PyTorchKR Papers

“Code Generation”

총 10,163편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2312.15224공식 코드 · Papers with Code

LLM-Powered Hierarchical Language Agent for Real-time Human-AI Coordination

AI agents powered by Large Language Models (LLMs) have made significant advances, enabling them to assist humans in diverse complex tasks and leading to a revolution in human-AI coordination. LLM-powered agents typically require invoking LLM APIs and employing artificially designed complex prompts, which results in high inference latency. While this paradigm works well in scenarios with minimal interactive demands, such as code generation, it is unsuitable for highly interactive and real-time applications, such as gaming. Traditional gaming AI often employs small models or reactive policies, enabling fast inference but offering limited task completion and interaction abilities. In this work, we consider Overcooked as our testbed where players could communicate with natural language and cooperate to serve orders. We propose a Hierarchical Language Agent (HLA) for human-AI coordination that provides both strong reasoning abilities while keeping real-time execution. In particular, HLA adopts a hierarchical framework and comprises three modules: a proficient LLM, referred to as Slow Mind, for intention reasoning and language interaction, a lightweight LLM, referred to as Fast Mind, for generating macro actions, and a reactive policy, referred to as Executor, for transforming macro actions into atomic actions. Human studies show that HLA outperforms other baseline agents, including slow-mind-only agents and fast-mind-only agents, with stronger cooperation abilities, faster responses, and more consistent language communications.

arXiv:2404.08495공식 코드 · Papers with Code

Dataset Reset Policy Optimization for RLHF

Reinforcement Learning (RL) from Human Preference-based feedback is a popular paradigm for fine-tuning generative models, which has produced impressive models such as GPT-4 and Claude3 Opus. This framework often consists of two steps: learning a reward model from an offline preference dataset followed by running online RL to optimize the learned reward model. In this work, leveraging the idea of reset, we propose a new RLHF algorithm with provable guarantees. Motivated by the fact that offline preference dataset provides informative states (i.e., data that is preferred by the labelers), our new algorithm, Dataset Reset Policy Optimization (DR-PO), integrates the existing offline preference dataset into the online policy training procedure via dataset reset: it directly resets the policy optimizer to the states in the offline dataset, instead of always starting from the initial state distribution. In theory, we show that DR-PO learns to perform at least as good as any policy that is covered by the offline dataset under general function approximation with finite sample complexity. In experiments, we demonstrate that on both the TL;DR summarization and the Anthropic Helpful Harmful (HH) dataset, the generation from DR-PO is better than that from Proximal Policy Optimization (PPO) and Direction Preference Optimization (DPO), under the metric of GPT4 win-rate. Code for this work can be found at https://github.com/Cornell-RL/drpo.

arXiv:2404.11317공식 코드 · Papers with Code

Improving Composed Image Retrieval via Contrastive Learning with Scaling Positives and Negatives

The Composed Image Retrieval (CIR) task aims to retrieve target images using a composed query consisting of a reference image and a modified text. Advanced methods often utilize contrastive learning as the optimization objective, which benefits from adequate positive and negative examples. However, the triplet for CIR incurs high manual annotation costs, resulting in limited positive examples. Furthermore, existing methods commonly use in-batch negative sampling, which reduces the negative number available for the model. To address the problem of lack of positives, we propose a data generation method by leveraging a multi-modal large language model to construct triplets for CIR. To introduce more negatives during fine-tuning, we design a two-stage fine-tuning framework for CIR, whose second stage introduces plenty of static representations of negatives to optimize the representation space rapidly. The above two improvements can be effectively stacked and designed to be plug-and-play, easily applied to existing CIR models without changing their original architectures. Extensive experiments and ablation analysis demonstrate that our method effectively scales positives and negatives and achieves state-of-the-art results on both FashionIQ and CIRR datasets. In addition, our method also performs well in zero-shot composed image retrieval, providing a new CIR solution for the low-resources scenario. Our code and data are released at https://github.com/BUAADreamer/SPN4CIR.

