Hitomi Yanaka

Asymmetric semantic relations are distinctly represented in language models, but the encoding of these relations reveals surprising limitations in their learning capabilities.

VLMs misinterpret spatial deictic expressions, failing to match human-like selection based on object proximity.

Even Japanese-specific LLMs fail to grasp kanji readings, revealing a significant shortcoming in their linguistic understanding.

LLMs can grasp negation to some extent, but their ability to recognize negation scope is surprisingly inconsistent and format-dependent.

LLMs may completely bypass the neglect-zero effect, challenging assumptions about their alignment with human cognitive biases.

Code-switching can degrade information retrieval performance by up to 27%, revealing a critical blind spot in current multilingual models.

VLMs already contain a rich latent space of aesthetic features that can be unlocked for personalized image ranking with just a linear readout, no fine-tuning needed.

Adapting Labovian narrative analysis to Japanese reveals the challenges and opportunities in cross-linguistic qualitative research, highlighting the need for language-specific guidelines.

NeuronMoE slashes multilingual LLM parameter counts by 40% without sacrificing performance, by cleverly allocating experts based on neuron-level language specialization rather than blunt layer-level assignments.