Deeper Learning

The authors identify a novel phenomenon, Boomerang Distillation, which occurs when distilling a large language model into a smaller one. In this blog post, they describe how Boomerang Distillation can be used to create entire families of LLMs of fine-grained sizes without any training from a single student-teacher pair.

• Preprint
• Code
• Models

The authorsintroduce LOTION, a framework that optimizes a continuous variant of the quantized loss surface while provably preserving all global minima of the original problem.

• Preprint

ToolUniverse is a framework for developing AI agents for science, often referred to as “AI scientists.” It provides an environment where LLMs interact with more than six hundred scientific tools, including machine learning models, databases, and simulators. ToolUniverse standardizes how AI models access and combine these tools, allowing researchers to develop, test, and evaluate AI agents for science.

• Project
• Paper
• Code

The authors use a leading cognitive model of value trade-off in polite speech to systematically examine how post-training choices like reasoning budget and alignment recipes might be affecting value trade-offs in language models.

• Preprint

The authors show that a type of synapses in the brain challenges a long-held assumption about synaptic plasticity rules. These synapses switch between more excitatory and more inhibitory in an experience-dependent manner, and contribute to online dopamine updates during reinforcement learning.

• Preprint

Gao and collaborators propose a new RL algorithm that estimates the optimal value function offline from the reference policy and performs on-policy updates using only one generation per prompt.

• Preprint
• GitHub

This work provides a sharp characterization of in-context learning (ICL) in an analytically-solvable model, which offers insights into the sample complexity and data quality requirements for ICL to happen. These insights can be applied to more complex, realistic architectures.

• Paper
• Code

The author introduces the first flow equivariant models that respect motion symmetries, leading to significantly improved generalization and sequence modeling.

• Preprint
• Code

Diffusion models are powerful generative frameworks that iteratively denoise white noise into structured data via learned score functions. Through theory and experiments, the authors demonstrate that these score functions are dominated by a linear Gaussian component.

• Preprint
• OpenReview

Kempner researchers present a new algorithm for offline reinforcement learning that features a key property: self-consistency.

• Preprint
• Code