Do LLMs Experience an Internal Polylogue? Investigating Reasoning through the Lens of Personas
Nils A. Herrmann, Leander Girrbach, Kirill Bykov, Zeynep Akata
Why It Matters
What makes this one worth your time
Understanding and manipulating reasoning processes in LLMs can lead to more accurate and reliable AI systems, which is crucial for practical applications.
This research explores dynamic persona vectors in LLMs to enhance reasoning accuracy through real-time interventions.
Summary
The paper investigates the concept of 'polylogue' in large language models (LLMs), treating persona vectors as dynamic signals that can be monitored and intervened upon during reasoning, demonstrating that these features can predict correctness and improve model accuracy through targeted interventions.
Key contributions
- Introduces the concept of 'polylogue' for reasoning-time monitoring in LLMs.
- Demonstrates that polylogue features can predict correctness on benchmarks like MMLU-Pro.
- Implements a paragraph-conditioned intervention that enhances accuracy across multiple models.
Notable insights
- The concept of 'polylogue' as a time series of persona vector alignments offers a novel approach to monitoring LLM reasoning.
- Dynamic steering of latent directions based on reasoning stages could provide a new methodology for improving model outputs.
Possible limitations
- Not stated in the abstract.
Abstract
arXiv:2605.09159v2 Announce Type: replace Abstract: Recent work shows that large language models (LLMs) encode behavioral traits ("personas") as linear directions in activation space, often called "persona vectors". Prior work has used such directions as static handles for behavioral steering. We instead treat them as dynamic signals: probes we can monitor and intervene on as reasoning unfolds. We use the term polylogue to denote the time series of alignments between persona vectors and hidden activations over the course of generation. Experiments across four open-weight models show that polylogue features contain predictive signal for correctness comparable to low-dimensional activation summaries, while remaining interpretable through their associated persona directions. They also suggest concrete steering targets, namely, which latent directions to modulate at different stages of a response. We instantiate this as a simple paragraph-conditioned intervention that improves accuracy on three of the four models but degrades the fourth, suggesting that stage-aware latent steering is possible but not yet robust. Together, this positions the polylogue as an interpretable tool for reasoning-time monitoring and intervention.