--- schema: 1 type: video title: >- Parameter Golf with AutoResearch — Vayum Arora, Zhengyao Jiang, Dixing Xu & Dhruv Srikanth, Weco AI description: >- An agent keeps the code changes that improve a language model's score and discards the rest. Weco AI turns that research loop into a workshop, using Mini Parameter Golf to make training experiments small enough to inspect as they run. Zhengyao Jiang introduces autoresearch as repeated proposals and evaluations, and Dixing Xu explains the team's Aiden system and its contributions to OpenAI's Parameter Golf challenge. Vayum Arora and Dhruv Srikanth then connect those results to a practical setup that combines a modifiable source file, a fixed evaluation command and a measurable goal. The walkthrough focuses on what determines whether the search is useful: fast and stable evaluations, a clear direction for the metric, time limits, initial guidance and protection against reward hacking. Weco's experiment tree makes each attempt inspectable, with its code changes and score visible in the dashboard. The presenters show how to branch from promising work and steer a running search toward a different idea instead of simply letting it continue. Internet problems lead to a guided setup demonstration, followed by questions about reinforcement learning and combining multiple objectives. The final exercise swaps model training for a Flappy Bird controller scored across normal play, moving gaps, wind and mixed conditions. A hidden validation set and limits on what the controller can edit make the competition another lesson in designing an evaluation that rewards the behavior you want. Speaker info: Vayum Arora: - https://www.linkedin.com/in/vayum-arora/ Zhengyao Jiang: - https://zhengyaojiang.github.io/ - https://www.linkedin.com/in/zhengyao-jiang-387b44145/ Dixing Xu: - https://dex.moe/ - https://linkedin.com/in/dex-xu Dhruv Srikanth: - https://www.linkedin.com/in/dhruv-srikanth Workshop resource: - https://github.com/WecoAI/mini-pg Timestamps: 0:00 - Presenters and the autoresearch loop 3:35 - Workshop plan and AIDE background 5:39 - Parameter Golf and Aiden 8:30 - Mini Parameter Golf 9:37 - Weco, Modal, and repository setup 11:10 - Anatomy of an autoresearch run 12:06 - Fast evaluations and reward hacking 13:02 - Define the metric and search budget 14:25 - Time limits and initial guidance 16:31 - Structure, observability, and steering 17:29 - Inspect the experiment tree 19:36 - Human ideas and research breakthroughs 20:47 - Guided setup and live dashboard 24:41 - Reinforcement learning and multiple objectives 26:18 - Evaluate and steer an experiment 28:46 - The Flappy Weco competition 30:08 - Environments, validation, and rules 32:30 - Exercise instructions and wrap up canonical: >- https://liquidgold.cc/channel/aidotengineer/explore/video/parameter-golf-with-autoresearch-vayum-arora-zhengy-1eb2f2785a59 markdown: >- https://liquidgold.cc/channel/aidotengineer/video/parameter-golf-with-autoresearch-vayum-arora-zhengy-1eb2f2785a59.md llms: null parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-23T22:19:06.258Z' language: null children: [] videoId: 7RFJRDYroic durationSeconds: 2053 publishedAt: '2026-10-11T21:00:16.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 23 endSeconds: 223 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=23' heldReason: rights - ordinal: 1 text: null startSeconds: 23 endSeconds: 223 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=23' heldReason: rights - ordinal: 2 text: null startSeconds: 234 endSeconds: 339 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=234' heldReason: rights - ordinal: 3 text: null startSeconds: 351 endSeconds: 513 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=351' heldReason: rights - ordinal: 4 text: null startSeconds: 351 endSeconds: 513 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=351' heldReason: rights - ordinal: 5 text: null startSeconds: 524 endSeconds: 581 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=524' heldReason: rights - ordinal: 6 text: null startSeconds: 681 endSeconds: 726 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=681' heldReason: rights - ordinal: 7 text: null startSeconds: 738 endSeconds: 782 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=738' heldReason: rights - ordinal: 8 text: null startSeconds: 792 endSeconds: 875 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=792' heldReason: rights - ordinal: 9 text: null startSeconds: 792 endSeconds: 875 