PyTorchKR Papers

“Code Generation”

총 10,163편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2609.28192v1cs.CV

변화 캡션에서 변화 탐지까지: 원격 탐사 변화 탐지를 위한 의미-외형 일치 프레임워크

From Change Captions to Change Detection: Semantic-Appearance Agreement Framework for Remote Sensing Change Detection

원격 탐사 변화 탐지(RSCD)는 토지 피복 변화 및 도시 개발을 모니터링하는 데 필수적입니다. 그러나 대부분의 방법은 픽셀 수준의 변화 마스크를 요구하며, 이는 주석 작업에 많은 비용과 시간이 소요됩니다. 약지도(weakly supervised) 방식은 이미지 수준의 변화 레이블을 사용하여 이러한 비용을 줄입니다. 하지만 이러한 레이블은 변화의 발생 여부만을 나타내므로, 모델이 추가적이고 복잡한 메커니즘을 통해 변화의 위치와 의미론적 의미를 복원해야 합니다. 이러한 누락된 정보는 무엇이 변하고 무엇으로 바뀌며 어디에서 발생하는지 설명하는 변화 캡션을 통해 직접 제공될 수 있습니다. 따라서 본 연구에서는 수동으로 주석을 단 변화 마스크 없이 변화 마스크를 학습하기 위한 유일한 태스크 특화 지도 정보로 변화 캡션을 활용하는 변화 캡션 유도 RSCD를 소개합니다. 제안하는 프레임워크는 두 가지 구성 요소로 이루어져 있습니다. 각 캡션과 일치하는 제어된 변화를 갖는 대규모 시계열 원격 탐사 이미지 쌍을 생성하는 캡션 기반 생성 파이프라인과, 캡션의 전이 의미론에 의해 유도되는 변화 탐지기입니다. 이 탐지기는 의미-외형 일치 프레임워크(SAAF)를 사용하여 캡션 기반 의미론적 반응과 RGB 차이를 결합해 변화 위치를 특정하며, 텍스트 조건화는 조밀 예측(dense prediction)을 유도합니다. 새롭게 구축된 Flair-RSGen 데이터셋 및 WHU-CDC에서의 실험 결과, SAAF가 평가 프로토콜 하에서 매크로 평균 IoU 및 F1 점수 모두 가장 유사하게 재현된 제한적 지도 베이스라인들을 능가함을 보여줍니다. 코드는 https://github.com/qianyuancs/SAAF 에서 공개되어 있습니다.

arXiv:2610.02959v1cs.CV

TerraVis: MLLM 워크플로우를 통한 텍스트-투-이미지 생성의 실세계 기반 시각적 일관성 평가를 향하여

TerraVis: Towards Evaluation of World-Grounded Visual Consistency in Text-to-Image Generation via MLLM Workflows

최근 텍스트-투-이미지(text-to-image) 모델은 사실성(photorealism), 심미성, 텍스트-이미지 정렬(alignment) 측면에서 상당한 발전을 이루었습니다. 그러나 시각적으로 매력적인 이미지라 하더라도 기형적인 객체 구조, 불가능한 해부학적 형태, 물리적으로 타당하지 않은 상호작용, 일관되지 않은 공간 관계를 보이며 실세계의 개연성(plausibility)을 위반할 수 있습니다. 이러한 오류는 기존의 충실도(fidelity), 심미성, 선호도 또는 정렬 지표로는 잘 포착되지 않습니다. 이러한 격차를 해결하기 위해, 본 연구에서는 생성된 이미지의 실세계 기반(world-grounded) 시각적 일관성을 평가하는 프레임워크인 TerraVis를 소개합니다. TerraVis는 객체, 상호작용 및 장면 수준의 오류를 포괄하는 실세계 일관성 위반에 대한 체계적인 분류 체계(taxonomy)를 정의하며, 이를 식별하고 정량화하기 위해 다단계 평가 프레임워크를 사용합니다. 이미지가 주어지면, TerraVis는 먼저 멀티모달 대규모 언어 모델(MLLM)을 사용하여 평가 적격성을 판단한 다음, 분류 체계에 정의된 18가지 유형에 걸쳐 위반 사항을 탐지하고 이를 경미(minor) 또는 중대(major)로 분류하여 전반적인 실세계 일관성 점수를 도출합니다. 널리 사용되는 두 가지 벤치마크에서 다양한 오픈소스 및 상용 텍스트-투-이미지 모델을 대상으로 평가한 결과, TerraVis는 기존 지표 중 실세계 일관성에 대한 인간의 판단과 가장 강력한 상관관계를 달성했습니다. 또한 벤치마크 결과는 기존 지표에서 우수한 성능을 보이는 모델이라도 상당한 실세계 일관성 오류를 나타낼 수 있음을 보여줍니다. 이러한 결과는 실세계 일관성이 상호보완적인 평가 차원임을 강조하며, TerraVis가 이러한 오류의 체계적인 정량화, 진단 및 비교를 가능하게 함을 입증합니다. 본 연구의 코드는 https://github.com/ShyFoo/TerraVis 에서 공개되어 있습니다.

