PyTorchKR Papers

“Code Generation”

총 10,186편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2609.02741v1cs.CR

SPADE: 커넥티드 차량 관점에서의 SPaT 공격 탐지

SPADE: SPaT Attack Detection from the Connected Vehicle's Perspective

신호 현시 및 타이밍(SPaT) 메시지는 커넥티드 차량(CV) 안전의 핵심 기반으로, 차량-인프라 간(V2I) 및 차량-차량 간(V2V) 통신을 통해 CV가 교차로 상태를 인지하고 대응할 수 있도록 한다. 노변 기지국(RSU)이나 주변 차량이 침해되었을 때 기존 인증 체계를 우회할 수 있는 다양한 응용 층 공격으로 인해 이러한 메시지의 무결성이 위협받고 있다. 기존 침입 탐지 연구는 인프라 측면을 방어하거나 V2V 기본 안전 메시지(BSM) / 협력 인식 메시지(CAM)의 오작동을 목표로 하여, SPaT 무결성에 대한 CV 차량 내부(onboard) 관점은 다루지 않은 채로 남아 있다. 이러한 공백을 해소하기 위해, 우리는 이 분야의 심층 학습 침입 탐지 시스템(IDS) 연구를 위해 특별히 설계된 라벨이 지정된 멀티모달 시뮬레이션 기반 데이터셋인 SPADE(SPaT Attack Detection and Evaluation dataset)를 소개한다. SPADE는 Eclipse MOSAIC을 활용하여 6가지 공격 클래스와 1가지 정상 클래스 전반에 걸쳐 SAE J2735 응용 층에서의 런타임 공격 주입을 통해 생성된다. 4개의 교차로 기하 구조, 6가지 주행 조건, 5개의 독립적인 무작위 시드 반복을 결합함으로써, SPADE는 180개의 고유한 기본 시나리오 실행으로 구성되어 라벨이 지정된 약 1,890,000개의 타임스텝 기록(클래스당 270,000개)을 생성한다. 각 기록은 40개 특징 전반에 걸쳐 SPaT 메시지 필드, 차량 탑재 카메라 신뢰도 점수, 협력적 V2V 피어 데이터를 융합하여, 의도적인 공격과 환경적 성능 저하를 구별하는 데 필요한 멀티모달 신호 공간을 반영한다. C-V2X 보안 분야에서 재현 가능하고 비교 가능한 IDS 연구를 지원하기 위해 데이터셋, 생성 코드 및 시나리오 설정이 공개적으로 배포된다. 개발된 툴박스, 사용 설명 및 데이터셋 링크는 GitHub(https://github.com/jdinovo/SPADE)에서 공개적으로 이용 가능하다.

arXiv:1804.05805공식 코드 · Papers with Code

Global Robustness Evaluation of Deep Neural Networks with Provable Guarantees for the $L_0$ Norm

Deployment of deep neural networks (DNNs) in safety- or security-critical systems requires provable guarantees on their correct behaviour. A common requirement is robustness to adversarial perturbations in a neighbourhood around an input. In this paper we focus on the $L_0$ norm and aim to compute, for a trained DNN and an input, the maximal radius of a safe norm ball around the input within which there are no adversarial examples. Then we define global robustness as an expectation of the maximal safe radius over a test data set. We first show that the problem is NP-hard, and then propose an approximate approach to iteratively compute lower and upper bounds on the network's robustness. The approach is \emph{anytime}, i.e., it returns intermediate bounds and robustness estimates that are gradually, but strictly, improved as the computation proceeds; \emph{tensor-based}, i.e., the computation is conducted over a set of inputs simultaneously, instead of one by one, to enable efficient GPU computation; and has \emph{provable guarantees}, i.e., both the bounds and the robustness estimates can converge to their optimal values. Finally, we demonstrate the utility of the proposed approach in practice to compute tight bounds by applying and adapting the anytime algorithm to a set of challenging problems, including global robustness evaluation, competitive $L_0$ attacks, test case generation for DNNs, and local robustness evaluation on large-scale ImageNet DNNs. We release the code of all case studies via GitHub.

