DexHoldem Agentic Perception - Loop Stage: leaderboard
DexHoldem agentic perception bench (Chen et al., 2026): an agent in its native harness (Codex, Claude Code or Gemini CLI) at a medium thinking budget parses one captured poker-table image, given earlier states, into the structured game state; 36 problems, mean of three runs. Accuracy in percent on the loop stage (whether the robot is acting, idle, settled between steps, in a win or loss, or needs recovery or help), on all 36 problems; higher is better.
Metric: Field accuracy (%; loop stage among seven values such as acting, idle, win or lose, all 36 problems; 36 tabletop states from real games, mean of three runs; the perceiver parses the image with its harness at a medium thinking budget). Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (Medium) | 72.2 | #48 (GPT-5.5) |
| 2 | GPT-5.4 (Medium) | 65.7 | #121 (GPT-5.4) |
| 3 | Gemini 3 Flash (Medium) | 63.9 | #159 (Gemini 3 Flash) |
| 4 | GPT-5.4 Mini (Medium) | 56.5 | #389 (GPT-5.4 Mini) |
| 5 | Claude Sonnet 4.6 (Medium) | 46.3 | #158 (Claude Sonnet 4.6) |
| 6 | Claude Opus 4.7 (Medium) | 43.5 | #75 (Claude Opus 4.7) |
| 7 | Gemini 3.1 Flash Lite (Medium) | 27.8 | #363 (Gemini 3.1 Flash Lite) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=dexholdem-agentic-perception-loop-stage · How It Works · Data refreshed daily, snapshot 2026-11-12.