PyTorchKR Papers

“Code Generation”

총 10,163편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2610.11620v1cs.AI

적응 위치의 중요성: 성능 유지를 위한 층 선택적 미세조정

Where to Adapt Matters: Layer-Selective Fine-Tuning for Capability Retention

매개변수 효율적 미세조정(PEFT)은 대규모 언어 모델(LLM)이 특화된 작업에 적응할 수 있도록 지원하지만, 사전학습 과정에서 획득한 일반적인 능력을 저하시키는 대가를 치르는 경우가 많습니다. 기존 접근법은 주로 데이터 재생(data replay)이나 정규화를 통해 이러한 트레이드오프를 완화하며, 추가 데이터나 명시적인 최적화 제약 조건에 의존합니다. 본 연구에서는 이 대신 '어느 위치에 적응을 적용해야 하는가?'라는 다른 질문에 집중합니다. 우리는 서로 다른 트랜스포머 층을 미세조정할 때 목표 작업에서의 성능 향상과 능력 저하의 정도가 서로 다르게 나타남을 발견했으며, 이는 모든 층이 적응에 동일하게 적합한 것은 아님을 시사합니다. 이러한 차이를 규명하기 위해 층별 경험적 피셔 정보(empirical Fisher information)를 사용하여 목표 작업 민감도를 측정합니다. 그러나 피셔 점수를 계산하려면 역전파 연산이 필요하며 대형 모델일수록 계산 비용이 급증합니다. 따라서 우리는 층별 민감도 순위를 매기기 위한 경량의 순전파 전용 대용 지표로 입출력 코사인 유사도(input--output cosine similarity)를 도입합니다. 다양한 모델과 작업 전반에 걸쳐 입출력 유사도가 낮은 층일수록 일관되게 더 높은 경험적 피셔 점수를 나타냈습니다. 이러한 관찰을 바탕으로, 우리는 입출력 유사도가 낮은 층에만 학습 가능한 LoRA 어댑터를 배치하는 층 선택적 LoRA(Layer-Selective LoRA, LS-LoRA)를 제안합니다. 수학적 추론 및 코드 생성 실험 결과, LS-LoRA는 표준적인 모든 층 LoRA(all-layer LoRA) 대비 평균 목표 작업 성능을 향상시키는 동시에 상식 추론 능력을 훨씬 더 많이 유지함을 보여주었으며, 이는 적응할 위치를 신중하게 선택하는 것이 목표 작업 적응과 일반 능력 유지 간의 균형을 맞추는 단순하고도 효과적인 방법을 제공할 수 있음을 입증합니다.

arXiv:2610.16301v1cs.AI

FON: A Draft-Revise Dual-Organ Language Model without Sequence-Level Quadratic Attention

Self-attention assigns two functions to one mechanism-where to read and how to mix-at the price of O(L^2) training, an O(T) decode state, and position signals bound to the training length. Existing attention-free alternatives all optimize the single function of generation. We propose FON (Fusion-Organ Network), which differentiates one weight stack into a drafting organ (multi-tap learnable delays + content-adaptive recursion; causal, O(S) training, O(1) decode state, no sequence-level quadratic attention) and a revision organ that re-processes the written prefix through multi-scale ripple kernels injected via weak gates, never generating. At 30M parameters, iso-parameter and iso-token budget, FON reaches perplexity 131.03 on WikiText-103, beating Transformer (197.73) and Mamba2 (152.53); the advantage persists under all depth-aligned baselines and at the 70M tier (-20.8%). At 8x length extrapolation FON degrades by ~8-9% without positional-encoding engineering, between ALiBi and RoPE. The revision channel provides a structural pathway for in-domain infilling that causal architectures lack in principle (+19% over on-demand coupling), at a perplexity cost indistinguishable within measurement precision. We report all costs and boundaries: impaired long-range retrieval, synthetic-task disadvantages, weaker out-of-domain code infilling than a same-recipe Transformer, and a 6.5-12.3x wall-clock gap.

arXiv:1812.01756공식 코드 · Papers with Code

Multi$^{\mathbf{3}}$Net: Segmenting Flooded Buildings via Fusion of Multiresolution, Multisensor, and Multitemporal Satellite Imagery