arXiv:2109.04145공식 코드 · Papers with Code

PIMNet: A Parallel, Iterative and Mimicking Network for Scene Text Recognition

Nowadays, scene text recognition has attracted more and more attention due to its various applications. Most state-of-the-art methods adopt an encoder-decoder framework with attention mechanism, which generates text autoregressively from left to right. Despite the convincing performance, the speed is limited because of the one-by-one decoding strategy. As opposed to autoregressive models, non-autoregressive models predict the results in parallel with a much shorter inference time, but the accuracy falls behind the autoregressive counterpart considerably. In this paper, we propose a Parallel, Iterative and Mimicking Network (PIMNet) to balance accuracy and efficiency. Specifically, PIMNet adopts a parallel attention mechanism to predict the text faster and an iterative generation mechanism to make the predictions more accurate. In each iteration, the context information is fully explored. To improve learning of the hidden layer, we exploit the mimicking learning in the training phase, where an additional autoregressive decoder is adopted and the parallel decoder mimics the autoregressive decoder with fitting outputs of the hidden layer. With the shared backbone between the two decoders, the proposed PIMNet can be trained end-to-end without pre-training. During inference, the branch of the autoregressive decoder is removed for a faster speed. Extensive experiments on public benchmarks demonstrate the effectiveness and efficiency of PIMNet. Our code will be available at https://github.com/Pay20Y/PIMNet.

arXiv:2203.17090공식 코드 · Papers with Code

PanGu-Bot: Efficient Generative Dialogue Pre-training from Pre-trained Language Model

In this paper, we introduce PanGu-Bot, a Chinese pre-trained open-domain dialogue generation model based on a large pre-trained language model (PLM) PANGU-alpha (Zeng et al.,2021). Different from other pre-trained dialogue models trained over a massive amount of dialogue data from scratch, we aim to build a powerful dialogue model with relatively fewer data and computation costs by inheriting valuable language capabilities and knowledge from PLMs. To this end, we train PanGu-Bot from the large PLM PANGU-alpha, which has been proven well-performed on a variety of Chinese natural language tasks. We investigate different aspects of responses generated by PanGu-Bot, including response quality, knowledge, and safety. We show that PanGu-Bot outperforms state-of-the-art Chinese dialogue systems (CDIALGPT (Wang et al., 2020), EVA (Zhou et al., 2021), EVA2.0 (Gu et al., 2022)) w.r.t. the above three aspects. We also demonstrate that PanGu-Bot can be easily deployed to generate emotional responses without further training. Throughout our empirical analysis, we also point out that the PanGu-Bot response quality, knowledge correctness, and safety are still far from perfect, and further explorations are indispensable to building reliable and smart dialogue systems. Our model and code will be available at https://github.com/huawei-noah/Pretrained-Language-Model/tree/master/PanGu-Bot soon.

arXiv:2204.08046공식 코드 · Papers with Code

Evaluating Mixed-initiative Conversational Search Systems via User Simulation

Clarifying the underlying user information need by asking clarifying questions is an important feature of modern conversational search system. However, evaluation of such systems through answering prompted clarifying questions requires significant human effort, which can be time-consuming and expensive. In this paper, we propose a conversational User Simulator, called USi, for automatic evaluation of such conversational search systems. Given a description of an information need, USi is capable of automatically answering clarifying questions about the topic throughout the search session. Through a set of experiments, including automated natural language generation metrics and crowdsourcing studies, we show that responses generated by USi are both inline with the underlying information need and comparable to human-generated answers. Moreover, we make the first steps towards multi-turn interactions, where conversational search systems asks multiple questions to the (simulated) user with a goal of clarifying the user need. To this end, we expand on currently available datasets for studying clarifying questions, i.e., Qulac and ClariQ, by performing a crowdsourcing-based multi-turn data acquisition. We show that our generative, GPT2-based model, is capable of providing accurate and natural answers to unseen clarifying questions in the single-turn setting and discuss capabilities of our model in the multi-turn setting. We provide the code, data, and the pre-trained model to be used for further research on the topic.

