Detached Skip-Links and $R$-Probe: Decoupling Feature Aggregation from Gradient Propagation for MLLM OCR
This addresses a specific optimization issue in MLLMs for OCR and multimodal tasks, offering an incremental but effective solution.
The paper tackles the problem of multimodal large language models (MLLMs) failing on OCR tasks due to gradient interference in skip pathways during multi-layer feature fusion, and proposes Detached Skip-Links to decouple feature aggregation from gradient propagation, resulting in consistent improvements on OCR-centric benchmarks and clear gains on general multimodal tasks across multiple ViT backbones and scales up to 7M training samples.
Multimodal large language models (MLLMs) excel at high-level reasoning yet fail on OCR tasks where fine-grained visual details are compromised or misaligned. We identify an overlooked optimization issue in multi-layer feature fusion. Skip pathways introduce direct back-propagation paths from high-level semantic objectives to early visual layers. This mechanism overwrites low-level signals and destabilizes training. To mitigate this gradient interference, we propose Detached Skip-Links, a minimal modification that reuses shallow features in the forward pass while stopping gradients through the skip branch during joint training. This asymmetric design reduces gradient interference, improving stability and convergence without adding learnable parameters. To diagnose whether fine-grained information is preserved and usable by an LLM, we introduce $R$-Probe, which measures pixel-level reconstructability of projected visual tokens using a shallow decoder initialized from the first quarter of the LLM layers. Across multiple ViT backbones and multimodal benchmarks, and at scales up to 7M training samples, our approach consistently improves OCR-centric benchmarks and delivers clear gains on general multimodal tasks.