We propose a novel approach for rapid segmentation of flooded buildings by fusing multiresolution, multisensor, and multitemporal satellite imagery in a convolutional neural network. Our model significantly expedites the generation of satellite imagery-based flood maps, crucial for first responders and local authorities in the early stages of flood events. By incorporating multitemporal satellite imagery, our model allows for rapid and accurate post-disaster damage assessment and can be used by governments to better coordinate medium- and long-term financial assistance programs for affected areas. The network consists of multiple streams of encoder-decoder architectures that extract spatiotemporal information from medium-resolution images and spatial information from high-resolution images before fusing the resulting representations into a single medium-resolution segmentation map of flooded buildings. We compare our model to state-of-the-art methods for building footprint segmentation as well as to alternative fusion approaches for the segmentation of flooded buildings and find that our model performs best on both tasks. We also demonstrate that our model produces highly accurate segmentation maps of flooded buildings using only publicly available medium-resolution data instead of significantly more detailed but sparsely available very high-resolution data. We release the first open-source dataset of fully preprocessed and labeled multiresolution, multispectral, and multitemporal satellite images of disaster sites along with our source code.

arXiv:2109.00894공식 코드 · Papers with Code

Generative Wind Power Curve Modeling Via Machine Vision: A Self-learning Deep Convolutional Network Based Method

This paper develops a novel self-training U-net (STU-net) based method for the automated WPC model generation without requiring data pre-processing. The self-training (ST) process of STU-net has two steps. First, different from traditional studies regarding the WPC modeling as a curve fitting problem, in this paper, we renovate the WPC modeling formulation from a machine vision aspect. To develop sufficiently diversified training samples, we synthesize supervisory control and data acquisition (SCADA) data based on a set of S-shape functions depicting WPCs. These synthesized SCADA data and WPC functions are visualized as images and paired as training samples(I_x, I_wpc). A U-net is then developed to approximate the model recovering I_wpc from I_x. The developed U-net is applied into observed SCADA data and can successfully generate the I_wpc. Moreover, we develop a pixel mapping and correction process to derive a mathematical form f_wpc representing I_wpcgenerated previously. The proposed STU-net only needs to train once and does not require any data preprocessing in applications. Numerical experiments based on 76 WTs are conducted to validate the superiority of the proposed method by benchmarking against classical WPC modeling methods. To demonstrate the repeatability of the presented research, we release our code at https://github.com/IkeYang/STU-net.

arXiv:2110.13719

Semi-supervised dry herbage mass estimation using automatic data and synthetic images

Monitoring species-specific dry herbage biomass is an important aspect of pasture-based milk production systems. Being aware of the herbage biomass in the field enables farmers to manage surpluses and deficits in herbage supply, as well as using targeted nitrogen fertilization when necessary. Deep learning for computer vision is a powerful tool in this context as it can accurately estimate the dry biomass of a herbage parcel using images of the grass canopy taken using a portable device. However, the performance of deep learning comes at the cost of an extensive, and in this case destructive, data gathering process. Since accurate species-specific biomass estimation is labor intensive and destructive for the herbage parcel, we propose in this paper to study low supervision approaches to dry biomass estimation using computer vision. Our contributions include: a synthetic data generation algorithm to generate data for a herbage height aware semantic segmentation task, an automatic process to label data using semantic segmentation maps, and a robust regression network trained to predict dry biomass using approximate biomass labels and a small trusted dataset with gold standard labels. We design our approach on a herbage mass estimation dataset collected in Ireland and also report state-of-the-art results on the publicly released Grass-Clover biomass estimation dataset from Denmark. Our code is available at https://git.io/J0L2a

arXiv:2203.13856공식 코드 · Papers with Code

Robust deep learning for eye fundus images: Bridging real and synthetic data for enhancing generalization

Deep learning applications for assessing medical images are limited because the datasets are often small and imbalanced. The use of synthetic data has been proposed in the literature, but neither a robust comparison of the different methods nor generalizability has been reported. Our approach integrates a retinal image quality assessment model and StyleGAN2 architecture to enhance Age-related Macular Degeneration (AMD) detection capabilities and improve generalizability. This work compares ten different Generative Adversarial Network (GAN) architectures to generate synthetic eye-fundus images with and without AMD. We combined subsets of three public databases (iChallenge-AMD, ODIR-2019, and RIADD) to form a single training and test set. We employed the STARE dataset for external validation, ensuring a comprehensive assessment of the proposed approach. The results show that StyleGAN2 reached the lowest Frechet Inception Distance (166.17), and clinicians could not accurately differentiate between real and synthetic images. ResNet-18 architecture obtained the best performance with 85% accuracy and outperformed the two human experts (80% and 75%) in detecting AMD fundus images. The accuracy rates were 82.8% for the test set and 81.3% for the STARE dataset, demonstrating the model's generalizability. The proposed methodology for synthetic medical image generation has been validated for robustness and accuracy, with free access to its code for further research and development in this field.

