PyTorchKR Papers

“Code Generation”

총 10,163편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2408.15914

CoRe: Context-Regularized Text Embedding Learning for Text-to-Image Personalization

Recent advances in text-to-image personalization have enabled high-quality and controllable image synthesis for user-provided concepts. However, existing methods still struggle to balance identity preservation with text alignment. Our approach is based on the fact that generating prompt-aligned images requires a precise semantic understanding of the prompt, which involves accurately processing the interactions between the new concept and its surrounding context tokens within the CLIP text encoder. To address this, we aim to embed the new concept properly into the input embedding space of the text encoder, allowing for seamless integration with existing tokens. We introduce Context Regularization (CoRe), which enhances the learning of the new concept's text embedding by regularizing its context tokens in the prompt. This is based on the insight that appropriate output vectors of the text encoder for the context tokens can only be achieved if the new concept's text embedding is correctly learned. CoRe can be applied to arbitrary prompts without requiring the generation of corresponding images, thus improving the generalization of the learned text embedding. Additionally, CoRe can serve as a test-time optimization technique to further enhance the generations for specific prompts. Comprehensive experiments demonstrate that our method outperforms several baseline methods in both identity preservation and text alignment. Code will be made publicly available.

arXiv:2410.18686

Hierarchical Multimodal LLMs with Semantic Space Alignment for Enhanced Time Series Classification

Leveraging large language models (LLMs) has garnered increasing attention and introduced novel perspectives in time series classification. However, existing approaches often overlook the crucial dynamic temporal information inherent in time series data and face challenges in aligning this data with textual semantics. To address these limitations, we propose HiTime, a hierarchical multi-modal model that seamlessly integrates temporal information into LLMs for multivariate time series classification (MTSC). Our model employs a hierarchical feature encoder to capture diverse aspects of time series data through both data-specific and task-specific embeddings. To facilitate semantic space alignment between time series and text, we introduce a dual-view contrastive alignment module that bridges the gap between modalities. Additionally, we adopt a hybrid prompting strategy to fine-tune the pre-trained LLM in a parameter-efficient manner. By effectively incorporating dynamic temporal features and ensuring semantic alignment, HiTime enables LLMs to process continuous time series data and achieves state-of-the-art classification performance through text generation. Extensive experiments on benchmark datasets demonstrate that HiTime significantly enhances time series classification accuracy compared to most competitive baseline methods. Our findings highlight the potential of integrating temporal features into LLMs, paving the way for advanced time series analysis. The code is publicly available for further research and validation. Our codes are publicly available1.

arXiv:2501.09927

IE-Bench: Advancing the Measurement of Text-Driven Image Editing for Human Perception Alignment

Recent advances in text-driven image editing have been significant, yet the task of accurately evaluating these edited images continues to pose a considerable challenge. Different from the assessment of text-driven image generation, text-driven image editing is characterized by simultaneously conditioning on both text and a source image. The edited images often retain an intrinsic connection to the original image, which dynamically change with the semantics of the text. However, previous methods tend to solely focus on text-image alignment or have not aligned with human perception. In this work, we introduce the Text-driven Image Editing Benchmark suite (IE-Bench) to enhance the assessment of text-driven edited images. IE-Bench includes a database contains diverse source images, various editing prompts and the corresponding results different editing methods, and total 3,010 Mean Opinion Scores (MOS) provided by 25 human subjects. Furthermore, we introduce IE-QA, a multi-modality source-aware quality assessment method for text-driven image editing. To the best of our knowledge, IE-Bench offers the first IQA dataset and model tailored for text-driven image editing. Extensive experiments demonstrate IE-QA's superior subjective-alignments on the text-driven image editing task compared with previous metrics. We will make all related data and code available to the public.

