PyTorchKR Papers

“Code Generation”

총 10,140편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2406.07935공식 코드 · Papers with Code

Defining and Detecting Vulnerability in Human Evaluation Guidelines: A Preliminary Study Towards Reliable NLG Evaluation

Human evaluation serves as the gold standard for assessing the quality of Natural Language Generation (NLG) systems. Nevertheless, the evaluation guideline, as a pivotal element ensuring reliable and reproducible human assessment, has received limited attention.Our investigation revealed that only 29.84% of recent papers involving human evaluation at top conferences release their evaluation guidelines, with vulnerabilities identified in 77.09% of these guidelines. Unreliable evaluation guidelines can yield inaccurate assessment outcomes, potentially impeding the advancement of NLG in the right direction. To address these challenges, we take an initial step towards reliable evaluation guidelines and propose the first human evaluation guideline dataset by collecting annotations of guidelines extracted from existing papers as well as generated via Large Language Models (LLMs). We then introduce a taxonomy of eight vulnerabilities and formulate a principle for composing evaluation guidelines. Furthermore, a method for detecting guideline vulnerabilities has been explored using LLMs, and we offer a set of recommendations to enhance reliability in human evaluation. The annotated human evaluation guideline dataset and code for the vulnerability detection method are publicly available online.

arXiv:2409.05294공식 코드 · Papers with Code

TERD: A Unified Framework for Safeguarding Diffusion Models Against Backdoors

Diffusion models have achieved notable success in image generation, but they remain highly vulnerable to backdoor attacks, which compromise their integrity by producing specific undesirable outputs when presented with a pre-defined trigger. In this paper, we investigate how to protect diffusion models from this dangerous threat. Specifically, we propose TERD, a backdoor defense framework that builds unified modeling for current attacks, which enables us to derive an accessible reversed loss. A trigger reversion strategy is further employed: an initial approximation of the trigger through noise sampled from a prior distribution, followed by refinement through differential multi-step samplers. Additionally, with the reversed trigger, we propose backdoor detection from the noise space, introducing the first backdoor input detection approach for diffusion models and a novel model detection algorithm that calculates the KL divergence between reversed and benign distributions. Extensive evaluations demonstrate that TERD secures a 100% True Positive Rate (TPR) and True Negative Rate (TNR) across datasets of varying resolutions. TERD also demonstrates nice adaptability to other Stochastic Differential Equation (SDE)-based models. Our code is available at https://github.com/PKU-ML/TERD.

arXiv:2410.18652공식 코드 · Papers with Code

$C^2$: Scalable Auto-Feedback for LLM-based Chart Generation

Generating high-quality charts with Large Language Models (LLMs) presents significant challenges due to limited data and the high cost of scaling through human curation. $\langle \text{instruction}, \text{data}, \text{code} \rangle$ triplets are scarce and expensive to manually curate as their creation demands technical expertise. To address this scalability challenge, we introduce a reference-free automatic feedback generator, which eliminates the need for costly human intervention. Our novel framework, C$^2$, consists of (1) an automatic feedback provider (ChartAF) and (2) a diverse, reference-free dataset (ChartUIE-8K). The results are compelling: in our first experiment, 74% of respondents strongly preferred, and 10% preferred, the results after feedback. The second post-feedback experiment demonstrates that ChartAF outperform nine baselines. Moreover, ChartUIE-8K significantly improves data diversity by increasing queries, datasets, and chart types by 5982%, 1936%, and 91%, respectively, over benchmarks. Finally, a study of LLM users revealed that 94% of participants preferred ChartUIE-8K's queries, with 93% deeming them aligned with real-world use cases. Core contributions are available as open-source at chartsquared.github.io, with ample qualitative examples.

arXiv:2410.21035공식 코드 · Papers with Code

Beyond Autoregression: Fast LLMs via Self-Distillation Through Time

Autoregressive (AR) Large Language Models (LLMs) have demonstrated significant success across numerous tasks. However, the AR modeling paradigm presents certain limitations; for instance, contemporary autoregressive LLMs are trained to generate one token at a time, which can result in noticeable latency. Recent advances have indicated that search and repeated sampling can enhance performance in various applications, such as theorem proving, code generation, and alignment, by utilizing greater computational resources during inference. In this study, we demonstrate that diffusion language models are capable of generating at least 32 tokens simultaneously, while exceeding the performance of AR models in text quality and on the LAMBADA natural language understanding benchmark. This outcome is achieved through a novel distillation method for discrete diffusion models, which reduces the number of inference steps by a factor of 32-64. Practically, our models, even without caching, can generate tokens at a rate that is up to 8 times faster than AR models employing KV caching, and we anticipate further improvements with the inclusion of caching. Moreover, we demonstrate the efficacy of our approach for diffusion language models with up to 860M parameters.