arXiv:1911.02247공식 코드 · Papers with Code

Unsupervised Opinion Summarization as Copycat-Review Generation

Opinion summarization is the task of automatically creating summaries that reflect subjective information expressed in multiple documents, such as product reviews. While the majority of previous work has focused on the extractive setting, i.e., selecting fragments from input reviews to produce a summary, we let the model generate novel sentences and hence produce abstractive summaries. Recent progress in summarization has seen the development of supervised models which rely on large quantities of document-summary pairs. Since such training data is expensive to acquire, we instead consider the unsupervised setting, in other words, we do not use any summaries in training. We define a generative model for a review collection which capitalizes on the intuition that when generating a new review given a set of other reviews of a product, we should be able to control the "amount of novelty" going into the new review or, equivalently, vary the extent to which it deviates from the input. At test time, when generating summaries, we force the novelty to be minimal, and produce a text reflecting consensus opinions. We capture this intuition by defining a hierarchical variational autoencoder model. Both individual reviews and the products they correspond to are associated with stochastic latent codes, and the review generator ("decoder") has direct access to the text of input reviews through the pointer-generator mechanism. Experiments on Amazon and Yelp datasets, show that setting at test time the review's latent code to its mean, allows the model to produce fluent and coherent summaries reflecting common opinions.

arXiv:2105.10026공식 코드 · Papers with Code

Improving Generation and Evaluation of Visual Stories via Semantic Consistency

Story visualization is an under-explored task that falls at the intersection of many important research directions in both computer vision and natural language processing. In this task, given a series of natural language captions which compose a story, an agent must generate a sequence of images that correspond to the captions. Prior work has introduced recurrent generative models which outperform text-to-image synthesis models on this task. However, there is room for improvement of generated images in terms of visual quality, coherence and relevance. We present a number of improvements to prior modeling approaches, including (1) the addition of a dual learning framework that utilizes video captioning to reinforce the semantic alignment between the story and generated images, (2) a copy-transform mechanism for sequentially-consistent story visualization, and (3) MART-based transformers to model complex interactions between frames. We present ablation studies to demonstrate the effect of each of these techniques on the generative power of the model for both individual images as well as the entire narrative. Furthermore, due to the complexity and generative nature of the task, standard evaluation metrics do not accurately reflect performance. Therefore, we also provide an exploration of evaluation metrics for the model, focused on aspects of the generated frames such as the presence/quality of generated characters, the relevance to captions, and the diversity of the generated images. We also present correlation experiments of our proposed automated metrics with human evaluations. Code and data available at: https://github.com/adymaharana/StoryViz

arXiv:2303.06869공식 코드 · Papers with Code

Adaptive Data-Free Quantization

Data-free quantization (DFQ) recovers the performance of quantized network (Q) without the original data, but generates the fake sample via a generator (G) by learning from full-precision network (P), which, however, is totally independent of Q, overlooking the adaptability of the knowledge from generated samples, i.e., informative or not to the learning process of Q, resulting into the overflow of generalization error. Building on this, several critical questions -- how to measure the sample adaptability to Q under varied bit-width scenarios? whether the largest adaptability is the best? how to generate the samples with adaptive adaptability to improve Q's generalization? To answer the above questions, in this paper, we propose an Adaptive Data-Free Quantization (AdaDFQ) method, which revisits DFQ from a zero-sum game perspective upon the sample adaptability between two players -- a generator and a quantized network. Following this viewpoint, we further define the disagreement and agreement samples to form two boundaries, where the margin is optimized to adaptively regulate the adaptability of generated samples to Q, so as to address the over-and-under fitting issues. Our AdaDFQ reveals: 1) the largest adaptability is NOT the best for sample generation to benefit Q's generalization; 2) the knowledge of the generated sample should not be informative to Q only, but also related to the category and distribution information of the training data for P. The theoretical and empirical analysis validate the advantages of AdaDFQ over the state-of-the-arts. Our code is available at https://github.com/hfutqian/AdaDFQ.

arXiv:2309.16289공식 코드 · Papers with Code

LawBench: Benchmarking Legal Knowledge of Large Language Models

Large language models (LLMs) have demonstrated strong capabilities in various aspects. However, when applying them to the highly specialized, safe-critical legal domain, it is unclear how much legal knowledge they possess and whether they can reliably perform legal-related tasks. To address this gap, we propose a comprehensive evaluation benchmark LawBench. LawBench has been meticulously crafted to have precise assessment of the LLMs' legal capabilities from three cognitive levels: (1) Legal knowledge memorization: whether LLMs can memorize needed legal concepts, articles and facts; (2) Legal knowledge understanding: whether LLMs can comprehend entities, events and relationships within legal text; (3) Legal knowledge applying: whether LLMs can properly utilize their legal knowledge and make necessary reasoning steps to solve realistic legal tasks. LawBench contains 20 diverse tasks covering 5 task types: single-label classification (SLC), multi-label classification (MLC), regression, extraction and generation. We perform extensive evaluations of 51 LLMs on LawBench, including 20 multilingual LLMs, 22 Chinese-oriented LLMs and 9 legal specific LLMs. The results show that GPT-4 remains the best-performing LLM in the legal domain, surpassing the others by a significant margin. While fine-tuning LLMs on legal specific text brings certain improvements, we are still a long way from obtaining usable and reliable LLMs in legal tasks. All data, model predictions and evaluation code are released in https://github.com/open-compass/LawBench/. We hope this benchmark provides in-depth understanding of the LLMs' domain-specified capabilities and speed up the development of LLMs in the legal domain.