arXiv:2302.10897공식 코드 · Papers with Code

Learning to Simulate Daily Activities via Modeling Dynamic Human Needs

Daily activity data that records individuals' various types of activities in daily life are widely used in many applications such as activity scheduling, activity recommendation, and policymaking. Though with high value, its accessibility is limited due to high collection costs and potential privacy issues. Therefore, simulating human activities to produce massive high-quality data is of great importance to benefit practical applications. However, existing solutions, including rule-based methods with simplified assumptions of human behavior and data-driven methods directly fitting real-world data, both cannot fully qualify for matching reality. In this paper, motivated by the classic psychological theory, Maslow's need theory describing human motivation, we propose a knowledge-driven simulation framework based on generative adversarial imitation learning. To enhance the fidelity and utility of the generated activity data, our core idea is to model the evolution of human needs as the underlying mechanism that drives activity generation in the simulation model. Specifically, this is achieved by a hierarchical model structure that disentangles different need levels, and the use of neural stochastic differential equations that successfully captures piecewise-continuous characteristics of need dynamics. Extensive experiments demonstrate that our framework outperforms the state-of-the-art baselines in terms of data fidelity and utility. Besides, we present the insightful interpretability of the need modeling. The code is available at https://github.com/tsinghua-fib-lab/SAND.

arXiv:2312.08365

An Invitation to Deep Reinforcement Learning

Training a deep neural network to maximize a target objective has become the standard recipe for successful machine learning over the last decade. These networks can be optimized with supervised learning, if the target objective is differentiable. For many interesting problems, this is however not the case. Common objectives like intersection over union (IoU), bilingual evaluation understudy (BLEU) score or rewards cannot be optimized with supervised learning. A common workaround is to define differentiable surrogate losses, leading to suboptimal solutions with respect to the actual objective. Reinforcement learning (RL) has emerged as a promising alternative for optimizing deep neural networks to maximize non-differentiable objectives in recent years. Examples include aligning large language models via human feedback, code generation, object detection or control problems. This makes RL techniques relevant to the larger machine learning audience. The subject is, however, time intensive to approach due to the large range of methods, as well as the often very theoretical presentation. In this introduction, we take an alternative approach, different from classic reinforcement learning textbooks. Rather than focusing on tabular problems, we introduce reinforcement learning as a generalization of supervised learning, which we first apply to non-differentiable objectives and later to temporal problems. Assuming only basic knowledge of supervised learning, the reader will be able to understand state-of-the-art deep RL algorithms like proximal policy optimization (PPO) after reading this tutorial.

논문 ID: document/1038169

Clustering-Guided Class Activation for Weakly Supervised Semantic Segmentation

Weakly-supervised semantic segmentation (WSSS) methods via transformer have been actively studied by leveraging their strong capability to capture the global context. However, since the activation function only highlights a few tokens in the self-attention mechanism of the transformer, these methods still suffer from the sparse attention map, which leads to the generation of incomplete pseudo labels. In this paper, we propose a novel class activation scheme that is able to uniformly highlight the whole object region. The key idea of the proposed method is to activate the object region by following the guide of clusters, which are formed by combining similar image features of the object. Specifically, the clustering-guided class activation map (ClusterCAM) is generated from the proposed clustering-based attention module, and highly responsive regions in this map are then adopted to activate target objects in the encoded feature space. This helps the model to explore the entire region of the target object by using the semantic proximity between patch tokens extracted from the same object. Based on this, we design an end-to-end WSSS framework that can simultaneously train classification and segmentation networks in a single-stage manner. Experimental results on benchmark datasets show that our proposed method significantly outperforms previous WSSS methods, including several multi-stage approaches. The code and model are publicly available at: https://github.com/DCVL-WSSS/ClusterCAM.

