Public experiment dashboard

简体中文

This dashboard brings together audited metrics committed to the repository for search, advertising, recommendation, foundation models, structured decisions, multimodal models, LLM post-training, and agents. It reads stable metrics under docs/ only. Local checkpoints, private data, and unaudited temporary runs are not included.

Agent results by evaluation tier

L1 mechanism diagnostics verify that mechanisms described in a paper can execute, including planning, memory, tool use, and RL credit assignment. Success metrics can saturate on a deterministic mini-suite, so they do not measure model capability and are not used to rank methods. Only L2/L3 results explicitly marked as formal comparisons can be used to compare performance.

—Matching results

—Methods

—Datasets

—Research domains

DomainAll domains Search SortDomain and methodMethod name

Loading public metrics…

How the data is presented

• scripts/generate_public_experiment_dashboard.py deterministically generates the dashboard data from committed metrics.
• Formal evaluation cards prioritize key metrics such as NDCG, recall, accuracy, AUC, success rate, loss, and latency.
• Agent L1 cards show the evaluation tier, benchmark, episode count, simulated cost, and mechanism counts. Saturated success scores remain in the expandable raw record only.
• Other fields are available under “View all metrics,” rather than crowding the page with raw JSON.
• Select “View metric artifact” to open the corresponding original JSON on GitHub and trace the result back to its evidence.