Search papers, labs, and topics across Lattice.
Institute of Foundation Models
7
0
11
Key contribution not extracted.
This paper introduces Magenta, a training-free agentic pipeline that, given only a natural-language problem, produces an answer, expresses it as a Lean 4 statement, and constructs a machine-checked proof, which achieves 100% accuracy across all evaluated olympiad benchmarks.
Auxiliary draft models are no longer necessary for speculative decoding: distilling lightweight diffusion heads directly into standard LLMs yields lossless 3脳 generation speedups even at peak batch sizes.
Training on compound reasoning traces yields better generalization than isolated atomic modules, reshaping our understanding of how LLMs can learn to reason.
Achieve SOTA reasoning performance with a 30B parameter model using up to 95% fewer tokens by explicitly controlling when and how deeply an LLM plans.
Today's best AI agents still fail more than half the time on real-world tasks combining vision, search, and coding, revealing critical gaps in reasoning and tool use.
Forget simple scaling laws: the compute-optimal number of parallel rollouts in LLM RL plateaus, revealing distinct mechanisms for easy vs. hard problems.