arXiv:2403.00932공식 코드 · Papers with Code

Differentially Private Knowledge Distillation via Synthetic Text Generation

Large Language models (LLMs) are achieving state-of-the-art performance in many different downstream tasks. However, the increasing urgency of data privacy puts pressure on practitioners to train LLMs with Differential Privacy (DP) on private data. Concurrently, the exponential growth in parameter size of LLMs necessitates model compression before deployment of LLMs on resource-constrained devices or latency-sensitive applications. Differential privacy and model compression generally must trade off utility loss to achieve their objectives. Moreover, simultaneously applying both schemes can compound the utility degradation. To this end, we propose DistilDP: a novel differentially private knowledge distillation algorithm that exploits synthetic data generated by a differentially private teacher LLM. The knowledge of a teacher LLM is transferred onto the student in two ways: one way from the synthetic data itself -- the hard labels, and the other way by the output distribution of the teacher evaluated on the synthetic data -- the soft labels. Furthermore, if the teacher and student share a similar architectural structure, we can further distill knowledge by aligning the hidden representations between both. Our experimental results demonstrate that DistilDP can substantially improve the utility over existing baselines, at least $9.0$ PPL on the Big Patent dataset, with strong privacy parameters, $\epsilon=2$. These promising results progress privacy-preserving compression of autoregressive LLMs. Our code can be accessed here: https://github.com/james-flemings/dp_compress.

arXiv:2406.14130공식 코드 · Papers with Code

ExVideo: Extending Video Diffusion Models via Parameter-Efficient Post-Tuning

Recently, advancements in video synthesis have attracted significant attention. Video synthesis models such as AnimateDiff and Stable Video Diffusion have demonstrated the practical applicability of diffusion models in creating dynamic visual content. The emergence of SORA has further spotlighted the potential of video generation technologies. Nonetheless, the extension of video lengths has been constrained by the limitations in computational resources. Most existing video synthesis models can only generate short video clips. In this paper, we propose a novel post-tuning methodology for video synthesis models, called ExVideo. This approach is designed to enhance the capability of current video synthesis models, allowing them to produce content over extended temporal durations while incurring lower training expenditures. In particular, we design extension strategies across common temporal model architectures respectively, including 3D convolution, temporal attention, and positional embedding. To evaluate the efficacy of our proposed post-tuning approach, we conduct extension training on the Stable Video Diffusion model. Our approach augments the model's capacity to generate up to $5\times$ its original number of frames, requiring only 1.5k GPU hours of training on a dataset comprising 40k videos. Importantly, the substantial increase in video length doesn't compromise the model's innate generalization capabilities, and the model showcases its advantages in generating videos of diverse styles and resolutions. We will release the source code and the enhanced model publicly.

arXiv:2407.09453

Weight Block Sparsity: Training, Compilation, and AI Engine Accelerators

Nowadays, increasingly larger Deep Neural Networks (DNNs) are being developed, trained, and utilized. These networks require significant computational resources, putting a strain on both advanced and limited devices. Our solution is to implement {\em weight block sparsity}, which is a structured sparsity that is friendly to hardware. By zeroing certain sections of the convolution and fully connected layers parameters of pre-trained DNN models, we can efficiently speed up the DNN's inference process. This results in a smaller memory footprint, faster communication, and fewer operations. Our work presents a vertical system that allows for the training of convolution and matrix multiplication weights to exploit 8x8 block sparsity on a single GPU within a reasonable amount of time. Compilers recognize this sparsity and use it for both data compaction and computation splitting into threads. Blocks like these take full advantage of both spatial and temporal locality, paving the way for fast vector operations and memory reuse. By using this system on a Resnet50 model, we were able to reduce the weight by half with minimal accuracy loss, resulting in a two-times faster inference speed. We will present performance estimates using accurate and complete code generation for AIE2 configuration sets (AMD Versal FPGAs) with Resnet50, Inception V3, and VGG16 to demonstrate the necessary synergy between hardware overlay designs and software stacks for compiling and executing machine learning applications.

