Study: Verifiable rewards boost video-language model accuracy but not time understanding

A new study published on arXiv explores the impact of reinforcement learning from verifiable rewards (RLVR) on video-language models' understanding of time. Researchers fine-tuned four open models, including Qwen3 VL 8B and Gemma-3-12B, using different data strategies. The findings indicate that while verified training significantly boosts in-domain accuracy, especially for less capable models, it does not necessarily improve out-of-domain performance and can even degrade it without a mix of real-world data. Crucially, the study found no evidence that RLVR or event-order rewards instill a genuine understanding of temporal order in these models. AI

IMPACT This research suggests that current training methods for video-language models may not effectively teach temporal understanding, despite improvements in benchmark accuracy.

RANK_REASON The cluster contains a research paper detailing a controlled multi-model study on video-language models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• arXiv
• Avyay Sadhu
• CLEVRER
• Gemma-3-12B
• Hugging Face
• MVBench
• Qwen2.5-VL-7B
• Qwen3 VL 8B

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →