AI
Where-OPD: Spatially Guided On-Policy Self-Distillation of MLLMs with Synthetic Scenes
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 30/09/2026, 20:00:00 UTC
Briefing created: 03/10/2026, 08:02:57 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/10/2026, 15:18:24 UTC.
Briefings
This study introduces Where-OPD, a method for on-policy self-distillation of multimodal large language models (MLLMs) using synthetic scenes. The approach provides the teacher with spatially grounded textual guidance identifying relevant visual elements, allowing the student to learn from image and question alone. Results show consistent improvements on counting, document, and chart benchmarks, with synthetic-to-real transfer yielding a 3.23-point average gain across CVBench, V*, ZoomBench, BLINK, HR-Bench, and MME-RealWorld. The method enables scalable, annotation-free post-training by using procedurally generated scenes with automatic object identities and coordinates.
Key points
- The Where-OPD method uses on-policy self-distillation to improve MLLMs by providing the teacher with spatially grounded textual guidance identifying relevant visual elements for a query.
- Post-training uses procedurally generated scenes with automatic object identities and spatial coordinates, enabling scalable and annotation-free performance improvements on counting, document, and chart understanding benchmarks.
Context and limitations
Results are limited to synthetic scenes; while performance transfers to real-world perception benchmarks, the specific transfer mechanism and generalizability to other real-world tasks remain unverified.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 03/10/2026, 08:02:57 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeieomgyaexicfhaha4ww64wehs6gax7t2carwyh7h6mtlkben65coy