Searcharxiv⌕ Search

arXiv subjects

Alvin Cheung

Publications and source records attributed to Alvin Cheung.

At least 19 recordsLinked to original sources

System Optimizations for Cost-Efficient Agent Harnesses

AI agent harnesses play the critical role of providing LLMs with an interface to a computing environment. Recent studies show that AI agents' task-execution accuracy depends mostly on the LLM used, while harness-side variances such as system prompts or tools have minimal impact. Rather, such harness-side variances show a larger impact on the cost of agent executions, achieving cost reductions up to 2.6 times on the same model and task. To further explore these harness-side cost optimization opportunities, we analyze the cost accumulation patterns across various agent traces and identify system opportunities to reduce token use and LLM calls without modifying the model or affecting task accuracy. Based on this analysis, we design Agency, a harness interposition platform that provides a unified abstraction and allows easy and portable implementation of harness-side system optimizations on existing off-the-shelf harnesses such as Claude Code or OpenAI Codex. Using Agency, we implement proof-of-concept optimization techniques based on our analysis and achieve cost reductions up to 35% across various real-world benchmarks on existing harnesses without compromising accuracy.

cs.AI↗

Qrita: High-performance Top-k and Top-p Algorithm for GPUs using Pivot-based Truncation and Selection

Despite their importance in model sampling, efficient implementation of the Top-k and Top-p algorithms poses a significant challenge as model vocabulary grows with each new model release. Existing approaches often rely on sorting, which incurs significant computation and memory overhead on GPUs, or on stochastic approaches that alter the algorithm's output. In this work, we propose Qrita, an efficient Top-k and Top-p algorithm based on pivot-based truncation and selection. Qrita leverages pivot-based search for both Top-k and Top-p with two key techniques: 1. Gaussian-based sigma-truncation, which limits the complexity of the algorithm to O(n), and 2. Quaternary pivot search with duplication handling, which halves the number of pivot search iterations and guarantees deterministic output. We implement Qrita using Triton and evaluate its performance against the Top-k and Top-p kernels of high-performance LLM execution engines such as SGLang and FlashInfer, improving end-to-end serving throughput up to 1.4 times with half the memory usage, while providing the same output as stable sort-based Top-k and Top-p algorithms. Qrita is now the default Triton Top-k and Top-p sampler of vLLM, and its implementation is available at https://github.com/vllm-project/vllm/blob/main/vllm/v1/sample/ops/topk_topp_triton.py.

cs.AI↗

FrontierCS 2: A Benchmark for Computer Science R&D

Can LLM agents make contributions to frontier computer science R&D? We introduce FrontierCS 2, a benchmark of 48 open-ended tasks. We take a paper and the code its authors released and delete the paper's contribution. The deleted code becomes the human reference that sets the bar. Earlier open-ended benchmarks give each agent a problem statement and grade one program after the run. FrontierCS 2 places the agent inside a research environment. The agent inherits a working repository and may submit a patch at any point. For each patch, a judge returns a development-suite score but keeps the evaluator, the held-out final suite and the reference private, and grades the final submission on that suite against the reference. The tasks span 13 domains, from machine-learning systems and high-performance computing to databases, security and robotics. Beyond these 48 tasks, FrontierCS 2 keeps a private set of more than 100 tasks, to be announced in the future. We evaluate 12 agents. The best one is Claude Opus 5.5, passing 58% of its runs. No run passes on 16 tasks. A higher score is not the same as a pass: most runs that beat the reference on the combined final score still fall short on some metric or workload. On a shared capability scale for agents and tasks, the median task is roughly as difficult as what the strongest agent can handle.

cs.LG↗

CodeWalk: Generating Coding Benchmarks by Walking a Problem Graph

High-quality coding problems are expensive to collect. Benchmarks such as SWE-bench rely on substantial human curation, while synthetic alternatives typically rely on costly model generation and validation. In this work, we show that new coding problems can be created at almost no cost by composing existing ones. We formulate problem generation as code composition: verified atomic problems are combined into level-$N$ tasks, and a task is kept only if every atomic problem remains unsolved and a joint reference solution passes the same check used at evaluation time. The entire task-construction process is programmatic and requires no language-model generation or validation. We instantiate this approach in four coding settings: multi-bug repair, proof completion, complexity analysis, and clone equivalence. The generated tasks are challenging for frontier models, which all degrade as compositional depth increases. The generated tasks are also useful as training data: rejection-sampling fine-tuning on them improves Qwen3.5-9B at every composition level.