arXiv:2501.18921공식 코드 · Papers with Code

Full-scale Representation Guided Network for Retinal Vessel Segmentation

The U-Net architecture and its variants have remained state-of-the-art (SOTA) for retinal vessel segmentation over the past decade. In this study, we introduce a Full Scale Guided Network (FSG-Net), where the feature representation network with modernized convolution blocks extracts full-scale information and the guided convolution block refines that information. Attention-guided filter is introduced to the guided convolution block under the interpretation that the filter behaves like the unsharp mask filter. Passing full-scale information to the attention block allows for the generation of improved attention maps, which are then passed to the attention-guided filter, resulting in performance enhancement of the segmentation network. The structure preceding the guided convolution block can be replaced by any U-Net variant, which enhances the scalability of the proposed approach. For a fair comparison, we re-implemented recent studies available in public repositories to evaluate their scalability and reproducibility. Our experiments also show that the proposed network demonstrates competitive results compared to current SOTA models on various public datasets. Ablation studies demonstrate that the proposed model is competitive with much smaller parameter sizes. Lastly, by applying the proposed model to facial wrinkle segmentation, we confirmed the potential for scalability to similar tasks in other domains. Our code is available on https://github.com/ZombaSY/FSG-Net-pytorch.

arXiv:2503.00564공식 코드 · Papers with Code

ToolDial: Multi-turn Dialogue Generation Method for Tool-Augmented Language Models

Tool-Augmented Language Models (TALMs) leverage external APIs to answer user queries across various domains. However, existing benchmark datasets for TALM research often feature simplistic dialogues that do not reflect real-world scenarios, such as the need for models to ask clarifying questions or proactively call additional APIs when essential information is missing. To address these limitations, we construct and release ToolDial, a dataset comprising 11,111 multi-turn dialogues, with an average of 8.95 turns per dialogue, based on APIs from RapidAPI. ToolDial has two key characteristics. First, the dialogues incorporate 16 user and system actions (e.g., "Request", "Clarify", "Fail inform") to capture the rich dynamics of real-world interactions. Second, we simulate dialogues where the system requests necessary information from the user based on API documentation and seeks additional APIs if the user fails to provide the required information. To facilitate this process, we introduce a method for generating an API graph that represents input and output compatibility between APIs. Using ToolDial, we evaluate a suite of language models on their ability to predict correct actions and extract input parameter values for API calls from the dialogue history. Modern language models achieve accuracy scores below 70%, indicating substantial room for improvement. We release our dataset and code at https://github.com/holi-lab/ToolDial.

arXiv:2503.06166공식 코드 · Papers with Code

Secure On-Device Video OOD Detection Without Backpropagation

Out-of-Distribution (OOD) detection is critical for ensuring the reliability of machine learning models in safety-critical applications such as autonomous driving and medical diagnosis. While deploying personalized OOD detection directly on edge devices is desirable, it remains challenging due to large model sizes and the computational infeasibility of on-device training. Federated learning partially addresses this but still requires gradient computation and backpropagation, exceeding the capabilities of many edge devices. To overcome these challenges, we propose SecDOOD, a secure cloud-device collaboration framework for efficient on-device OOD detection without requiring device-side backpropagation. SecDOOD utilizes cloud resources for model training while ensuring user data privacy by retaining sensitive information on-device. Central to SecDOOD is a HyperNetwork-based personalized parameter generation module, which adapts cloud-trained models to device-specific distributions by dynamically generating local weight adjustments, effectively combining central and local information without local fine-tuning. Additionally, our dynamic feature sampling and encryption strategy selectively encrypts only the most informative feature channels, largely reducing encryption overhead without compromising detection performance. Extensive experiments across multiple datasets and OOD scenarios demonstrate that SecDOOD achieves performance comparable to fully fine-tuned models, enabling secure, efficient, and personalized OOD detection on resource-limited edge devices. To enhance accessibility and reproducibility, our code is publicly available at https://github.com/Dystopians/SecDOOD.

arXiv:2503.10687공식 코드 · Papers with Code

Context-guided Responsible Data Augmentation with Diffusion Models

Generative diffusion models offer a natural choice for data augmentation when training complex vision models. However, ensuring reliability of their generative content as augmentation samples remains an open challenge. Despite a number of techniques utilizing generative images to strengthen model training, it remains unclear how to utilize the combination of natural and generative images as a rich supervisory signal for effective model induction. In this regard, we propose a text-to-image (T2I) data augmentation method, named DiffCoRe-Mix, that computes a set of generative counterparts for a training sample with an explicitly constrained diffusion model that leverages sample-based context and negative prompting for a reliable augmentation sample generation. To preserve key semantic axes, we also filter out undesired generative samples in our augmentation process. To that end, we propose a hard-cosine filtration in the embedding space of CLIP. Our approach systematically mixes the natural and generative images at pixel and patch levels. We extensively evaluate our technique on ImageNet-1K,Tiny ImageNet-200, CIFAR-100, Flowers102, CUB-Birds, Stanford Cars, and Caltech datasets, demonstrating a notable increase in performance across the board, achieving up to $\sim 3\%$ absolute gain for top-1 accuracy over the state-of-the-art methods, while showing comparable computational overhead. Our code is publicly available at https://github.com/khawar-islam/DiffCoRe-Mix

