Do LLMs Experience an Internal Polylogue? Investigating Reasoning through the Lens of Personas

Researchers demonstrate that large language models encode behavioral traits as linear directions in activation space called "persona vectors," which can be monitored and manipulated during reasoning. By treating these vectors as dynamic signals over generation time—termed "polylogue"—they achieve competitive accuracy prediction on MMLU-Pro while enabling stage-aware latent steering that improves model performance.

This research advances mechanistic interpretability of LLM reasoning by moving beyond static behavioral analysis to dynamic monitoring of how persona-aligned activations evolve during generation. The concept of polylogue—tracking how hidden states align with persona directions over time—provides a novel window into model cognition that bridges interpretability and performance optimization. Rather than treating persona vectors as fixed behavioral handles, the authors leverage them as temporal probes, revealing how different aspects of model behavior activate at different reasoning stages.

The work builds on established findings that neural networks encode semantic and behavioral information as exploitable linear directions, extending this to temporal analysis. By demonstrating that polylogue features predict correctness competitively with low-dimensional activation baselines while maintaining interpretability, the research validates an approach that most AI safety and alignment researchers prioritize: understanding model internals without sacrificing performance diagnostics. The ability to identify "stage-aware latent steering targets" represents a meaningful step toward fine-grained control over reasoning processes.

For AI development, this suggests that model behavior is more systematically controllable than previously thought, with intervention points identifiable through persona vector alignment patterns. The demonstrated improvements in accuracy across three of four tested models indicate practical utility beyond theoretical insights. This capability to steer reasoning at specific generation stages could accelerate progress in alignment, fact-checking, and error-correction mechanisms. The work positions future development toward interpretable, controllable reasoning systems rather than black-box scaling approaches.

• →Persona vectors can be monitored as dynamic time-series signals (polylogue) throughout model generation, not just static behavioral handles.
• →Polylogue features predict answer correctness competitively with standard activation analysis while remaining interpretable.
• →Stage-aware latent steering identifies optimal intervention points in reasoning sequences to improve accuracy.
• →Paragraph-conditioned interventions improved accuracy on three of four tested models, demonstrating practical steering capability.
• →The approach bridges AI interpretability and performance optimization, enabling both understanding and control of LLM reasoning.