arXiv:2310.12708공식 코드 · Papers with Code

Generating Robust Adversarial Examples against Online Social Networks (OSNs)

Online Social Networks (OSNs) have blossomed into prevailing transmission channels for images in the modern era. Adversarial examples (AEs) deliberately designed to mislead deep neural networks (DNNs) are found to be fragile against the inevitable lossy operations conducted by OSNs. As a result, the AEs would lose their attack capabilities after being transmitted over OSNs. In this work, we aim to design a new framework for generating robust AEs that can survive the OSN transmission; namely, the AEs before and after the OSN transmission both possess strong attack capabilities. To this end, we first propose a differentiable network termed SImulated OSN (SIO) to simulate the various operations conducted by an OSN. Specifically, the SIO network consists of two modules: 1) a differentiable JPEG layer for approximating the ubiquitous JPEG compression and 2) an encoder-decoder subnetwork for mimicking the remaining operations. Based upon the SIO network, we then formulate an optimization framework to generate robust AEs by enforcing model outputs with and without passing through the SIO to be both misled. Extensive experiments conducted over Facebook, WeChat and QQ demonstrate that our attack methods produce more robust AEs than existing approaches, especially under small distortion constraints; the performance gain in terms of Attack Success Rate (ASR) could be more than 60%. Furthermore, we build a public dataset containing more than 10,000 pairs of AEs processed by Facebook, WeChat or QQ, facilitating future research in the robust AEs generation. The dataset and code are available at https://github.com/csjunjun/RobustOSNAttack.git.

arXiv:2406.12832공식 코드 · Papers with Code

LaMDA: Large Model Fine-Tuning via Spectrally Decomposed Low-Dimensional Adaptation

Low-rank adaptation (LoRA) has become the default approach to fine-tune large language models (LLMs) due to its significant reduction in trainable parameters. However, trainable parameter demand for LoRA increases with increasing model embedding dimensions, leading to high compute costs. Additionally, its backward updates require storing high-dimensional intermediate activations and optimizer states, demanding high peak GPU memory. In this paper, we introduce large model fine-tuning via spectrally decomposed low-dimensional adaptation (LaMDA), a novel approach to fine-tuning large language models, which leverages low-dimensional adaptation to achieve significant reductions in trainable parameters and peak GPU memory footprint. LaMDA freezes a first projection matrix (PMA) in the adaptation path while introducing a low-dimensional trainable square matrix, resulting in substantial reductions in trainable parameters and peak GPU memory usage. LaMDA gradually freezes a second projection matrix (PMB) during the early fine-tuning stages, reducing the compute cost associated with weight updates to enhance parameter efficiency further. We also present an enhancement, LaMDA++, incorporating a ``lite-weight" adaptive rank allocation for the LoRA path via normalized spectrum analysis of pre-trained model weights. We evaluate LaMDA/LaMDA++ across various tasks, including natural language understanding with the GLUE benchmark, text summarization, natural language generation, and complex reasoning on different LLMs. Results show that LaMDA matches or surpasses the performance of existing alternatives while requiring up to 17.7x fewer parameter updates and up to 1.32x lower peak GPU memory usage during fine-tuning. Code will be publicly available.