arXiv:2411.00341공식 코드 · Papers with Code

A Survey on Bundle Recommendation: Methods, Applications, and Challenges

In recent years, bundle recommendation systems have gained significant attention in both academia and industry due to their ability to enhance user experience and increase sales by recommending a set of items as a bundle rather than individual items. This survey provides a comprehensive review on bundle recommendation, beginning by a taxonomy for exploring product bundling. We classify it into two categories based on bundling strategy from various application domains, i.e., discriminative and generative bundle recommendation. Then we formulate the corresponding tasks of the two categories and systematically review their methods: 1) representation learning from bundle and item levels and interaction modeling for discriminative bundle recommendation; 2) representation learning from item level and bundle generation for generative bundle recommendation. Subsequently, we survey the resources of bundle recommendation including datasets and evaluation metrics, and conduct reproducibility experiments on mainstream models. Lastly, we discuss the main challenges and highlight the promising future directions in the field of bundle recommendation, aiming to serve as a useful resource for researchers and practitioners. Our code and datasets are publicly available at https://github.com/WUT-IDEA/bundle-recommendation-survey.

arXiv:2412.08685

ChatDyn: Language-Driven Multi-Actor Dynamics Generation in Street Scenes

Generating realistic and interactive dynamics of traffic participants according to specific instruction is critical for street scene simulation. However, there is currently a lack of a comprehensive method that generates realistic dynamics of different types of participants including vehicles and pedestrians, with different kinds of interactions between them. In this paper, we introduce ChatDyn, the first system capable of generating interactive, controllable and realistic participant dynamics in street scenes based on language instructions. To achieve precise control through complex language, ChatDyn employs a multi-LLM-agent role-playing approach, which utilizes natural language inputs to plan the trajectories and behaviors for different traffic participants. To generate realistic fine-grained dynamics based on the planning, ChatDyn designs two novel executors: the PedExecutor, a unified multi-task executor that generates realistic pedestrian dynamics under different task plannings; and the VehExecutor, a physical transition-based policy that generates physically plausible vehicle dynamics. Extensive experiments show that ChatDyn can generate realistic driving scene dynamics with multiple vehicles and pedestrians, and significantly outperforms previous methods on subtasks. Code and model will be available at https://vfishc.github.io/chatdyn.

arXiv:2412.14672

FiVL: A Framework for Improved Vision-Language Alignment

Large Vision Language Models (LVLMs) have achieved significant progress in integrating visual and textual inputs for multimodal reasoning. However, a recurring challenge is ensuring these models utilize visual information as effectively as linguistic content when both modalities are necessary to formulate an accurate answer. We hypothesize that hallucinations arise due to the lack of effective visual grounding in current LVLMs. This issue extends to vision-language benchmarks, where it is difficult to make the image indispensable for accurate answer generation, particularly in vision question-answering tasks. In this work, we introduce FiVL, a novel method for constructing datasets designed to train LVLMs for enhanced visual grounding and to evaluate their effectiveness in achieving it. These datasets can be utilized for both training and assessing an LVLM's ability to use image content as substantive evidence rather than relying solely on linguistic priors, providing insights into the model's reliance on visual information. To demonstrate the utility of our dataset, we introduce an innovative training task that outperforms baselines alongside a validation method and application for explainability. The code is available at https://github.com/IntelLabs/fivl.

arXiv:2503.22876공식 코드 · Papers with Code

VizFlyt: Perception-centric Pedagogical Framework For Autonomous Aerial Robots

Autonomous aerial robots are becoming commonplace in our lives. Hands-on aerial robotics courses are pivotal in training the next-generation workforce to meet the growing market demands. Such an efficient and compelling course depends on a reliable testbed. In this paper, we present VizFlyt, an open-source perception-centric Hardware-In-The-Loop (HITL) photorealistic testing framework for aerial robotics courses. We utilize pose from an external localization system to hallucinate real-time and photorealistic visual sensors using 3D Gaussian Splatting. This enables stress-free testing of autonomy algorithms on aerial robots without the risk of crashing into obstacles. We achieve over 100Hz of system update rate. Lastly, we build upon our past experiences of offering hands-on aerial robotics courses and propose a new open-source and open-hardware curriculum based on VizFlyt for the future. We test our framework on various course projects in real-world HITL experiments and present the results showing the efficacy of such a system and its large potential use cases. Code, datasets, hardware guides and demo videos are available at https://pear.wpi.edu/research/vizflyt.html

