AI
Learning Native Reflection in Unified Models with Interleaved Reinforcement Learning
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 27/09/2026, 20:00:00 UTC
Briefing created: 29/09/2026, 11:27:09 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 29/09/2026, 06:18:24 UTC.
Briefings
Unified Multimodal Models (UMM) can self-repair image generations by diagnosing errors, rendering revisions, and re-evaluating, but this requires joint learning of reflection text and image generation. The proposed UMM-Reflection method uses reinforcement learning to complete reflection trajectories within a single model, leveraging sibling trajectories to compare strategies and update both reflection tokens and flow-based revisions. This approach avoids combinatorial credit assignment issues and eliminates the need for external verifiers at inference. On the BAGEL benchmark, UMM-Reflection improved GenEval by 12.05 points compared to supervised fine-tuning, with gains transferring to WISE, OneIG-Bench, and T2I-CompBench++ benchmarks.
Key points
- UMM-Reflection applies reinforcement learning to complete reflection trajectories inside a unified model, using sibling trajectories to compare strategies and update both reflection tokens and flow-based revisions.
- UMM-Reflection improves GenEval by 12.05 points on BAGEL compared to supervised fine-tuning, with gains transferring to WISE, OneIG-Bench, and T2I-CompBench++ benchmarks.
Context and limitations
Credit assignment flows across rounds and to both roles of the same model, and no verifier is needed at inference. Gains are measured on external benchmarks not used in training.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 29/09/2026, 11:27:09 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeia5xtfcginqdy3o5l4zv5kt3d7u22djilkcobgdvn3misa6aze6iq