arXiv:2406.13977

Similarity-aware Syncretic Latent Diffusion Model for Medical Image Translation with Representation Learning

Non-contrast CT (NCCT) imaging may reduce image contrast and anatomical visibility, potentially increasing diagnostic uncertainty. In contrast, contrast-enhanced CT (CECT) facilitates the observation of regions of interest (ROI). Leading generative models, especially the conditional diffusion model, demonstrate remarkable capabilities in medical image modality transformation. Typical conditional diffusion models commonly generate images with guidance of segmentation labels for medical modal transformation. Limited access to authentic guidance and its low cardinality can pose challenges to the practical clinical application of conditional diffusion models. To achieve an equilibrium of generative quality and clinical practices, we propose a novel Syncretic generative model based on the latent diffusion model for medical image translation (S$^2$LDM), which can realize high-fidelity reconstruction without demand of additional condition during inference. S$^2$LDM enhances the similarity in distinct modal images via syncretic encoding and diffusing, promoting amalgamated information in the latent space and generating medical images with more details in contrast-enhanced regions. However, syncretic latent spaces in the frequency domain tend to favor lower frequencies, commonly locate in identical anatomic structures. Thus, S$^2$LDM applies adaptive similarity loss and dynamic similarity to guide the generation and supplements the shortfall in high-frequency details throughout the training process. Quantitative experiments confirm the effectiveness of our approach in medical image translation. Our code will release lately.

arXiv:2408.10537공식 코드 · Papers with Code

Subspace Prototype Guidance for Mitigating Class Imbalance in Point Cloud Semantic Segmentation

Point cloud semantic segmentation can significantly enhance the perception of an intelligent agent. Nevertheless, the discriminative capability of the segmentation network is influenced by the quantity of samples available for different categories. To mitigate the cognitive bias induced by class imbalance, this paper introduces a novel method, namely subspace prototype guidance (\textbf{SPG}), to guide the training of segmentation network. Specifically, the point cloud is initially separated into independent point sets by category to provide initial conditions for the generation of feature subspaces. The auxiliary branch which consists of an encoder and a projection head maps these point sets into separate feature subspaces. Subsequently, the feature prototypes which are extracted from the current separate subspaces and then combined with prototypes of historical subspaces guide the feature space of main branch to enhance the discriminability of features of minority categories. The prototypes derived from the feature space of main branch are also employed to guide the training of the auxiliary branch, forming a supervisory loop to maintain consistent convergence of the entire network. The experiments conducted on the large public benchmarks (i.e. S3DIS, ScanNet v2, ScanNet200, Toronto-3D) and collected real-world data illustrate that the proposed method significantly improves the segmentation performance and surpasses the state-of-the-art method. The code is available at \url{https://github.com/Javion11/PointLiBR.git}.

arXiv:2412.06141공식 코드 · Papers with Code

MMedPO: Aligning Medical Vision-Language Models with Clinical-Aware Multimodal Preference Optimization

The advancement of Large Vision-Language Models (LVLMs) has propelled their application in the medical field. However, Medical LVLMs (Med-LVLMs) encounter factuality challenges due to modality misalignment, where the models prioritize textual knowledge over visual input, leading to hallucinations that contradict information in medical images. Previous attempts to enhance modality alignment in Med-LVLMs through preference optimization have inadequately mitigated clinical relevance in preference data, making these samples easily distinguishable and reducing alignment effectiveness. To address this challenge, we propose MMedPO, a novel multimodal medical preference optimization approach that considers the clinical relevance of preference samples to enhance Med-LVLM alignment. MMedPO curates multimodal preference data by introducing two types of dispreference: (1) plausible hallucinations injected through target Med-LVLMs or GPT-4o to produce medically inaccurate responses, and (2) lesion region neglect achieved through local lesion-noising, disrupting visual understanding of critical areas. We then calculate clinical relevance for each sample based on scores from multiple Med-LLMs and visual tools, and integrate these scores into the preference optimization process as weights, enabling effective alignment. Our experiments demonstrate that MMedPO significantly enhances factual accuracy in Med-LVLMs, achieving substantial improvements over existing preference optimization methods by averaging 14.2% and 51.7% across the Med-VQA and report generation tasks. Our code are available in https://github.com/aiming-lab/MMedPO.

