Current benchmarks for evaluating the chemical reasoning capabilities of Large Language Models (LLMs) are limited by oversimplified tasks, lack of process-level evaluation, and misalignment with expert-level chemistry skills. To address these issues, we introduce SUPERChem, a benchmark of 500 expert-curated reasoning-intensive chemistry problems, covering diverse subfields and provided in both multimodal and text-only formats. Original content and an iterative curation pipeline eliminate flawed items and mitigate data contamination. Each problem is paired with an expert-authored solution path, enabling Reasoning Path Fidelity (RPF) scoring to evaluate reasoning quality beyond final-answer accuracy. Evaluations against a human baseline of 40.3% accuracy show that even the best-performing model, GPT-5 (High), reaches only 38.5%, followed closely by Gemini 2.5 Pro (37.9%) and DeepSeek-V3.1-Think (37.3%). SUPERChem elicits multi-step, multimodal reasoning, reveals model-dependent effects of visual information, and distinguishes high-fidelity reasoners from heuristic ones. By providing a challenging benchmark and a reliable evaluation framework, SUPERChem aims to facilitate the advancement of LLMs toward expert-level chemical intelligence. The dataset of the benchmark is available at https://huggingface.co/datasets/ZehuaZhao/SUPERChem.

This work explores the challenge of building "Machines that Can Remember", framing long-term memory as the problem of efficient ultra-long context modeling. We argue that this requires three key properties: sparsity, random-access flexibility, and length generalization. To address ultra-long-context modeling, we leverage Hierarchical Sparse Attention (HSA), a novel attention mechanism that satisfies all three properties. We integrate HSA into Transformers to build HSA-UltraLong, which is an 8B-parameter MoE model trained on over 8 trillion tokens and is rigorously evaluated on different tasks with in-domain and out-of-domain context lengths to demonstrate its capability in handling ultra-long contexts. Results show that our model performs comparably to full-attention baselines on in-domain lengths while achieving over 90% accuracy on most in-context retrieval tasks with contexts up to 16M. This report outlines our experimental insights and open problems, contributing a foundation for future research in ultra-long context modeling. Figure 1: Despite being pre-trained with an 8K context window and mid-trained up to 32K, HSA-UltraLong achieves near-perfect accuracy on S-NIAH even at a 16M-token context length. The red dashed line at 32K marks the boundary between in-domain (left) and out-of-domain (right).

Despite massive investments in scale, deep models for click-through rate (CTR) prediction often exhibit rapidly diminishing returns - a stark contrast to the smooth, predictable gains seen in large language models. We identify the root cause as a structural misalignment: Transformers assume sequential compositionality, while CTR data demand combinatorial reasoning over high-cardinality semantic fields. Unstructured attention spreads capacity indiscriminately, amplifying noise under extreme sparsity and breaking scalable learning. To restore alignment, we introduce the Field-Aware Transformer (FAT), which embeds field-based interaction priors into attention through decomposed content alignment and cross-field modulation. This design ensures model complexity scales with the number of fields F, not the total vocabulary size n >> F, leading to tighter generalization and, critically, observed power-law scaling in AUC as model width increases. We present the first formal scaling law for CTR models, grounded in Rademacher complexity, that explains and predicts this behavior. On large-scale benchmarks, FAT improves AUC by up to +0.51% over state-of-the-art methods. Deployed online, it delivers +2.33% CTR and +0.66% RPM. Our work establishes that effective scaling in recommendation arises not from size, but from structured expressivity-architectural coherence with data semantics.

Failure cases of GET . It is worth noting that the Gaussian equivalence property (Theorem 3) may no longer hold if we train the features longer. In particular, because of our mean-field parameteri-zation, the first-layer weight W needs to travel sufficiently far away from initialization to achieve small training loss (see Figure 2). Hence in our experimental simulations (where n,d,N are large but finite), as the number of steps t increases, we expect the Gaussian equivalence predictions to become inaccurate at some point. This transition is empirically demonstrated in Figure 4(a).

We study the gradient-based training of large-depth residual networks (ResNets) from standard random initializations. We show that with a diverging depth $L$, a fixed embedding dimension $D$, and an arbitrary hidden width $M$, the training dynamics converges to a Neural Mean ODE training dynamics. Remarkably, the limit is independent of the scaling of $M$, covering practical cases of, say, Transformers, where $M$ (the number of hidden units or attention heads per layer) is typically of the order of $D$. For a residual scale $Θ_D\big(\fracα{LM}\big)$, we obtain the error bound $O_D\big(\frac{1}{L}+ \fracα{\sqrt{LM}}\big)$ between the model's output and its limit after a fixed number gradient of steps, and we verify empirically that this rate is tight. When $α=Θ(1)$, the limit exhibits complete feature learning, i.e. the Mean ODE is genuinely non-linearly parameterized. In contrast, we show that $α\to \infty$ yields a \lazy ODE regime where the Mean ODE is linearly parameterized. We then focus on the particular case of ResNets with two-layer perceptron blocks, for which we study how these scalings depend on the embedding dimension $D$. We show that for this model, the only residual scale that leads to complete feature learning is $Θ\big(\frac{\sqrt{D}}{LM}\big)$. In this regime, we prove the error bound $O\big(\frac{1}{L}+ \frac{\sqrt{D}}{\sqrt{LM}}\big)$ between the ResNet and its limit after a fixed number of gradient steps, which is also empirically tight. Our convergence results rely on a novel mathematical perspective on ResNets : (i) due to the randomness of the initialization, the forward and backward pass through the ResNet behave as the stochastic approximation of certain mean ODEs, and (ii) by propagation of chaos (that is, asymptotic independence of the units) this behavior is preserved through the training dynamics.

