#DeeperLearningBlog
LightNews LightNews
kempnerinstitute.bsky.social
New in the #DeeperLearningBlog: Kempner researchers Nikhil Anand and Chloe Su discuss new work on how numerical precision can impact the accuracy and stability of #LLMs.

kempnerinstitute.harvard.edu/research/dee...

#AI

(1/2)
Characterization and Mitigation of Training Instabilities in Microscaling Formats - Kempner Institute
This research uncovers consistent training instabilities when using new, highly efficient low-precision formats, which has implications for the development of next-generation AI. By pinpointing the ro...
kempnerinstitute.harvard.edu
June 27, 2025 at 2:31 PM
kempnerinstitute.bsky.social
New in the #DeeperLearningBlog: @shunnnli.bsky.social ‬ describes the discovery of a type of synapse that switches between more excitatory and more inhibitory in an experience-dependent manner.

bit.ly/40OYYw4

#neuroscience #dopamine #RL @blsabatini.bsky.social
ANN-like Synapses in the Brain Mediate Online Reinforcement Learning - Kempner Institute
Synaptic plasticity rules in the brain are normally thought of as changing synaptic weights but not signs, unlike artificial neural networks. We show that a type of synapse in the […]
bit.ly
August 4, 2025 at 7:54 PM
kempnerinstitute.bsky.social
New in the #DeeperLearningBlog: Zhaolin Gao and collaborators including the #KempnerInstitute's Kianté Brantley presents a powerful new #RL algorithm tailored for reasoning tasks with #LLMs that updates using only one generation per prompt.

bit.ly/44US1Mt

@xkianteb.bsky.social #AI
Accelerating RL for LLM Reasoning with Optimal Advantage Regression - Kempner Institute
Recent LLM advances show the effectiveness of RL with rule-based rewards, but methods like GRPO and PPO are costly due to critics or multiple generations per prompt. We propose a […]
bit.ly
July 30, 2025 at 8:05 PM
kempnerinstitute.bsky.social
New in the #DeeperLearningBlog: #KempnerInstitute research fellow @andykeller.bsky.social introduces the first flow equivariant neural networks, which reflect motion symmetries, greatly enhancing generalization and sequence modeling.

bit.ly/451fQ48

#AI #NeuroAI
Flow Equivariant Recurrent Neural Networks - Kempner Institute
Sequence transformations, like visual motion, dominate the world around us, but are poorly handled by current models. We introduce the first flow equivariant models that respect these motion symmetrie...
bit.ly
July 22, 2025 at 1:21 PM
kempnerinstitute.bsky.social
New in the #DeeperLearningBlog: #KempnerInstitute researchers @binxuwang.bsky.social and John J. Vastola explain their work uncovering the linear Gaussian structure in diffusion models and the potential to use it to enhance performance.
bit.ly/4lCauDv
#AI #DiffusionModels
The Hidden Linear Structure in Diffusion Models and its Application in Analytical Teleportation - Kempner Institute
Diffusion models are powerful generative frameworks that iteratively denoise white noise into structured data via learned score functions. Through theory and experiments, we demonstrate that these sco...
bit.ly
July 18, 2025 at 6:59 PM
kempnerinstitute.bsky.social
New in the #DeeperLearningBlog: the #KempnerInstitute's Mary Letey presents work recently published in PNAS that offers generalizable insights into in-context learning (ICL) in an analytically-solvable model architecture.

bit.ly/4lPK15p

#AI @pnas.org

(1/2)
Solvable Model of In-Context Learning Using Linear Attention - Kempner Institute
Attention-based architectures are a powerful force in modern AI. In particular, the emergence of in-context learning enables these models to perform tasks far beyond the original next-token prediction...
kempnerinstitute.harvard.edu
July 28, 2025 at 7:24 PM
kempnerinstitute.bsky.social
New in the #DeeperLearningBlog: Researchers from the #KempnerInstitute, Cornell, and Carnegie Mellon introduce a new method for improving offline RL by scaling-up test-time compute.

kempnerinstitute.harvard.edu/research/dee...

#AI #RL

(1/2)
Scaling Offline Reinforcement Learning at Test Time - Kempner Institute
This research introduces a novel approach to scaling reinforcement learning (RL) during training and inference. Inspired by the recent work on LLM test-time scaling, we demonstrate how greater test-ti...
kempnerinstitute.harvard.edu
July 14, 2025 at 7:47 PM