PulseAugur
EN
LIVE 06:15:25

MLLMs show reasoning gap with image text; self-distillation bridges divide

Researchers have identified a significant performance gap in multimodal large language models (MLLMs) when processing text presented as images compared to standard text tokens. This "modality gap" is primarily driven by the models' reduced reasoning effort when input is visual, leading to shorter, less computational outputs. A new self-distillation fine-tuning method, which pairs image inputs with the models' own reasoning traces from text mode, effectively closes this gap, improving accuracy and transferring gains to new benchmarks. AI

IMPACT Identifies a key limitation in MLLMs and proposes a method to improve their reasoning capabilities on visual text inputs.

RANK_REASON Academic paper detailing a new finding and method for multimodal LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

COVERAGE [1]

  1. arXiv cs.CL TIER_1 English(EN) · Kaiser Sun, Xiaochuang Yuan, Hongjun Liu, Chen Zhao, Cheng Zhang, Mark Dredze, Fan Bai ·

    Reading, Not Thinking: Understanding and Bridging the Modality Gap When Text Becomes Pixels in Multimodal LLMs

    arXiv:2603.09095v2 Announce Type: replace Abstract: Multimodal large language models (MLLMs) can process text presented as images, yet they often perform worse than when the same content is provided as textual tokens. We systematically diagnose this "modality gap" by evaluating s…