arXiv:2412.11599공식 코드 · Papers with Code

3D$^2$-Actor: Learning Pose-Conditioned 3D-Aware Denoiser for Realistic Gaussian Avatar Modeling

Advancements in neural implicit representations and differentiable rendering have markedly improved the ability to learn animatable 3D avatars from sparse multi-view RGB videos. However, current methods that map observation space to canonical space often face challenges in capturing pose-dependent details and generalizing to novel poses. While diffusion models have demonstrated remarkable zero-shot capabilities in 2D image generation, their potential for creating animatable 3D avatars from 2D inputs remains underexplored. In this work, we introduce 3D$^2$-Actor, a novel approach featuring a pose-conditioned 3D-aware human modeling pipeline that integrates iterative 2D denoising and 3D rectifying steps. The 2D denoiser, guided by pose cues, generates detailed multi-view images that provide the rich feature set necessary for high-fidelity 3D reconstruction and pose rendering. Complementing this, our Gaussian-based 3D rectifier renders images with enhanced 3D consistency through a two-stage projection strategy and a novel local coordinate representation. Additionally, we propose an innovative sampling strategy to ensure smooth temporal continuity across frames in video synthesis. Our method effectively addresses the limitations of traditional numerical solutions in handling ill-posed mappings, producing realistic and animatable 3D human avatars. Experimental results demonstrate that 3D$^2$-Actor excels in high-fidelity avatar modeling and robustly generalizes to novel poses. Code is available at: https://github.com/silence-tang/GaussianActor.

arXiv:2502.01814공식 코드 · Papers with Code

PolyhedronNet: Representation Learning for Polyhedra with Surface-attributed Graph

Ubiquitous geometric objects can be precisely and efficiently represented as polyhedra. The transformation of a polyhedron into a vector, known as polyhedra representation learning, is crucial for manipulating these shapes with mathematical and statistical tools for tasks like classification, clustering, and generation. Recent years have witnessed significant strides in this domain, yet most efforts focus on the vertex sequence of a polyhedron, neglecting the complex surface modeling crucial in real-world polyhedral objects. This study proposes \textbf{PolyhedronNet}, a general framework tailored for learning representations of 3D polyhedral objects. We propose the concept of the surface-attributed graph to seamlessly model the vertices, edges, faces, and their geometric interrelationships within a polyhedron. To effectively learn the representation of the entire surface-attributed graph, we first propose to break it down into local rigid representations to effectively learn each local region's relative positions against the remaining regions without geometric information loss. Subsequently, we propose PolyhedronGNN to hierarchically aggregate the local rigid representation via intra-face and inter-face geometric message passing modules, to obtain a global representation that minimizes information loss while maintaining rotation and translation invariance. Our experimental evaluations on four distinct datasets, encompassing both classification and retrieval tasks, substantiate PolyhedronNet's efficacy in capturing comprehensive and informative representations of 3D polyhedral objects. Code and data are available at {https://github.com/dyu62/3D_polyhedron}.