arXiv:2407.03550공식 코드 · Papers with Code

CoMix: A Comprehensive Benchmark for Multi-Task Comic Understanding

The comic domain is rapidly advancing with the development of single-page analysis and synthesis models. However, evaluation metrics and datasets lag behind, often limited to small-scale or single-style test sets. We introduce a novel benchmark, CoMix, designed to evaluate the multi-task capabilities of models in comic analysis. Unlike existing benchmarks that focus on isolated tasks such as object detection or text recognition, CoMix addresses a broader range of tasks including object detection, speaker identification, character re-identification, reading order, and multi-modal reasoning tasks like character naming and dialogue generation. Our benchmark comprises three existing datasets with expanded annotations to support multi-task evaluation. To mitigate the over-representation of manga-style data, we have incorporated a new dataset of carefully selected American comic-style books, thereby enriching the diversity of comic styles. CoMix is designed to assess pre-trained models in zero-shot and limited fine-tuning settings, probing their transfer capabilities across different comic styles and tasks. The validation split of the benchmark is publicly available for research purposes, and an evaluation server for the held-out test split is also provided. Comparative results between human performance and state-of-the-art models reveal a significant performance gap, highlighting substantial opportunities for advancements in comic understanding. The dataset, baseline models, and code are accessible at https://github.com/emanuelevivoli/CoMix-dataset. This initiative sets a new standard for comprehensive comic analysis, providing the community with a common benchmark for evaluation on a large and varied set.

arXiv:2407.09174공식 코드 · Papers with Code

DART: An Automated End-to-End Object Detection Pipeline with Data Diversification, Open-Vocabulary Bounding Box Annotation, Pseudo-Label Review, and Model Training

Accurate real-time object detection is vital across numerous industrial applications, from safety monitoring to quality control. Traditional approaches, however, are hindered by arduous manual annotation and data collection, struggling to adapt to ever-changing environments and novel target objects. To address these limitations, this paper presents DART, an innovative automated end-to-end pipeline that revolutionizes object detection workflows from data collection to model evaluation. It eliminates the need for laborious human labeling and extensive data collection while achieving outstanding accuracy across diverse scenarios. DART encompasses four key stages: (1) Data Diversification using subject-driven image generation (DreamBooth with SDXL), (2) Annotation via open-vocabulary object detection (Grounding DINO) to generate bounding box and class labels, (3) Review of generated images and pseudo-labels by large multimodal models (InternVL-1.5 and GPT-4o) to guarantee credibility, and (4) Training of real-time object detectors (YOLOv8 and YOLOv10) using the verified data. We apply DART to a self-collected dataset of construction machines named Liebherr Product, which contains over 15K high-quality images across 23 categories. The current instantiation of DART significantly increases average precision (AP) from 0.064 to 0.832. Its modular design ensures easy exchangeability and extensibility, allowing for future algorithm upgrades, seamless integration of new object categories, and adaptability to customized environments without manual labeling and additional data collection. The code and dataset are released at https://github.com/chen-xin-94/DART.

arXiv:2407.13511공식 코드 · Papers with Code

Can Open-Source LLMs Compete with Commercial Models? Exploring the Few-Shot Performance of Current GPT Models in Biomedical Tasks

Commercial large language models (LLMs), like OpenAI's GPT-4 powering ChatGPT and Anthropic's Claude 3 Opus, have dominated natural language processing (NLP) benchmarks across different domains. New competing Open-Source alternatives like Mixtral 8x7B or Llama 3 have emerged and seem to be closing the gap while often offering higher throughput and being less costly to use. Open-Source LLMs can also be self-hosted, which makes them interesting for enterprise and clinical use cases where sensitive data should not be processed by third parties. We participated in the 12th BioASQ challenge, which is a retrieval augmented generation (RAG) setting, and explored the performance of current GPT models Claude 3 Opus, GPT-3.5-turbo and Mixtral 8x7b with in-context learning (zero-shot, few-shot) and QLoRa fine-tuning. We also explored how additional relevant knowledge from Wikipedia added to the context-window of the LLM might improve their performance. Mixtral 8x7b was competitive in the 10-shot setting, both with and without fine-tuning, but failed to produce usable results in the zero-shot setting. QLoRa fine-tuning and Wikipedia context did not lead to measurable performance gains. Our results indicate that the performance gap between commercial and open-source models in RAG setups exists mainly in the zero-shot setting and can be closed by simply collecting few-shot examples for domain-specific use cases. The code needed to rerun these experiments is available through GitHub.