arXiv:2504.11143

Taming Consistency Distillation for Accelerated Human Image Animation

Recent advancements in human image animation have been propelled by video diffusion models, yet their reliance on numerous iterative denoising steps results in high inference costs and slow speeds. An intuitive solution involves adopting consistency models, which serve as an effective acceleration paradigm through consistency distillation. However, simply employing this strategy in human image animation often leads to quality decline, including visual blurring, motion degradation, and facial distortion, particularly in dynamic regions. In this paper, we propose the DanceLCM approach complemented by several enhancements to improve visual quality and motion continuity at low-step regime: (1) segmented consistency distillation with an auxiliary light-weight head to incorporate supervision from real video latents, mitigating cumulative errors resulting from single full-trajectory generation; (2) a motion-focused loss to centre on motion regions, and explicit injection of facial fidelity features to improve face authenticity. Extensive qualitative and quantitative experiments demonstrate that DanceLCM achieves results comparable to state-of-the-art video diffusion models with a mere 2-4 inference steps, significantly reducing the inference burden without compromising video quality. The code and models will be made publicly available.

arXiv:2504.10106공식 코드 · Papers with Code

SoccerNet-v3D: Leveraging Sports Broadcast Replays for 3D Scene Understanding

Sports video analysis is a key domain in computer vision, enabling detailed spatial understanding through multi-view correspondences. In this work, we introduce SoccerNet-v3D and ISSIA-3D, two enhanced and scalable datasets designed for 3D scene understanding in soccer broadcast analysis. These datasets extend SoccerNet-v3 and ISSIA by incorporating field-line-based camera calibration and multi-view synchronization, enabling 3D object localization through triangulation. We propose a monocular 3D ball localization task built upon the triangulation of ground-truth 2D ball annotations, along with several calibration and reprojection metrics to assess annotation quality on demand. Additionally, we present a single-image 3D ball localization method as a baseline, leveraging camera calibration and ball size priors to estimate the ball's position from a monocular viewpoint. To further refine 2D annotations, we introduce a bounding box optimization technique that ensures alignment with the 3D scene representation. Our proposed datasets establish new benchmarks for 3D soccer scene understanding, enhancing both spatial and temporal analysis in sports analytics. Finally, we provide code to facilitate access to our annotations and the generation pipelines for the datasets.

arXiv:2504.18423

LLMpatronous: Harnessing the Power of LLMs For Vulnerability Detection

Despite the transformative impact of Artificial Intelligence (AI) across various sectors, cyber security continues to rely on traditional static and dynamic analysis tools, hampered by high false positive rates and superficial code comprehension. While generative AI offers promising automation capabilities for software development, leveraging Large Language Models (LLMs) for vulnerability detection presents unique challenges. This paper explores the potential and limitations of LLMs in identifying vulnerabilities, acknowledging inherent weaknesses such as hallucinations, limited context length, and knowledge cut-offs. Previous attempts employing machine learning models for vulnerability detection have proven ineffective due to limited real-world applicability, feature engineering challenges, lack of contextual understanding, and the complexities of training models to keep pace with the evolving threat landscape. Therefore, we propose a robust AI-driven approach focused on mitigating these limitations and ensuring the quality and reliability of LLM based vulnerability detection. Through innovative methodologies combining Retrieval-Augmented Generation (RAG) and Mixtureof-Agents (MoA), this research seeks to leverage the strengths of LLMs while addressing their weaknesses, ultimately paving the way for dependable and efficient AI-powered solutions in securing the ever-evolving software landscape.

arXiv:2505.12890

ORQA: A Benchmark and Foundation Model for Holistic Operating Room Modeling

The real-world complexity of surgeries necessitates surgeons to have deep and holistic comprehension to ensure precision, safety, and effective interventions. Computational systems are required to have a similar level of comprehension within the operating room. Prior works, limited to single-task efforts like phase recognition or scene graph generation, lack scope and generalizability. In this work, we introduce ORQA, a novel OR question answering benchmark and foundational multimodal model to advance OR intelligence. By unifying all four public OR datasets into a comprehensive benchmark, we enable our approach to concurrently address a diverse range of OR challenges. The proposed multimodal large language model fuses diverse OR signals such as visual, auditory, and structured data, for a holistic modeling of the OR. Finally, we propose a novel, progressive knowledge distillation paradigm, to generate a family of models optimized for different speed and memory requirements. We show the strong performance of ORQA on our proposed benchmark, and its zero-shot generalization, paving the way for scalable, unified OR modeling and significantly advancing multimodal surgical intelligence. We will release our code and data upon acceptance.

