Tsinghua's AV-FABLE benchmark shows AV-LLMs near-chance at long-video deepfake forensics
Tsinghua · hf · 2026-10-27
Tsinghua and Peking University introduce AV-FABLE, a dual-timeline benchmark for sparse audio-visual manipulations in open-scene long-form videos: 16,650 pristine–manipulated pairs (998 hours) across 27 editing configurations, with separate audio and visual timelines for modality-specific temporal localization.
Key findings:
- Zero-shot prompting of audio-visual LLMs yields near-chance detection and poor localization
- Frozen perception features of AV-LLMs do encode strong forensic signals, but language outputs fail to expose them
- The proposed NFSR method reads frozen features with separate audio/visual temporal decoders, achieving the best in-domain and out-of-distribution localization, including transfer to an external long-video forensic benchmark
Benchmark, code, and model weights will be publicly released.
More from Research
- EdgeRock architecture keeps KV cache at 72 MiB for 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- Equivariant Music Transformer accepted to ISMIR 2026 with Best Paper Award nomination — umpedronosapato · 2026-10-27
- Deedy shares links to the OpenAI and Claude ETH-refutation documents — deedydas · 2026-10-27
- Claims: OpenAI and Claude models independently found subexponential 3-SAT algorithms, refuting ETH — deedydas · 2026-10-27
- ETH humanoid uses head-mounted LiDAR for onboard terrain mapping in 30ms — ChongZzZhang · 2026-10-27
- NYU and UT Austin study finds AI models overuse language patterns linked to reading difficulty — andrewaltair · 2026-10-27