20260422:更新

raw/papers/zhang-hyperagents-2026.md

@@ -0,0 +1,89 @@
+# Hyperagents (arXiv:2603.19461)
+
+## Metadata
+- **Title**: Hyperagents
+- **Authors**: Jenny Zhang, Bingchen Zhao, Wannan Yang, Jakob Foerster, Jeff Clune, Minqi Jiang, Sam Devlin, Tatiana Shavrina
+- **arXiv ID**: 2603.19461
+- **Submission Date**: 19 Mar 2026
+- **Subjects**: Artificial Intelligence (cs.AI)
+- **DOI**: https://doi.org/10.48550/arXiv.2603.19461
+- **Code**: https://github.com/facebookresearch/Hyperagents
+- **License**: Creative Commons Attribution 4.0 International
+
+## Abstract
+
+Self-improving AI systems aim to reduce reliance on human engineering by learning to improve their own learning and problem-solving processes. Existing approaches to self-improvement rely on fixed, handcrafted meta-level mechanisms, fundamentally limiting how fast such systems can improve. The Darwin Gödel Machine (DGM) demonstrates open-ended self-improvement in coding by repeatedly generating and evaluating self-modified variants. Because both evaluation and self-modification are coding tasks, gains in coding ability can translate into gains in self-improvement ability. However, this alignment does not generally hold beyond coding domains.
+
+We introduce **hyperagents**, self-referential agents that integrate a task agent (which solves the target task) and a meta agent (which modifies itself and the task agent) into a single editable program. Crucially, the meta-level modification procedure is itself editable, enabling metacognitive self-modification, improving not only the task-solving behavior, but also the mechanism that generates future improvements.
+
+We instantiate this framework by extending DGM to create DGM-Hyperagents (DGM-H), eliminating the assumption of domain-specific alignment between task performance and self-modification skill to potentially support self-accelerating progress on any computable task. Across diverse domains, the DGM-H improves performance over time and outperforms baselines without self-improvement or open-ended exploration, as well as prior self-improving systems. Furthermore, the DGM-H improves the process by which it generates new agents (e.g., persistent memory, performance tracking), and these meta-level improvements transfer across domains and accumulate across runs.
+
+DGM-Hyperagents offer a glimpse of open-ended AI systems that do not merely search for better solutions, but continually improve their search for how to improve.
+
+## Key Concepts
+
+### 1. Hyperagents
+Self-referential agents that integrate task-solving and self-modification capabilities into a single editable program. The meta-level modification procedure is itself editable, enabling metacognitive self-modification.
+
+### 2. Darwin Gödel Machine (DGM)
+A framework for open-ended self-improvement in coding domains, where both evaluation and self-modification are coding tasks, creating a natural alignment between task performance and self-improvement ability.
+
+### 3. DGM-Hyperagents (DGM-H)
+Extension of DGM that eliminates the domain-specific alignment assumption, enabling self-accelerating progress on any computable task.
+
+### 4. Metacognitive Self-Modification
+The ability to not only improve task-solving behavior but also improve the mechanism that generates future improvements.
+
+### 5. Self-Accelerating Progress
+The property where improvements in problem-solving ability lead to improvements in self-improvement ability, creating a positive feedback loop.
+
+## Methodology
+
+### Framework Architecture
+1. **Integrated Program**: Single editable program containing both task agent and meta agent
+2. **Editable Meta-Level**: The modification procedure itself can be modified
+3. **Self-Referential Loop**: Improvements in task-solving → improvements in self-modification → further improvements in task-solving
+
+### DGM-H Implementation
+- Extends the original DGM framework
+- Removes domain-specific alignment requirement
+- Supports persistent memory and performance tracking
+- Enables meta-level improvements to transfer across domains
+
+## Results
+
+### Performance Improvements
+- DGM-H improves performance over time across diverse domains
+- Outperforms baselines without self-improvement
+- Outperforms prior self-improving systems
+
+### Meta-Level Improvements
+- Improves the process of generating new agents
+- Improvements transfer across domains
+- Improvements accumulate across runs
+
+## Significance
+
+### Theoretical Contribution
+- Introduces the concept of hyperagents as a general framework for self-improving AI
+- Demonstrates metacognitive self-modification as a key capability
+- Provides a path toward self-accelerating progress on arbitrary computable tasks
+
+### Practical Implications
+- Potential for creating AI systems that continuously improve their own improvement processes
+- Reduces reliance on human engineering for meta-level design
+- Enables open-ended progress beyond fixed meta-level mechanisms
+
+## Related Work
+- Darwin Gödel Machine (DGM)
+- Self-improving AI systems
+- Meta-learning and meta-reinforcement learning
+- Program synthesis and genetic programming
+
+## References
+- arXiv:2603.19461 [cs.AI]
+- GitHub: https://github.com/facebookresearch/Hyperagents
+- DOI: https://doi.org/10.48550/arXiv.2603.19461
+
+## Tags
+#hyperagents #self-improving-ai #darwin-godel-machine #metacognitive-self-modification #self-accelerating-progress #ai-research #meta-learning