arXiv:2506.15981공식 코드 · Papers with Code

Double Entendre: Robust Audio-Based AI-Generated Lyrics Detection via Multi-View Fusion

The rapid advancement of AI-based music generation tools is revolutionizing the music industry but also posing challenges to artists, copyright holders, and providers alike. This necessitates reliable methods for detecting such AI-generated content. However, existing detectors, relying on either audio or lyrics, face key practical limitations: audio-based detectors fail to generalize to new or unseen generators and are vulnerable to audio perturbations; lyrics-based methods require cleanly formatted and accurate lyrics, unavailable in practice. To overcome these limitations, we propose a novel, practically grounded approach: a multimodal, modular late-fusion pipeline that combines automatically transcribed sung lyrics and speech features capturing lyrics-related information within the audio. By relying on lyrical aspects directly from audio, our method enhances robustness, mitigates susceptibility to low-level artifacts, and enables practical applicability. Experiments show that our method, DE-detect, outperforms existing lyrics-based detectors while also being more robust to audio perturbations. Thus, it offers an effective, robust solution for detecting AI-generated music in real-world scenarios. Our code is available at https://github.com/deezer/robust-AI-lyrics-detection.

arXiv:2507.04036공식 코드 · Papers with Code

PresentAgent: Multimodal Agent for Presentation Video Generation

We present PresentAgent, a multimodal agent that transforms long-form documents into narrated presentation videos. While existing approaches are limited to generating static slides or text summaries, our method advances beyond these limitations by producing fully synchronized visual and spoken content that closely mimics human-style presentations. To achieve this integration, PresentAgent employs a modular pipeline that systematically segments the input document, plans and renders slide-style visual frames, generates contextual spoken narration with large language models and Text-to-Speech models, and seamlessly composes the final video with precise audio-visual alignment. Given the complexity of evaluating such multimodal outputs, we introduce PresentEval, a unified assessment framework powered by Vision-Language Models that comprehensively scores videos across three critical dimensions: content fidelity, visual clarity, and audience comprehension through prompt-based evaluation. Our experimental validation on a curated dataset of 30 document-presentation pairs demonstrates that PresentAgent approaches human-level quality across all evaluation metrics. These results highlight the significant potential of controllable multimodal agents in transforming static textual materials into dynamic, effective, and accessible presentation formats. Code will be available at https://github.com/AIGeeksGroup/PresentAgent.

arXiv:2507.05146공식 코드 · Papers with Code

VERITAS: Verification and Explanation of Realness in Images for Transparency in AI Systems

The widespread and rapid adoption of AI-generated content, created by models such as Generative Adversarial Networks (GANs) and Diffusion Models, has revolutionized the digital media landscape by allowing efficient and creative content generation. However, these models also blur the difference between real images and AI-generated synthetic images, raising concerns regarding content authenticity and integrity. While many existing solutions to detect fake images focus solely on classification and higher-resolution images, they often lack transparency in their decision-making, making it difficult for users to understand why an image is classified as fake. In this paper, we present VERITAS, a comprehensive framework that not only accurately detects whether a small (32x32) image is AI-generated but also explains why it was classified that way through artifact localization and semantic reasoning. VERITAS produces human-readable explanations that describe key artifacts in synthetic images. We show that this architecture offers clear explanations of the basis of zero-shot synthetic image detection tasks. Code and relevant prompts can be found at https://github.com/V-i-g-n-e-s-h-N/VERITAS .

arXiv:2609.03079v2cs.LG

LeanStream: 효율적인 온디바이스 LLM 추론을 위한 추측-정제(Speculate-and-Refine) 스트리밍 프레임워크

LeanStream: A Speculate-and-Refine Streaming Framework for Efficient on-Device LLM Inference