arXiv:2412.10047공식 코드 · Papers with Code

Large Action Models: From Inception to Implementation

As AI continues to advance, there is a growing demand for systems that go beyond language-based assistance and move toward intelligent agents capable of performing real-world actions. This evolution requires the transition from traditional Large Language Models (LLMs), which excel at generating textual responses, to Large Action Models (LAMs), designed for action generation and execution within dynamic environments. Enabled by agent systems, LAMs hold the potential to transform AI from passive language understanding to active task completion, marking a significant milestone in the progression toward artificial general intelligence. In this paper, we present a comprehensive framework for developing LAMs, offering a systematic approach to their creation, from inception to deployment. We begin with an overview of LAMs, highlighting their unique characteristics and delineating their differences from LLMs. Using a Windows OS-based agent as a case study, we provide a detailed, step-by-step guide on the key stages of LAM development, including data collection, model training, environment integration, grounding, and evaluation. This generalizable workflow can serve as a blueprint for creating functional LAMs in various application domains. We conclude by identifying the current limitations of LAMs and discussing directions for future research and industrial deployment, emphasizing the challenges and opportunities that lie ahead in realizing the full potential of LAMs in real-world applications. The code for the data collection process utilized in this paper is publicly available at: https://github.com/microsoft/UFO/tree/main/dataflow, and comprehensive documentation can be found at https://microsoft.github.io/UFO/dataflow/overview/.

arXiv:2503.20047공식 코드 · Papers with Code

Med3DVLM: An Efficient Vision-Language Model for 3D Medical Image Analysis

Vision-language models (VLMs) have shown promise in 2D medical image analysis, but extending them to 3D remains challenging due to the high computational demands of volumetric data and the difficulty of aligning 3D spatial features with clinical text. We present Med3DVLM, a 3D VLM designed to address these challenges through three key innovations: (1) DCFormer, an efficient encoder that uses decomposed 3D convolutions to capture fine-grained spatial features at scale; (2) SigLIP, a contrastive learning strategy with pairwise sigmoid loss that improves image-text alignment without relying on large negative batches; and (3) a dual-stream MLP-Mixer projector that fuses low- and high-level image features with text embeddings for richer multi-modal representations. We evaluate our model on the M3D dataset, which includes radiology reports and VQA data for 120,084 3D medical images. Results show that Med3DVLM achieves superior performance across multiple benchmarks. For image-text retrieval, it reaches 61.00% R@1 on 2,000 samples, significantly outperforming the current state-of-the-art M3D model (19.10%). For report generation, it achieves a METEOR score of 36.42% (vs. 14.38%). In open-ended visual question answering (VQA), it scores 36.76% METEOR (vs. 33.58%), and in closed-ended VQA, it achieves 79.95% accuracy (vs. 75.78%). These results highlight Med3DVLM's ability to bridge the gap between 3D imaging and language, enabling scalable, multi-task reasoning across clinical applications. Our code is publicly available at https://github.com/mirthAI/Med3DVLM.

arXiv:2506.10391공식 코드 · Papers with Code

ReconMOST: Multi-Layer Sea Temperature Reconstruction with Observations-Guided Diffusion

Accurate reconstruction of ocean is essential for reflecting global climate dynamics and supporting marine meteorological research. Conventional methods face challenges due to sparse data, algorithmic complexity, and high computational costs, while increasing usage of machine learning (ML) method remains limited to reconstruction problems at the sea surface and local regions, struggling with issues like cloud occlusion. To address these limitations, this paper proposes ReconMOST, a data-driven guided diffusion model framework for multi-layer sea temperature reconstruction. Specifically, we first pre-train an unconditional diffusion model using a large collection of historical numerical simulation data, enabling the model to attain physically consistent distribution patterns of ocean temperature fields. During the generation phase, sparse yet high-accuracy in-situ observational data are utilized as guidance points for the reverse diffusion process, generating accurate reconstruction results. Importantly, in regions lacking direct observational data, the physically consistent spatial distribution patterns learned during pre-training enable implicitly guided and physically plausible reconstructions. Our method extends ML-based SST reconstruction to a global, multi-layer setting, handling over 92.5% missing data while maintaining reconstruction accuracy, spatial resolution, and superior generalization capability. We pre-train our model on CMIP6 numerical simulation data and conduct guided reconstruction experiments on CMIP6 and EN4 analysis data. The results of mean squared error (MSE) values achieve 0.049 on guidance, 0.680 on reconstruction, and 0.633 on total, respectively, demonstrating the effectiveness and robustness of the proposed framework. Our source code is available at https://github.com/norsheep/ReconMOST.

