On Computation and Reinforcement Learning

How does the amount of compute available to a reinforcement learning (RL) policy affect its learning? Can policies using a fixed amount of parameters, still benefit from additional compute? The standard RL framework does not provide a language to answer these questions formally. Empirically, deep RL policies are often parameterized as neural networks with static architectures, conflating the amount of compute and the number of parameters. In this paper, we formalize compute bounded policies and prove that policies which use more compute can solve problems and generalize to longer-horizon tasks that are outside the scope of policies with less compute. Building on prior work in algorithmic learning and model-free planning, we propose a minimal architecture that can use a variable amount of compute. Our experiments complement our theory. On a set 31 different tasks spanning online and offline RL, we show that $(1)$ this architecture achieves stronger performance simply by using more compute, and $(2)$ stronger generalization on longer-horizon test tasks compared to standard feedforward networks or deep residual network using up to 5 times more parameters.

Parity, Sensitivity, and Transformers

The transformer architecture is almost a decade old. Despite that, we still have a limited understanding of what this architecture can or cannot compute. For instance, can a 1-layer transformer solve PARITY -- or more generally -- which kinds of transformers can do it? Known constructions for PARITY have at least 2 layers and employ impractical features: either a length-dependent positional encoding, or hardmax, or layernorm without the regularization parameter, or they are not implementable with causal masking. We give a new construction of a transformer for PARITY with softmax, length-independent and polynomially bounded positional encoding, no layernorm, working both with and without causal masking. We also give the first lower bound for transformers solving PARITY -- by showing that it cannot be done with only one layer and one head.

Cost-Efficient RAG for Entity Matching with LLMs: A Blocking-based Exploration

Retrieval-augmented generation (RAG) enhances LLM reasoning in knowledge-intensive tasks, but existing RAG pipelines incur substantial retrieval and generation overhead when applied to large-scale entity matching. To address this limitation, we introduce CE-RAG4EM, a cost-efficient RAG architecture that reduces computation through blocking-based batch retrieval and generation. We also present a unified framework for analyzing and evaluating RAG systems for entity matching, focusing on blocking-aware optimizations and retrieval granularity. Extensive experiments suggest that CE-RAG4EM can achieve comparable or improved matching quality while substantially reducing end-to-end runtime relative to strong baselines. Our analysis further reveals that key configuration parameters introduce an inherent trade-off between performance and overhead, offering practical guidance for designing efficient and scalable RAG systems for entity matching and data integration.

FastVMT: Eliminating Redundancy in Video Motion Transfer

Video motion transfer aims to synthesize videos by generating visual content according to a text prompt while transferring the motion pattern observed in a reference video. Recent methods predominantly use the Diffusion Transformer (DiT) architecture. To achieve satisfactory runtime, several methods attempt to accelerate the computations in the DiT, but fail to address structural sources of inefficiency. In this work, we identify and remove two types of computational redundancy in earlier work: motion redundancy arises because the generic DiT architecture does not reflect the fact that frame-to-frame motion is small and smooth; gradient redundancy occurs if one ignores that gradients change slowly along the diffusion trajectory. To mitigate motion redundancy, we mask the corresponding attention layers to a local neighborhood such that interaction weights are not computed unnecessarily distant image regions. To exploit gradient redundancy, we design an optimization scheme that reuses gradients from previous diffusion steps and skips unwarranted gradient computations. On average, FastVMT achieves a 3.43x speedup without degrading the visual fidelity or the temporal consistency of the generated videos.

Low-Cost Underwater In-Pipe Centering and Inspection Using a Minimal-Sensing Robot

Autonomous underwater inspection of submerged pipelines is challenging due to confined geometries, turbidity, and the scarcity of reliable localization cues. This paper presents a minimal-sensing strategy that enables a free-swimming underwater robot to center itself and traverse a flooded pipe of known radius using only an IMU, a pressure sensor, and two sonars: a downward-facing single-beam sonar and a rotating 360 degree sonar. We introduce a computationally efficient method for extracting range estimates from single-beam sonar intensity data, enabling reliable wall detection in noisy and reverberant conditions. A closed-form geometric model leverages the two sonar ranges to estimate the pipe center, and an adaptive, confidence-weighted proportional-derivative (PD) controller maintains alignment during traversal. The system requires no Doppler velocity log, external tracking, or complex multi-sensor arrays. Experiments in a submerged 46 cm-diameter pipe using a Blue Robotics BlueROV2 heavy remotely operated vehicle demonstrate stable centering and successful full-pipe traversal despite ambient flow and structural deformations. These results show that reliable in-pipe navigation and inspection can be achieved with a lightweight, computationally efficient sensing and processing architecture, advancing the practicality of autonomous underwater inspection in confined environments.

Learning Compact Boolean Networks

Floating-point neural networks dominate modern machine learning but incur substantial inference cost, motivating interest in Boolean networks for resource-constrained settings. However, learning compact and accurate Boolean networks is challenging due to their combinatorial nature. In this work, we address this challenge from three different angles: learned connections, compact convolutions and adaptive discretization. First, we propose a novel strategy to learn efficient connections with no additional parameters and negligible computational overhead. Second, we introduce a novel convolutional Boolean architecture that exploits the locality with reduced number of Boolean operations than existing methods. Third, we propose an adaptive discretization strategy to reduce the accuracy drop when converting a continuous-valued network into a Boolean one. Extensive results on standard vision benchmarks demonstrate that the Pareto front of accuracy vs. computation of our method significantly outperforms prior state-of-the-art, achieving better accuracy with up to 37x fewer Boolean operations.

