KAIST's VETO suppresses text-only shortcuts hidden behind VideoLLM SFT gains
Youngwoo Shin · hf · 2026-11-20
A KAIST/PYLER paper shows VideoLLMs learn a shortcut during SFT — answering from textual regularities while bypassing the video — which inflates in-distribution accuracy without real gains. VETO (Video-Evidence Training via Orthogonal projection) calibrates the text-only activation subspace of the base model and projects SFT updates onto its orthogonal complement, suppressing the shortcut as it forms. Across seven VideoLLMs it preserves SFT gains and beats SFT on six out-of-shift benchmarks.
More from Research
- ICLR 2027 desk-rejects papers after reviewers fail to file AI-use disclosure forms — algo_diver · 2026-11-20
- Chorus paper: shared n-gram tables cut memory 87% while beating DeepSeek's Engram — brown2green · 2026-11-20
- Synthetic benchmarks for concept bottleneck models isolate failure modes, NeurIPS 2026 — berkustun · 2026-11-20
- Inhabitant: training-free LLM agent completes 74/75 long-horizon household tasks — Yao Zhang · 2026-11-20
- EPFL team unveils unified 3D generative model for structure-based drug design and docking — HannesStaerk · 2026-11-20
- ICLR 2027 Review Score Distribution: 40,160 Submissions, a 3.0 Average Only Ranks Top 14% — May_F1_ · 2026-11-20