Qizheng Zhang

Yuhan Liu, Hanchen Li, Yihua Cheng et al. · stanford

As large language models (LLMs) take on complex tasks, their inputs are supplemented with longer contexts that incorporate domain knowledge. Yet using long contexts is challenging, as nothing can be generated until the whole context is processed by the LLM. While the context-processing delay can be reduced by reusing the KV cache of a context across different inputs, fetching the KV cache, which contains large tensors, over the network can cause high extra network delays. CacheGen is a fast context-loading module for LLM systems. First, CacheGen uses a custom tensor encoder, leveraging KV cache's distributional properties to encode a KV cache into more compact bitstream representations with negligible decoding overhead, to save bandwidth usage. Second, CacheGen adapts the compression level of different parts of a KV cache to cope with changes in available bandwidth, in order to maintain low context-loading delay and high generation quality. % When available bandwidth drops, CacheGen may raise the compression level for a part of the context or recompute its KV cache on the fly. We test CacheGen on popular LLMs and datasets. Compared to the recent systems that reuse the KV cache, CacheGen reduces the KV cache size by 3.5-4.3x and the total delay in fetching and processing contexts by 3.2-3.7x with negligible impact on the LLM response quality. Our code is at: https://github.com/UChi-JCL/CacheGen.

Kuntai Du, Qizheng Zhang, Anton Arapin et al. · stanford

With more videos being recorded by edge sensors (cameras) and analyzed by computer-vision deep neural nets (DNNs), a new breed of video streaming systems has emerged, with the goal to compress and stream videos to remote servers in real time while preserving enough information to allow highly accurate inference by the server-side DNNs. An ideal design of the video streaming system should simultaneously meet three key requirements: (1) low latency of encoding and streaming, (2) high accuracy of server-side DNNs, and (3) low compute overheads on the camera. Unfortunately, despite many recent efforts, such video streaming system has hitherto been elusive, especially when serving advanced vision tasks such as object detection or semantic segmentation. This paper presents AccMPEG, a new video encoding and streaming system that meets all the three requirements. The key is to learn how much the encoding quality at each (16x16) macroblock can influence the server-side DNN accuracy, which we call accuracy gradient. Our insight is that these macroblock-level accuracy gradient can be inferred with sufficient precision by feeding the video frames through a cheap model. AccMPEG provides a suite of techniques that, given a new server-side DNN, can quickly create a cheap model to infer the accuracy gradient on any new frame in near realtime. Our extensive evaluation of AccMPEG on two types of edge devices (one Intel Xeon Silver 4100 CPU or NVIDIA Jetson Nano) and three vision tasks (six recent pre-trained DNNs) shows that AccMPEG (with the same camera-side compute resources) can reduce the end-to-end inference delay by 10-43% without hurting accuracy compared to the state-of-the-art baselines

Kuntai Du, Yuhan Liu, Yitian Hao et al. · stanford

Deep learning inference on streaming media data, such as object detection in video or LiDAR feeds and text extraction from audio waves, is now ubiquitous. To achieve high inference accuracy, these applications typically require significant network bandwidth to gather high-fidelity data and extensive GPU resources to run deep neural networks (DNNs). While the high demand for network bandwidth and GPU resources could be substantially reduced by optimally adapting the configuration knobs, such as video resolution and frame rate, current adaptation techniques fail to meet three requirements simultaneously: adapt configurations (i) with minimum extra GPU or bandwidth overhead; (ii) to reach near-optimal decisions based on how the data affects the final DNN's accuracy, and (iii) do so for a range of configuration knobs. This paper presents OneAdapt, which meets these requirements by leveraging a gradient-ascent strategy to adapt configuration knobs. The key idea is to embrace DNNs' differentiability to quickly estimate the accuracy's gradient to each configuration knob, called AccGrad. Specifically, OneAdapt estimates AccGrad by multiplying two gradients: InputGrad (i.e. how each configuration knob affects the input to the DNN) and DNNGrad (i.e. how the DNN input affects the DNN inference output). We evaluate OneAdapt across five types of configurations, four analytic tasks, and five types of input data. Compared to state-of-the-art adaptation schemes, OneAdapt cuts bandwidth usage and GPU usage by 15-59% while maintaining comparable accuracy or improves accuracy by 1-5% while using equal or fewer resources.