arXiv:2503.13964공식 코드 · Papers with Code

MDocAgent: A Multi-Modal Multi-Agent Framework for Document Understanding

Document Question Answering (DocQA) is a very common task. Existing methods using Large Language Models (LLMs) or Large Vision Language Models (LVLMs) and Retrieval Augmented Generation (RAG) often prioritize information from a single modal, failing to effectively integrate textual and visual cues. These approaches struggle with complex multi-modal reasoning, limiting their performance on real-world documents. We present MDocAgent (A Multi-Modal Multi-Agent Framework for Document Understanding), a novel RAG and multi-agent framework that leverages both text and image. Our system employs five specialized agents: a general agent, a critical agent, a text agent, an image agent and a summarizing agent. These agents engage in multi-modal context retrieval, combining their individual insights to achieve a more comprehensive understanding of the document's content. This collaborative approach enables the system to synthesize information from both textual and visual components, leading to improved accuracy in question answering. Preliminary experiments on five benchmarks like MMLongBench, LongDocURL demonstrate the effectiveness of our MDocAgent, achieve an average improvement of 12.1% compared to current state-of-the-art method. This work contributes to the development of more robust and comprehensive DocQA systems capable of handling the complexities of real-world documents containing rich textual and visual information. Our data and code are available at https://github.com/aiming-lab/MDocAgent.

arXiv:2504.05815

Parasite: A Steganography-based Backdoor Attack Framework for Diffusion Models

Recently, the diffusion model has gained significant attention as one of the most successful image generation models, which can generate high-quality images by iteratively sampling noise. However, recent studies have shown that diffusion models are vulnerable to backdoor attacks, allowing attackers to enter input data containing triggers to activate the backdoor and generate their desired output. Existing backdoor attack methods primarily focused on target noise-to-image and text-to-image tasks, with limited work on backdoor attacks in image-to-image tasks. Furthermore, traditional backdoor attacks often rely on a single, conspicuous trigger to generate a fixed target image, lacking concealability and flexibility. To address these limitations, we propose a novel backdoor attack method called "Parasite" for image-to-image tasks in diffusion models, which not only is the first to leverage steganography for triggers hiding, but also allows attackers to embed the target content as a backdoor trigger to achieve a more flexible attack. "Parasite" as a novel attack method effectively bypasses existing detection frameworks to execute backdoor attacks. In our experiments, "Parasite" achieved a 0 percent backdoor detection rate against the mainstream defense frameworks. In addition, in the ablation study, we discuss the influence of different hiding coefficients on the attack results. You can find our code at https://anonymous.4open.science/r/Parasite-1715/.

arXiv:2505.15781공식 코드 · Papers with Code

dKV-Cache: The Cache for Diffusion Language Models

Diffusion Language Models (DLMs) have been seen as a promising competitor for autoregressive language models. However, diffusion language models have long been constrained by slow inference. A core challenge is that their non-autoregressive architecture and bidirectional attention preclude the key-value cache that accelerates decoding. We address this bottleneck by proposing a KV-cache-like mechanism, delayed KV-Cache, for the denoising process of DLMs. Our approach is motivated by the observation that different tokens have distinct representation dynamics throughout the diffusion process. Accordingly, we propose a delayed and conditioned caching strategy for key and value states. We design two complementary variants to cache key and value step-by-step: (1) dKV-Cache-Decode, which provides almost lossless acceleration, and even improves performance on long sequences, suggesting that existing DLMs may under-utilise contextual information during inference. (2) dKV-Cache-Greedy, which has aggressive caching with reduced lifespan, achieving higher speed-ups with quadratic time complexity at the cost of some performance degradation. dKV-Cache, in final, achieves from 2-10x speedup in inference, largely narrowing the gap between ARs and DLMs. We evaluate our dKV-Cache on several benchmarks, delivering acceleration across general language understanding, mathematical, and code-generation benchmarks. Experiments demonstrate that cache can also be used in DLMs, even in a training-free manner from current DLMs.

