arXiv subjects

Brian Zhang

Publications and source records attributed to Brian Zhang.

12 recordsLinked to original sources

arXiv2026

Loss-Cone-Limited Dark Matter Accretion onto Early Black Hole Seeds

The rapid appearance of supermassive black holes at high redshift motivates a reassessment of non-baryonic growth channels. We develop a loss-cone framework for collisionless dark-matter (DM) capture by early black-hole seeds, with particular attention to phase-space depletion and refilling. The calculation combines Eddington-inverted NFW-like halos, a relativistic direct-capture boundary, and an orbit-averaged Fokker-Planck treatment of angular-momentum transport. Primordial black holes (PBHs) are treated as massive perturbers whose refilling strength depends on both their abundance and individual mass. Collisionless refilling by triaxial or chaotic centrophilic orbits is included as a phenomenological upper-envelope channel. We show that ordinary stellar relaxation gives negligible DM-driven growth for the fiducial high-redshift seeds. PBH-driven granularity can yield order-of-magnitude growth in sufficiently compact halos, while triaxial or full-loss-cone supply can produce a rapid early burst. In the self-consistent calculations, however, the evolution generally becomes supply-limited. Once the accessible low-angular-momentum phase-space reservoir is depleted, the capture rate collapses and the black-hole mass saturates. Fixed-background NFW calculations therefore overestimate sustained growth, especially in the full-loss-cone limit. A TNG50-calibrated NFW benchmark gives negligible growth even under optimistic refilling assumptions. Collisionless DM capture is therefore unlikely to solve early SMBH growth in generic NFW-like halos, but it can provide a radiatively dark upper envelope in rare compact environments with efficient angular-momentum refilling.

astro-ph.CO↗

arXiv2026

Test-Time Strategies for More Efficient and Accurate Agentic RAG

Retrieval-Augmented Generation (RAG) systems face challenges with complex, multihop questions, and agentic frameworks such as Search-R1 (Jin et al., 2025), which operates iteratively, have been proposed to address these complexities. However, such approaches can introduce inefficiencies, including repetitive retrieval of previously processed information and challenges in contextualizing retrieved results effectively within the current generation prompt. Such issues can lead to unnecessary retrieval turns, suboptimal reasoning, inaccurate answers, and increased token consumption. In this paper, we investigate test-time modifications to the Search-R1 pipeline to mitigate these identified shortcomings. Specifically, we explore the integration of two components and their combination: a contextualization module to better integrate relevant information from retrieved documents into reasoning, and a de-duplication module that replaces previously retrieved documents with the next most relevant ones. We evaluate our approaches using the HotpotQA (Yang et al., 2018) and the Natural Questions (Kwiatkowski et al., 2019) datasets, reporting the exact match (EM) score, an LLM-as-a-Judge assessment of answer correctness, and the average number of turns. Our best-performing variant, utilizing GPT-4.1-mini for contextualization, achieves a 5.6% increase in EM score and reduces the number of turns by 10.5% compared to the Search-R1 baseline, demonstrating improved answer accuracy and retrieval efficiency.

cs.IR↗

arXiv2025

OpenAI GPT-5 System Card

This is the system card published alongside the OpenAI GPT-5 launch, August 2025. GPT-5 is a unified system with a smart and fast model that answers most questions, a deeper reasoning model for harder problems, and a real-time router that quickly decides which model to use based on conversation type, complexity, tool needs, and explicit intent (for example, if you say 'think hard about this' in the prompt). The router is continuously trained on real signals, including when users switch models, preference rates for responses, and measured correctness, improving over time. Once usage limits are reached, a mini version of each model handles remaining queries. This system card focuses primarily on gpt-5-thinking and gpt-5-main, while evaluations for other models are available in the appendix. The GPT-5 system not only outperforms previous models on benchmarks and answers questions more quickly, but -- more importantly -- is more useful for real-world queries. We've made significant advances in reducing hallucinations, improving instruction following, and minimizing sycophancy, and have leveled up GPT-5's performance in three of ChatGPT's most common uses: writing, coding, and health. All of the GPT-5 models additionally feature safe-completions, our latest approach to safety training to prevent disallowed content. Similarly to ChatGPT agent, we have decided to treat gpt-5-thinking as High capability in the Biological and Chemical domain under our Preparedness Framework, activating the associated safeguards. While we do not have definitive evidence that this model could meaningfully help a novice to create severe biological harm -- our defined threshold for High capability -- we have chosen to take a precautionary approach.

