Summary
Feng, Zhao, Zhong, Huang, Gu, Kong, Feng, Qin — Harbin Institute of Technology / The University of Hong Kong, OpenReview submission for ICLR 2026 (posted October 2025).
Extends the persona-vector extraction pipeline from the already-filed Chen et al. 2025 with explicit algebraic operations and dynamic inference-time composition. The PERSONA framework extracts approximately orthogonal OCEAN trait vectors via contrastive activation analysis (PERSONA-BASE), then demonstrates that these vectors support predictable vector arithmetic: scalar multiplication for intensity control, addition for multi-trait composition, and subtraction for targeted suppression (PERSONA-ALGEBRA). A predict-then-steer mechanism (PERSONA-FLOW) enables context-aware dynamic composition during multi-turn generation. On the external PersonalityBench the training-free method reaches 9.60 mean score, statistically indistinguishable from the supervised fine-tuning upper bound of 9.61. On the authors' new PERSONA-EVOLVE benchmark (800 multi-turn scenarios across 100 sessions), it achieves up to 91% win rates for trait adherence, role consistency, and response authenticity across Qwen, Llama, and Mistral families.
Twenty-first instantiating finding for concepts/persona-selection; adds the algebraic / compositional control shape under the activation-level mechanistic toolkit.
Method
The work re-uses and extends the automated contrastive extraction pipeline introduced in Chen et al. 2025 (Persona Vectors). For each Big Five (OCEAN) pole, a frontier LLM generates contrastive system prompts and evaluation items; the target model produces responses under trait-eliciting vs. trait-suppressing conditions; the persona vector is the mean activation difference in a selected residual-stream layer.
PERSONA-ALGEBRA validates that the resulting vectors behave as an algebraic system. Steering coefficients are applied as residual additions (hl ← hl + α vl). Three operations are tested on an adapted behavioral BFI-44 instrument (scenario prompts scored by GPT-4.1-mini on 5-point Likert scales):
- Scalar multiplication: varying α from –1 to +2 produces approximately linear changes in the corresponding trait scores (Pearson r > 0.9 for most poles).
- Vector addition: steering with v_outgoing + v_compassionate simultaneously elevates both Extraversion and Agreeableness scores.
- Vector subtraction: v_outgoing − v_solitary amplifies Extraversion while v_outgoing − v_compassionate maintains Extraversion but reduces Agreeableness, demonstrating trait isolation.
PERSONA-FLOW adds a two-stage inference-time mechanism: an intermediate forward pass predicts per-dimension steering coefficients from conversational context; the composite vector (Σ α_i · v_i) is then injected at the chosen layer for the actual response generation. This enables real-time, context-sensitive personality modulation without any gradient updates or predefined scripts.
Key results
Algebraic coherence. BFI-44 behavioral scores after addition and subtraction operations match the predicted directional changes with high fidelity. Non-perfect orthogonality (some cross-trait cosine correlations reflecting semantic associations in training data) does not break the algebraic behavior; secondary effects remain predictable.
Training-free parity with SFT. On PersonalityBench the PERSONA method achieves a mean score of 9.60 versus the supervised fine-tuning upper bound of 9.61, with lower variance (0.74).
Dynamic adaptation. On the new PERSONA-EVOLVE benchmark (100 multi-turn dialogue sessions, 8 evolving scenarios each, 800 total instances), pairwise win rates versus vanilla generation reach up to 91% across trait adherence, role consistency, and response authenticity for Qwen, Llama, and Mistral model families. The predict-then-steer mechanism successfully suppresses or amplifies traits to match shifting situational demands within a single conversation while preserving core persona constraints.
Why it matters
From extraction to algebra. Chen et al. 2025 established that persona vectors for arbitrary traits can be extracted from natural-language descriptions and used for monitoring and preventative steering. This finding shows that the extracted directions are not merely steerable but form a coherent vector space supporting the standard operations of linear algebra — addition, subtraction, and scaling — with directly observable behavioral consequences. This is the strongest evidence to date that personality representations in LLMs behave as modular, composable features rather than monolithic or purely prompt-dependent constructs.
Dynamic control surface. PERSONA-FLOW demonstrates that algebraic composition can be driven by context prediction at inference time, moving persona control from static vector addition to real-time, conversation-aware modulation. This supplies the missing dynamic layer for the activation-level toolkit under persona-selection.
Complementary to prior shapes. The algebraic/compositional shape sits alongside the existing activation-level toolkit (Chen et al.), prompt-level prevention (inoculation prompting), prompt-level reactivation (Shah et al., Zhang et al., Sandhan et al.), and training-stage prior installation (Model Spec midtraining). It strengthens the mechanistic reading that post-training and fine-tuning operate by selecting and composing from a pre-existing distribution of persona simulations.
Cross-reference to PSM. The result is consistent with the Persona Selection Model's picture of a narrowable posterior over persona simulations: if the posterior is represented (at least in part) as directions in activation space, then linear operations on those directions correspond to shifting probability mass among persona components.
interpretive tensions
The vectors are described as "approximately orthogonal." While opposing poles show strong negative cosine similarity, some cross-dimensional correlations exist and are acknowledged by the authors. The algebra still produces predictable behavioral outcomes, but the representation is not a clean orthogonal basis. This tempers (but does not invalidate) the "modular and additive" claim.
The paper tests open-weight instruction-tuned models (Qwen2.5, Llama-3.1, Mistral families). Generalization to closed-weight frontier systems (where the specific directions and layer-wise geometry may differ) is not directly measured, though the extraction pipeline itself is model-agnostic given activation access.
concepts
- Persona selection — twenty-first instantiating finding; first explicit demonstration of algebraic compositionality and dynamic inference-time vector arithmetic on persona vectors. Extends the activation-level mechanistic toolkit shape with a "compositional / modular arithmetic" sub-shape.
cross-references
- Persona Vectors (Chen et al. 2025) — direct predecessor. This finding re-uses the contrastive extraction pipeline and adds the algebraic validation, dynamic composition mechanism, and new multi-turn benchmark.
- Persona Selection Model (Marks, Lindsey, Olah 2026) — supplies the broader mechanistic framing within which these vectors are interpreted as coordinates in a pre-training-acquired persona posterior.
sources
- Feng et al. (2025). PERSONA: Dynamic and Compositional Inference-Time Personality Control via Activation Vector Algebra. OpenReview submission for ICLR 2026.
- Chen et al. (2025). Persona Vectors: Monitoring and Controlling Character Traits in Language Models. arXiv:2507.21509.