arXiv:2505.20321공식 코드 · Papers with Code

BiomedSQL: Text-to-SQL for Scientific Reasoning on Biomedical Knowledge Bases

Biomedical researchers increasingly rely on large-scale structured databases for complex analytical tasks. However, current text-to-SQL systems often struggle to map qualitative scientific questions into executable SQL, particularly when implicit domain reasoning is required. We introduce BiomedSQL, the first benchmark explicitly designed to evaluate scientific reasoning in text-to-SQL generation over a real-world biomedical knowledge base. BiomedSQL comprises 68,000 question/SQL query/answer triples grounded in a harmonized BigQuery knowledge base that integrates gene-disease associations, causal inference from omics data, and drug approval records. Each question requires models to infer domain-specific criteria, such as genome-wide significance thresholds, effect directionality, or trial phase filtering, rather than rely on syntactic translation alone. We evaluate a range of open- and closed-source LLMs across prompting strategies and interaction paradigms. Our results reveal a substantial performance gap: GPT-o3-mini achieves 59.0% execution accuracy, while our custom multi-step agent, BMSQL, reaches 62.6%, both well below the expert baseline of 90.0%. BiomedSQL provides a new foundation for advancing text-to-SQL systems capable of supporting scientific discovery through robust reasoning over structured biomedical knowledge bases. Our dataset is publicly available at https://huggingface.co/datasets/NIH-CARD/BiomedSQL, and our code is open-source at https://github.com/NIH-CARD/biomedsql.

arXiv:2506.03234v2cs.LG

The Alignment Illusion: How BadReward Corrupts Text-to-Image Generation

Modern text-to-image generative models increasingly rely on Reinforcement Learning from Human Feedback (RLHF) to align generated content with human preferences and safety standards. This post-training stage is critical for optimizing models toward high-fidelity and value-aligned behaviors using human preference data. However, while previous works have extensively investigated poisoning attacks against the supervised fine-tuning (SFT) stage of text-to-image training, the security vulnerabilities introduced by the RLHF process remain unexplored. In this paper, we present the first exploration into the feasibility of poisoning the RLHF process as a novel attack surface for compromising text-to-image models. We introduce BadReward, an indirect clean-label poisoning attack that manipulates reward signals to embed malicious behaviors. Our approach leverages a feature collision mechanism to produce poison preference image pairs that are visually benign, but aligned with a malicious target concept in the reward feature space. Our experiments demonstrate that this manipulation effectively corrupts the reward signal. Online RLHF further amplifies this corruption through its reward-model feedback loop, whereas offline preference optimization learns directly from the poisoned preferences. Finally, we demonstrate the real-world effectiveness of BadReward on modern text-to-image architectures, showing high attack success rates and visual stealthiness across different RLHF algorithms. Our code is available at https://github.com/ZJUICSR/BadReward.

arXiv:2605.31494v2cs.CL

Consolidating Rewarded Perturbations for LLM Post-Training

Post-training of language models is commonly framed as a sample-score-update loop implemented by gradient descent. A recent line of work, exemplified by RandOpt, relocates this loop to weight space, sampling Gaussian perturbations around a pretrained model and ensembling the top-K rewarded specialists at inference. While competitive with PPO and GRPO under matched training compute, this prediction-level ensemble incurs K forward passes per test example and does not extend cleanly to free-form generation. We ask whether the rewarded population can instead be folded into a single deployable model, replacing the inference-time ensemble with one consolidated update. A split-half analysis over 25 model-task pairs reveals reproducible low-rank structure in every case. We turn this geometry into CoRP (Consolidating Rewarded Perturbations), a gradient-free operator that combines reward-weighted aggregation, compatibility-aware reweighting, and a held-out validation gate, with no gradient flowing through the language model. Across five language models from 0.5B to 8B and five tasks covering math, code, and creative writing, CoRP improves the base model by 8.1 points on average. Using one tenth of RandOpt's perturbation budget, CoRP exceeds single-inference RandOpt by 6.5 points and recovers more than half of the gain of the 50-pass majority-vote ensemble, at one forward pass per test example.