arXiv:2502.19610공식 코드 · Papers with Code

Program Synthesis Dialog Agents for Interactive Decision-Making

Many real-world eligibility problems, ranging from medical diagnosis to tax planning, can be mapped to decision problems expressed in natural language, wherein a model must make a binary choice based on user features. Large-scale domains such as legal codes or frequently updated funding opportunities render human annotation (e.g., web forms or decision trees) impractical, highlighting the need for agents that can automatically assist in decision-making. Since relevant information is often only known to the user, it is crucial that these agents ask the right questions. As agents determine when to terminate a conversation, they face a trade-off between accuracy and the number of questions asked, a key metric for both user experience and cost. To evaluate this task, we propose BeNYfits, a new benchmark for determining user eligibility for multiple overlapping social benefits opportunities through interactive decision-making. Our experiments show that current language models struggle with frequent hallucinations, with GPT-4o scoring only 35.7 F1 using a ReAct-style chain-of-thought. To address this, we introduce ProADA, a novel approach that leverages program synthesis to assist in decision-making by mapping dialog planning to a code generation problem and using gaps in structured data to determine the best next action. Our agent, ProADA, improves the F1 score to 55.6 while maintaining nearly the same number of dialog turns.

arXiv:2505.11907

Are Multimodal Large Language Models Ready for Omnidirectional Spatial Reasoning?

The 180x360 omnidirectional field of view captured by 360-degree cameras enables their use in a wide range of applications such as embodied AI and virtual reality. Although recent advances in multimodal large language models (MLLMs) have shown promise in visual-spatial reasoning, most studies focus on standard pinhole-view images, leaving omnidirectional perception largely unexplored. In this paper, we ask: Are MLLMs ready for omnidirectional spatial reasoning? To investigate this, we introduce OSR-Bench, the first benchmark specifically designed for this setting. OSR-Bench includes over 153,000 diverse question-answer pairs grounded in high-fidelity panoramic indoor scene maps. It covers key reasoning types including object counting, relative distance, and direction. We also propose a negative sampling strategy that inserts non-existent objects into prompts to evaluate hallucination and grounding robustness. For fine-grained analysis, we design a two-stage evaluation framework assessing both cognitive map generation and QA accuracy using rotation-invariant matching and a combination of rule-based and LLM-based metrics. We evaluate eight state-of-the-art MLLMs, including GPT-4o, Gemini 1.5 Pro, and leading open-source models under zero-shot settings. Results show that current models struggle with spatial reasoning in panoramic contexts, highlighting the need for more perceptually grounded MLLMs. OSR-Bench and code will be released at: https://huggingface.co/datasets/UUUserna/OSR-Bench

arXiv:2505.20003공식 코드 · Papers with Code

TabPFN: One Model to Rule Them All?

Hollmann et al. (Nature 637 (2025) 319-326) recently introduced TabPFN, a transformer-based deep learning model for regression and classification on tabular data, which they claim "outperforms all previous methods on datasets with up to 10,000 samples by a wide margin, using substantially less training time." Furthermore, they have called TabPFN a "foundation model" for tabular data, as it can support "data generation, density estimation, learning reusable embeddings and fine-tuning". If these statements are well-supported, TabPFN may have the potential to supersede existing modeling approaches on a wide range of statistical tasks, mirroring a similar revolution in other areas of artificial intelligence that began with the advent of large language models. In this paper, we provide a tailored explanation of how TabPFN works for a statistics audience, by emphasizing its interpretation as approximate Bayesian inference. We also provide more evidence of TabPFN's "foundation model" capabilities: We show that an out-of-the-box application of TabPFN vastly outperforms specialized state-of-the-art methods for semi-supervised parameter estimation, prediction under covariate shift, and heterogeneous treatment effect estimation. We further show that TabPFN can outperform LASSO at sparse regression and can break a robustness-efficiency trade-off in classification. All experiments can be reproduced using the code provided at https://github.com/qinglong-tian/tabpfn_study (https://github.com/qinglong-tian/tabpfn_study).