Shaoting Feng, Yuhan Liu, Hanchen Li et al. · stanford

Reusing KV cache is essential for high efficiency of Large Language Model (LLM) inference systems. With more LLM users, the KV cache footprint can easily exceed GPU memory capacity, so prior work has proposed to either evict KV cache to lower-tier storage devices, or compress KV cache so that more KV cache can be fit in the fast memory. However, prior work misses an important opportunity: jointly optimizing the eviction and compression decisions across all KV caches to minimize average generation latency without hurting quality. We propose EVICPRESS, a KV-cache management system that applies lossy compression and adaptive eviction to KV cache across multiple storage tiers. Specifically, for each KV cache of a context, EVICPRESS considers the effect of compression and eviction of the KV cache on the average generation quality and delay across all contexts as a whole. To achieve this, EVICPRESS proposes a unified utility function that quantifies the effect of quality and delay of the lossy compression or eviction. To this end, EVICPRESS's profiling module periodically updates the utility function scores on all possible eviction-compression configurations for all contexts and places KV caches using a fast heuristic to rearrange KV caches on all storage tiers, with the goal of maximizing the utility function scores on each storage tier. Compared to the baselines that evict KV cache or compress KV cache, EVICPRESS achieves higher KV-cache hit rates on fast devices, i.e., lower delay, while preserving high generation quality by applying conservative compression to contexts that are sensitive to compression errors. Evaluation on 12 datasets and 5 models demonstrates that EVICPRESS achieves up to 2.19x faster time-to-first-token (TTFT) at equivalent generation quality.

Hanchen Li, Qiuyang Mang, Runyuan He et al.

Agentic LLM applications interleave LLM generation requests with tool calls. These tool calls break the continuity of the workflow by creating pauses between LLM requests, bringing many challenges for the serving system, especially under multi-turn scenarios. Each pause potentially causes KV cache eviction and extra waiting time before entering the continuous batch for the following LLM request. Since these pauses happen for each call, this problem becomes increasingly severe as turn number grow for agentic programs. Previous works either fail to incorporate information from the tool call, evicting KV cache that leads to repetitive prefill or loading, or ignore the continuity of a multi-turn program, creating waiting time between turns that increases per-request latency. We present Continuum, a serving system to optimize job completion time for multi-turn agent workloads by combining tool-aware KV cache timeout with program-level scheduling. By predicting tool call durations in agentic workflows, Continuum selectively pins the KV cache in GPU memory with a time-to-live value based on total turn number. When combined with program-level first-come-first-serve, Continuum prevents scheduling bubbles, preserves multi-turn continuity, and optimizes for throughput for complex agentic workflows. By modeling the variability of tool call and agent program continuity, Continuum outperforms state-of-the-art baselines. Our evaluation on real-world agentic workloads (SWE-Bench and BFCL) with Llama-3.1 8B/70B models shows that Continuum significantly improves the average job completion times, and remains performant across different hardware setups and DRAM offloading schemes. Preview code is available at: https://github.com/Hanchenli/vllm-continuum

Pengcheng Jiang, Jiacheng Lin, Zhiyi Shi et al. · stanford

Cutting-edge agentic AI systems are built on foundation models that can be adapted to plan, reason, and interact with external tools to perform increasingly complex and specialized tasks. As these systems grow in capability and scope, adaptation becomes a central mechanism for improving performance, reliability, and generalization. In this paper, we unify the rapidly expanding research landscape into a systematic framework that spans both agent adaptations and tool adaptations. We further decompose these into tool-execution-signaled and agent-output-signaled forms of agent adaptation, as well as agent-agnostic and agent-supervised forms of tool adaptation. We demonstrate that this framework helps clarify the design space of adaptation strategies in agentic AI, makes their trade-offs explicit, and provides practical guidance for selecting or switching among strategies during system design. We then review the representative approaches in each category, analyze their strengths and limitations, and highlight key open challenges and future opportunities. Overall, this paper aims to offer a conceptual foundation and practical roadmap for researchers and practitioners seeking to build more capable, efficient, and reliable agentic AI systems.