Determining Energy Efficiency Sweet Spots in Production LLM Inference

Large Language Models (LLMs) inference is central in modern AI applications, making it critical to understand their energy footprint. Existing approaches typically estimate energy consumption through simple linear functions of input and output sequence lengths, yet our observations reveal clear Energy Efficiency regimes: peak efficiency occurs with short-to-moderate inputs and medium-length outputs, while efficiency drops sharply for long inputs or very short outputs, indicating a non-linear dependency. In this work, we propose an analytical model derived from the computational and memory-access complexity of the Transformer architecture, capable of accurately characterizing the efficiency curve as a function of input and output lengths. To assess its accuracy, we evaluate energy consumption using TensorRT-LLM on NVIDIA H100 GPUs across a diverse set of LLMs ranging from 1B to 9B parameters, including OPT, LLaMA, Gemma, Falcon, Qwen2, and Granite, tested over input and output lengths from 64 to 4096 tokens, achieving a mean MAPE of 1.79%. Our results show that aligning sequence lengths with these efficiency "Sweet Spots" can substantially reduce energy usage, supporting informed truncation, summarization, and adaptive generation strategies in production systems.

Context Forcing: Consistent Autoregressive Video Generation with Long Context

Recent approaches to real-time long video generation typically employ streaming tuning strategies, attempting to train a long-context student using a short-context (memoryless) teacher. In these frameworks, the student performs long rollouts but receives supervision from a teacher limited to short 5-second windows. This structural discrepancy creates a critical \textbf{student-teacher mismatch}: the teacher's inability to access long-term history prevents it from guiding the student on global temporal dependencies, effectively capping the student's context length. To resolve this, we propose \textbf{Context Forcing}, a novel framework that trains a long-context student via a long-context teacher. By ensuring the teacher is aware of the full generation history, we eliminate the supervision mismatch, enabling the robust training of models capable of long-term consistency. To make this computationally feasible for extreme durations (e.g., 2 minutes), we introduce a context management system that transforms the linearly growing context into a \textbf{Slow-Fast Memory} architecture, significantly reducing visual redundancy. Extensive results demonstrate that our method enables effective context lengths exceeding 20 seconds -- 2 to 10 times longer than state-of-the-art methods like LongLive and Infinite-RoPE. By leveraging this extended context, Context Forcing preserves superior consistency across long durations, surpassing state-of-the-art baselines on various long video evaluation metrics.

Horizon-LM: A RAM-Centric Architecture for LLM Training

The rapid growth of large language models (LLMs) has outpaced the evolution of single-GPU hardware, making model scale increasingly constrained by memory capacity rather than computation. While modern training systems extend GPU memory through distributed parallelism and offloading across CPU and storage tiers, they fundamentally retain a GPU-centric execution paradigm in which GPUs host persistent model replicas and full autograd graphs. As a result, scaling large models remains tightly coupled to multi-GPU clusters, complex distributed runtimes, and unpredictable host memory consumption, creating substantial barriers for node-scale post-training workloads such as instruction tuning, alignment, and domain adaptation. We present Horizon-LM, a memory-centric training system that redefines the roles of CPU and GPU for large-model optimization. Horizon-LM treats host memory as the authoritative parameter store and uses GPUs solely as transient compute engines through a CPU-master, GPU-template execution model. By eliminating persistent GPU-resident modules and autograd graphs, employing explicit recomputation with manual gradient propagation, and introducing a pipelined double-buffered execution engine, Horizon-LM decouples model scale from GPU count and bounds memory usage to the theoretical parameter footprint. On a single H200 GPU with 1.5\,TB host RAM, Horizon-LM reliably trains models up to 120B parameters. On a standard single A100 machine, Horizon-LM achieves up to 12.2$\times$ higher training throughput than DeepSpeed ZeRO-3 with CPU offloading while preserving numerical correctness. Across platforms and scales, Horizon-LM sustains high device utilization and predictable memory growth, demonstrating that host memory, not GPU memory, defines the true feasibility boundary for node-scale large-model training.

Broken neural scaling laws in materials science

In materials science, data are scarce and expensive to generate, whether computationally or experimentally. Therefore, it is crucial to identify how model performance scales with dataset size and model capacity to distinguish between data- and model-limited regimes. Neural scaling laws provide a framework for quantifying this behavior and guide the design of materials datasets and machine learning architectures. Here, we investigate neural scaling laws for a paradigmatic materials science task: predicting the dielectric function of metals, a high-dimensional response that governs how solids interact with light. Using over 200,000 dielectric functions from high-throughput ab initio calculations, we study two multi-objective graph neural networks trained to predict the frequency-dependent complex interband dielectric function and the Drude frequency. We observe broken neural scaling laws with respect to dataset size, whereas scaling with the number of model parameters follows a simple power law that rapidly saturates.