arXiv:2507.00081

State and Memory is All You Need for Robust and Reliable AI Agents

Large language models (LLMs) have enabled powerful advances in natural language understanding and generation. Yet their application to complex, real-world scientific workflows remain limited by challenges in memory, planning, and tool integration. Here, we introduce SciBORG (Scientific Bespoke Artificial Intelligence Agents Optimized for Research Goals), a modular agentic framework that allows LLM-based agents to autonomously plan, reason, and achieve robust and reliable domain-specific task execution. Agents are constructed dynamically from source code documentation and augmented with finite-state automata (FSA) memory, enabling persistent state tracking and context-aware decision-making. This approach eliminates the need for manual prompt engineering and allows for robust, scalable deployment across diverse applications via maintaining context across extended workflows and to recover from tool or execution failures. We validate SciBORG through integration with both physical and virtual hardware, such as microwave synthesizers for executing user-specified reactions, with context-aware decision making and demonstrate its use in autonomous multi-step bioassay retrieval from the PubChem database utilizing multi-step planning, reasoning, agent-to-agent communication and coordination for execution of exploratory tasks. Systematic benchmarking shows that SciBORG agents achieve reliable execution, adaptive planning, and interpretable state transitions. Our results show that memory and state awareness are critical enablers of agentic planning and reliability, offering a generalizable foundation for deploying AI agents in complex environments.

arXiv:2611.10005v1cs.CV

EarthJev: Efficient Visual Evidence Acquisition and Decision for UHR Remote Sensing Understanding

Ultra-high-resolution (UHR) remote sensing visual question answering (VQA) requires locating fine-grained evidence in large scenes under limited visual input budgets. Active observation addresses this challenge through selective exploration, but repeated generative judgments can make region search costly. Its decisions also concern distinct prediction targets: spatial pointer validity, evidence presence, and reasoning correctness. We propose EarthJev, a framework that separates region scoring from answer generation and uses target-specific decision signals while keeping the answering model frozen. A spatial pointer verifier uses visual and geometric cues to guide exploration or bypass search. An evidence-aware decision mechanism scores compact candidate views in batches using a lightweight head on frozen representations, without generating textual region judgments. Format-specific confidence calibration then guides selection between candidate answers generated from memory and overview presentations of the acquired evidence. Across five benchmarks, EarthJev achieves a mean accuracy of 53.06% with Qwen3.5-9B, outperforming ZoomSearch with the same answering model by 2.33 to 14.73 percentage points per benchmark while reducing mean latency from 11.30 - 1.67 seconds per question. With four additional frozen answering models, model-specific fitting of the decision modules and calibration mappings yields mean accuracy gains of 3.41 - 6.53 percentage points over the corresponding single-pass baselines. Code is available at https://github.com/Da1daidaidai/EarthJev

arXiv:2606.12828v1cs.AI

인공지능 연구의 주제적 상전이: 대규모 증거와 신흥 주제에 대한 조기 경보 시그너처

Topical Phase Transitions in Artificial Intelligence Research: Large-Scale Evidence and an Early-Warning Signature for Emerging Topics

인공지능 연구 주제는 점진적으로 성장하는가, 아니면 급격하고 탐지 가능한 도약으로 전진하는가? 2017년부터 2025년까지 5개 주요 AI 학회(ACL, CVPR, ICLR, ICML, NeurIPS)의 수락된 메인 트랙 논문 80,814편을 분석한 결과, 주요 AI 주제들이 주제적 상전이(topical phase transitions)를 통해 발전함을 보여준다. 수년간 주변적 상태를 유지하다가 1~3년 안에 여러 학회에서 급증하는 패턴이다. 대규모 언어 모델(LLM)은 2025년까지 지배적인 교차 학회(cross-venue) 주제가 되었고, 확산 모델(diffusion models)은 유사한 급격성으로 부상했으며, 언어 모델 방법은 비전-언어(vision-language) 모델을 통해 컴퓨터 비전(CV)으로 확장되었다. 반면 강화학습(RL)은 완만하게 누적 성장하여, 진정한 상전이를 일반적 성장과 구별한다. 이러한 구조가 우리의 핵심 기여이다: AI 연구가 재편되는 방식에 대한 대규모 교차 학회 특성화이다. 그다음 우리는 상전이가 정점에 이르기 전에 탐지 가능한 흔적을 남기는지 묻는다. 조기 경보 시그너처를 정의하고, 2017-2021 데이터에 고정된 네 가지 출판 역학(publication-dynamics) 기준을 사용하여 2023-2025 상전이에 대해 표본 외(out of sample)로 평가하여, 13.5% 기저율(base rate) 대비 정밀도(precision) 27%, 재현율(recall) 63%를 얻었다. 2025년 데이터에 적용하면, 이 시그너처는 추론(reasoning) 및 테스트 시점 계산(test-time compute), 에이전트(agent) AI, 멀티모달 LLM, 검색-증강 생성(RAG), 월드 모델(world models)을 2026-2028년 동안 모니터링할 주제로 표시한다. 소스 코드는 https://github.com/KurbanIntelligenceLab/ai-phase-transitions GitHub에서 공개적으로 이용 가능하다.