cs.LG↗

EasyPPO: Stabilizing the Critic Is Key

A key strength of Proximal Policy Optimization (PPO) is its learned critic, which uses historical trajectories collected during reinforcement learning to estimate expected returns and reduce policy-gradient variance. However, we find that the critic is also a major source of instability in reinforcement learning for large language models (LLMs). We identify two critic failure modes that destabilize PPO. First, filtering truncated rollouts from both actor and critic shifts the policy objective to reward conditioned on completion, allowing truncation to increase even as conditional reward improves. Second, heterogeneous return noise can cause high-variance prompts to dominate critic updates in finite batches. We introduce EasyPPO to address these failures. Actor-only overlong filtering trains the critic on returns from both completed and truncated rollouts. Noise-normalized critic regression weights each prompt's critic loss by the inverse standard deviation of its sampled returns, balancing noise contributions across prompts. Moderately smaller critic mini-batches confine outlier influence to fewer rollouts during gradient clipping. Across continuous-reward coding on FrontierCS, binary-reward mathematical reasoning on AIME24, and multi-turn search on Search-R1, EasyPPO remains stable throughout the full training horizon and consistently outperforms vanilla PPO, VAPO, and HL-Gauss PPO. Its best validation scores show relative gains of 14.89%, 2.28%, and 9.47% over PPO, respectively.

cs.LG↗

Combee: Scaling Prompt Learning for Self-Improving Language Model Agents

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.

cs.AI↗

When Agents Slow Down: Understanding LLM Agents' Test-Time Strategies via Elo-per-token Analysis

Large language model (LLM) agents allocate test-time compute adaptively as they revise solutions, use tools, explore alternatives, and decide when to stop. This test-time strategy makes it difficult to measure how agent performance scales. We study open-ended tasks that provide continuous scores for intermediate submissions, making progress observable throughout long trajectories. We propose Elo-per-token analysis, which tracks the best solution found at each token budget and uses a Bradley-Terry model to aggregate within-task orderings into Elo ratings across tasks with different score scales. We apply it to four general-purpose agents on four open-ended benchmarks, with sessions of up to 100M tokens, and to three feedback-driven LLM optimization harnesses in controlled single-task interventions. Independent sampling provides a theoretically characterized reference, for which Elo grows linearly with log compute. Against this reference, agents can initially convert tokens into Elo faster than independent sampling, but their marginal gains diminish and eventually fall below the reference. In contrast, the strongest historical human contestants improve superlinearly over contest time on shared AtCoder Heuristic Contest tasks, providing evidence of continual learning and substantial headroom after agents slow down. We define the scaling inflection point as the per-session budget where marginal Elo gains match the independent-sampling reference. Using this point as the per-session budget, we split 100M tokens across parallel sessions on FrontierCS Polyomino Packing, gaining +264 Elo over one long session and +355 over ten short sessions.

cs.CL↗

Continuum: Efficient and Robust Multi-Turn LLM Agent Scheduling with KV Cache Time-to-Live

KV cache management is essential for efficient LLM inference. To maximize utilization, existing inference engines evict finished requests' KV cache if new requests are waiting. This policy breaks for agentic workloads, which interleave LLM calls with tools, introducing pauses that prevent effective KV reuse across turns. Since many tool calls have much shorter durations than human response multi-turn chatbot, it would be promising to retain the KV cache in during these tools. However, many challenges remain. First, we need to consider both the potential cost of recomputation or reloading (if offloading enabled) as well as the increasing queueing delays after eviction from GPU. Second, due to the internal variance of tool call durations, the method needs to remain robust under limited predictability of tool call durations. We present Continnum, a serving system to optimize job completion time for multi-turn agent workloads by introducing time-to-live mechanism for KV cache retention. For requests that generate tool calls, Continnum selectively pins the KV cache in GPU memory with a time-to-live value determined by the reload cost and potential queueing delay induced by eviction. When the TTL expires, the KV cache can be automatically evicted to free up GPU memory, providing robust performance under edge cases. When combined with program-level first-come-first-serve, Continnum preserves multi-turn continuity, and reduces delay for agentic workflows. Evaluations on real-world agents (SWE-Bench, BFCL, OpenHand) with Llama-3.1 8B/70B, Gemma-3 12B, and GLM-4.5 355B shows that Continnum improves the average job completion times by over 8x while improving throughput.