arXiv:2210.09890공식 코드 · Papers with Code

IntTower: the Next Generation of Two-Tower Model for Pre-Ranking System

Scoring a large number of candidates precisely in several milliseconds is vital for industrial pre-ranking systems. Existing pre-ranking systems primarily adopt the \textbf{two-tower} model since the ``user-item decoupling architecture'' paradigm is able to balance the \textit{efficiency} and \textit{effectiveness}. However, the cost of high efficiency is the neglect of the potential information interaction between user and item towers, hindering the prediction accuracy critically. In this paper, we show it is possible to design a two-tower model that emphasizes both information interactions and inference efficiency. The proposed model, IntTower (short for \textit{Interaction enhanced Two-Tower}), consists of Light-SE, FE-Block and CIR modules. Specifically, lightweight Light-SE module is used to identify the importance of different features and obtain refined feature representations in each tower. FE-Block module performs fine-grained and early feature interactions to capture the interactive signals between user and item towers explicitly and CIR module leverages a contrastive interaction regularization to further enhance the interactions implicitly. Experimental results on three public datasets show that IntTower outperforms the SOTA pre-ranking models significantly and even achieves comparable performance in comparison with the ranking models. Moreover, we further verify the effectiveness of IntTower on a large-scale advertisement pre-ranking system. The code of IntTower is publicly available\footnote{https://github.com/archersama/IntTower}

arXiv:2211.06408공식 코드 · Papers with Code

Physically-Based Face Rendering for NIR-VIS Face Recognition

Near infrared (NIR) to Visible (VIS) face matching is challenging due to the significant domain gaps as well as a lack of sufficient data for cross-modality model training. To overcome this problem, we propose a novel method for paired NIR-VIS facial image generation. Specifically, we reconstruct 3D face shape and reflectance from a large 2D facial dataset and introduce a novel method of transforming the VIS reflectance to NIR reflectance. We then use a physically-based renderer to generate a vast, high-resolution and photorealistic dataset consisting of various poses and identities in the NIR and VIS spectra. Moreover, to facilitate the identity feature learning, we propose an IDentity-based Maximum Mean Discrepancy (ID-MMD) loss, which not only reduces the modality gap between NIR and VIS images at the domain level but encourages the network to focus on the identity features instead of facial details, such as poses and accessories. Extensive experiments conducted on four challenging NIR-VIS face recognition benchmarks demonstrate that the proposed method can achieve comparable performance with the state-of-the-art (SOTA) methods without requiring any existing NIR-VIS face recognition datasets. With slightly fine-tuning on the target NIR-VIS face recognition datasets, our method can significantly surpass the SOTA performance. Code and pretrained models are released under the insightface (https://github.com/deepinsight/insightface/tree/master/recognition).

arXiv:2309.01274공식 코드 · Papers with Code

Diffusion Models with Deterministic Normalizing Flow Priors

For faster sampling and higher sample quality, we propose DiNof ($\textbf{Di}$ffusion with $\textbf{No}$rmalizing $\textbf{f}$low priors), a technique that makes use of normalizing flows and diffusion models. We use normalizing flows to parameterize the noisy data at any arbitrary step of the diffusion process and utilize it as the prior in the reverse diffusion process. More specifically, the forward noising process turns a data distribution into partially noisy data, which are subsequently transformed into a Gaussian distribution by a nonlinear process. The backward denoising procedure begins with a prior created by sampling from the Gaussian distribution and applying the invertible normalizing flow transformations deterministically. To generate the data distribution, the prior then undergoes the remaining diffusion stochastic denoising procedure. Through the reduction of the number of total diffusion steps, we are able to speed up both the forward and backward processes. More importantly, we improve the expressive power of diffusion models by employing both deterministic and stochastic mappings. Experiments on standard image generation datasets demonstrate the advantage of the proposed method over existing approaches. On the unconditional CIFAR10 dataset, for example, we achieve an FID of 2.01 and an Inception score of 9.96. Our method also demonstrates competitive performance on CelebA-HQ-256 dataset as it obtains an FID score of 7.11. Code is available at https://github.com/MohsenZand/DiNof.