arXiv:2606.24679v1cs.LG

FlowPipe: LLM-Enhanced Conditional Generative Flow Networks for Data Preparation Pipeline Construction

데이터 준비 파이프라인 자동 구성의 조합적 탐색과 비용 높은 end-to-end 평가 한계를 FlowPipe가 조건부 생성 흐름 네트워크(C-GFlowNets)의 Trajectory Balance, LLM 기반 FiLM 의미 조절, 실패 인식 흐름 목표로 해결하도록 제안한다. 74개 실데이터셋 두 벤치마크에서 기존 SOTA 대비 평균 정확도 11.96% 향상과 학습 수렴 12.5배 가속을 달성했다.

arXiv:2607.01232v2cs.LG

Is One Layer Enough? Training A Single Transformer Layer Can Match Full-Parameter RL Training

대규모 언어 모델(LLM)의 강화학습(RL) 사후학습에서 성능 향상이 트랜스포머 레이어 전체에 균등하게 분포한다는 기존 가정에 도전하며, 단일 레이어만 학습해도 전체 파라미터 RL 학습 성능의 대부분을 회복하거나 능가할 수 있음을 체계적 실험으로 규명했다. Qwen3, Qwen2.5 등 7개 모델과 GRPO, GiGPO, Dr. GRPO 세 가지 알고리즘에 걸쳐 고기여 레이어가 일관되게 트랜스포머 스택 중간부(전체 깊이의 40~60%)에 집중된다는 구조적 패턴을 발견하고, 이를 활용한 레이어 인식 학습 전략이 표준 전체 파라미터 RL을 지속적으로 능가함을 보였다.

arXiv:2603.16307v3cs.AI

NeSy-Route: 제약 조건 하 원격 탐사 경로 계획을 위한 신경-기호 벤치마크

NeSy-Route: A Neuro-Symbolic Benchmark for Constrained Route Planning in Remote Sensing

원격 탐사는 재난 구호, 생태 현장 조사 등 중요한 응용의 기반이 되며, 이러한 분야에서는 시스템이 복잡한 장면과 제약 조건을 이해하고 신뢰할 수 있는 결정을 내려야 한다. 현재 원격 탐사 벤치마크는 주로 멀티모달 대규모 언어 모델(MLLM)의 인식(perception) 및 추론 능력 평가에 초점을 맞춘다. 이들은 계획(planning) 역량을 평가하지 못하는데, 이는 대규모로 계획 과제를 큐레이션하고 검증하기 어렵거나, 부정확하고 불충분한 평가 프로토콜에서 기인한다. 이러한 한계를 해결하기 위해, 우리는 원격 탐사에서의 제약 경로 계획을 위한 대규모 신경-기호 벤치마크인 NeSy-Route를 제안한다. 이 벤치마크 내에서, 우리는 고충실도 의미 마스크(semantic masks)를 휴리스틱 탐색(heuristic search)과 통합하여 최적성이 증명 가능한(provably optimal) 다양한 경로 계획 과제를 생성하는 자동화 데이터 생성 프레임워크를 도입한다. 이를 통해 NeSy-Route는 10,821개의 경로 계획 샘플에 걸쳐 계획을 포괄적으로 평가할 수 있으며, 이는 기존 최대 규모 벤치마크보다 거의 10배 크다. 또한, 정확한 평가를 가능하게 하고 인식, 추론, 계획에 대한 세밀한 분석을 동시에 지원하기 위해 3단계 계층적 신경-기호 평가 프로토콜을 개발한다. 다양한 최신(state-of-the-art) MLLM에 대한 포괄적 평가 결과, 기존 MLLM들은 인식 및 계획 역량에서 상당한 결함을 보인다. 우리는 NeSy-Route가 원격 탐사를 위한 더 강력한 MLLM의 추가 연구 및 개발을 지원할 수 있기를 기대한다. 데이터셋과 코드는 https://mingyang1010.github.io/NeSy-Route/에서 이용할 수 있다.