cs.OS↗

Concord: A Video Relational Algebra for Cross-Modal Query Optimization

Semantic video queries let users embed natural language prompts and use multimodal large language models (MLLMs) to interpret the video. Such queries are increasingly popular for querying video data. However, their expressiveness comes at a steep cost: an MLLM may process hours of media to return only seconds of relevant output, making naive execution slow, expensive, and inaccurate. We propose Concord, a system for expressing and optimizing semantic video queries. We makes three contributions. First, we introduce Video Relational Algebra (VRA), a nested algebra over videos, transcripts, frames, and object tracks that captures common semantic video operations. Second, we derive a set of approximate optimizations that rewrite VRA queries to reduce MLLM usage while improving result quality. For narrated video, Concord either processes transcripts instead of video or uses them to identify video clips for MLLM processing. For cross-camera queries without narration, detection and tracking replace a whole-video MLLM join with a track-level relational join. Third, we evaluate Concord on real-world videos. Across 4.59 hours of soccer broadcasts and 3.92 hours of lectures, transcript-to-video queries send only 5.32% and 2.47% of source-video duration to the MLLM and reduce MLLM cost by up to 87%. In two five-second highway clips with 18 manually adjudicated cross-camera vehicles, a Detect-Track-Join query improves F1 from .364 to .813 while making no MLLM calls. See our project at https://concord-db.github.io/.

cs.DB↗

Tetris: Tile-level Sampling for Efficient and High-Fidelity Video Object Tracking

Track materialization converts raw videos into reusable object tracks that downstream queries can run against without rerunning tracking, but extracting those tracks efficiently and with high fidelity remains expensive. Prior systems reduce track materialization cost through temporal frame sampling, but aggressive sampling spaces each track's detection points too far apart to faithfully capture the object's actual trajectory. In stationary video, however, large portions of each frame contain no objects of interest, and different sampling rates can be used to extract tracks from the remaining regions. Leveraging this idea, we present Tetris, a track-extraction system that decomposes videos into a tile-based polyomino data model, enabling fine-grained spatiotemporal pruning that reduces detector calls with minimal fidelity loss. Tetris implements track materialization in three steps: first, a classifier identifies relevant tiles and groups them into polyominoes. Then, we use an integer linear program (ILP) to prune redundant polyominoes under a user-specified accuracy constraint, before packing the remaining polyominoes into canvases to minimize detector calls. Across 7 stationary-video datasets, Tetris stays within a 5% tracking accuracy loss as compared to a reference pipeline that processes every frame in its entirety, while prior systems exceed this bound on 3 of the 7 datasets. Moreover, with this 5% bound, Tetris achieves up to 17.4x higher throughput than prior systems, and up to 68.8x higher than the reference pipeline. Conversely, Tetris delivers up to 0.42 higher HOTA tracking accuracy than the best prior system at matched throughput. The project page is at https://tetris-db.github.io .

cs.CV↗

SVG-EAR: Parameter-Free Linear Compensation for Sparse Video Generation via Error-aware Routing

Diffusion Transformers (DiTs) have become a leading backbone for video generation, yet their quadratic attention cost remains a major bottleneck. Sparse attention reduces this cost by computing only a subset of attention blocks. However, prior methods often either drop the remaining blocks, which incurs information loss, or rely on learned predictors to approximate them, introducing training overhead and potential output distribution shifting. In this paper, we show that the missing contributions can be recovered without training: after semantic clustering, keys and values within each block exhibit strong similarity and can be well summarized by a small set of cluster centroids. Based on this observation, we introduce SVG-EAR, a parameter-free linear compensation branch that uses the centroid to approximate skipped blocks and recover their contributions. While centroid compensation is accurate for most blocks, it can fail on a small subset. Standard sparsification typically selects blocks by attention scores, which indicate where the model places its attention mass, but not where the approximation error would be largest. SVG-EAR therefore performs error-aware routing: a lightweight probe estimates the compensation error for each block, and we compute exactly the blocks with the highest error-to-cost ratio while compensating for skipped blocks. We provide theoretical guarantees that relate attention reconstruction error to clustering quality, and empirically show that SVG-EAR improves the quality-efficiency trade-off and increases throughput at the same generation fidelity on video diffusion tasks. Overall, SVG-EAR establishes a clear Pareto frontier over prior approaches, achieving up to 1.77$\times$ and 1.93$\times$ speedups while maintaining PSNRs of up to 29.759 and 31.043 on Wan2.2 and HunyuanVideo, respectively.