arXiv:2506.20024공식 코드 · Papers with Code

Elucidated Rolling Diffusion Models for Probabilistic Weather Forecasting

Diffusion models are a powerful tool for probabilistic forecasting, yet most applications in high-dimensional chaotic systems predict future snapshots one-by-one. This common approach struggles to model complex temporal dependencies and fails to explicitly account for the progressive growth of uncertainty inherent to such systems. While rolling diffusion frameworks, which apply increasing noise to forecasts at longer lead times, have been proposed to address this, their integration with state-of-the-art, high-fidelity diffusion techniques remains a significant challenge. We tackle this problem by introducing Elucidated Rolling Diffusion Models (ERDM), the first framework to successfully unify a rolling forecast structure with the principled, performant design of Elucidated Diffusion Models (EDM). To do this, we adapt the core EDM components-its noise schedule, network preconditioning, and Heun sampler-to the rolling forecast setting. The success of this integration is driven by three key contributions: (i) a novel loss weighting scheme that focuses model capacity on the mid-range forecast horizons where determinism gives way to stochasticity; (ii) an efficient initialization strategy using a pre-trained EDM for the initial window; and (iii) a bespoke hybrid sequence architecture for robust spatiotemporal feature extraction under progressive denoising. On 2D Navier-Stokes simulations and ERA5 global weather forecasting at 1.5^\circ resolution, ERDM consistently outperforms key diffusion-based baselines, including conditional autoregressive EDM. ERDM offers a flexible and powerful general framework for tackling diffusion-based sequence generation problems where modeling escalating uncertainty is paramount. Code is available at: https://github.com/salvaRC/erdm

arXiv:2611.14492v1cs.AI

ReTrace: 도구 사용 에이전트를 플래너 없는 워크플로로 증류하기

ReTrace: Distilling Tool-Using Agents into Planner-Free Workflows

비즈니스 프로세스 자동화는 동일한 표준 운영 절차(SOP)를 새로운 입력에 대해 반복적으로 실행한다. 거대 언어 모델(LLM) 에이전트는 자연어로 작성된 SOP를 따를 수 있지만, 매 실행의 모든 단계를 선택하기 위해 LLM을 호출하므로 대규모에서 지연 시간과 비용이 증가한다. 본 논문은 SOP와 소수의 기록된 에이전트 실행으로부터, LLM이 아닌 코드가 다음 단계를 선택하는 재사용 가능한 워크플로를 만들어 내는 ReTrace를 제시한다. 에이전트의 추론, 즉 각 도구 호출과 함께 작성하는 설명은 행동을 취한 이유를 명시하는 경우가 많지만, 도구 호출만으로는 그렇지 않다. ReTrace는 각 단계의 근거가 되는 조건을 추출하고, 이 조건을 여러 실행에 걸쳐 병합하여 간결한 규칙 문서를 만든 다음, 이로부터 실행 가능한 워크플로를 생성한다. 마지막 단계만 플랫폼에 의존하므로, 규칙 문서는 서로 다른 오케스트레이션 플랫폼용으로 컴파일될 수 있다. 자사 프로덕션 플랫폼의 $79$개 절차에서 ReTrace 워크플로는 에이전트의 $96.8\%$에 대해 $93.8\%$의 완료율을 달성하며 과제당 비용은 $24\%$ 낮다. $τ^2$-Bench에서는 $75.8\%$의 과제 성공률(에이전트: $80.6\%$)을 달성하고 $170$초 대신 $1.3$초에 실행되며, 학습 실패 사례에 대한 세 차례의 수리는 $k{=}2$ 워크플로를 $66.7\%$에서 $81.2\%$로 끌어올린다. 8개의 SOP-Bench 절차에서 수리 없는 ReTrace는 퓨샷 생성보다 뒤처지지만($80.6\%$ 대 $83.6\%$), 라벨이 있는 학습 실패 사례로부터의 수리를 통해 과제당 $0.40$초에 $87.8\%$까지 향상되며, 이는 에이전트의 $46.6$초에 $85.5\%$와 비교된다.