A key property of deep neural networks (DNNs) is their ability to learn new features during training. This intriguing aspect of deep learning stands out most clearly in recently reported Grokking phenomena. While mainly reflected as a sudden increase in test accuracy, Grokking is also believed to be a beyond lazy-learning/Gaussian Process (GP) phenomenon involving feature learning. Here we apply a recent development in the theory of feature learning, the adaptive kernel approach, to two teacher-student models with cubic-polynomial and modular addition teachers. We provide analytical predictions on feature learning and Grokking properties of these models and demonstrate a mapping between Grokking and the theory of phase transitions. We show that after Grokking, the state of the DNN is analogous to the mixed phase following a first-order phase transition. In this mixed phase, the DNN generates useful internal representations of the teacher that are sharply distinct from those before the transition.

We investigate the capabilities of transformer large language models (LLMs) on relational reasoning tasks involving abstract symbols. Such tasks have long been studied in the neuroscience literature as fundamental building blocks for more complex abilities in programming, mathematics, and verbal reasoning. For (i) regression tasks, we prove that transformers generalize when trained, but require astonishingly large quantities of training data. For (ii) next-token-prediction tasks with symbolic labels, we show an "inverse scaling law": transformers fail to generalize as their embedding dimension increases. For both settings (i) and (ii), we propose subtle transformer modifications which can reduce the amount of data needed by adding two trainable parameters per head.

Deep learning has emerged as the standard approach to exploiting massive high-dimensional datasets. At the core of its success lies its capability to learn effective features with fairly blackbox architectures without suffering from the curse of dimensionality. To explain this success, two structural properties of data are commonly conjectured: (i) a low-dimensional structure that SGD-trained neural networks are able to adapt to; (ii) a hierarchical structure that neural networks can leverage with SGD training. In particular, From a statistical viewpoint: A line of work [Bac17, SH20, KK16, BK19] has investigated the sample complexity of learning with deep neural networks, decoupled from computational considerations. By directly considering global solutions of empirical risk minimization (ERM) problems over arbitrarily large neural networks and sparsity inducing norms, they showed that deep neural networks can overcome the curse of dimensionality on classes of functions with low-dimensional and hierarchical structures. However, this approach does not provide efficient algorithms: instead, a number of works have shown computational hardness of ERM problems [BR88, KS09, DLSS14] and it is unclear how much this line of work can inform practical neural networks, which are trained using SGD and variants.

These lecture notes are based on two series of lectures given by Andrea Montanari, first at the "Deep Learning Theory Summer School" at Princeton from July 27th to August 4th 2021, and then at the summer school "Statistical Physics & Machine Learning", that took place at Les Houches School of Physics in France from 4th to 29th July 2022. Theodor Misiakiewicz was Teaching Assistant to AM's lectures at the "Deep Learning Theory Summer School". We thank Boris Hanin for organizing the summer school in Princeton, and Florent Krzakala and Lenka Zdeborová from EPFL for organizing the summer school in Les Houches.

Understanding how the dynamics in biological and artificial neural networks implement the computations required for a task is a salient open question in machine learning and neuroscience. In particular, computations requiring complex memory storage and retrieval pose a significant challenge for these networks to implement or learn. Recently, a family of models described by neural ordinary differential equations (nODEs) has emerged as powerful dynamical neural network models capable of capturing complex dynamics. Here, we extend nODEs by endowing them with adaptive timescales using gating interactions. We refer to these as gated neural ODEs (gnODEs). Using a task that requires memory of continuous quantities, we demonstrate the inductive bias of the gnODEs to learn (approximate) continuous attractors. We further show how reduced-dimensional gnODEs retain their modeling power while greatly improving interpretability, even allowing explicit visualization of the structure of learned attractors. We introduce a novel measure of expressivity which probes the capacity of a neural network to generate complex trajectories. Using this measure, we explore how the phase-space dimension of the nODEs and the complexity of the function modeling the flow field contribute to expressivity. We see that a more complex function for modeling the flow field allows a lower-dimensional nODE to capture a given target dynamics. Finally, we demonstrate the benefit of gating in nODEs on several real-world tasks.