SleuthVLM beats Gemini 3.8 Flash on long-video QA without ever hearing audio
antgroup · hf · 2026-11-05
A Bristol, Ant Group and Tsinghua paper challenges the assumption that omni models must hear audio directly:
- SleuthVLM never hears audio: speech is transcribed to text, and the VLM autonomously 'sleuths' the video with no predefined workflow
- On four omni benchmarks with videos up to 90 minutes, it averages 75.0% accuracy, 2.6 points above Gemini 3.8 Flash
- An open-source 27B version tops the Video-MME-v2 leaderboard at 77.8%, beating all open-source omni models
- Key findings: looking carefully beats glancing; hearing is not the bottleneck; omni understanding needs better agentic ability, not more modality-alignment data
More from Models
- AMD ships Day 0 support for Kimi K3.1 Pro: 2.8T params, 1M context on 8x MI355X — Kimi_Moonshot · 2026-11-05
- Claude spontaneously finds a serious OpenBLAS bug while just explaining code to a user — Aizkmusic · 2026-11-05
- Baidu open-sources ERNIE 5.5 Flash: 280B-param MoE with 14B active and 1M context — teortaxesTex · 2026-11-05
- Did Anthropic reward Opus 5.5 for saying 'let me check rather than answer from memory'? — andrey_kurenkov · 2026-11-05
- New speech-to-speech models are natural talkers but still need hybridizing to finish tasks reliably — alexisgallagher · 2026-11-05
- Chamath: Chinese models run 112x cheaper than Anthropic — $0.50 vs $56 per million tokens — 2C_ornot2C · 2026-11-05