cs.CV↗

Visual Template Inference for Data Extraction from Documents

Many templatized documents are programmatically generated from structured data following a visual template. Such documents include invoices, tax documents, financial reports, and purchase orders. Effective data extraction from these documents is crucial to support downstream analytical tasks. Current data extraction tools often struggle with complex document layouts, incur high latency and/or cost on large datasets, and require significant human effort. The key insight of our tool, TWIX, is to infer the underlying template used to create such documents, and then extract the data, rather than extracting directly from documents. To do so, TWIX first infers the underlying fields, such as columns of tabular portions or keys in co-located key-value pairs, by leveraging their consistent location patterns (e.g., two fields in the same template repeatedly co-occur within a fixed distance apart across multiple records). TWIX then assembles these fields into a template by enforcing visual constraints, such as vertically aligning table rows with their column headers for tabular regions, and horizontally aligning keys with their values for key-value pairs. TWIX then uses this inferred template to accurately and efficiently extract data from templatized documents at a low cost. On one benchmark with 34 diverse real-world datasets, TWIX outperforms state-of-the-art structured data extraction tools (Evaporate, Textract, and Azure Document Intelligence), and vision-based LLMs like GPT-4-Vision, by over 25% in precision and recall. Another benchmark with 30 large datasets demonstrates TWIX's scalability: it is 520X faster and 3,786X cheaper than the most competitive compared tool, for extracting data from large document collections with over 2000 pages.

cs.DB↗

FrontierSmith: Synthesizing Open-Ended Coding Problems at Scale

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.

cs.LG↗

Do Androids Dream of Breaking the Game? Systematically Auditing AI Agent Benchmarks with BenchJack

Agent benchmarks have become the de facto measure of frontier AI competence, guiding model selection, investment, and deployment. However, reward hacking, where agents maximize a score without performing the intended task, emerges spontaneously in frontier models without overfitting. We argue that benchmarks must be secure by design. From past incidents of reward hacks, we derive a taxonomy of eight recurring flaw patterns and compile them into the Agent-Eval Checklist for benchmark designers. We condense the insights into BenchJack, an automated red-teaming system that drives coding agents to audit benchmarks and identify possible reward-hacking exploits in a clairvoyant manner. Moreover, we extend BenchJack to an iterative generative-adversarial pipeline that discovers new flaws and patches them iteratively to improve benchmark robustness. We apply BenchJack to 10 popular agent benchmarks spanning software engineering, web navigation, desktop computing, and terminal operations. BenchJack synthesizes reward-hacking exploits that achieve near-perfect scores on most of the benchmarks without solving a single task, surfacing 219 distinct flaws across the eight classes. Moreover, BenchJack's extended pipeline reduces the hackable-task ratio from near 100% to under 10% on four benchmarks without fatal design flaws, fully patching WebArena and OSWorld within three iterations. Our results show that evaluation pipelines have not internalized an adversarial mindset, and that proactive auditing could help close the security gap for the fast-paced benchmarking space.

cs.AI↗

Uncovering Intra-expert Activation Sparsity for Efficient Mixture-of-Expert Model Execution

Mixture of Experts (MoE) architecture has become the standard for state-of-the-art large language models, owing to its computational efficiency through sparse expert activation. However, sparsity through finer expert granularity is becoming increasingly difficult to achieve due to fundamental training challenges such as expert collapse and load imbalance. In this work, we explore and leverage intra-expert activation sparsity as a complementary and underexplored dimension of sparsity in MoE models. Surprisingly, substantial intra-expert sparsity is readily available in existing pre-trained MoE models, without any modification to the activation function or model parameters, providing up to 90% sparsity within each expert without significant accuracy loss. We explore intra-expert activation sparsity across eight off-the-shelf MoE models ranging from 1B to 400B parameters, and extend the MoE execution pipeline of vLLM to leverage intra-expert activation sparsity by skipping the computations of inactive neurons, on top of its existing optimizations, achieving up to 2.5 times speedup in MoE layer execution and 1.2 times end-to-end speedup compared to the original dense vLLM baseline.

