Search papers, labs, and topics across Lattice.
11
1
11
6
Natural paper revisions can be harnessed to train AI agents for precise and context-aware editing of complex scientific diagrams.
C8 and C12 carbon rings challenge our understanding of electronic structure by violating Hund's rule, previously thought unique to graphene.
Current LLMs only achieve 27.3% accuracy in reasoning about scientific lineage, revealing a critical gap in their compositional capabilities.
Transforming multimodal resources into executable agent skills boosts performance by nearly 12 percentage points, showcasing the power of diverse learning materials.
LLMs struggle to connect identified root causes to their causal paths, achieving only 61.5% success in grounding diagnoses despite a 76% identification rate.
Forget blindly chasing teacher-student disagreement in on-policy distillation – focusing on *learnable* disagreement, where the teacher nudges the student within its existing possibilities, unlocks surprisingly efficient learning.
Model-generated skills can actually hurt agent performance, and bigger models don't necessarily make for better skill extractors or consumers.
SkillOpt transforms agent skill development into a reproducible optimization process, achieving state-of-the-art results by treating skills as trainable parameters.
Hierarchical planning and self-reflection can finally wrangle AIGC tools into producing coherent, visually consistent webpages.
Today's best text-to-audio-video models may look and sound impressive, but they still struggle with basic physics, coherent speech, and even rendering text correctly.
Current image generation models fall far short of the mark when it comes to the structured and multi-constraint demands of real-world commercial design, as revealed by a new systematic benchmark.
