Throughput-Optimal Scheduling Algorithms for LLM Inference and AI Agents

As demand for Large Language Models (LLMs) and AI agents rapidly grows, optimizing systems for efficient LLM inference becomes critical. While significant efforts have targeted system-level engineering, little is explored through a mathematical modeling and queuing perspective. In this paper, we aim to develop the queuing fundamentals for LLM inference, bridging the gap between queuing and LLM system communities. In particular, we study the throughput aspect in LLM inference systems. We prove that a large class of 'work-conserving' scheduling algorithms can achieve maximum throughput for both individual requests and AI agent workloads, highlighting 'work-conserving' as a key design principle in practice. Evaluations of real-world systems show that Orca and Sarathi-serve are throughput-optimal, reassuring practitioners, while FastTransformer and vanilla vLLM are not maximally stable and should be used with caution. Our results highlight the substantial benefits queuing community can offer in improving LLM inference systems and call for more interdisciplinary developments.

Data Mixture Optimization: A Multi-fidelity Multi-scale Bayesian Framework

Careful curation of data sources can significantly improve the performance of LLM pre-training, but predominant approaches rely heavily on intuition or costly trial-and-error, making them difficult to generalize across different data domains and downstream tasks. Although scaling laws can provide a principled and general approach for data curation, standard deterministic extrapolation from small-scale experiments to larger scales requires strong assumptions on the reliability of such extrapolation, whose brittleness has been highlighted in prior works. In this paper, we introduce a $\textit{probabilistic extrapolation framework}$ for data mixture optimization that avoids rigid assumptions and explicitly models the uncertainty in performance across decision variables. We formulate data curation as a sequential decision-making problem$\unicode{x2013}$multi-fidelity, multi-scale Bayesian optimization$\unicode{x2013}$where $\{$data mixtures, model scale, training steps$\}$ are adaptively selected to balance training cost and potential information gain. Our framework naturally gives rise to algorithm prototypes that leverage noisy information from inexpensive experiments to systematically inform costly training decisions. To accelerate methodological progress, we build a simulator based on 472 language model pre-training runs with varying data compositions from the SlimPajama dataset. We observe that even simple kernels and acquisition functions can enable principled decisions across training models from 20M to 1B parameters and achieve $\textbf{2.6x}$ and $\textbf{3.3x}$ speedups compared to multi-fidelity BO and random search baselines. Taken together, our framework underscores potential efficiency gains achievable by developing principled and transferable data mixture optimization methods.

LLM Generated Persona is a Promise with a Catch

The use of large language models (LLMs) to simulate human behavior has gained significant attention, particularly through personas that approximate individual characteristics. Persona-based simulations hold promise for transforming disciplines that rely on population-level feedback, including social science, economic analysis, marketing research, and business operations. Traditional methods to collect realistic persona data face significant challenges. They are prohibitively expensive and logistically challenging due to privacy constraints, and often fail to capture multi-dimensional attributes, particularly subjective qualities. Consequently, synthetic persona generation with LLMs offers a scalable, cost-effective alternative. However, current approaches rely on ad hoc and heuristic generation techniques that do not guarantee methodological rigor or simulation precision, resulting in systematic biases in downstream tasks. Through extensive large-scale experiments including presidential election forecasts and general opinion surveys of the U.S. population, we reveal that these biases can lead to significant deviations from real-world outcomes. Our findings underscore the need to develop a rigorous science of persona generation and outline the methodological innovations, organizational and institutional support, and empirical foundations required to enhance the reliability and scalability of LLM-driven persona simulations. To support further research and development in this area, we have open-sourced approximately one million generated personas, available for public access and analysis at https://huggingface.co/datasets/Tianyi-Lab/Personas.

Differences-in-Neighbors for Network Interference in Experiments

Experiments in online platforms frequently suffer from network interference, in which a treatment applied to a given unit affects outcomes for other units connected via the platform. This SUTVA violation biases naive approaches to experiment design and estimation. A common solution is to reduce interference by clustering connected units, and randomizing treatments at the cluster level, typically followed by estimation using one of two extremes: either a simple difference-in-means (DM) estimator, which ignores remaining interference; or an unbiased Horvitz-Thompson (HT) estimator, which eliminates interference at great cost in variance. Even combined with clustered designs, this presents a limited set of achievable bias variance tradeoffs. We propose a new estimator, dubbed Differences-in-Neighbors (DN), designed explicitly to mitigate network interference. Compared to DM estimators, DN achieves bias second order in the magnitude of the interference effect, while its variance is exponentially smaller than that of HT estimators. When combined with clustered designs, DN offers improved bias-variance tradeoffs not achievable by existing approaches. Empirical evaluations on a large-scale social network and a city-level ride-sharing simulator demonstrate the superior performance of DN in experiments at practical scale.

How Well do LLMs Compress Their Own Chain-of-Thought? A Token Complexity Approach

Chain-of-thought prompting has emerged as a powerful technique for enabling large language models (LLMs) to solve complex reasoning tasks. However, these reasoning chains can be verbose, raising concerns about efficiency. In response, recent works have sought to decrease response lengths through simple prompting strategies (e.g. 'be concise'). In this work, we conduct the first systematic study of the relationship between reasoning length and model performance across a diverse range of compression instructions (e.g. 'use 10 words or less' or 'remove all punctuation'). In doing so, we discover a universal tradeoff between reasoning length and accuracy that persists across even very distinct reasoning chains. We demonstrate that this tradeoff emerges from a sharp threshold behavior at the question level: each task has an intrinsic 'token complexity' - a minimal number of tokens required for successful problem-solving. We show how token complexity enables us to compute information-theoretic limits on the accuracy-compression tradeoff, and find that prompt-based compression strategies operate far from these theoretical limits. This suggests there may be significant room for improvement and our framework provides a benchmark to help researchers evaluate progress in reasoning efficiency. Our work also highlights the importance of adaptive compression -- giving shorter responses for easier questions -- and we show that token complexity is a useful tool for measuring this capability.