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=792' heldReason: rights - ordinal: 10 text: null startSeconds: 875 endSeconds: 887 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=875' heldReason: rights - ordinal: 11 text: null startSeconds: 887 endSeconds: 991 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=887' heldReason: rights - ordinal: 12 text: null startSeconds: 1002 endSeconds: 1051 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=1002' heldReason: rights - ordinal: 13 text: null startSeconds: 1064 endSeconds: 1180 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=1064' heldReason: rights - ordinal: 14 text: null startSeconds: 1064 endSeconds: 1180 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=1064' heldReason: rights - ordinal: 15 text: null startSeconds: 1191 endSeconds: 1247 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=1191' heldReason: rights - ordinal: 16 text: null startSeconds: 1491 endSeconds: 1585 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=1491' heldReason: rights - ordinal: 17 text: null startSeconds: 1598 endSeconds: 1734 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=1598' heldReason: rights - ordinal: 18 text: null startSeconds: 1746 endSeconds: 1814 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=1746' heldReason: rights - ordinal: 19 text: null startSeconds: 1814 endSeconds: 1825 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=1814' heldReason: rights - ordinal: 20 text: null startSeconds: 1825 endSeconds: 1952 watch: 'https://liquidgold.cc/watch/7RFJRDYroic?t=1825' heldReason: rights --- # Parameter Golf with AutoResearch — Vayum Arora, Zhengyao Jiang, Dixing Xu & Dhruv Srikanth, Weco AI An agent keeps the code changes that improve a language model's score and discards the rest. Weco AI turns that research loop into a workshop, using Mini Parameter Golf to make training experiments small enough to inspect as they run. Zhengyao Jiang introduces autoresearch as repeated proposals and evaluations, and Dixing Xu explains the team's Aiden system and its contributions to OpenAI's Parameter Golf challenge. Vayum Arora and Dhruv Srikanth then connect those results to a practical setup that combines a modifiable source file, a fixed evaluation command and a measurable goal. The walkthrough focuses on what determines whether the search is useful: fast and stable evaluations, a clear direction for the metric, time limits, initial guidance and protection against reward hacking. Weco's experiment tree makes each attempt inspectable, with its code changes and score visible in the dashboard. The presenters show how to branch from promising work and steer a running search toward a different idea instead of simply letting it continue. Internet problems lead to a guided setup demonstration, followed by questions about reinforcement learning and combining multiple objectives. The final exercise swaps model training for a Flappy Bird controller scored across normal play, moving gaps, wind and mixed conditions. A hidden validation set and limits on what the controller can edit make the competition another lesson in designing an evaluation that rewards the behavior you want. Speaker info: Vayum Arora: - https://www.linkedin.com/in/vayum-arora/ Zhengyao Jiang: - https://zhengyaojiang.github.io/ - https://www.linkedin.com/in/zhengyao-jiang-387b44145/ Dixing Xu: - https://dex.moe/ - https://linkedin.com/in/dex-xu Dhruv Srikanth: - https://www.linkedin.com/in/dhruv-srikanth Workshop resource: - https://github.com/WecoAI/mini-pg Timestamps: 0:00 - Presenters and the autoresearch loop 3:35 - Workshop plan and AIDE background 5:39 - Parameter Golf and Aiden 8:30 - Mini Parameter Golf 9:37 - Weco, Modal, and repository setup 11:10 - Anatomy of an autoresearch run 12:06 - Fast evaluations and reward hacking 13:02 - Define the metric and search budget 14:25 - Time limits and initial guidance 16:31 - Structure, observability, and steering 17:29 - Inspect the experiment tree 19:36 - Human ideas and research breakthroughs 20:47 - Guided setup and live dashboard 24:41 - Reinforcement learning and multiple objectives 26:18 - Evaluate and steer an experiment 28:46 - The Flappy Weco competition 30:08 - Environments, validation, and rules 32:30 - Exercise instructions and wrap up Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/parameter-golf-with-autoresearch-vayum-arora-zhengy-1eb2f2785a59 · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks > Missing: no stored document chunks are available.