Qiuyang Mang, Wenhao Chai, Zhifei Li et al.

We introduce FrontierCS, a benchmark of 156 open-ended problems across diverse areas of computer science, designed and reviewed by experts, including CS PhDs and top-tier competitive programming participants and problem setters. Unlike existing benchmarks that focus on tasks with known optimal solutions, FrontierCS targets problems where the optimal solution is unknown, but the quality of a solution can be objectively evaluated. Models solve these tasks by implementing executable programs rather than outputting a direct answer. FrontierCS includes algorithmic problems, which are often NP-hard variants of competitive programming problems with objective partial scoring, and research problems with the same property. For each problem we provide an expert reference solution and an automatic evaluator. Combining open-ended design, measurable progress, and expert curation, FrontierCS provides a benchmark at the frontier of computer-science difficulty. Empirically, we find that frontier reasoning models still lag far behind human experts on both the algorithmic and research tracks, that increasing reasoning budgets alone does not close this gap, and that models often over-optimize for generating merely workable code instead of discovering high-quality algorithms and system designs.

Yoonho Lee, Roshen Nair, Qizheng Zhang et al.

The performance of large language model (LLM) systems depends not only on model weights, but also on their harness: the code that determines what information to store, retrieve, and present to the model. Yet harnesses are still designed largely by hand, and existing text optimizers are poorly matched to this setting because they compress feedback too aggressively. We introduce Meta-Harness, an outer-loop system that searches over harness code for LLM applications. It uses an agentic proposer that accesses the source code, scores, and execution traces of all prior candidates through a filesystem. On online text classification, Meta-Harness improves over a state-of-the-art context management system by 7.7 points while using 4x fewer context tokens. On retrieval-augmented math reasoning, a single discovered harness improves accuracy on 200 IMO-level problems by 4.7 points on average across five held-out models. On agentic coding, discovered harnesses surpass the best hand-engineered baselines on TerminalBench-2. Together, these results show that richer access to prior experience can enable automated harness engineering.

Shubhangi Upasani, Chen Wu, Jay Rainton et al.

Test-time adaptation enables large language models (LLMs) to modify their behavior at inference without updating model parameters. A common approach is many-shot prompting, where large numbers of in-context learning (ICL) examples are injected as an input-space test-time update. Although performance can improve as more demonstrations are added, the reliability and limits of this update mechanism remain poorly understood, particularly for open-source models. We present an empirical study of many-shot prompting across tasks and model backbones, analyzing how performance varies with update magnitude, example ordering, and selection policy. We further study Dynamic and Reinforced ICL as alternative test-time update strategies that control which information is injected and how it constrains model behavior. We find that many-shot prompting is effective for structured tasks where demonstrations provide high information gain, but is highly sensitive to selection strategy and often shows limited benefits for open-ended generation tasks. Overall, we characterize the practical limits of prompt-based test-time adaptation and outline when input-space updates are beneficial versus harmful.

Zhuohan Gu, Qizheng Zhang, Omar Khattab et al.

