Research map / Learning theory and algorithms / Prediction and decision losses
Procedural Generation Of Algorithm Discovery Tasks in Machine Learning
Alexander David Goldie, Zilin Wang, Adrian Hayler, Deepak Nathani, Edan Toledo, Ken Thampiratwong, Aleksandra Kalisz, Michael Beukman, Alistair Letcher, Shashank Reddy Chirra, Clarisse Wibault, Theo Wolf and 7 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase tasks used for prompt optimisation from 1 to 30 (ablation) | ADA prompt tuning (Deepseek-v3.2) · In-distribution domains | meta-test Elo: 965 -> 1067 | Improved |
| increase tasks used for prompt optimisation from 1 to 30 (ablation) | ADA prompt tuning (Deepseek-v3.2) · Held-out domains | meta-test Elo: 983 -> 1024 | Improved |
| increase tasks used for prompt optimisation from 1 to 30 (ablation) | ADA prompt tuning (Deepseek-v3.2) · In-distribution domains | meta-train Elo: 955 -> 1047 | No clear effect |
| start modules from baseline implementations instead of empty (ablation) | MLGym ADA (Deepseek-v3.2) · Quick-to-run domains | success rate: 75.4% -> 92.4% | Improved |
| start modules from baseline implementations instead of empty (ablation) | MLGym ADA (Deepseek-v3.2) · Quick-to-run domains | Elo: 899 -> 1101 | Improved |
| increase number of editable modules in On-Policy RL (ablation) | MLGym ADA · On-Policy RL module sweep (63 tasks) | success rate | Hurt |
| edit all modules instead of a single module (ablation) | MLGym ADA (GPT-OSS 120B) · DiscoBench | success rate: 59.5% -> 17.8% | Hurt |
| edit all modules instead of a single module (ablation) | MLGym ADA (Deepseek-v3.2) · DiscoBench | success rate: 71.3% -> 40.5% | Hurt |
| evaluate on meta-test instead of meta-train for LLM ADAs versus fixed baseline | MLGym ADA · DiscoBench | Elo gap to baseline | Mixed |
| use Success@3 instead of aggregated success rate | MLGym ADA · DiscoBench | success rate: 10-30 percentage points higher | Improved |
Closest papers by meaning
- Discovering Preference Optimization Algorithms with and for Large Language Models (NeurIPS 2024 · similarity 0.84)
- Understanding the Challenges in Iterative Generative Optimization with LLMs (NeurIPS 2026 · similarity 0.84)
- DisCo-DSO: Coupling Discrete and Continuous Optimization for Efficient Generative Design in Hybrid Spaces (AAAI 2025 · similarity 0.84)
- Discovering Temporally-Aware Reinforcement Learning Algorithms (ICLR 2024 · similarity 0.84)
- DeepSearch: Overcome the Bottleneck of Reinforcement Learning with Verifiable Rewards via Monte Carlo Tree Search (ICLR 2026 · similarity 0.83)
- Trace is the Next AutoDiff: Generative Optimization with Rich Feedback, Execution Traces, and LLMs (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