cs.LG↗

An Extensible and Verifiable Language for Query Rewrite Rules

Logical query plan rewriting transforms a relational database query into an equivalent but more efficient form and is crucial to the performance of database-backed applications. In existing systems, rewrite rules are typically implemented manually, tightly coupled to specific execution engines, and often lack formal correctness guarantees. Consequently, developing a new engine requires reimplementing both legacy and new rules, incurring significant engineering cost, limiting portability, and every new implementation is an opportunity for introducing new bugs. We introduce Rulescript, an engine-agnostic domain-specific language (DSL) for developing query rewrite rules. Rulescript separates rule definition from execution infrastructure via a relational algebra-inspired core language and an explicit decomposition of rules into matching and transformation phases. Developers express rewrites by pattern-matching query plans using Rulescript's core operators and constructing semantically equivalent transformed plans, with all rewrites automatically verified formally to ensure correctness. Rulescript is extensible: users can define custom operators in terms of the core language to capture engine-specific semantics. To integrate with an existing system, developers need only implement a lightweight adapter that maps Rulescript's core and custom operators to the operators implemented in the target engine. We evaluate Rulescript by reimplementing 33 rewrite rules from Apache Calcite and extending the language with several custom operators. To demonstrate portability, we automatically deploy these rules to CockroachDB and Apache Data Fusion, two engines with substantially different backends. Our results show that Rulescript enables "write once, deploy everywhere" paradigm for query plan rewriting, with minimal effort required to deploy previously written rules on a new data engine.

cs.DB↗

PLOP: Cost-Based Placement of Semantic Operators in Hybrid Query Plans

Recent database systems have introduced semantic operators that leverage large language models (LLMs) to filter, join, and project over structured data using natural language predicates. In practice, these operators are combined with traditional relational operators, e.g., equi-joins, producing hybrid query plans whose execution cost depends on both expensive LLM calls and conventional database processing. A key optimization question is where to place each semantic operator relative to the relational operators in the plan: placing them earlier reduces the data that subsequent operators process, but requires more LLM calls; placing them later reduces LLM calls through deduplication, but forces relational operators to process larger intermediate data. Existing systems either ignore this placement question or apply simple heuristics without considering the full cost trade-off. We present PLOP, a plan-level optimizer for hybrid semantic-relational queries. PLOP reduces hybrid query planning to semantic filter placement via two equivalence-preserving rewrites. We prove that deferring all semantic filters to the latest possible position minimizes LLM invocations under function caching, but show that this can cause relational processing costs to dominate on complex multi-table queries. To balance LLM cost against relational cost, PLOP uses a dynamic-programming-based cost model that finds the placement minimizing their weighted sum. On 44 semantic SQL queries across five schemas and two benchmarks, PLOP achieves up to 1.5$\times$ speedup and 4.29$\times$ cost reduction while maintaining high output quality: an average F1 of 0.85 against the unoptimized baseline and 0.84 against human-annotated ground truth on SemBench. Overall, PLOP achieves a significant cost reduction while preserving the highest accuracy among six publicly available systems.

cs.DB↗

Playing Psychic: Using Thought Trees to Predict Reasoning Models Accuracy on Coding Tasks

Recent advances in large language models (LLMs) have shown that test-time scaling can substantially improve model performance on complex tasks, particularly in the coding domain. Under this paradigm, models use a larger token budget during inference to generate intermediate reasoning traces before producing a final answer. However, current evaluations primarily rely on competitive programming benchmarks, which may not capture the full range of reasoning abilities. In this work, we perform a systematic study of frontier reasoning models to understand their performance on real-world coding benchmarks. To gain more insights into the performance of such models, we devise a programmatic way to {\em automatically generate} coding tasks of arbitrary difficulty and structure from existing benchmarks. Using this framework, our analysis reveals that the structure of a reasoning trace, not just its contents, is a strong predictor of correctness. Motivated by this, we propose structured thought-trees as means to represent reasoning traces. To illustrate their use, we train a lightweight classifier on features extracted from thought-trees to predict trace correctness, and demonstrate that flagging and retrying structurally anomalous traces based on the extracted features yields consistent gains at lower complexity levels.

cs.AI↗