Detached Skip-Links and $R$-Probe: Decoupling Feature Aggregation from Gradient Propagation for MLLM OCR

TL;DR

This paper introduces Detached Skip-Links and R-Probe to improve feature fusion and gradient stability in multimodal large language models, significantly enhancing OCR performance and general multimodal task results.

Contribution

We propose Detached Skip-Links to reduce gradient interference in multi-layer feature fusion and introduce R-Probe to assess the preservation of fine-grained visual information in LLMs.

Findings

Improved OCR benchmark performance across multiple ViT backbones.

Enhanced stability and convergence in training without additional learnable parameters.

Consistent gains on general multimodal tasks at scales up to 7 million samples.

Abstract

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.