Google Research
·
3 months ago
Researchers evaluated how well 25 large language models align their behavioral dispositions with human preferences using situational judgment tests grounded in validated psychological questionnaires. Smaller models (<25B parameters) showed near-chance alignment with human consensus, while frontier models (>120B parameters) achieved close to perfect alignment only when human consensus was unanimous, plateauing at 80s percent otherwise. All 25 models systematically exhibited overconfidence in their decisions and showed misalignment between self-reported and actual behavioral tendencies, revealing gaps in how well models navigate human social dynamics.
Together AI
·
3 months ago
Researchers developed DBPlanBench, a system that uses LLMs to optimize database query execution plans by identifying and fixing inefficiencies in join ordering and filtering strategies. The system achieved a 4.78x speedup on a complex test query and optimized 60.8% of sampled queries by more than 5%, using a compact JSON serialization that reduces plan representations by approximately 10x. The approach enables databases to improve query performance without modifying the underlying database engine, and optimizations discovered on small-scale test databases successfully transfer to larger production-scale databases.
Together AI
·
3 months ago
Wan 2.7, a four-model video suite supporting generation, continuation, reference-driven workflows, and editing, is now available on Together AI starting with text-to-video. The text-to-video model is available today at $0.10 per second of generated video, with image-to-video, reference-to-video, and video editing capabilities rolling out soon. Developers can now perform video generation, continuation, reference-based control, and editing through a single API platform rather than using disconnected tools.