cs.CL↗

arXiv2025

Dark Matter and the Early Formation of Supermassive Black Holes

We investigate the growth of supermassive black holes (SMBHs) at high redshift ($z \ge 10$) from a combination of dark matter capture, black-hole mergers, and gas accretion. It has previously been shown that SMBHs can form by $z \approx 10$ via black-hole mergers, Eddington-limited Bondi gas accretion and tidal disruption events with stars within dense nuclear clusters. Here, we examine the degree to which the capture of collisionless dark matter by a growing SMBH may also contribute. We first consider models deduced from cosmological simulations of galaxy formation and central BH formation. We show that in the case that the dense nuclear star cluster forms by cooling and collapse of gas, while the DM remains in a standard NFW profile, the contribution from cold dark matter accretion is insignificant. However, we suggest models for which dark matter clustering can occur (possibly by self interaction). We show that such clustering may affect SMBH growth. In such cases, a small seed stellar-remnant black hole can more easily reach $> 10^7$ M$_{\odot}$ by $z = 10$ in the core of dense nuclear star clusters. This remains true for either cold dark matter or ultralight dark matter with the observationally inferred mass of $\sim 10^{-22}$ eV. We highlight the unique possible evolution of ULDM capture by the growing SMBH due to the fact that the ULDM de Broglie wavelength exceeds the initial nuclear star cluster half-mass radius.

astro-ph.HE↗

arXiv2025

TrajFlow: Multi-modal Motion Prediction via Flow Matching

Efficient and accurate motion prediction is crucial for ensuring safety and informed decision-making in autonomous driving, particularly under dynamic real-world conditions that necessitate multi-modal forecasts. We introduce TrajFlow, a novel flow matching-based motion prediction framework that addresses the scalability and efficiency challenges of existing generative trajectory prediction methods. Unlike conventional generative approaches that employ i.i.d. sampling and require multiple inference passes to capture diverse outcomes, TrajFlow predicts multiple plausible future trajectories in a single pass, significantly reducing computational overhead while maintaining coherence across predictions. Moreover, we propose a ranking loss based on the Plackett-Luce distribution to improve uncertainty estimation of predicted trajectories. Additionally, we design a self-conditioning training technique that reuses the model's own predictions to construct noisy inputs during a second forward pass, thereby improving generalization and accelerating inference. Extensive experiments on the large-scale Waymo Open Motion Dataset (WOMD) demonstrate that TrajFlow achieves state-of-the-art performance across various key metrics, underscoring its effectiveness for safety-critical autonomous driving applications. The code and other details are available on the project website https://traj-flow.github.io/.

cs.CV↗

arXiv2025

LithOS: An Operating System for Efficient Machine Learning on GPUs

The surging demand for GPUs in datacenters for machine learning (ML) has made efficient GPU utilization crucial. However, meeting the diverse needs of ML models while optimizing resource usage is challenging. To enable transparent, fine-grained GPU management that maximizes utilization and energy efficiency while maintaining strong isolation, an operating system (OS) approach is needed. This paper introduces LithOS, a first step toward a GPU OS. LithOS includes the following new abstractions and mechanisms for efficient GPU resource management: (i) a novel TPC Scheduler that supports spatial scheduling at the granularity of individual TPCs, unlocking efficient TPC stealing between workloads; (ii) transparent kernel atomization to reduce head-of-line blocking and enable dynamic resource reallocation mid-execution; (iii) a lightweight hardware right-sizing mechanism that determines the minimal TPC resources needed per atom; and (iv) a transparent power management mechanism that reduces power consumption based on in-flight work behavior. We implement LithOS in Rust and evaluate its performance across extensive ML environments, comparing it to state-of-the-art solutions from NVIDIA and prior research. For inference stacking, LithOS reduces tail latencies by 13x compared to MPS; compared to the best SotA, it reduces tail latencies by 3x while improving aggregate throughput by 1.6x. In hybrid inference-training stacking, LithOS reduces tail latencies by 4.7x compared to MPS; compared to the best SotA, it reduces tail latencies 1.18x while improving aggregate throughput by 1.35x. Finally, for a modest performance hit under 4%, LithOS's right-sizing provides a quarter of GPU capacity savings on average, while for a 7% hit, its power management yields a quarter of a GPU's energy savings. Overall, LithOS increases GPU efficiency, establishing a foundation for future OS research on GPUs.

cs.OS↗

arXiv2024

Iteratively Regularized Gradient Tracking Methods for Optimal Equilibrium Seeking

