Stargazer - Medium: leaderboard

Stargazer (2026) is an agent environment for physics-grounded model fitting: given a radial-velocity time series and the host star's mass, the agent works in a ReAct loop with a Python REPL and a submit tool to infer the planetary system (number of planets, periods, amplitudes, eccentricities). A submission passes only if it clears all four criteria: statistical detection (delta BIC), residual RMS, a match score against the true orbits and the correct planet count. The 100 synthetic tasks are stratified into Easy, Medium and Hard tiers by a physics-based difficulty score (SNR, planet multiplicity, period coverage). The boards are the per-tier pass rates of Table 1, averaged over three independent runs; the Hard tier is not published because no model exceeds 6% there, and the 20 real archival tasks are not published because every model scores 0. This board is the medium tier.

Metric: Pass rate (%) on the Medium tier of synthetic radial-velocity model-fitting tasks (all four criteria: delta BIC, RMS, orbit match, planet count), mean of three runs. Source: arxiv.org. Saturation forecast: Around April 2028. 8 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)35#99
2GPT-5 Mini33.6#364
3GPT-5.3 Codex30.8#124
4GPT-5.230#185
5Qwen 3.5 Plus25#242
6O3 Mini24.2#546
7Claude Sonnet 4.622.5#173
8Kimi K2.517.9#266

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=stargazer-medium · How It Works · Data refreshed daily, snapshot 2026-11-21.