Information Ricoh paper on information-theoretic method for steering LLM behavior accepted at AACL 2026 Steering LLM behavior for different purposes without additional training while maintaining natural language quality

TOKYO, November 5, 2026 – Ricoh announced today that its research paper on an information-theoretic method for steering large language model (LLM) behavior has been accepted for presentation at The 5th Asia-Pacific Chapter of the Association for Computational Linguistics & the 15th International Joint Conference on Natural Language Processing (AACL 2026), an international conference in the field of natural language processing. The paper is scheduled to be presented at the conference, which will be held in China from November 6 to 10, 2026.

The research proposes a method for steering LLM behavior for different purposes through lightweight intervention in the model's internal processing during inference, without requiring additional model training. Using an information-theoretic framework, the method formulates the design of steering interventions as an optimization problem. Experiments presented in the paper demonstrated that the method can steer LLM behavior while maintaining natural language quality. The approach could help reduce the cost and operational burden of adapting LLMs for different purposes, with potential applications ranging from controlling text generation behavior to enhancing the safety of generative AI.

Ricoh will continue to evaluate the technology and explore its potential applications in enterprise generative AI systems. Through continued research and development, Ricoh aims to advance technologies that enhance the safety and reliability of generative AI and contribute to AI services tailored to customers' industries and business needs, supporting Fulfillment through Work.