Deception Detection on Real-life Trial (Honest Leave-One-Speaker-Out)
67.8AccuracySelf-hosted features → gradient-boosting
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Self-hosted features → gradient-boostingTrained?=Yes, LLM?=No, Video sent?=No2026.06 | 67.8 | 74.1 | |
| Self-hosted + LLM (trained late fusion)Trained?=Yes, LLM?=Yes, Video sent?=Yes2026.06 | 67.8 | 75.2 | |
| LLM judges our features — Gemini 2.5 ProTrained?=No, LLM?=Yes, Video sent?=No2026.06 | 66.9 | 70.4 | |
| LLM watches raw video — Gemini 2.5 ProTrained?=No, LLM?=Yes, Video sent?=Yes2026.06 | 66.9 | 74.9 | |
| LLM judges our features — Claude Opus 4.8Trained?=No, LLM?=Yes, Video sent?=No2026.06 | 62 | 75.5 | |
| Majority baselineVideo sent?=No2026.06 | 50.4 | 50 |