arXiv:2606.24267v3cs.CL

Pigeonholing: how bad prompts hurt models, causing collapse and mistakes

While in-context learning is generally shown to be effective in Large Language Models (LLMs), bad contexts can cause performance degradation and mode collapse, a phenomenon we call "pigeonholing." **Unintentionally bad** contexts can happen without malicious jailbreaking intents: For example, a user asks the model to justify an incorrect math theorem or fails to correct the model's buggy code. Specifically, we investigate ``pigeonholing" in two scenarios: (1) when the user suggests a solution, and (2) when the conversation context includes the assistant's previous (incorrect) responses. Our experiments across 10 verifiable and open-ended tasks with 10 different models show that pigeonholing manifests in several ways: (1) repeating the incorrect answers from context (leading to 38-40% performance drop), (2) converging on a narrow set of answers in coding and text generation without exploring alternatives, and (3) flipping stance on controversial topics to align with the user or the assistant's previous claims. We find that pigeonholing worsens almost monotonically with the number of conversation turns (performance drops by additional 14+% as repeated mistakes increase from 1 to 5), and pigeonholing-induced mode collapse can happen even when the provided example is correct. As a step toward mitigation, we propose RLVR with synthetic errors which improves models by 43-60% under bad contexts compared to vanilla RLVR baselines.

arXiv:2607.00798v1cs.CV

ClinRAG-GRAPH: 유방암 pCR 예측을 위한 도메인 적대적 학습 기반 임상 사전 검색-증강 그래프 모델

ClinRAG-GRAPH: Clinical-prior Retrieval-Augmented Graph Model with Domain Adversarial Learning for Breast pCR Prediction

DCE-MRI, 구조화된 임상 변수, 생검 기반 병리 바이오마커를 통합한 임상 사전 그래프(clinical-prior graph)와 관계 인식 그래프 합성곱 네트워크(R-GCN), 도메인 적대적 학습(domain-adversarial learning), LLM 기반 서브그래프 RAG 모듈을 결합하여 유방암 신보조 화학요법(NAC)의 치료 전 병리학적 완전 관해(pCR) 예측 프레임워크 ClinRAG-GRAPH를 제안한다. 다기관 대규모 코호트 실험에서 내부 테스트셋 AUC 0.815, 두 외부 테스트셋에서 각각 0.774/0.712를 달성하며 기관 간 강건한 예측 성능을 입증하였다.

arXiv:2607.01248v1cs.CY

A Practice Auditing Framework for Large Language Model Use: Collective Empiricism, Pseudo-Rational Cognition, and Governance of AI-Generated Content

대규모 언어 모델(LLM) 사용에서 발생하는 인지적 위험을 다루기 위해, 이 연구는 LLM이 대규모 인간 경험을 압축·재구성하는 방식을 설명하는 집단적 경험주의(collective empiricism)와 사용자가 AI 생성 구조화 표현을 자신의 이성적 이해로 오인하는 유사 이성적 인지(pseudo-rational cognition) 개념을 제안한다. 이를 기반으로 요구사항 정의, 증거 출처 감사, 실천적 검증, 역질문, 로깅 및 롤백 메커니즘을 포함한 실천 감사(practice auditing) 프레임워크를 제시하며, AI 생성 콘텐츠가 미래 컨텍스트·메모리 시스템·검색 공간에 유입될 때 발생하는 템플릿 루프, 메모리 오염 등의 거버넌스 위험도 분석한다.