Large language model (LLM) agents increasingly operate over long and recurring external contexts, like document corpora and code repositories. Across invocations, existing approaches preserve either the agent's trajectory, passive access to raw material, or task-level strategies. None of them preserves what we argue is most needed for repeated same-context workloads: reusable orientation knowledge (e.g., what the context contains, how it is organized, and which entities, constants, and schemas have historically been useful) about the recurring context itself. We introduce PEEK, a system that caches and maintains this orientation knowledge as a context map: a small, constant-sized artifact in the agent's prompt that gives it a persistent peek into the external context. The map is maintained by a programmable cache policy with three modules: a Distiller that extracts transferable knowledge from inference-time signals, a Cartographer that translates it into structured edits, and a priority-based Evictor that enforces a fixed token budget. On long-context reasoning and information aggregation, PEEK improves over strong baselines by 6.3-34.0% while using 93-145 fewer iterations and incurring 1.7-5.8x lower cost than the state-of-the-art prompt-learning framework, ACE. On context learning, PEEK improves solving rate and rubric accuracy by 6.0-14.0% and 7.8-12.1%, respectively, at 1.4x lower cost than ACE. These gains generalize across LMs and agent architectures, including OpenAI Codex, a production-grade coding agent. Together, these results show that a context map helps long-context LLM agents interact with recurring external contexts more accurately and efficiently.

Qizheng Zhang, Changran Hu, Shubhangi Upasani et al. · stanford

Large language model (LLM) applications such as agents and domain-specific reasoning increasingly rely on context adaptation -- modifying inputs with instructions, strategies, or evidence, rather than weight updates. Prior approaches improve usability but often suffer from brevity bias, which drops domain insights for concise summaries, and from context collapse, where iterative rewriting erodes details over time. Building on the adaptive memory introduced by Dynamic Cheatsheet, we introduce ACE (Agentic Context Engineering), a framework that treats contexts as evolving playbooks that accumulate, refine, and organize strategies through a modular process of generation, reflection, and curation. ACE prevents collapse with structured, incremental updates that preserve detailed knowledge and scale with long-context models. Across agent and domain-specific benchmarks, ACE optimizes contexts both offline (e.g., system prompts) and online (e.g., agent memory), consistently outperforming strong baselines: +10.6% on agents and +8.6% on finance, while significantly reducing adaptation latency and rollout cost. Notably, ACE could adapt effectively without labeled supervision and instead by leveraging natural execution feedback. On the AppWorld leaderboard, ACE matches the top-ranked production-level agent on the overall average and surpasses it on the harder test-challenge split, despite using a smaller open-source model. These results show that comprehensive, evolving contexts enable scalable, efficient, and self-improving LLM systems with low overhead.

Runyuan He, Qiuyang Mang, Shang Zhou et al.

Many real-world coding challenges are open-ended and admit no known optimal solution. Yet, recent progress in LLM coding has focused on well-defined tasks such as feature implementation, bug fixing, and competitive programming. Open-ended coding remains a weak spot for LLMs, largely because open-ended training problems are scarce and expensive to construct. Our goal is to synthesize open-ended coding problems at scale to train stronger LLM coders. We introduce FrontierSmith, an automated system for iteratively evolving open-ended problems from existing closed-ended coding tasks. Starting from competitive programming problems, FrontierSmith generates candidate open-ended variants by changing the problems'goals, restricting outputs, and generalizing inputs. It then uses a quantitative idea divergence metric to select problems that elicit genuinely diverse approaches from different solvers. Agents then generate test cases and verifiers for the surviving candidates. On two open-ended coding benchmarks, training on our synthesized data yields substantial gains over the base models: Qwen3.5-9B improves by +8.82 score on FrontierCS and +306.36 (Elo-rating-based performance) on ALE-bench; Qwen3.5-27B improves by +12.12 and +309.12, respectively. The synthesized problems also make agents take more turns and use more tokens, similar to human-curated ones, suggesting that closed-ended seeds can be a practical starting point for long-horizon coding data.

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Jungwoo Kim, Rubens Lacouture, Genghan Zhang et al.

