AI
Self-Evolving Coding Agents: From Digital Programs to Physical-World Intelligence
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 27/09/2026, 20:00:00 UTC
Briefing created: 29/09/2026, 04:21:23 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 29/09/2026, 04:18:25 UTC.
Briefings
This article proposes Physical Coding as a framework for self-evolving coding agents, where task states and execution are represented as code to enable generalization and long-horizon planning. The authors build HexaAnything, which integrates vision-language-action (VLA) and world-action (WAM) models to perform in-the-loop decisions from external feedback. On RoboCasa365, HexaAnything outperforms existing VLA models on unseen tasks, while its Harness-trained variant beats base models on every split. Experiments on PhyBench and an AgileX dual-arm robot show autonomous completion of physics experiments and tabletop tasks, often surpassing published results. The research aims to move from tool-based evolution to model weight internalization and hardware deployment.
Key points
- The authors propose Physical Coding, which represents task state and execution as code to enable generalization and long-horizon planning in physical-world agents.
- HexaAnything improves Composite-Unseen and overall success over existing VLA models on RoboCasa365, with its Harness-trained variant beating the base on every split.
Context and limitations
Data and model performance are limited to specific benchmarks including RoboCasa365, PhyBench, and AgileX dual-arm robot experiments. The study focuses on physical-world intelligence and does not cover all aspects of the broader field.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 29/09/2026, 04:21:23 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeieiv5vqyjn6oq6kgt3lrbcbdbykk2lyo6kketagz2vwuh54s4gc6q