QGym: Scalable Simulation and Benchmarking of Queuing Network Controllers

Queuing network control determines the allocation of scarce resources to manage congestion, a fundamental problem in manufacturing, communications, and healthcare. Compared to standard RL problems, queueing problems are distinguished by unique challenges: i) a system operating in continuous time, ii) high stochasticity, and iii) long horizons over which the system can become unstable (exploding delays). To spur methodological progress tackling these challenges, we present an open-sourced queueing simulation framework, QGym, that benchmark queueing policies across realistic problem instances. Our modular framework allows the researchers to build on our initial instances, which provide a wide range of environments including parallel servers, criss-cross, tandem, and re-entrant networks, as well as a realistically calibrated hospital queuing system. QGym makes it easy to compare multiple policies, including both model-free RL methods and classical queuing policies. Our testbed complements the traditional focus on evaluating algorithms based on mathematical guarantees in idealized settings, and significantly expands the scope of empirical benchmarking in prior work. QGym code is open-sourced at https://github.com/namkoong-lab/QGym.

Dark Experience for Incremental Keyword Spotting

Spoken keyword spotting (KWS) is crucial for identifying keywords within audio inputs and is widely used in applications like Apple Siri and Google Home, particularly on edge devices. Current deep learning-based KWS systems, which are typically trained on a limited set of keywords, can suffer from performance degradation when encountering new domains, a challenge often addressed through few-shot fine-tuning. However, this adaptation frequently leads to catastrophic forgetting, where the model's performance on original data deteriorates. Progressive continual learning (CL) strategies have been proposed to overcome this, but they face limitations such as the need for task-ID information and increased storage, making them less practical for lightweight devices. To address these challenges, we introduce Dark Experience for Keyword Spotting (DE-KWS), a novel CL approach that leverages dark knowledge to distill past experiences throughout the training process. DE-KWS combines rehearsal and distillation, using both ground truth labels and logits stored in a memory buffer to maintain model performance across tasks. Evaluations on the Google Speech Command dataset show that DE-KWS outperforms existing CL baselines in average accuracy without increasing model size, offering an effective solution for resource-constrained edge devices. The scripts are available on GitHub for the future research.

Speeding up Policy Simulation in Supply Chain RL

Simulating a single trajectory of a dynamical system under some state-dependent policy is a core bottleneck in policy optimization algorithms. The many inherently serial policy evaluations that must be performed in a single simulation constitute the bulk of this bottleneck. To wit, in applying policy optimization to supply chain optimization (SCO) problems, simulating a single month of a supply chain can take several hours. We present an iterative algorithm for policy simulation, which we dub Picard Iteration. This scheme carefully assigns policy evaluation tasks to independent processes. Within an iteration, a single process evaluates the policy only on its assigned tasks while assuming a certain 'cached' evaluation for other tasks; the cache is updated at the end of the iteration. Implemented on GPUs, this scheme admits batched evaluation of the policy on a single trajectory. We prove that the structure afforded by many SCO problems allows convergence in a small number of iterations, independent of the horizon. We demonstrate practical speedups of 400x on large-scale SCO problems even with a single GPU, and also demonstrate practical efficacy in other RL environments.

Correcting for Interference in Experiments: A Case Study at Douyin

Interference is a ubiquitous problem in experiments conducted on two-sided content marketplaces, such as Douyin (China's analog of TikTok). In many cases, creators are the natural unit of experimentation, but creators interfere with each other through competition for viewers' limited time and attention. "Naive" estimators currently used in practice simply ignore the interference, but in doing so incur bias on the order of the treatment effect. We formalize the problem of inference in such experiments as one of policy evaluation. Off-policy estimators, while unbiased, are impractically high variance. We introduce a novel Monte-Carlo estimator, based on "Differences-in-Qs" (DQ) techniques, which achieves bias that is second-order in the treatment effect, while remaining sample-efficient to estimate. On the theoretical side, our contribution is to develop a generalized theory of Taylor expansions for policy evaluation, which extends DQ theory to all major MDP formulations. On the practical side, we implement our estimator on Douyin's experimentation platform, and in the process develop DQ into a truly "plug-and-play" estimator for interference in real-world settings: one which provides robust, low-bias, low-variance treatment effect estimates; admits computationally cheap, asymptotically exact uncertainty quantification; and reduces MSE by 99\% compared to the best existing alternatives in our applications.

Markovian Interference in Experiments

We consider experiments in dynamical systems where interventions on some experimental units impact other units through a limiting constraint (such as a limited inventory). Despite outsize practical importance, the best estimators for this `Markovian' interference problem are largely heuristic in nature, and their bias is not well understood. We formalize the problem of inference in such experiments as one of policy evaluation. Off-policy estimators, while unbiased, apparently incur a large penalty in variance relative to state-of-the-art heuristics. We introduce an on-policy estimator: the Differences-In-Q's (DQ) estimator. We show that the DQ estimator can in general have exponentially smaller variance than off-policy evaluation. At the same time, its bias is second order in the impact of the intervention. This yields a striking bias-variance tradeoff so that the DQ estimator effectively dominates state-of-the-art alternatives. From a theoretical perspective, we introduce three separate novel techniques that are of independent interest in the theory of Reinforcement Learning (RL). Our empirical evaluation includes a set of experiments on a city-scale ride-hailing simulator.