Hedge-Bench - Theme Coverage: leaderboard
Hedge-Bench 1.0 (Trata, 2026) tests agents on 102 open-ended financial-analyst research tasks built from recorded hedge fund analyst discussions, run with the Terminus 2 agent over a sandboxed document corpus and graded by an LLM judge against the expert analysts rubric moves. This board is the share (%) of rubric themes an answer covers with grounded, non-hallucinated moves, averaged over tasks; higher is better.
Metric: Themes covered (%), macro-averaged over tasks: a theme counts as covered once enough of its grounded expert moves are hit (max(1, min(n-1, 3)) of n); 102 open-ended financial-analyst research tasks derived from recorded expert analyst discussions, run in the Harbor harness with the Terminus 2 agent over a sandboxed document corpus, eight trials per task, graded by a Gemini-3.1-Pro judge against expert rubric moves with hallucination-supported moves discounted; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 56.4 | #161 |
| 2 | Claude Opus 4.7 | 53.6 | #79 |
| 3 | GPT-5.5 | 48.1 | #49 |
| 4 | Gemini 3.5 Flash | 48 | #101 |
| 5 | Claude Opus 4.8 | 47.3 | #56 |
| 6 | Claude Haiku 4.5 | 32.5 | #568 |
| 7 | Gemini 3.1 Pro (Preview) | 30.5 | #94 |
| 8 | GPT-5.4 Mini | 19.8 | #389 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=hedge-bench-theme-coverage · How It Works · Data refreshed daily, snapshot 2026-11-13.