arXiv:2608.12511v1cs.CR

SoK: 소프트웨어 시스템에서 개인정보 문서의 생성부터 소비까지

SoK: From Generation to Consumption of Privacy Documents in Software Systems

개인정보 문서(예: 개인정보 처리방침)는 디지털 서비스가 데이터 처리 관행을 공개하고 사용자 동의를 구하는 핵심 수단이다. 지난 수십 년간 개인정보 문서에 대한 연구는 전통적인 개인정보 처리방침뿐만 아니라 짧은 고지(예: 개인정보 라벨) 및 인터페이스 수준의 투명성 메커니즘까지 포괄하며 크게 확대되었다. 이 연구 분야가 계속 성장함에 따라, 개인정보 문서가 수명주기 전반에 걸쳐 어떻게 생성·분석·평가·유지되는지에 대한 일관된 관점을 얻기가 점점 더 어려워졌다. 본 SoK는 소프트웨어 공학 관점에서 개인정보 문서에 대한 통합적이고 수명주기 중심의 관점을 제시한다. 우리는 2010년부터 2025년까지 발표된 290편의 논문을 체계적으로 검토·분석하고, 개인정보 문서가 (1) 어떻게 정의되고 범위가 설정되는지, (2) 어떻게 생성되는지, (3) 어떻게 분석·추출되는지, (4) 불일치 및 비준수 여부가 어떻게 점검되는지, (5) 사용성 측면에서 어떻게 평가·개선되는지를 살펴 다섯 가지 연구 질문을 중심으로 이를 구조화했다. 분석 결과를 바탕으로 15가지 핵심 연구 동향과 21가지 미해결 연구 기회를 도출했다. 또한 (i) AI 중심 플랫폼에서의 새로운 과제, (ii) 다양하고 최신화된 데이터 기반의 필요성, (iii) LLM 기반 통합 정책-코드 분석, (iv) 최종 사용자와 개발자를 위한 이중 사용성을 강조하는 네 가지 광범위한 연구 방향을 정리했다. 본 SoK가 개인정보 처리방침 및 개인정보 문서에 대한 향후 연구의 공통 기반이 되기를 기대한다.

arXiv:2608.24855v1cs.CV

LeFlow: 월드 모델을 위한 생성적 잠재 플로우 계획

LeFlow: Generative Latent Flow Planning for World Models

잠재 월드 모델(latent world model)은 이미지 픽셀을 잠재 임베딩으로 변환하는 본질적으로 강력한 인코더이지만, 기존 월드 모델은 행동 계획(action planning)에 여전히 온라인 궤적 최적화(trajectory optimization)에 의존한다. 모든 상태-목표(state-goal) 쌍마다 반복 옵티마이저를 처음부터 실행해 최적 행동 시퀀스를 탐색하며, 월드 모델을 블랙박스 시뮬레이터로 취급한다. 이 방식은 재계획(replanning) 단계마다 반복 최적화 비용을 전부 새로 지불하고, 쿼리 간 계획 경험을 재사용하지 않는다. 본 연구에서는 잠재 월드 모델이 학습된 뒤 계획(planning) 자체를 상각(amortize)할 수 있는지 묻는다. LeFlow를 제시하는데, 이는 월드 모델로부터 잠재 역학(latent dynamics) 공간에서 직접 작동하는 재사용 가능한 잠재 궤적 사전(latent trajectory prior)을 학습한다. LeFlow는 계획을 조건부 잠재 궤적 생성으로 재구성한다. rectified-flow 모델이 현재와 목표 임베딩 사이의 미래 잠재 경로를 상상하고, 역역학(inverse dynamics) 디코더가 잠재 전이를 행동 청크(action chunk)로 바꾸며, 고정(frozen) 월드 모델이 각 후보를 자회귀(autoregressive) 롤아웃으로 검증한다. 네 가지 주요 목표 조건(goal-conditioned) 픽셀 제어 벤치마크에서 LeFlow는 반복적 행동 공간 최적화를 상각된(amortized) 잠재 계획과 고정 예산 롤아웃 선택으로 대체하여, 계획 시간을 한 자릿수(order-of-magnitude) 줄이면서도 일관된 성공률 향상을 달성한다. 결과는 잠재 월드 모델이 예측뿐 아니라 재사용 가능한 계획 사전(planning prior)도 지원해야 함을 시사한다. 코드는 https://github.com/hsiangwei0903/LeFlow 에서 제공한다.

arXiv:2604.14572v4cs.IR

Corpus2Skill: 질의응답 및 RAG를 위해 엔터프라이즈 지식을 탐색 가능한 에이전트 스킬로 증류하기

Corpus2Skill: Distilling Enterprise Knowledge into Navigable Agent Skills for QA and RAG

