'cruel' sits next to 残酷: embeddings explain VLM script confusions
VikParuchuri · x · 2026-10-16
Vik Paruchuri's team's best explanation for VLM script confusions: words with the same meaning in different scripts end up with nearly identical embeddings—"cruel" sits close to 残酷. When the page is hard to read, the model's prior wins over the evidence: at one failure point, Chinese "cruel" scored 22.2% next-token probability, beating the English "cruel" at 10.5%, even though the page said a different word entirely. The fixed Chandra model now has only correct-script candidates in its top 5 at that position.
Related event: Datalab founder explains why VLMs confuse scripts during OCR(5 posts)→
More from Models
- NEURA launches Neuron AI, a multimodal PLA model giving robots touch, force feedback and memory — CyberRobooo · 2026-10-17
- Sentience claims per-user models: train, continually update and serve one model per person — soleio · 2026-10-17
- Four Gemini 4 Argon agents beat Minecraft in 27 hours, writing 2,783 programs — GoogleDeepMind · 2026-10-17
- Meta open-sources Muse Spark Open to mixed reception: 'trash' license, nerfed benchmark scores — gabriberton · 2026-10-17
- If new diffusion models are open-weight only, what's left for commercial use? — blackpanda8484 · 2026-10-17
- Reflection Beam OSS architecture: sliding window attention, gated attention, QK Norm — A_K_Nain · 2026-10-17