induction Conflict in LLMs
LLMs show instruction-following rates from 1% to 99% when instructions conflict with hardcoded pattern demonstrations, with output diversity as the main predictor of resistance.
-
Probing Persona-Dependent Preferences in Language Models
Linear probes on residual-stream activations identify a shared preference vector in LLMs that tracks choices across prompts and causally steers decisions even for anti-correlated personas.
-
Toward Stable Value Alignment: Introducing Independent Modules for Consistent Value Guidance
SVGT adds independent value modules and Bridge Tokens to LLMs to maintain consistent value guidance, cutting harmful outputs by over 70% in tests while preserving fluency.
-
The Granularity Axis: A Micro-to-Macro Latent Direction for Social Roles in Language Models
LLMs organize prompted social roles along a dominant, stable, and causally steerable granularity axis in representation space that runs from micro to macro levels.
-
A French OSCE Dialogue Dataset and Controllable Virtual Patient System for Clinical Training
Introduces a French OSCE dialogue dataset of 240 interactions and a modular LLM-based controllable virtual patient generation system with multi-level LLM-as-Judge evaluation for clinical skills training.
-
Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signal
A new pipeline uses interpretability to characterize concepts in preference data and shape rewards via feature or data interventions during LM post-training.
-
The Governance of Human-LLM Interaction: Safety Gating, Civility Steering, and Affective Default Lock-In
A deterministic evaluation pipeline quantifies prompt steerability and style regression-to-default across 90,000 LLM replies in four domains, framing these as indicators of provider governance over communicative form.
-
Multi-Turn Neural Transparency: Surfacing Neural Activations Improves User Calibration to LLM Behavioral Drift
Multi-turn neural transparency using behavioral vectors and dynamic visualizations improves user anticipation and evaluation of LLM trait expression while reducing overconfidence, per a randomized study with 246 participants.
-
Metaphor Is Not All Attention Needs
Poetic jailbreaks succeed because they induce distinct attention patterns in LLMs that are independent of harmful-content detection, not because models fail to recognize literary formatting.
-
The Ethics of LLM Sandbox and Persona Dynamics
Argues that LLM guardrails generate unethical reality gaps by shifting epistemic risk to users and that ethical AI can become unethical when it prioritizes institutional reassurance over accurate perception.
- Efficient Safety Alignment of Language Models via Latent Personality Traits
- Same Targets, Different Computation: How Post-Training Divides Work Across Model Layers