arXiv:2409.03525

FrozenSeg: Harmonizing Frozen Foundation Models for Open-Vocabulary Segmentation

Open-vocabulary segmentation poses significant challenges, as it requires segmenting and recognizing objects across an open set of categories in unconstrained environments. Building on the success of powerful vision-language (ViL) foundation models, such as CLIP, recent efforts sought to harness their zero-short capabilities to recognize unseen categories. Despite notable performance improvements, these models still encounter the critical issue of generating precise mask proposals for unseen categories and scenarios, resulting in inferior segmentation performance eventually. To address this challenge, we introduce a novel approach, FrozenSeg, designed to integrate spatial knowledge from a localization foundation model (e.g., SAM) and semantic knowledge extracted from a ViL model (e.g., CLIP), in a synergistic framework. Taking the ViL model's visual encoder as the feature backbone, we inject the space-aware feature into the learnable queries and CLIP features within the transformer decoder. In addition, we devise a mask proposal ensemble strategy for further improving the recall rate and mask quality. To fully exploit pre-trained knowledge while minimizing training overhead, we freeze both foundation models, focusing optimization efforts solely on a lightweight transformer decoder for mask proposal generation-the performance bottleneck. Extensive experiments demonstrate that FrozenSeg advances state-of-the-art results across various segmentation benchmarks, trained exclusively on COCO panoptic data, and tested in a zero-shot manner. Code is available at https://github.com/chenxi52/FrozenSeg.

arXiv:2410.10724

Large Language Models Are Active Critics in NLG Evaluation

The conventional paradigm of using large language models (LLMs) for evaluating natural language generation (NLG) systems typically relies on two key inputs: (1) a clear definition of the NLG task to be evaluated and (2) a list of pre-defined evaluation criteria. This process treats LLMs as ''passive critics,'' strictly following human-defined criteria for evaluation. However, as new NLG tasks emerge, the criteria for assessing text quality can vary greatly. Consequently, these rigid evaluation methods struggle to adapt to diverse NLG tasks without extensive prompt engineering customized for each specific task. To address this limitation, we introduce Active-Critic, a novel LLM-based NLG evaluation protocol that enables LLMs to function as ''active critics.'' Specifically, our protocol comprises two key stages. In the first stage, the LLM is instructed to infer the target NLG task and establish relevant evaluation criteria from the data. Building on this self-inferred information, the second stage dynamically optimizes the prompt to guide the LLM toward more human-aligned scoring decisions, while also generating detailed explanations to justify its evaluations. Experiments across four NLG evaluation tasks show that our approach achieves stronger alignment with human judgments than state-of-the-art evaluation methods. Our comprehensive analysis further highlights the effectiveness and explainability of Active-Critic with only a small amount of labeled data. We will share our code and data on GitHub.

arXiv:2410.21329공식 코드 · Papers with Code

CloudCast -- Total Cloud Cover Nowcasting with Machine Learning

Cloud cover plays a critical role in weather prediction and impacts several sectors, including agriculture, solar power generation, and aviation. Despite advancements in numerical weather prediction (NWP) models, forecasting total cloud cover remains challenging due to the small-scale nature of cloud formation processes. In this study, we introduce CloudCast, a convolutional neural network (CNN) based on the U-Net architecture, designed to predict total cloud cover (TCC) up to five hours ahead. Trained on five years of satellite data, CloudCast significantly outperforms traditional NWP models and optical flow methods. Compared to a reference NWP model, CloudCast achieves a 24% lower mean absolute error and reduces multi-category prediction errors by 46%. The model demonstrates strong performance, particularly in capturing the large-scale structure of cloud cover in the first few forecast hours, though later predictions are subject to blurring and underestimation of cloud formation. An ablation study identified the optimal input features and loss functions, with MAE-based models performing the best. CloudCast has been integrated into the Finnish Meteorological Institute's operational nowcasting system, where it improves cloud cover forecasts used by public and private sector clients. While CloudCast is limited by a relatively short skillful lead time of about three hours, future work aims to extend this through more complex network architectures and higher-resolution data. CloudCast code is available at https://github.com/fmidev/cloudcast.

arXiv:2501.08545

Comprehensive Subjective and Objective Evaluation Method for Text-generated Video

