Search papers, labs, and topics across Lattice.
National Institute of Informatics, Japan
7
0
12
3
Frontier LLMs excel in social deduction games, but most fail to sustain deception, with retention rates plummeting below 50%.
Dynamic data mixing via loss trajectories boosts performance across tasks while using just 25% of the proxy compute budget.
Adding just one spatial word can lead MLLMs to consistently choose the wrong answer, revealing a critical vulnerability in their reasoning processes.
Multimodal LLMs encode chart information but fail to route it effectively for predictions, revealing a critical gap in scientific claim verification.
Clinically-focused NER for prion diseases is now possible with PrionNER, a new dataset that exposes the limitations of existing models in extracting fine-grained, complex information from biomedical literature.
RL models trained with verifiable rewards exhibit a surprising deductive-over-abductive reasoning asymmetry, even in controlled environments, suggesting a fundamental challenge in current RLVR approaches.
LLMs alone can't capture the nuances of mathematical research, but injecting aspect-aware information into a heterogeneous GNN unlocks surprisingly effective paper recommendations.