arXiv:2404.17063공식 코드 · Papers with Code

WheelPose: Data Synthesis Techniques to Improve Pose Estimation Performance on Wheelchair Users

Existing pose estimation models perform poorly on wheelchair users due to a lack of representation in training data. We present a data synthesis pipeline to address this disparity in data collection and subsequently improve pose estimation performance for wheelchair users. Our configurable pipeline generates synthetic data of wheelchair users using motion capture data and motion generation outputs simulated in the Unity game engine. We validated our pipeline by conducting a human evaluation, investigating perceived realism, diversity, and an AI performance evaluation on a set of synthetic datasets from our pipeline that synthesized different backgrounds, models, and postures. We found our generated datasets were perceived as realistic by human evaluators, had more diversity than existing image datasets, and had improved person detection and pose estimation performance when fine-tuned on existing pose estimation models. Through this work, we hope to create a foothold for future efforts in tackling the inclusiveness of AI in a data-centric and human-centric manner with the data synthesis techniques demonstrated in this work. Finally, for future works to extend upon, we open source all code in this research and provide a fully configurable Unity Environment used to generate our datasets. In the case of any models we are unable to share due to redistribution and licensing policies, we provide detailed instructions on how to source and replace said models.

arXiv:2405.13722공식 코드 · Papers with Code

LightningDrag: Lightning Fast and Accurate Drag-based Image Editing Emerging from Videos

Accuracy and speed are critical in image editing tasks. Pan et al. introduced a drag-based image editing framework that achieves pixel-level control using Generative Adversarial Networks (GANs). A flurry of subsequent studies enhanced this framework's generality by leveraging large-scale diffusion models. However, these methods often suffer from inordinately long processing times (exceeding 1 minute per edit) and low success rates. Addressing these issues head on, we present LightningDrag, a rapid approach enabling high quality drag-based image editing in ~1 second. Unlike most previous methods, we redefine drag-based editing as a conditional generation task, eliminating the need for time-consuming latent optimization or gradient-based guidance during inference. In addition, the design of our pipeline allows us to train our model on large-scale paired video frames, which contain rich motion information such as object translations, changing poses and orientations, zooming in and out, etc. By learning from videos, our approach can significantly outperform previous methods in terms of accuracy and consistency. Despite being trained solely on videos, our model generalizes well to perform local shape deformations not presented in the training data (e.g., lengthening of hair, twisting rainbows, etc.). Extensive qualitative and quantitative evaluations on benchmark datasets corroborate the superiority of our approach. The code and model will be released at https://github.com/magic-research/LightningDrag.

arXiv:2408.01571공식 코드 · Papers with Code

Counterfactual Explanations for Medical Image Classification and Regression using Diffusion Autoencoder

Counterfactual explanations (CEs) aim to enhance the interpretability of machine learning models by illustrating how alterations in input features would affect the resulting predictions. Common CE approaches require an additional model and are typically constrained to binary counterfactuals. In contrast, we propose a novel method that operates directly on the latent space of a generative model, specifically a Diffusion Autoencoder (DAE). This approach offers inherent interpretability by enabling the generation of CEs and the continuous visualization of the model's internal representation across decision boundaries. Our method leverages the DAE's ability to encode images into a semantically rich latent space in an unsupervised manner, eliminating the need for labeled data or separate feature extraction models. We show that these latent representations are helpful for medical condition classification and the ordinal regression of severity pathologies, such as vertebral compression fractures (VCF) and diabetic retinopathy (DR). Beyond binary CEs, our method supports the visualization of ordinal CEs using a linear model, providing deeper insights into the model's decision-making process and enhancing interpretability. Experiments across various medical imaging datasets demonstrate the method's advantages in interpretability and versatility. The linear manifold of the DAE's latent space allows for meaningful interpolation and manipulation, making it a powerful tool for exploring medical image properties. Our code is available at https://doi.org/10.5281/zenodo.13859266.

arXiv:2409.04005공식 코드 · Papers with Code

Qihoo-T2X: An Efficient Proxy-Tokenized Diffusion Transformer for Text-to-Any-Task