arXiv:2607.17060v1cs.LG

베이즈 필터링된 트랜스포머는 무엇을 믿는가? 예측적 몬테카를로 접근법

What does a Bayes-filtered transformer believe? A predictive Monte Carlo approach

베이즈 필터링된 트랜스포머(BFT, Bayes-filtered transformer)는 두 단계로 생성된 시퀀스에서 학습된 트랜스포머로, 먼저 잠재 과제(latent task)가 사전분포(prior)에서 추출되고, 그다음 해당 과제를 조건으로 관측이 추출된다. 자기회귀(autoregressive) 로그 손실 하에서 학습된 BFT의 다음 토큰 예측은, 이상화된 극한에서, 그 사전분포와 조건부 법칙에 의해 유도되는 베이즈 사후 예측 분포(PPD, posterior predictive distribution)이다. 실제로 학습된 BFT는 이 이상적인 PPD의 근사에 불과하며, 이는 해석적 질문을 제기한다: 학습된 BFT가 실제로 내재화한 잠재 과제에 대한 사전분포와 사후분포는 무엇인가? 기존 연구는 학습된 BFT의 예측을 다양한 "참조(reference)" 사후분포의 예측과 비교하여 이 질문에 답한다. 각 참조 사후분포는 BFT가 구현하고 있을 수 있는 서로 다른 후보 알고리즘이나 계산을 대표한다. 이 예측 공간 비교는 취약하다: 서로 다른 사후분포가 동일한 사후 평균 예측을 공유할 수 있다. 우리는 예측적 몬테카를로(PMC, predictive Monte Carlo)를 임의의 BFT에 대한 일반적인 해석 가능성(interpretability) 도구로 사용한다. 다음 토큰 생성만을 사용하여 PMC는 잠재 과제에 대한 암묵적 사전분포와 사후분포의 근사를 반환하며, 해석적 질문에 잠재 공간에서 직접 답한다. 우리는 0-Markov 및 1-Markov 교환 가능성(exchangeability)에 걸친 세 가지 양식화된(stylized) 과제 계열에 PMC를 적용한다. 이러한 설정에서 이전에 보고된 현상들은 잠재 공간에서도 여전히 관찰된다. 코드는 https://github.com/afiq-aswadi/bft-pmc 에서 이용 가능하다.

arXiv:2607.20427v1cs.CL

MoE 라우팅은 허프만 코딩인가? 사고의 연쇄에서 발견한 빈도-다양성 법칙

Is MoE Routing a Huffman Code? Discovering the Frequency-Diversity Law in Chain-of-Thought

