This curated list addresses the fragmentation and lack of practical examples in autoresearch discussions by aggregating public use cases across various industries. It serves as a high-signal field guide for engineers and researchers seeking to understand real-world autoresearch applications, identify transferable patterns, and evaluate adoption potential. The primary benefit is providing concrete evidence of autoresearch's utility beyond theoretical discussions.

How It Works

The core of autoresearch, as exemplified in this list, revolves around an iterative loop: modify, verify, keep/discard, and repeat. An agent or system autonomously modifies a component (e.g., code, prompts, parameters), evaluates the change against a fixed benchmark or metric, and retains only improvements while reverting regressions. This automated, self-correcting cycle drives incremental optimization and discovery across diverse domains.

Quick Start & Requirements

This repository is a curated list of examples and does not have a direct installation or execution command. Each listed autoresearch project has its own specific requirements, which may include Python, ML frameworks (PyTorch, TensorFlow, JAX), specific libraries, GPUs, or datasets, as detailed within the individual project's documentation.

Highlighted Details

• Extensive coverage across domains including Scientific Research (19 entries), Software/Systems Optimization (23 entries), Evaluation/Red Teaming (7 entries), and Finance/Trading (11 entries).
• Numerous entries showcase quantifiable improvements, such as lifting AUC from 0.902892 to 0.916721, improving parse+render performance by 53%, or reducing mean rank from 344.68 to 157.43.
• Demonstrates autoresearch's adaptability beyond ML training, including applications in GPU kernel optimization, SAT solvers, chip design, prompt engineering, and even genealogy research.
• Features infrastructure and framework adaptations like n-autoresearch for parallelism and crash recovery, and MLX ports for Apple Silicon.

Maintenance & Community

The README does not specify maintainers, community channels (e.g., Discord, Slack), or a public roadmap. Contributions are guided by a CONTRIBUTING.md file, suggesting a community-driven curation process.

Licensing & Compatibility

The repository is licensed under the MIT License. This permissive license generally allows for broad compatibility with commercial and closed-source projects.

Limitations & Caveats

This list is intentionally selective, serving as a "high-signal, fast-scanning field guide" rather than a comprehensive database. Inclusion requires public, citable evidence explicitly demonstrating the autoresearch loop. The effectiveness of any autoresearch system critically depends on the quality, robustness, and non-gamability of its evaluation mechanism, which can be a significant design challenge. Some discussions highlight potential issues like wasted compute cycles if usefulness-aware stop criteria are absent.