/gaia-leaderboard

name: gaia-leaderboard
description: Fetch and display current HAL GAIA leaderboard scores and our positioning
argument-hint: "[--level=1] [--top=20]"

/gaia leaderboard

Display the current HAL GAIA leaderboard and compare with stored ruflo runs.

Usage

/gaia leaderboard
/gaia leaderboard --level=1 --top=20
/gaia leaderboard --level=2

Options

| Flag | Default | Description | |------|---------|-------------| | `--level` | `1` | Show scores for this GAIA level (1, 2, or 3) | | `--top` | `20` | How many leaderboard entries to display | | `--show-ours` | on | Overlay our stored run results in the table |

What this does

1. Fetches the GAIA leaderboard from the HAL HuggingFace space API: `https://huggingface.co/spaces/gaia-benchmark/leaderboard` 2. Parses the top N entries by L1/L2/L3 pass-rate. 3. Loads our stored runs from the `gaia-runs` memory namespace. 4. Displays a comparison table:

Rank  System                  L1%    L2%    L3%    Overall
----  ----------------------  -----  -----  -----  -------
  1   HAL (Sonnet 4.5)        74.6   55.2   31.4   60.1
  2   GPT-4o (OpenAI)         71.3   51.8   28.9   56.6
...
 --   ruflo (this session)    20.8    --     --     20.8*

`*` denotes partial run (L1 only, 53/300 questions).

Known baselines

| System | L1 | Source | |--------|----|--------| | HAL Sonnet 4.5 | 74.6% | Princeton HAL reference, 300 Q | | ruflo iter 23 | 20.8% | 53 Q, post-SOTA web_search | | ruflo iter 15 | 9.4% | 53 Q, broken web_search |

Steps Claude should follow

1. Use the WebFetch tool to retrieve leaderboard data from `https://huggingface.co/spaces/gaia-benchmark/leaderboard`. 2. If the API returns JSON parse it directly; if HTML, extract the table rows. 3. Load local runs: `npx @claude-flow/cli@latest memory list --namespace gaia-runs` 4. Render a Markdown table comparing leaderboard entries with local runs. 5. Highlight the gap between our best run and the top-10 median. 6. Suggest which configuration changes would close the gap most efficiently.