Hugging Face Releases DuoDiscover: Revolutionizing Autonomous Discovery Efficiency

Summary:Hugging Face has released DuoDiscover, a novel search system designed to overcome the evaluation bottleneck in autonomous discovery. The system employs three key mechanisms: 1) Collaborative discovery of optimization targets and fast, tailored proxy evaluators (yardsticks) by coding agents; 2) Continuous fidelity tests to ensure yardsticks' accuracy; 3) A decentralized, lightweight orchestration approach for agent management. Experimental results demonstrate DuoDiscover's effectiveness, includin

---

Technical Mechanism Analysis

The core innovation of DuoDiscover lies in its novel approach to autonomous optimization and evaluation:

1. Collaborative Discovery and Yardstick Evaluators:

   • Each agent is responsible for discovering optimization targets and generating a fast, tailored proxy evaluator (yardstick) for rapid progress assessment. This reduces evaluation costs from hours to seconds, significantly boosting search efficiency.
   • The yardstick is not static but dynamically adjusted to adapt to task changes, ensuring minimal deviation from the true objective.

2. Faithfulness Tests:

   • To prevent discrepancies between the yardstick and the true objective, the system employs faithfulness tests to continuously verify the yardstick's accuracy.
   • The system promptly updates or replaces the yardstick if deviations are detected, ensuring the reliability of the search process.

3. Decentralized Coordination Mechanism:

   • The system uses a lightweight orchestrator to manage agents, rather than relying on centralized task decomposition. This design enhances scalability and robustness.
   • The orchestrator handles agent spawning and termination, task allocation, and sharing of work results, ensuring the system's coordinated operation.

Engineering Trade-offs and Performance

DuoDiscover demonstrates strong performance across multiple benchmarks:

• SWE-fficiency: In the software engineering efficiency benchmark, DuoDiscover doubles the average speedup of human experts, showcasing its powerful capabilities in optimizing software code.
• SOL-Exec: In the SOL-Exec benchmark, DuoDiscover ranks first in 121 out of 235 kernels and in the top three in 212 kernels, highlighting its leadership in GPU and TPU kernel optimization.
• vLLM and SQLite Engines: On the heavily optimized vLLM and SQLite engines, DuoDiscover achieves 1.92x and 1.38x performance gains, respectively.
• CPU Microarchitecture Optimization: In the CPU microarchitecture optimization task, DuoDiscover outperforms AlphaEvolve-based ArchAgent v2 and the best human design, achieving a 1.65x performance gain in 16.7 days, compared to over 40 days for the human design.

Developer Deployment and Recommendations

The release of DuoDiscover offers the following opportunities for developers:

• Software Optimization: Developers can leverage DuoDiscover to quickly optimize code and improve software performance.
• Chip Design: In chip design, DuoDiscover can accelerate the kernel optimization process and shorten design cycles.
• Resource Management: DuoDiscover's decentralized architecture makes it suitable for large-scale distributed systems, optimizing resource allocation.

Developers are advised to review DuoDiscover's documentation and sample code to better understand its working mechanisms and apply it to their projects.

---

Source: Hugging Face Daily Papers (2026-10-12)

— END —

Tags: #DuoDiscover Follow #Autonomous Discovery Follow #AI Optimization Follow #Hugging Face Follow #Chip Design Follow