The global self-attention mechanism in diffusion transformers involves redundant computation due to the sparse and redundant nature of visual information, and the attention map of tokens within a spatial window shows significant similarity. To address this redundancy, we propose the Proxy-Tokenized Diffusion Transformer (PT-DiT), which employs sparse representative token attention (where the number of representative tokens is much smaller than the total number of tokens) to model global visual information efficiently. Specifically, within each transformer block, we compute an averaging token from each spatial-temporal window to serve as a proxy token for that region. The global semantics are captured through the self-attention of these proxy tokens and then injected into all latent tokens via cross-attention. Simultaneously, we introduce window and shift window attention to address the limitations in detail modeling caused by the sparse attention mechanism. Building on the well-designed PT-DiT, we further develop the Qihoo-T2X family, which includes a variety of models for T2I, T2V, and T2MV tasks. Experimental results show that PT-DiT achieves competitive performance while reducing the computational complexity in both image and video generation tasks (e.g., a 49% reduction compared to DiT and a 34% reduction compared to PixArt-$\alpha$). The visual exhibition and source code of Qihoo-T2X is available at https://360cvgroup.github.io/Qihoo-T2X/.

arXiv:2409.19911공식 코드 · Papers with Code

Replace Anyone in Videos

The field of controllable human-centric video generation has witnessed remarkable progress, particularly with the advent of diffusion models. However, achieving precise and localized control over human motion in videos, such as replacing or inserting individuals while preserving desired motion patterns, still remains a formidable challenge. In this work, we present the ReplaceAnyone framework, which focuses on localized human replacement and insertion featuring intricate backgrounds. Specifically, we formulate this task as an image-conditioned video inpainting paradigm with pose guidance, utilizing a unified end-to-end video diffusion architecture that facilitates image-conditioned video inpainting within masked regions. To prevent shape leakage and enable granular local control, we introduce diverse mask forms involving both regular and irregular shapes. Furthermore, we implement an enriched visual guidance mechanism to enhance appearance alignment, a hybrid inpainting encoder to further preserve the detailed background information in the masked video, and a two-phase optimization methodology to simplify the training difficulty. ReplaceAnyone enables seamless replacement or insertion of characters while maintaining the desired pose motion and reference appearance within a single framework. Extensive experimental results demonstrate the effectiveness of our method in generating realistic and coherent video content. The proposed ReplaceAnyone can be seamlessly applied not only to traditional 3D-UNet base models but also to DiT-based video models such as Wan2.1. The code will be available at https://github.com/ali-vilab/UniAnimate-DiT.

arXiv:2502.04829

Optimistic Gradient Learning with Hessian Corrections for High-Dimensional Black-Box Optimization

Black-box algorithms are designed to optimize functions without relying on their underlying analytical structure or gradient information, making them essential when gradients are inaccessible or difficult to compute. Traditional methods for solving black-box optimization (BBO) problems predominantly rely on non-parametric models and struggle to scale to large input spaces. Conversely, parametric methods that model the function with neural estimators and obtain gradient signals via backpropagation may suffer from significant gradient errors. A recent alternative, Explicit Gradient Learning (EGL), which directly learns the gradient using a first-order Taylor approximation, has demonstrated superior performance over both parametric and non-parametric methods. In this work, we propose two novel gradient learning variants to address the robustness challenges posed by high-dimensional, complex, and highly non-linear problems. Optimistic Gradient Learning (OGL) introduces a bias toward lower regions in the function landscape, while Higher-order Gradient Learning (HGL) incorporates second-order Taylor corrections to improve gradient accuracy. We combine these approaches into the unified OHGL algorithm, achieving state-of-the-art (SOTA) performance on the synthetic COCO suite. Additionally, we demonstrate OHGLs applicability to high-dimensional real-world machine learning (ML) tasks such as adversarial training and code generation. Our results highlight OHGLs ability to generate stronger candidates, offering a valuable tool for ML researchers and practitioners tackling high-dimensional, non-linear optimization challenges

