analiza aiqbench
This research introduces SIGMA, a novel framework that enables large language models to autonomously improve their own safety alignment using only a behavioral specification and single-turn chat data. By leveraging the model's reasoning capabilities to synthesize stress-test scenarios and acting as its own reward model, the method effectively mitigates misalignment risks in agentic environments without relying on external supervision bottlenecks. The approach demonstrates significant reductions in harmfulness and misalignment while preserving general capabilities, offering a promising path toward scalable self-improving safety systems.
SIGMA: Self-Improving Alignment Generalization from a Model Spec
arXiv:2610.07935v1 Announce Type: new Abstract: LLM agents are increasingly capable of executing complex tasks and of recursively improving themselves on easy-to-verify objectives such as software engineering and mathematics. Since alignment is much harder to verify, this creates a growing risk of capabilities increasing without appropriate safety alignment, especially as capabilities expand to auto-research and cybersecurity. Existing approaches focus on capability self-improvement using verifiable feedback or on alignment training with supervision from stronger models or curated data, creating an external supervision bottleneck for alignment. We ask whether current models can improve their own safety alignment, and propose SIGMA, a data generation and training pipeline enabling alignment self-improvement that generalizes to out-of-distribution settings. Given only a "Model Spec" stating the model's desired behavior, SIGMA leverages a model's reasoning capabilities to strengthen its own safety reasoning. SIGMA first performs spec-guided task synthesis, using the candidate model as a task designer agent to generate diverse alignment dilemma scenarios and convert them into training tasks that stress-test its understanding of the Model Spec. Next, SIGMA conducts self-judged alignment training through supervised fine-tuning and rubric-based reinforcement learning with the model itself as the reward model. Despite training only on single-turn chat data, SIGMA improves safety alignment in multi-turn agentic environments (AgentHarm harmfulness decreases from 22.6 to 14.8; Agentic Misalignment decreases from 79.1 to 3.8), outperforms Deliberative Alignment and Constitutional AI baselines, and retains general capability. Analyses show that a Model Spec balancing harmlessness and helpfulness, test-time reasoning for safety deliberation, and high-quality rubrics from SIGMA's task designer agent are crucial for effective self-improvement.
Abstract:LLM agents are increasingly capable of executing complex tasks and of recursively improving themselves on easy-to-verify objectives such as software engineering and mathematics. Since alignment is much harder to verify, this creates a growing risk of capabilities increasing without appropriate safety alignment, especially as capabilities expand to auto-research and cybersecurity. Existing approaches focus on capability self-improvement using verifiable feedback or on alignment training with supervision from stronger models or curated data, creating an external supervision bottleneck for alignment. We ask whether current models can improve their own safety alignment, and propose SIGMA, a data generation and training pipeline enabling alignment self-improvement that generalizes to out-of-distribution settings. Given only a "Model Spec" stating the model's desired behavior, SIGMA leverages a model's reasoning capabilities to strengthen its own safety reasoning. SIGMA first performs spec-guided task synthesis, using the candidate model as a task designer agent to generate diverse alignment dilemma scenarios and convert them into training tasks that stress-test its understanding of the Model Spec. Next, SIGMA conducts self-judged alignment training through supervised fine-tuning and rubric-based reinforcement learning with the model itself as the reward model. Despite training only on single-turn chat data, SIGMA improves safety alignment in multi-turn agentic environments (AgentHarm harmfulness decreases from 22.6 to 14.8; Agentic Misalignment decreases from 79.1 to 3.8), outperforms Deliberative Alignment and Constitutional AI baselines, and retains general capability. Analyses show that a Model Spec balancing harmlessness and helpfulness, test-time reasoning for safety deliberation, and high-quality rubrics from SIGMA's task designer agent are crucial for effective self-improvement.
Artificial Intelligence (cs.AI)
arXiv:2610.07935 [cs.AI]
(or arXiv:2610.07935v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2610.07935 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
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 · 346 cuvinte
Subiecte conexe
Projection sampling preserves capabilities but diminishes gains from alignment fine-tuning
LessWrong · 16 oct. 2026
Models Can't Remember Their Training. Neither Can You.
LessWrong · 19 iul. 2026
Coverage-driven alignment - What ‘Teaching Claude Why’ can borrow from AV verification
LessWrong · 8 iun. 2026
Thousand-dimensional structure
AI Alignment Forum · 30 iul. 2026