Mixture-of-Experts (MoE) has become a dominant architecture for scaling large language models (LLMs). However, the execution characteristics of MoE inference are changing rapidly and increasingly mismatch the assumptions underlying existing Processing-in-Memory (PIM) systems. Prior PIM systems for LLMs rely on static rules to offload memory-bound operations to PIM, without accounting for the combined effects of load imbalance and inter-GPU communication. Meanwhile, modern MoE models activate fewer experts out of increasingly many, creating a bimodal expert distribution: a small set of experts receives many tokens, while a long tail of experts receives only one or a few. We identify a trend in modern MoE models toward increasingly bimodal token-to-expert distributions, quantify the resulting disparity in arithmetic intensity across experts, and show that this disparity dramatically reduces the efficiency of state-of-the-art PIM systems for LLMs. To address this problem, we propose a scheduler for serving MoE models on multi-GPU systems with attached HBM-PIM stacks. Our scheduler partitions expert execution between GPU and PIM based on runtime token-to-expert distributions, while jointly considering interconnect overhead, memory bandwidth, GPU throughput, and PIM throughput. Moreover, we propose Sieve, a runtime framework that employs the scheduler to coordinate execution across GPUs and their attached HBM-PIM stacks. Sieve overlaps GPU computation, PIM computation, and intra- and inter-device communication while preserving cross-device dependencies induced by expert parallelism. Sieve is evaluated on our cycle-accurate simulator based on Ramulator 2.0. Compared to state-of-the-art PIM systems for MoE, Sieve improves both throughput and interactivity by 1.3x, 1.3x, and 1.6x on Qwen3.5-397B-A17B, GPT-OSS-120B, and Qwen3-30B-A3B, respectively.

Sharad Agarwal, Pooria Namyar, Alec Wolman et al.

Agentic AI systems that invoke external tools are powerful but costly, leading developers to default to large models and overspend inference budgets. Model routing can mitigate this, but existing routers are designed for chat completion rather than tool use. We present Switchcraft, the first (to the best of our knowledge) model router optimized for agentic tool calling. Switchcraft operates inline, selecting the lowest-cost model subject to correctness. We construct an evaluation framework on five function-calling benchmarks and train a DistilBERT-based classifier, deployed under a latency budget. Switchcraft achieves 82.9% accuracy -- matching or exceeding the best individual model -- while reducing inference cost by 84%, saving over $3,600 per million queries. We find that larger models do not consistently outperform smaller ones on tool-use tasks, and that nominally cheaper models can incur higher total cost due to token-intensive reasoning. Our work enables cost-aware agentic AI deployment without sacrificing correctness.

Hanchen Li, Runyuan He, Qizheng Zhang et al.

Recent advances in prompt learning allow large language model agents to acquire task-relevant knowledge from inference-time context without parameter changes. For example, existing methods (like ACE or GEPA) can learn system prompts to improve accuracy based on previous agent runs. However, these methods primarily focus on single-agent or low-parallelism settings. This fundamentally limits their ability to efficiently learn from a large set of collected agentic traces. It would be efficient and beneficial to run prompt learning in parallel to accommodate the growing trend of learning from many agentic traces or parallel agent executions. Yet without a principled strategy for scaling, current methods suffer from quality degradation with high parallelism. To improve both the efficiency and quality of prompt learning, we propose Combee, a novel framework to scale parallel prompt learning for self-improving agents. Combee speeds up learning and enables running many agents in parallel while learning from their aggregate traces without quality degradation. To achieve this, Combee leverages parallel scans and employs an augmented shuffle mechanism; Combee also introduces a dynamic batch size controller to balance quality and delay. Evaluations on AppWorld, Terminal-Bench, Formula, and FiNER demonstrate that Combee achieves up to 17x speedup over previous methods with comparable or better accuracy and equivalent cost.