검색-증강 생성(RAG)은 대규모 언어 모델(LLM)의 응답을 외부 근거에 그라운딩(grounding)하지만, 코퍼스가 어떻게 구성되어 있는지 또는 모델이 아직 보지 못한 내용이 무엇인지에 대한 조망 없이 모델을 검색 결과의 수동적인 소비자로 취급합니다. 본 논문에서는 엔터프라이즈 지식 베이스와 같이 범위가 한정되고 구조적 일관성을 갖춘 코퍼스를 위한 시스템 수준의 검색 아키텍처인 Corpus2Skill을 제안합니다. 오프라인 컴파일러가 코퍼스를 계층적 스킬 디렉터리로 증류하고, 서빙 시점에는 LLM 에이전트가 이를 탐색하여 조감도(bird's-eye view) 수준에서 점진적으로 더 세부적인 요약을 거쳐 문서에 이르기까지 드릴다운하며 특정 브랜치에서 성과가 없을 경우 백트래킹을 수행합니다. 엔터프라이즈 고객 지원 벤치마크에서 Corpus2Skill은 적당한 비용 트레이드오프 하에 싱글샷 밀집(dense), 하이브리드, 계층적 검색 및 에이전틱(agentic) RAG 베이스라인 대비 답변 품질과 그라운딩을 모두 향상시키며, 인코더를 일치시킨 통제 조건과 대응 표본 유의성 검정에서도 우위가 유지됩니다. 11개 데이터셋에 대한 연구는 코퍼스 탐색이 기존 검색을 완전히 대체하는 만능 해결책은 아님을 보여줍니다. 즉, 5개 데이터셋에서는 유의미하게 우세하고 3개에서는 동등하며 3개에서는 뒤처집니다. 이는 복원 가능한 주제별 분류 체계를 가진 단일 도메인 코퍼스에는 도움이 되지만, 상위 수준 클러스터링을 무력화하는 오픈 도메인 단문형 사실(factoid) 풀이나 동질적 표 형태의 코퍼스에서는 여전히 평탄형(flat) 검색이 선호됩니다. 우리는 이러한 적용 범위의 구분을 지식 기반 그라운딩 시스템을 위한 설계 가이드라인으로 규정합니다. 코드는 https://github.com/dukesun99/Corpus2Skill 에서 이용할 수 있습니다.

arXiv:2609.23880v1cs.CL

Q-TIE: 시간 정보 검색을 위한 경량·범용 재순위화 프레임워크

Q-TIE: A Lightweight and Generalizable Re-ranking Framework for Temporal Information Retrieval

검색-증강 생성(RAG)의 부상에 따라 시간 정보 검색(TIR)의 중요성이 점점 커지고 있다. 시간적으로 맞지 않는 근거는 매우 오해를 불러일으킬 수 있으므로, TIR은 쿼리에 대해 의미적으로 및 시간적으로 관련 있는 문서를 검색하는 것을 목표로 한다. 두 가지 TIR 패러다임—시간 검색기(temporal retriever)와 시간 재순위화기(temporal re-ranker)—가 등장했으며, 시간적 관련성을 모델링하는 방식에서 차이가 난다. 이 패러다igm들은 상호 보완적인 강점을 제공하지만, 우리의 분석에 따르면 각각만으로는 견고한 TIR에 미치지 못한다. 시간 검색기는 학습된 표현을 통해 유연한 쿼리 이해를 제공하지만, 종종 시간 제약을 명시적으로 반영하지 못하고, 시간 재순위화기는 그러한 제약을 더 명시적으로 강제할 수 있지만, 종종 사전 정의된 재순위화 규칙에 의존한다. 이를 해결하기 위해, 학습 기반 시간 의도 추출(TIE)에 기반한 재순위화 프레임워크인 Q-TIE를 제안한다. 각 쿼리의 시간 제약을 통합된 구간 표현(즉, $\langle t_{start}, t_{end} \rangle$)으로 매핑하는 TIE 모델을 도입함으로써, Q-TIE는 모델 기반 학습을 통해 사전 정의된 규칙을 넘어 일반화하면서 시간 제약을 별도의 신호로 명시적으로 모델링한다—각 패러다igm이 일반적으로 상충(trade-off)하던 바를 함께 달성한다. 실험 결과 Q-TIE는 시간적 쿼리 유형 전반에 걸쳐 더 강한 일반화 가능성과 함께 기존 TIR 방법을 일관되게 능가하며, 시간 인식 RAG 파이프라인을 위한 경량이면서도 효과적인 부가 모듈을 제공함을 보여준다. Code: https://github.com/ssoy0701/Q-TIE.