전문가 혼합(MoE) 아키텍처는 스케일링을 혁신했지만, 라우팅의 근본적 논리는 여전히 블랙박스로 남아 있다. 본 논문에서는 MoE 라우팅이 단순한 선택이 아니라 허프만 코딩(Huffman Coding)의 현현임을 밝히는 근본적 지배 원칙을 발견한다. 빈도-다양성 법칙(Frequency-Diversity Law)을 제안하며, Phi-3.5-MoE, Gemma-4-27B-A4B와 같은 최신 모델이 자발적으로 정보이론적 엔진으로 작동함을 보인다. 이러한 모델은 흔한 토큰에는 희소 전문가 자원을 배분하는 반면, 사고의 연쇄(Chain-of-Thought) 경로에서 발견되는 드물고 복잡한 과제에는 고다양성 전문가 위원회를 호출한다. 그러나 Qwen3.5-35B-A3B에서 중요한 중복 함정을 확인한다. 유효 희소성(k/E_eff)이 충분히 낮을 때, 로드 밸런싱(load-balancing)이 의도치 않게 기능적 중복을 강요하여 근본적인 허프만 효율 신호를 가린다. 이 격차를 메우기 위해 기능적 중복을 제거하는 정밀 전략인 Subset Difference Pruning을 제안한다. 가지치기가 추론을 저하시키지 않음을 보이며, 오히려 모델의 잠재적 허프만 효율을 해방시켜 논리가 간결하고 고밀도 경로로 수렴하도록 한다. 우리의 발견은 차세대 MoE가 강제적 로드 밸런싱을 넘어 최소 기술 길이(MDL, Minimum Description Length) 최적성을 향해 나아가야 함을 시사한다. 고빈도 정보에는 더 짧은 전문가 라우팅 코드를, 저빈도 정보에는 더 길고 다양한 코드를 할당함으로써 라우팅을 휴리스틱에서 원칙적 압축 엔진으로 전환할 수 있다.

arXiv:2607.04607v2cs.CV

G2VD: 반사실적 개입 및 인과적 얽힘 해제를 통한 일반화 가능한 AI 생성 비디오 탐지

G2VD: Generalizable AI-Generated Video Detection via Counterfactual Intervention and Causal Disentanglement

AI 비디오 생성의 급격한 발전은 점증하는 보안 위협을 야기하며 강력한 교차 도메인 일반화 능력을 갖춘 신뢰할 수 있는 탐지기를 요구하고 있습니다. 기존 방법들은 도메인 내 평가에서는 우수한 성능을 보이지만, 접해보지 못한 생성기에 대해서는 성능이 크게 저하됩니다. 주요 원인은 탐지기가 본질적인 포렌식 단서 대신 도메인 특화 바이어스에 의존하는 지름길 학습(shortcut learning)입니다. 이 문제를 해결하기 위해 반사실적 개입(counterfactual intervention)과 인과적 얽힘 해제(causal disentanglement)에 기반한 일반화 가능한 AI 생성 비디오 탐지 프레임워크인 G2VD를 제안합니다. 먼저, G2VD는 VAE 기반 재구성과 후속 주파수 도메인 및 픽셀 도메인 정렬을 통해 반사실적 샘플을 구성하는 반사실적 개입 파이프라인(CFIPipeline)을 도입하여, 도메인 특화 바이어스와 진위성 레이블 간의 허위 상관관계를 약화합니다. 이러한 개입을 바탕으로, 상호보완적인 목표를 가진 두 개의 도메인 앵커 브랜치와 힐베르트-슈미트 독립성 기준(HSIC) 기반 제약 조건을 결합한 인과적 얽힘 해제 분류기를 추가로 설계하여 인과적 표현과 비인과적 표현이 각각 본질적인 포렌식 단서와 도메인 특화 바이어스를 포착하도록 유도합니다. 4개의 공개 데이터셋에 걸친 실험은 강력한 교차 도메인 성능과 기준선 방법 대비 일관된 성능 향상을 입증합니다. 까다로운 GenVidBench 설정에서 G2VD는 사용 가능한 학습 데이터의 10%만을 사용하면서도 90% 이상의 전체 정확도(ACC)를 달성했으며, 비교 가능한 최첨단 방법 대비 F1에서 0.194, AUC에서 0.104의 향상을 기록했습니다. 코드는 https://github.com/DMOSCAR-98/G2VD 에서 이용할 수 있습니다.