Xuekai Zhu, Daixuan Cheng, Dinghuai Zhang et al. · stanford, tsinghua

We propose FlowRL: matching the full reward distribution via flow balancing instead of maximizing rewards in large language model (LLM) reinforcement learning (RL). Recent advanced reasoning models adopt reward-maximizing methods (\eg, PPO and GRPO), which tend to over-optimize dominant reward signals while neglecting less frequent but valid reasoning paths, thus reducing diversity. In contrast, we transform scalar rewards into a normalized target distribution using a learnable partition function, and then minimize the reverse KL divergence between the policy and the target distribution. We implement this idea as a flow-balanced optimization method that promotes diverse exploration and generalizable reasoning trajectories. We conduct experiments on math and code reasoning tasks: FlowRL achieves a significant average improvement of $10.0\%$ over GRPO and $5.1\%$ over PPO on math benchmarks, and performs consistently better on code reasoning tasks. These results highlight reward distribution-matching as a key step toward efficient exploration and diverse reasoning in LLM reinforcement learning.

Qizheng Zhang, Michael Wornow, Kunle Olukotun · stanford

LLM-based agentic applications have shown increasingly remarkable capabilities in complex workflows but incur substantial costs due to extensive planning and reasoning requirements. Existing LLM caching techniques (like context caching and semantic caching), primarily designed for serving chatbots, are insufficient for agentic applications where outputs depend on external data or environmental contexts. We propose agentic plan caching, a novel approach that extracts, stores, adapts, and reuses structured plan templates from planning stages of agentic applications across semantically similar tasks to reduce the cost of serving. Unlike traditional semantic caching, our system extracts plan templates from completed agent executions at test-time, employs keyword extraction to match new requests against cached plans, and utilizes lightweight models to adapt these templates to task-specific plans with contexts. Evaluation across multiple real-world agentic applications shows that our system can reduce costs by 46.62% on average while maintaining performance, offering a more efficient solution for serving LLM-based agents that complements existing LLM serving infrastructures.

Zikai Zhou, Qizheng Zhang, Hermann Kumbong et al.

Fine-tuning large language models (LLMs) is increasingly costly as models scale to hundreds of billions of parameters, and even parameter-efficient fine-tuning (PEFT) methods like LoRA remain resource-intensive. We introduce LowRA, the first framework to enable LoRA fine-tuning below 2 bits per parameter with minimal performance loss. LowRA optimizes fine-grained quantization - mapping, threshold selection, and precision assignment - while leveraging efficient CUDA kernels for scalable deployment. Extensive evaluations across 4 LLMs and 4 datasets show that LowRA achieves a superior performance-precision trade-off above 2 bits and remains accurate down to 1.15 bits, reducing memory usage by up to 50%. Our results highlight the potential of ultra-low-bit LoRA fine-tuning for resource-constrained environments.

Yihua Cheng, Ziyi Zhang, Hanchen Li et al.

In real-time video communication, retransmitting lost packets over high-latency networks is not viable due to strict latency requirements. To counter packet losses without retransmission, two primary strategies are employed -- encoder-based forward error correction (FEC) and decoder-based error concealment. The former encodes data with redundancy before transmission, yet determining the optimal redundancy level in advance proves challenging. The latter reconstructs video from partially received frames, but dividing a frame into independently coded partitions inherently compromises compression efficiency, and the lost information cannot be effectively recovered by the decoder without adapting the encoder. We present a loss-resilient real-time video system called GRACE, which preserves the user's quality of experience (QoE) across a wide range of packet losses through a new neural video codec. Central to GRACE's enhanced loss resilience is its joint training of the neural encoder and decoder under a spectrum of simulated packet losses. In lossless scenarios, GRACE achieves video quality on par with conventional codecs (e.g., H.265). As the loss rate escalates, GRACE exhibits a more graceful, less pronounced decline in quality, consistently outperforming other loss-resilient schemes. Through extensive evaluation on various videos and real network traces, we demonstrate that GRACE reduces undecodable frames by 95% and stall duration by 90% compared with FEC, while markedly boosting video quality over error concealment methods. In a user study with 240 crowdsourced participants and 960 subjective ratings, GRACE registers a 38% higher mean opinion score (MOS) than other baselines.