On-device LLM inference is attractive for privacy and responsiveness, but remains challenging on mobile and embedded devices because model weights far exceed available DRAM. Prior systems exploit activation sparsity and offload weights to SSD or flash storage, but face a fundamental systems trade-off: accurate sparse execution decisions require the latest context, whereas efficient computation-I/O overlap requires early prediction. As a result, existing designs either serialize execution or incur redundant weight fetches, extra computation, and large cache overheads. We present LeanStream, a streaming speculate-and-refine framework for efficient on-device LLM inference. LeanStream progressively refines computation, loading, and cache-retention priorities using partial GPU results, enabling fine-grained overlap between GPU execution and storage I/O. We implement LeanStream on both mobile and embedded platforms. Compared with prior on-device LLM inference systems, LeanStream reduces memory usage by 4.8$\times$ to 7.5$\times$ at the best throughput achieved by prior work, while further improving token generation throughput by 1.6$\times$ to 2.1$\times$. Source code is available at https://github.com/RenyuanLIU/LeanStream.

arXiv:2606.06915v1cs.CL

ThinkBooster: 대규모 언어 모델(LLM) 추론의 매끄러운 테스트 타임 스케일링을 위한 통합 프레임워크

ThinkBooster: A Unified Framework for Seamless Test-Time Scaling of LLM Reasoning

테스트 타임 컴퓨트(TTC) 스케일링은 추론(inference) 과정에서 추가적인 연산을 할당함으로써 대규모 언어 모델(LLM) 추론을 개선하는 강력한 패러다임으로 등장했습니다(예: 다중 샘플 생성 및 검증 기반 재순위화). 기존의 TTC 스케일링 전략과 추론 채점기(reasoning scorer)는 여전히 파편화되어 있고, 일관되지 않은 프로토콜 하에서 평가되며, 품질-비용(trade-off) 관점에서 거의 분석되지 않았습니다. 우리는 LLM 추론의 원활한 테스트 타임 컴퓨트 스케일링을 위한 통합 프레임워크인 ThinkBooster를 소개합니다. 이는 (i) 최신 TTC 스케일링 전략과 채점기 패밀리(families)를 구현하는 모듈식 Python 라이브러리, (ii) 성능과 계산 효율성을 공동으로 평가하는 벤치마크(benchmark), (iii) 실제 애플리케이션에 적응형 추론(adaptive reasoning)을 드롭인(drop-in) 방식으로 통합할 수 있는 배포 가능한 OpenAI 호환 프록시 서비스로 구성됩니다. 또한 우리는 추론 궤적(reasoning trajectories), 중간 선택 결정, 대안적 추론 경로를 검사하기 위한 데모 시각화 디버거를 제공합니다. 수학 및 코딩 작업에 대한 실증적 결과는 TTC 스케일링 전략과 채점 방법의 성능-연산(performance-compute) 트레이드오프를 보여주며, ThinkBooster가 실제 작업에서 실용적인 이득을 제공함을 입증합니다. 코드는 MIT 라이선스 하에 온라인에서 제공됩니다.

arXiv:2606.11201v1cs.LG

개입할 것인가 말 것인가: 확률적 모델 혼합을 통한 추론 시점 정렬 안내

To Intervene or Not: Guiding Inference-time Alignment with Probabilistic Model Blending

대규모 언어 모델(LLM)의 널리 퍼진 배포로 인해, 새로 학습된 모델이 사용자 지시에 안전하고 효과적으로 응답하도록 하기 위해 모델 정렬이 필수적이 되었습니다. 여러 방법 중에서, 추론 시점 정렬(inference-time alignment)은 출력 생성 중에만 개입(즉, 가이던스를 제공)하므로 종종 더 비용 효율적입니다. 기존 제안들은 특정 정렬된 모델에서 추출한 가이던스를 적용하지만, 그 신뢰성을 적절히 평가하지 못합니다. 그럼에도 불구하고, 우리의 체계적인 평가에 따르면 가이던스의 효과성은 모델마다 크게 다릅니다. 비효과적인 가이던스는 추가적인 혼란을 야기하고 결과적으로 더 많은 개입을 유발하므로, 이러한 과도한 개입은 일반적으로 낮은 성능을 나타냅니다. 개입을 더 효과적이고 효율적으로 만들기 위해, 우리는 이진 결정에서 벗어나 두 모델의 지식을 통합하는 하이브리드 분포를 생성하는 BlendIn이라는 추론 시점 정렬 프레임워크를 소개합니다. BlendIn은 품질 인식 정렬(quality-aware alignment)을 수행하고 신뢰성을 기반으로 각 모델의 기여에 비례하여 가중치를 부여함으로써 추론 시점 정렬을 안정화합니다. 기존 연구와 비교할 때, BlendIn은 신뢰할 수 없는 제안의 가중치를 낮추면서 유익한 가이던스를 보존합니다. BlendIn은 오정렬된 가이던스에 대한 진단 신호와 완화 전략을 모두 제공하며, 도전적인 모델 쌍에서 최대 50%까지 일관된 성능 향상을 달성했습니다. 우리의 코드는 다음에서 확인할 수 있습니다: https://github.com/DecayingSeart/BlendIn.

