Towards Safe and Honest AI Agents with Neural Self-Other Overlap

As AI systems increasingly make critical decisions, deceptive AI poses a significant challenge to trust and safety. We present Self-Other Overlap (SOO) fine-tuning, a promising approach in AI Safety that could substantially improve our ability to build honest artificial intelligence. Inspired by cognitive neuroscience research on empathy, SOO aims to align how AI models represent themselves and others. Our experiments on LLMs with 7B, 27B, and 78B parameters demonstrate SOO's efficacy: deceptive responses of Mistral-7B-Instruct-v0.2 dropped from 73.6% to 17.2% with no observed reduction in general task performance, while in Gemma-2-27b-it and CalmeRys-78B-Orpo-v0.1 deceptive responses were reduced from 100% to 9.3% and 2.7%, respectively, with a small impact on capabilities. In reinforcement learning scenarios, SOO-trained agents showed significantly reduced deceptive behavior. SOO's focus on contrastive self and other-referencing observations offers strong potential for generalization across AI architectures. While current applications focus on language models and simple RL environments, SOO could pave the way for more trustworthy AI in broader domains. Ethical implications and long-term effects warrant further investigation, but SOO represents a significant step forward in AI safety research.

SAR: Generalization of Physiological Agility and Dexterity via Synergistic Action Representation

Learning effective continuous control policies in high-dimensional systems, including musculoskeletal agents, remains a significant challenge. Over the course of biological evolution, organisms have developed robust mechanisms for overcoming this complexity to learn highly sophisticated strategies for motor control. What accounts for this robust behavioral flexibility? Modular control via muscle synergies, i.e. coordinated muscle co-contractions, is considered to be one putative mechanism that enables organisms to learn muscle control in a simplified and generalizable action space. Drawing inspiration from this evolved motor control strategy, we use physiologically accurate human hand and leg models as a testbed for determining the extent to which a Synergistic Action Representation (SAR) acquired from simpler tasks facilitates learning more complex tasks. We find in both cases that SAR-exploiting policies significantly outperform end-to-end reinforcement learning. Policies trained with SAR were able to achieve robust locomotion on a wide set of terrains with high sample efficiency, while baseline approaches failed to learn meaningful behaviors. Additionally, policies trained with SAR on a multiobject manipulation task significantly outperformed (>70% success) baseline approaches (<20% success). Both of these SAR-exploiting policies were also found to generalize zero-shot to out-of-domain environmental conditions, while policies that did not adopt SAR failed to generalize. Finally, we establish the generality of SAR on broader high-dimensional control problems using a robotic manipulation task set and a full-body humanoid locomotion task. To the best of our knowledge, this investigation is the first of its kind to present an end-to-end pipeline for discovering synergies and using this representation to learn high-dimensional continuous control across a wide diversity of tasks.