arXiv:2503.07602

DreamRelation: Relation-Centric Video Customization

Relational video customization refers to the creation of personalized videos that depict user-specified relations between two subjects, a crucial task for comprehending real-world visual content. While existing methods can personalize subject appearances and motions, they still struggle with complex relational video customization, where precise relational modeling and high generalization across subject categories are essential. The primary challenge arises from the intricate spatial arrangements, layout variations, and nuanced temporal dynamics inherent in relations; consequently, current models tend to overemphasize irrelevant visual details rather than capturing meaningful interactions. To address these challenges, we propose DreamRelation, a novel approach that personalizes relations through a small set of exemplar videos, leveraging two key components: Relational Decoupling Learning and Relational Dynamics Enhancement. First, in Relational Decoupling Learning, we disentangle relations from subject appearances using relation LoRA triplet and hybrid mask training strategy, ensuring better generalization across diverse relationships. Furthermore, we determine the optimal design of relation LoRA triplet by analyzing the distinct roles of the query, key, and value features within MM-DiT's attention mechanism, making DreamRelation the first relational video generation framework with explainable components. Second, in Relational Dynamics Enhancement, we introduce space-time relational contrastive loss, which prioritizes relational dynamics while minimizing the reliance on detailed subject appearances. Extensive experiments demonstrate that DreamRelation outperforms state-of-the-art methods in relational video customization. Code and models will be made publicly available.

arXiv:2504.11850

ACE: Attentional Concept Erasure in Diffusion Models

Large text-to-image diffusion models have demonstrated remarkable image synthesis capabilities, but their indiscriminate training on Internet-scale data has led to learned concepts that enable harmful, copyrighted, or otherwise undesirable content generation. We address the task of concept erasure in diffusion models, i.e., removing a specified concept from a pre-trained model such that prompting the concept (or related synonyms) no longer yields its depiction, while preserving the model's ability to generate other content. We propose a novel method, Attentional Concept Erasure (ACE), that integrates a closed-form attention manipulation with lightweight fine-tuning. Theoretically, we formulate concept erasure as aligning the model's conditional distribution on the target concept with a neutral distribution. Our approach identifies and nullifies concept-specific latent directions in the cross-attention modules via a gated low-rank adaptation, followed by adversarially augmented fine-tuning to ensure thorough erasure of the concept and its synonyms. Empirically, we demonstrate on multiple benchmarks, including object classes, celebrity faces, explicit content, and artistic styles, that ACE achieves state-of-the-art concept removal efficacy and robustness. Compared to prior methods, ACE better balances generality (erasing concept and related terms) and specificity (preserving unrelated content), scales to dozens of concepts, and is efficient, requiring only a few seconds of adaptation per concept. We will release our code to facilitate safer deployment of diffusion models.

arXiv:2505.11080공식 코드 · Papers with Code

BLEUBERI: BLEU is a surprisingly effective reward for instruction following

Reward models are central to aligning LLMs with human preferences, but they are costly to train, requiring large-scale human-labeled preference data and powerful pretrained LLM backbones. Meanwhile, the increasing availability of high-quality synthetic instruction-following datasets raises the question: can simpler, reference-based metrics serve as viable alternatives to reward models during RL-based alignment? In this paper, we show first that BLEU, a basic string-matching metric, surprisingly matches strong reward models in agreement with human preferences on general instruction-following datasets. Based on this insight, we develop BLEUBERI, a method that first identifies challenging instructions and then applies Group Relative Policy Optimization (GRPO) using BLEU directly as the reward function. We demonstrate that BLEUBERI-trained models are competitive with models trained via reward model-guided RL across four challenging instruction-following benchmarks and three different base language models. A human evaluation further supports that the quality of BLEUBERI model outputs is on par with those from reward model-aligned models. Moreover, BLEUBERI models generate outputs that are more factually grounded than competing methods. Overall, we show that given access to high-quality reference outputs (easily obtained via existing instruction-following datasets or synthetic data generation), string matching-based metrics are cheap yet effective proxies for reward models during alignment. We release our code and data at https://github.com/lilakk/BLEUBERI.

arXiv:2605.17291v2cs.LG

Step-wise Rubric Rewards for LLM Reasoning

Reinforcement Learning with Verifiable Rewards (RLVR) improves language-model reasoning using final-answer correctness, but lacks supervision for intermediate steps. Rubric-based methods such as Rubrics as Rewards (RaR) provide finer-grained criteria, yet collapse rubric scores into a response-level scalar. This loses criterion-level structure, rewards flawed steps in correct responses, penalizes valid steps in incorrect responses, and can encourage unbounded self-correction. In 1,000 sampled problems, 18.2% of steps in answer-correct responses are wrong, while 49.9% of steps in answer-incorrect responses are correct. We propose Step-wise Rubrics as Rewards (SRaR), which (i) uses an LLM judge to assign rubric items to reasoning steps, (ii) normalizes step scores across rollouts to focus learning on quality differences, and (iii) combines step-wise and outcome rewards through a decoupled advantage estimator that preserves a stable outcome baseline. We also construct a 16K-problem rubric dataset by contrastively distilling criteria from verified correct and flawed reasoning paths. Across six math benchmarks, SRaR improves average accuracy over RaR by 3.57 and 2.75 points on Qwen3-8B-Non-Thinking and Qwen3-32B-Non-Thinking, respectively. On AIME 2025, it increases the Faithful Reasoning Rate from 34.5% to 46.7% and reduces self-correction looping from 48.1% to 26.5%. SRaR also achieves the best results on three of four code-generation benchmarks.

