Hexo Labs introduces SIA (Self-Improving AI), a framework that treats agents as editable systems by co-optimizing their scaffolds and internal weights in a unified loop.
SIA is an open-source framework developed by Hexo Labs that allows an AI meta-agent loop to autonomously edit and improve all facets of a target agent, including its external scaffold (such as tools, parsers, and verifiers) and its model weights. By treating both code and weights as optimization targets, SIA achieves significant performance gains across tasks, including 70.1% accuracy on the LawBench legal benchmark, optimization of Triton CUDA kernels down to 1,017 µs, and a 0.289 mse_norm reduction in scientific denoising tasks.
Co-optimizing an agent's harness and internal model weights elevates autonomous self-improvement from basic prompt tuning to a unified system optimization paradigm.
- –Traditional approaches optimize either model weights (via RL) or the scaffolding (via prompt engineering), leaving synergistic gains on the table; SIA's unified loop solves this division.
- –Achieving state-of-the-art results on LawBench and custom CUDA kernel optimization demonstrates strong generalization across heterogeneous domain types (legal reasoning, system programming, and scientific computing).
- –Letting agents modify their own code/harness alongside weights introduces substantial security risks, requiring robust sandboxing and verification to prevent unintended side effects.
DISCOVERED
1h ago
2026-06-14
PUBLISHED
2h ago
2026-06-14
RELEVANCE
AUTHOR
AlphaSignalAI