Reinforcement learning (GRPO) on non-chain-of-thought egocentric QA data lifts Qwen2.5-VL-3B to 73.7% on EgoSchema, but the result is clouded by unverified train/test separation and a supervised baseline that scores 74.0%.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
EgoVLM: Policy Optimization for Egocentric Video Understanding
Reinforcement learning (GRPO) on non-chain-of-thought egocentric QA data lifts Qwen2.5-VL-3B to 73.7% on EgoSchema, but the result is clouded by unverified train/test separation and a supervised baseline that scores 74.0%.