AI
VideoLoop: Looped Working Memory Against Semantic Thrashing in Long-Form Video Agents
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 28/09/2026, 20:00:00 UTC
Briefing created: 30/09/2026, 04:36:19 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 30/09/2026, 04:18:25 UTC.
Briefings
VideoLoop addresses semantic thrashing in long-form video agents by coupling reasoning loops with an unbounded filesystem for artifacts and a bounded working memory rewritten after each step. Experiments show VideoLoop improves four popular LVLM backbones by an average of 4.2% on VideoMME (long), with Gemini 3.1 Pro reaching 88.3% on VideoMME (long). A blind judge correctly answered 81.1% of hardest questions with VideoLoop context versus 60.9% for append-only agents, indicating effective mitigation of semantic thrashing.
Key points
- VideoLoop introduces a multimodal agent architecture with an outer reasoning loop and an inner loop that retrieves artifacts from an unbounded filesystem after each step, rewriting a bounded working memory.
- VideoLoop improves four popular LVLM backbones by an average of 4.2% on VideoMME (long), with Gemini 3.1 Pro achieving 88.3% on VideoMME (long) and 88.8% on VideoMMMU.
Context and limitations
Results are based on experiments with four popular LVLM backbones; specific agent configurations for other backbones are not reported. Blind judge evaluation was conducted on the hardest quarter of VideoMME (long) questions only.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 30/09/2026, 04:36:19 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeifpg6codsde3h6qpaqqqqmlmh5pqjj4xls5sz25ovsijvqpofgwnu