arXiv:2607.16038v1cs.AI

SciForge: 과학적 발견을 위한 AI 네이티브 멀티모달 워크벤치

SciForge: An AI-Native, Multimodal Workbench for Scientific Discovery

과학 연구는 점점 논문, 코드, 데이터셋, 과학 파일 형식, 모델 출력, 그림, 원고, 팀 의사결정 등 이질적인 산출물에 걸쳐 이루어지지만, 범용 AI 어시스턴트는 이러한 객체들을 일관되고 감사 가능한 연구 상태로 보존하는 경우가 드물다. 본 논문에서는 SciForge를 제안한다. SciForge는 멀티모달 연구 네이티브 AI 워크벤치로, 그래픽 인터페이스는 인간의 판단을 위해 남겨두면서 검색, 파싱, 모델 라우팅, 워크플로 실행, 플로팅, 작성, 프레젠테이션 생성은 에이전트가 접근 가능한 모듈형 서비스로 동작한다. SciForge는 다섯 가지 기둥을 중심으로 구축된다: (i) \emph{목표 범위 과학적 의사결정 거버넌스(goal-scoped scientific decision governance)}를 통한 \textbf{목표 지향적(goal-oriented)} 연구, 검토 게이트와 공유 검토 표면; (ii) \emph{번역 후 추론(translate-then-reason)}을 통한 \textbf{멀티모달(multimodal)} 입력, 에이전트가 추론하기 전에 과학 객체를 도메인 번역기를 통해 라우팅; (iii) \emph{증거 거버넌스(evidence governance)}를 통한 \textbf{감사 가능한(auditable)} 추적 가능성, 주장을 프로비넌스 체인과 감사 결과에 연결; (iv) \emph{협력적 팀 과학(collaborative team science)}을 통한 \textbf{협력적(collaborative)} 연구, 다중 역할 의사결정 거버넌스 지원, 공유 팀 워크스페이스는 향후 릴리스 예정; (v) \emph{실세계 응용 시나리오(real-world application scenarios)}를 통한 \textbf{실용적(practical)} 영향, 8개의 end-to-end 사용자 사례로 시연, 대표 시연으로 유전자 발견, AI 안내 de novo 단백질 설계, 분자 최적화, genome-to-BGC 발견을 위한 다일 에이전트 연구 스프린트 포함. 이 시스템은 얇은 상호작용 층, 맥락적 연구 역량 패턴, 에이전트 런타임 및 워크플로 엔진, Evidence-DAG 감사 사이드카, 과학 모델 라우터를 결합한다. SciForge는 현재 데스크톱 애플리케이션으로 실행되며 모바일 감독 지원을 제공한다. 향후 릴리스에서는 팀 협업을 심화할 예정이다. 이 시스템은 오픈소스이며 https://github.com/AGI4Sci/SciForge 에서 이용할 수 있다.

arXiv:2603.02259v3cs.MA

정렬 플라이휠: 아키텍처에 구애받지 않는 안전성을 위한 거버넌스 중심 하이브리드 MAS

The Alignment Flywheel: A Governance-Centric Hybrid MAS for Architecture-Agnostic Safety

다중 에이전트 시스템(MAS)은 역할 분해, 조정, 규범적 거버넌스를 위한 성숙한 추상화를 제공하지만, 학습된 컴포넌트의 성능이 점차 향상됨에 따라 배포 후 안전성(safety)을 점검, 감사 및 업데이트하기가 더 어려워지고 있습니다. 안전성 행동이 의사결정 컴포넌트에 흡수되면, 국소적인 실패만으로도 전체 컴포넌트의 재학습(retraining)이나 롤백이 필요할 수 있습니다. 본 연구는 의사결정 생성과 안전성 거버넌스를 분리하는 거버넌스 중심 하이브리드 MAS 아키텍처로서 정렬(alignment) 플라이휠에 대한 비전을 구체화합니다. 후보 궤적을 생성하는 에이전트 또는 정책을 제안자(Proposer)로 정의하며, 이는 출력을 거버넌스된 안전성 오라클(Safety Oracle) 스택으로 전달합니다. 이 오라클 스택은 안정적인 인터페이스를 통해 안전성 점수, 예측 불확실성, 감사 커버리지 불확실성, 증거 훅을 반환합니다. 집행 층(Enforcement layer)은 런타임에 명시적인 위험 정책을 적용합니다. 이 루프를 중심으로 거버넌스 MAS는 모니터링, 레드팀 테스팅(red-teaming), 검증, 트리아지, 개선 및 버전 관리되는 릴리스 관리를 수행합니다. 핵심적인 엔지니어링 원칙은 패치 국소성(patch locality)입니다. 새로 관찰된 많은 안전성 실패는 제안자를 재학습하거나 철회하는 대신, 오라클 스택과 해당 감사 상태에 대한 소규모 거버넌스 배치를 통해 완화될 수 있습니다. 이 아키텍처는 제안자와 오라클 모두에 대해 구현에 구애받지 않습니다(implementation-agnostic). 또한 런타임 게이팅, 감사 수집, 서명된 업데이트, 단계적 롤아웃 및 롤백에 필요한 역할, 산출물, 프로토콜, 릴리스 시맨틱을 정의합니다. 우리는 회귀 검증된 거버넌스 업데이트를 통해 패치된 학습 기반 공간 오라클과 구조화된 규범, 에스컬레이션, 감사 커버리지를 보여주는 임상 GenAI 프록시 환경의 두 가지 시나리오에서 실행 가능성을 입증합니다. 구현 코드와 문서는 https://github.com/decide-ugent/Alignment-Flywheel 에서 오픈 소스로 제공됩니다.

