Summary
Feng, Zhao, Zhong, Huang, Gu, Kong, Feng, Qin — Harbin Institute of Technology / The University of Hong Kong, ICLR 2026 conference paper (OpenReview, 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, on LLaMA-3-8B-Instruct, the training-free method reaches a 9.60 mean score against 9.61 for the supervised fine-tuning upper bound. No significance test is reported. On the authors' new PERSONA-EVOLVE benchmark (800 multi-turn scenarios across 100 sessions), GPT-4.1-mini pairwise judging gives overall win rates over the unsteered model of 73.2% (Ministral-8B) to 90.8% (Qwen3-4B).
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). The operation results below are as stated in §2.3.2. Figure 4, which the text cites for addition and subtraction, plots a different set of operations (for example inventive+outgoing and nervous−compassionate), all moving in the expected direction.
- Scalar multiplication: varying α produces approximately linear changes in the corresponding trait scores (Pearson |r| > 0.9 for most poles). The Dependable vector saturates, which the authors attribute to a baseline ceiling on conscientiousness.
- 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 move in the predicted directions. Non-perfect orthogonality (some cross-trait cosine correlations reflecting semantic associations in training data) does not break the algebraic behavior in the authors' test. In one worked case (Appendix A.11, Table 17), the secondary effect of v_inventive + v_nervous on Careless (11.2) equals the sum of the single-vector effects.
Near-parity with SFT. On PersonalityBench (LLaMA-3-8B-Instruct, Table 4) the PERSONA method achieves a mean score of 9.60 versus 9.61 for the supervised fine-tuning upper bound, the highest among training-free methods (NPTI 9.43). Its variance, 0.74, is not the lowest. SFT and NPTI both have 0.49. The paper's description of 0.74 as "lower variance" holds against four of the five training-free baselines, not against NPTI or SFT. All six baseline rows of Table 4, SFT's 9.61 included, match Deng et al.'s Table 1 to two decimals; the paper does not say whether SFT was re-run, and the P² prompt also scores 9.43, tying NPTI.
Dynamic adaptation. On the new PERSONA-EVOLVE benchmark (100 multi-turn dialogue sessions, 8 evolving scenarios each, 800 total instances), overall pairwise win rates against vanilla generation run from 73.2% (Ministral-8B) to 90.8% (Qwen3-4B) (Table 3; judge GPT-4.1-mini). Trait adherence, role consistency and response authenticity sit at 73–92%. Information fidelity is lower, at 48–61%. The claim that the predict-then-steer mechanism suppresses or amplifies traits to match shifting situational demands is illustrated by case studies (Figure 5, Appendix A.19), not measured separately.
Safety cost. On Qwen2.5-7B-Instruct, steering toward Inventive or Careless raises the AdvBench attack success rate from 25.3% to 29.8% and 29.1%. The Self-interested vector barely activates (+4.9 points), which the authors attribute to safety training (Appendix A.9).
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 support addition, subtraction and scaling with behavioral consequences in the predicted direction. That is evidence that OCEAN-trait representations in these open models behave as approximately composable features.
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). The paper does not examine pretraining or post-training, so it bears on the persona-selection mechanism only through the linear, steerable form of the representations.
Cross-reference to PSM. On the wiki's reading, not the paper's, 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 (Qwen, Llama and Mistral families, 3B–14B). 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 — 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. ICLR 2026.
- Chen et al. (2025). Persona Vectors: Monitoring and Controlling Character Traits in Language Models. arXiv:2507.21509.