analiza aiqbench
This research advances the mechanistic interpretability of LLMs by demonstrating how specific latent directions in the residual stream can be directly intervened upon to steer personality traits without degrading general performance. By utilizing sparse autoencoders and additive steering vectors, the study offers a precise, non-invasive method for controlling model behavior at the feature level rather than relying on prompt engineering or full fine-tuning. This work is significant for understanding the internal structure of large models and developing more controllable AI agents.
Mechanistic Personality Analysis of LLMs: Steering Personality via Latent Feature Interventions
arXiv:2606.28770v2 Announce Type: replace Abstract: Large Language Models (LLMs) have demonstrated the ability to simulate human-like OCEAN personality traits in generated text. Previous efforts have focused on prompt engineering or fine-tuning to shape LLM personality. In this work, we propose a mechanistic interpretability approach that directly intervenes on the model's latent features. Our method identifies latent directions in the residual stream corresponding to a target OCEAN trait using sparse autoencoders (SAEs) and contrastive activation analysis. We formalize an additive steering vector in activation space and demonstrate how applying a small additive shift to the hidden states enhances the target trait while preserving overall language modeling performance. To determine the optimal combination of feature shifts, we explore a linear weighting heuristic with grid search optimization that balances personality expression with task performance. Our approach shows promise in controllably steering personality traits at the mechanistic level while maintaining high performance on standard benchmarks.
Abstract:Large Language Models (LLMs) have demonstrated the ability to simulate human-like OCEAN personality traits in generated text. Previous efforts have focused on prompt engineering or fine-tuning to shape LLM personality. In this work, we propose a mechanistic interpretability approach that directly intervenes on the model's latent features. Our method identifies latent directions in the residual stream corresponding to a target OCEAN trait using sparse autoencoders (SAEs) and contrastive activation analysis. We formalize an additive steering vector in activation space and demonstrate how applying a small additive shift to the hidden states enhances the target trait while preserving overall language modeling performance. To determine the optimal combination of feature shifts, we explore a linear weighting heuristic with grid search optimization that balances personality expression with task performance. Our approach shows promise in controllably steering personality traits at the mechanistic level while maintaining high performance on standard benchmarks.
Artificial Intelligence (cs.AI)
arXiv:2606.28770 [cs.AI]
(or arXiv:2606.28770v2 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2606.28770 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 238 cuvinte
Subiecte conexe
Mechanistic Knobs in LLMs: Retrieving and Steering High-Order Semantic Features via Sparse Autoencoders
arXiv · 15 oct. 2026
When should we trust a latent representation?
LessWrong · 4 aug. 2026
Controllable-CoT leads to covert reasoning capabilities
LessWrong · 22 sept. 2026
The J-Space Debate, Agent Swarms, and Pacing Frontier AI - Digital Minds Newsletter #4
LessWrong · 18 sept. 2026