In noncooperative Nash games, equilibria are often inefficient. This is exemplified by the Prisoner's Dilemma and was first provably shown in the 1980s. Since then, understanding the quality of Nash equilibrium (NE) received considerable attention, leading to the emergence of inefficiency measures characterized by the best or the worst equilibrium. Traditionally, computing an optimal NE in monotone regimes is done through two-loop schemes which lack scalability and provable performance guarantees. The goal in this work lies in the development of among the first single-timescale distributed gradient tracking optimization methods for optimal NE seeking over networks. Our main contributions are as follows. By employing a regularization-based relaxation approach within two existing distributed gradient tracking methods, namely Push-Pull and DSGT, we devise and analyze two single-timescale iteratively regularized gradient tracking algorithms. The first method addresses computing the optimal NE over directed networks, while the second method addresses a stochastic variant of this problem over undirected networks. For both methods, we establish the convergence to the optimal NE and derive new convergence rate statements for the consensus error of the generated iterates. We provide preliminary numerical results on a Nash-Cournot game.

math.OC↗

arXiv2022

Optimal Correlated Equilibria in General-Sum Extensive-Form Games: Fixed-Parameter Algorithms, Hardness, and Two-Sided Column-Generation

We study the problem of finding optimal correlated equilibria of various sorts in extensive-form games: normal-form coarse correlated equilibrium (NFCCE), extensive-form coarse correlated equilibrium (EFCCE), and extensive-form correlated equilibrium (EFCE). We make two primary contributions. First, we introduce a new algorithm for computing optimal equilibria in all three notions. Its runtime depends exponentially only on a parameter related to the information structure of the game. We also prove a fundamental complexity gap: while our size bounds for NFCCE are similar to those achieved in the case of team games by Zhang et al., this is impossible to achieve for the other two concepts under standard complexity assumptions. Second, we propose a two-sided column generation approach for use when the runtime or memory usage of the previous algorithm is prohibitive. Our algorithm improves upon the one-sided approach of Farina et al. by means of a new decomposition of correlated strategies which allows players to re-optimize their sequence-form strategies with respect to correlation plans which were previously added to the support. Experiments show that our techniques outperform the prior state of the art for computing optimal general-sum correlated equilibria.

cs.GT↗

arXiv2019

Cache-Friendly Search Trees; or, In Which Everything Beats std::set

While a lot of work in theoretical computer science has gone into optimizing the runtime and space usage of data structures, such work very often neglects a very important component of modern computers: the cache. In doing so, very often, data structures are developed that achieve theoretically-good runtimes but are slow in practice due to a large number of cache misses. In 1999, Frigo et al. introduced the notion of a cache-oblivious algorithm: an algorithm that uses the cache to its advantage, regardless of the size or structure of said cache. Since then, various authors have designed cache-oblivious algorithms and data structures for problems from matrix multiplication to array sorting. We focus in this work on cache-oblivious search trees; i.e. implementing an ordered dictionary in a cache-friendly manner. We will start by presenting an overview of cache-oblivious data structures, especially cache-oblivious search trees. We then give practical results using these cache-oblivious structures on modern-day machinery, comparing them to the standard std::set and other cache-friendly dictionaries such as B-trees.

cs.DS↗

arXiv2019

Coconuts and Islanders: A Statistics-First Guide to the Boltzmann Distribution

The Boltzmann distribution is one of the key equations of thermal physics and is widely used in machine learning as well. Here I derive a Boltzmann distribution in a simple pedagogical example using only tools from a first-year probability course. The example is called "coconuts and islanders" and was taught to me by my father, Shoucheng Zhang (1963 - 2018), to whom these notes are dedicated. By focusing on this simple story, which can be easily simulated on a computer, I aim to provide a more accessible and intuitive presentation of the Boltzmann distribution. Yet I hope this exposition also inspires deep thinking about statistical physics. For instance, I show that the coconuts and islanders story illuminates a connection between the "fundamental assumption of statistical mechanics"---all microstates are equally probable---and the statistical property of detailed balance.

cond-mat.stat-mech↗

arXiv2018

Zap: Making Predictions Based on Online User Behavior

This paper introduces Zap, a generic machine learning pipeline for making predictions based on online user behavior. Zap combines well known techniques for processing sequential data with more obscure techniques such as Bloom filters, bucketing, and model calibration into an end-to-end solution. The pipeline creates website- and task-specific models without knowing anything about the structure of the website. It is designed to minimize the amount of website-specific code, which is realized by factoring all website-specific logic into example generators. New example generators can typically be written up in a few lines of code.

cs.LG↗

arXiv2017

The Kinetics Human Action Video Dataset

We describe the DeepMind Kinetics human action video dataset. The dataset contains 400 human action classes, with at least 400 video clips for each action. Each clip lasts around 10s and is taken from a different YouTube video. The actions are human focussed and cover a broad range of classes including human-object interactions such as playing instruments, as well as human-human interactions such as shaking hands. We describe the statistics of the dataset, how it was collected, and give some baseline performance figures for neural network architectures trained and tested for human action classification on this dataset. We also carry out a preliminary analysis of whether imbalance in the dataset leads to bias in the classifiers.

cs.CV↗