Peng et al.
MLLMs miss text readable by squinting; diagnosis and training method recovers it
Peng et al. show that spatial selection and high-frequency interference cause the failure, and their Squint OPSD method reaches 88.61% ANLS on a production-derived benchmark, closing the gap without auxiliary views at inference.
Chinese Tech
Chen Peng · Shan Huang · Jianbo Huang · Linrui Qin · Qibo Sun · Mengwei Xu
Beijing University of Posts and Telecommunications · Tencent
Research Digest··3 min read
The authors define Squint Text: text embedded in an image's coarse structure that humans can read by squinting but multimodal LLMs (MLLMs) often report as absent.
Why this paper
From Tencent and Beijing University of Posts and Telecommunications
In one line
Multimodal LLMs fail to read squint text because glyph identity is weakly encoded; training with downscaled views and transcriptions lifts ANLS from 26.50 to 88.61.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§