arXiv:2610.09804v1cs.LG

BoT-GRPO: 토큰 주머니 집계를 통한 추론용 고효율 프로세스 보상 강화학습

BoT-GRPO: Efficient Process-Reward RL for Reasoning via Bag-of-Token Aggregation

강화학습(RL)은 이제 대규모 언어 모델(LLM)에서 추론 능력을 이끌어내는 데 핵심적인 역할을 하지만, 널리 사용되는 그룹 상대 정책 최적화(GRPO) 알고리즘에서는 롤아웃(rollout) 내의 모든 토큰이 동일한 어드밴티지(advantage)를 부여받습니다. 본 연구에서는 가치 네트워크(value network)의 비용을 들이지 않고 수렴을 가속화하며 최종 품질을 향상시키는, 효율적인 프로세스 감독(process supervision) 방안을 모색합니다. 이를 위해 길이 불변 '토큰 주머니(bag of tokens)' 집계를 통해 GRPO를 토큰 수준 보상 모델로 확장하는 BoT-GRPO(Bag-of-Tokens Group Relative Policy Optimization)를 제안합니다. 이 방법은 여러 롤아웃에 걸쳐 모든 토큰 수준 보상을 수집하고, 각 보상에 원래 시퀀스 길이의 역수를 가중치로 부여한 뒤 가중치가 적용된 그룹 통계를 기준으로 토큰별 어드밴티지를 계산합니다. BoT-GRPO는 크리틱이 필요 없는(critic-free) 방식이며, 토큰 수준 보상을 활용할 수 있는 환경이라면 기존 GRPO를 완벽히 대체(drop-in replacement)할 수 있습니다. React 프론트엔드 코드 생성 작업에서 BoT-GRPO는 GRPO보다 최대 $1.9\times$ 빠르게 $80\%$ 컴파일 비율에 도달하며, 최신 GRPO 변형 모델(GSPO, DAPO, PURE)보다 더 빠르게 수렴하는 동시에 더 높은 최종 컴파일 비율과 VLM 평가 승률을 달성합니다. 두 번째 작업인 AIME 수학적 추론에서는 BoT-GRPO가 절반의 스텝만으로 GRPO 대비 최대 $8.1\%$의 절대적 Pass@$k$ 향상을 보여줍니다. 두 작업 모두에서 추론 계열과 비추론 계열의 베이스 모델 패밀리(Qwen2.5-3B, SmolLM3-3B, Phi-4-mini-reasoning)를 대상으로 알고리즘 성능을 비교합니다. 또한 본 실험은 보상 모델 자체에 대한 실용적인 지침을 제시합니다. 즉, 보상의 풍부함보다 안정성이 더 중요하며, 노이즈가 있는 대안들이 학습을 정체시키는 반면 정제되고 경계가 명확하며(bounded) 안정적인 세분화된 신호는 일관되게 학습을 가속화합니다.

arXiv:1711.09561

HP-GAN: Probabilistic 3D human motion prediction via GAN

Predicting and understanding human motion dynamics has many applications, such as motion synthesis, augmented reality, security, and autonomous vehicles. Due to the recent success of generative adversarial networks (GAN), there has been much interest in probabilistic estimation and synthetic data generation using deep neural network architectures and learning algorithms. We propose a novel sequence-to-sequence model for probabilistic human motion prediction, trained with a modified version of improved Wasserstein generative adversarial networks (WGAN-GP), in which we use a custom loss function designed for human motion prediction. Our model, which we call HP-GAN, learns a probability density function of future human poses conditioned on previous poses. It predicts multiple sequences of possible future human poses, each from the same input sequence but a different vector z drawn from a random distribution. Furthermore, to quantify the quality of the non-deterministic predictions, we simultaneously train a motion-quality-assessment model that learns the probability that a given skeleton sequence is a real human motion. We test our algorithm on two of the largest skeleton datasets: NTURGB-D and Human3.6M. We train our model on both single and multiple action types. Its predictive power for long-term motion estimation is demonstrated by generating multiple plausible futures of more than 30 frames from just 10 frames of input. We show that most sequences generated from the same input have more than 50\% probabilities of being judged as a real human sequence. We will release all the code used in this paper to Github.