Deeper Learning

Model-free RL agents trained on simple tasks can be evaluated by whether or not they achieve their simple rewards, but complex open-ended environments demand richer analysis. ForageWorld introduces a naturalistic foraging benchmark where agents must explore, remember food patches, and survive across procedurally generated arenas. In our experiments, naive RNN agents develop structured exploration strategies, multi-factor patch memory, and implicit future planning. These behaviors mirror those documented in foraging insects and rodents and are recoverable only using neuroscience-inspired analysis tools. The approach generalizes across arena sizes, transfers readily to other recurrent architectures, and opens a path toward behavioral and neural transparency in increasingly complex AI systems.

• Code
• Paper

The authors introduce a Energy-Based Fine-Tuning (EBFT), a method that matches the long-range statistics of model generations to ground-truth sequences in high-dimensional feature spaces. EBFT corrects the error amplification caused by standard teacher-forced next token training while improving downstream task performance. Experiments show that EBFT matches the accuracy improvements of RLVR while requiring no external reward signal or verifier. In contrast to RLVR, EBFT simultaneously improves the validation cross-entropy.

• Preprint
• Project website
• Code

The authors introduce InputDSA, a new method to measure the similarity between two complex systems when they are driven by external inputs, like biological neural circuits or reinforcement learning agents. The method disentangles each systems’ intrinsic dynamics from its input-driven effects, enabling highly accurate, robust, and efficient comparisons of those components.

• Preprint
• Code

The authors develop a mean-field theory of task-trained recurrent networks that continuously interpolates between these regimes, and find evidence that macaque motor cortex is best captured by an intermediate level of task-specific recurrent restructuring.

• Paper
• Code
• Data

Predicting future neural activity is a critical step toward achieving real-time, closed-loop neurotechnologies. To this end, we introduce POCO, a unified forecasting model trained on…

• Paper
• Code

The authors provide a theoretical analysis demonstrating the existence of anytime learning schedules for overparameterized linear regression, and highlight the central role of weight averaging—also known as model merging—in achieving the optimal convergence rates of stochastic gradient descent.

• Preprint
• Code

Despite reaching equal performance success when trained on the same task, artificial neural networks can develop dramatically different internal solutions, much like different students solving the same math problem using completely different approaches. Our study introduces a unified framework to quantify this variability across Recurrent Neural Network (RNN) solutions, which we term solution degeneracy, and analyze what factors shape it across thousands of recurrent networks trained on memory and decision-making tasks.

• Paper

This work introduces a relational foundation model for neurological discovery and evaluates it through discovery loops that connect AI predictions to experiments in Parkinson’s disease, bipolar disorder, and Alzheimer’s disease.

• PROTON website
• Paper
• Code
• Model

This work explores using video as the primary modality for robot foundation models. Unlike static images, videos naturally encode physical dynamics and semantics of the world, providing a rich prior for physical decision-making.

• Preprint
• Hugging Face
• Project website

From Linear Directions to Convex Geometry

The authors ask the fundamental question: is the linear view of DINOv2 under the Linear Representation Hypothesis (LRH) sufficient to describe how deep vision models organize information? The authors examine the geometry and statistics of the learned concepts themselves and the results suggest that representations are organized beyond linear sparsity alone.

• Interactive demo
• Preprint