arXiv:2601.03792v2cs.CL

VietMed-MCQ: 베트남 전통 의학 평가를 위한 일관성 필터링 데이터 합성 프레임워크

VietMed-MCQ: A Consistency-Filtered Data Synthesis Framework for Vietnamese Traditional Medicine Evaluation

대규모 언어 모델(LLM)은 일반 의학 분야에서 탁월한 능력을 보여주었습니다. 그러나 고품질의 구조화된 벤치마크가 부족하여, 베트남 전통 의학(VTM)과 같은 전문적이고 문화적으로 특정한 분야에서는 성능이 현저히 저하됩니다. 본 논문에서는 자동화된 일관성 검사 메커니즘을 갖춘 검색-증강 생성(RAG) 파이프라인을 통해 생성된 새로운 객관식 질문 데이터셋인 VietMed-MCQ를 소개합니다. 기존 합성 데이터셋과 달리, 본 프레임워크는 독립적인 답변 검증을 통해 추론 일관성을 보장하기 위해 이중 모델 검증 접근 방식을 통합하며, 부분 문자열 기반 증거 검사는 알려진 한계가 있습니다. 총 3,190개의 질문으로 구성된 전체 데이터셋은 세 가지 난이도 수준에 걸쳐 있으며, 한 명의 의학 전문가와 네 명의 학생에 의해 검증을 거쳐 94.2%의 승인율과 상당한 평가자 간 일치도(Fleiss' kappa = 0.82)를 달성했습니다. 우리는 VietMed-MCQ에서 7개의 오픈 소스 모델을 평가했습니다. 결과에 따르면 강력한 중국어 사전 지식을 가진 범용 모델이 베트남어 중심 모델보다 우수한 성능을 보여 언어 간 개념 전이를 강조하는 반면, 모든 모델은 여전히 복잡한 진단 추론에 어려움을 겪고 있습니다. 저자원 의학 분야의 연구를 촉진하기 위해 코드와 데이터셋을 공개합니다.

arXiv:2606.13496v1cs.CV

예산 제약 기반 스텝 수준 확산 모델 캐싱

Budget-Constrained Step-Level Diffusion Caching

스텝 수준 캐싱(step-level caching)은 디노이징 스텝 간의 시간적 중복성을 활용하여 확산 모델(diffusion model)을 가속화한다. 기존 방법들은 최종 출력 품질을 직접 최적화하지 않고 임계값 기반 휴리스틱을 사용하여 스텝별 캐시 결정을 내린다. 그 결과 추론(inference) 지연 시간이 입력에 따라 달라지며 배포 시 제어하기 어렵다. 본 연구에서는 이 공식을 역전시킨 BudCache를 제안한다: 스텝별 오류 임계값이 런타임 비용을 결정하도록 하는 대신, 연산 예산을 사전에 고정하고 최종 출력을 가장 잘 보존하는 캐시 정책을 탐색한다. 스텝 선택의 조합적 복잡성을 해결하기 위해 모의 담금질(Simulated Annealing)과 결정론적 힐 클라이밍(Hill Climbing)을 결합한다. 이 오프라인 탐색은 몇 분 내에 고품질 캐시 정책을 식별하며 추론 중 온라인 탐색이나 임계값 처리 오버헤드를 추가하지 않는다. 연산 예산이 매우 제한적일 때, 캐시로 인한 궤적 불일치를 줄이기 위해 선택된 캐시 정책에 시간 이산화를 적응시키는 캐시 인식 스케줄 정렬(cache-aware schedule alignment)을 추가로 도입한다. FLUX.1-dev와 Wan2.1에 대한 실험에서 BudCache는 동일한 추론 예산 하에서 휴리스틱 캐싱 기준선보다 더 나은 생성 품질을 달성함을 보여준다. 코드는 https://github.com/Westlake-AGI-Lab/BudCache 에서 확인할 수 있다.