arXiv:2606.02955v3cs.CL

Fast-dLLM++: Fr\'echet Profile Decoding for Faster Diffusion LLM Inference

Diffusion large language models promise parallel token generation, yet inference remains bottlenecked by deciding which masked tokens can be safely committed together. Fast-dLLM addressed this with KV caching and confidence-guided parallel decoding, but its decoding theory uses a homogeneous high-confidence assumption that effectively reduces each candidate set to its weakest selected token. We argue that this leaves speed on the table because real decoding steps exhibit heterogeneous confidence profiles. We propose Fast-dLLM++, a training-free extension that introduces Fr\'echet profile decoding: selecting parallel commit sets from the full sorted confidence profile rather than a single worst-case confidence. The resulting rule is a heterogeneous-confidence generalization of Fast-dLLM's factor selector and it recovers the previous rule exactly in the equal-confidence case and adds a provable heterogeneity bonus when the selected tokens have uneven confidences. Fast-dLLM++ leaves the model, diffusion process, and cache implementation entirely unchanged, making it a drop-in replacement for existing Fast-dLLM decoding. Experiments on GSM8K, MATH, HumanEval, and MBPP with the LLaDA-8B model show that the theoretical improvement translates directly into empirical gains: profile-aware selection improves the accuracy-throughput frontier by exploiting safe parallelism that weakest-token rules miss, achieving up to 37% higher throughput at comparable accuracy. Our code release is at https://github.com/Ringo-Star/FastdLLM_plusplus.

arXiv:2608.25992v1cs.AI

ProgRouter: 품질-비용 절충 하 다중 에이전트 LLM 워크플로를 위한 온라인 진행률 기반 오케스트레이션

ProgRouter: Online Progress-Guided Orchestration for Multi-Agent LLM Workflows under Quality-Cost Tradeoffs

다중 에이전트 대규모 언어 모델(LLM) 워크플로는 전문화된 LLM 에이전트 간의 협력적 추론을 통해 복잡하고 개방형인 과제를 해결하는 강력한 패러다임으로 부상했으나, 반복적인 LLM 호출과 장기 컨텍스트 축적으로 인해 상당한 운영 비용이 발생한다. 기존 캐스케이드 라우팅 방법은 일회성 쿼리 수준의 결정을 내리며, 각 단계에서 적절한 LLM이 진행 중인 과제 진행 상황, 남은 과제 난이도, 비용 효율 요구사항에 따라 달라지는 다단계 워크플로의 동적이고 상태 의존적인 특성에 적응할 수 없다. 우리는 과제 해결 품질을 유지하면서 시간 및 비용 예산을 준수하기 위해 워크플로 단계마다 LLM 에이전트를 적응적으로 선택하는 온라인 진행률 기반 라우팅 프레임워크인 ProgRouter를 제안한다. ProgRouter는 거친 워크플로 결과 체제와 하위 과제 완료, 진행 추세, 워크플로 상태 품질에 대한 세밀한 신호를 결합하는 다중 관점 과제 진행률 점수화기를 도입한다. 이어서 이중 경로 과제 진행률 예측기와 적응형 메타 게이팅(meta-gating) 메커니즘이 각 후보 라우팅 LLM에 대한 진행 이득을 추정한다. ProgRouter는 진행 이득, 과제 시간 예산, 장기 운영 비용 효율을 균형 있게 고려한 온라인 단계별 라우팅 결정을 수행한다. 에이전틱 코드 생성, 수학적 추론, 검색-증강 생성(RAG) 기반 장문 질의응답을 아우르는 HumanEval Plus, MBPP, MATH-500, ASQA에서의 실험은 ProgRouter가 강력한 과제 해결 성능을 유지하면서 주요 베이스라인 대비 운영 비용을 줄임을 보여준다.