RIVER Bench - Retro-Memory (Open-Ended): leaderboard
Metric: Open-ended (OE) accuracy (%) on RIVER Bench Retro-Memory (about 1.5k questions on an event earlier in the stream, 15 seconds to an hour back) questions, each free-form response judged for consistency with the reference answer by Qwen2.5-72B, each model at its recommended frame sampling (GPT-4o and Gemini 1.5 Pro 50 frames, the offline open models 16 frames, Flash-VStream its native streaming input), the question asked at its timestamp in the stream; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 39.09 | #333 |
| 2 | Gemini 1.5 Pro | 24.24 | #442 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=river-bench-retro-memory-open-ended · How It Works · Data refreshed daily, snapshot 2026-10-11.