Recent text-to-video (T2V) technology advancements, as demonstrated by models such as Gen3, Pika, and Sora, have significantly broadened its applicability and popularity. This progress has created a growing demand for accurate quality assessment metrics to evaluate the perceptual quality of text-generated videos and optimize video generation models. However, assessing the quality of text-generated videos remains challenging due to the presence of highly complex distortions, such as unnatural actions and phenomena that defy human cognition. To address these challenges, we constructed a large-scale benchmark dataset for \textbf{T}ext-generated \textbf{V}ideo \textbf{eval}uation, \textbf{T2VEval-Bench}, comprising 148 textual words and 1,783 videos generated by 12 models. During the subjective evaluation, we collected five key scores: overall impression, video quality, aesthetic quality, realness, and text-video consistency. For objective evaluation, we developed the \textbf{T2VEval} model, which assesses videos across three branches: quality, authenticity, and consistency. Using an attention-based fusion module, T2VEval effectively integrates features from each branch and predicts scores with the aid of a large oracle model. Additionally, we implemented a progressive training strategy, enabling each branch to learn targeted knowledge while maintaining synergy with the others. Experimental results demonstrate that T2VEval achieves state-of-the-art performance across multiple metrics. The dataset and code will be open-sourced upon completion of the follow-up work.

arXiv:2501.13898공식 코드 · Papers with Code

PointOBB-v3: Expanding Performance Boundaries of Single Point-Supervised Oriented Object Detection

With the growing demand for oriented object detection (OOD), recent studies on point-supervised OOD have attracted significant interest. In this paper, we propose PointOBB-v3, a stronger single point-supervised OOD framework. Compared to existing methods, it generates pseudo rotated boxes without additional priors and incorporates support for the end-to-end paradigm. PointOBB-v3 functions by integrating three unique image views: the original view, a resized view, and a rotated/flipped (rot/flp) view. Based on the views, a scale augmentation module and an angle acquisition module are constructed. In the first module, a Scale-Sensitive Consistency (SSC) loss and a Scale-Sensitive Feature Fusion (SSFF) module are introduced to improve the model's ability to estimate object scale. To achieve precise angle predictions, the second module employs symmetry-based self-supervised learning. Additionally, we introduce an end-to-end version that eliminates the pseudo-label generation process by integrating a detector branch and introduces an Instance-Aware Weighting (IAW) strategy to focus on high-quality predictions. We conducted extensive experiments on the DIOR-R, DOTA-v1.0/v1.5/v2.0, FAIR1M, STAR, and RSAR datasets. Across all these datasets, our method achieves an average improvement in accuracy of 3.56% in comparison to previous state-of-the-art methods. The code will be available at https://github.com/ZpyWHU/PointOBB-v3.

arXiv:2502.14922공식 코드 · Papers with Code

SIFT: Grounding LLM Reasoning in Contexts via Stickers

This paper identifies the misinterpretation of the context can be a significant issue during the reasoning process of large language models, spanning from smaller models like Llama3.2-3B-Instruct to cutting-edge ones like DeepSeek-R1. For example, in the phrase "10 dollars per kilo," LLMs might not recognize that "per" means "for each," leading to calculation errors. We introduce a novel, post-training approach called **Stick to the Facts (SIFT)** to tackle this. SIFT leverages increasing inference-time compute to ground LLM reasoning in contexts. At the core of SIFT lies the *Sticker*, which is generated by the model itself to explicitly emphasize the key information within the context. Given the curated Sticker, SIFT generates two predictions -- one from the original query and one from the query augmented with the Sticker. If they differ, the Sticker is sequentially refined via *forward* optimization (to better align the extracted facts with the query) and *inverse* generation (to conform with the model's inherent tendencies) for more faithful reasoning outcomes. Studies across diverse models (from 3B to 100B+) and benchmarks (e.g., GSM8K, MATH-500) reveal consistent performance improvements. Notably, SIFT improves the pass@1 accuracy of DeepSeek-R1 on AIME2024 from 78.33% to **85.67**%, establishing a new state-of-the-art in the open-source community. The code is available at https://github.com/zhijie-group/SIFT.

arXiv:2504.15362

LongPerceptualThoughts: Distilling System-2 Reasoning for System-1 Perception

Recent reasoning models through test-time scaling have demonstrated that long chain-of-thoughts can unlock substantial performance boosts in hard reasoning tasks such as math and code. However, the benefit of such long thoughts for system-2 reasoning is relatively less explored in other domains such as perceptual tasks where shallower, system-1 reasoning seems sufficient. In this paper, we introduce LongPerceptualThoughts, a new synthetic dataset with 30K long-thought traces for perceptual tasks. The key challenges in synthesizing elaborate reasoning thoughts for perceptual tasks are that off-the-shelf models are not yet equipped with such thinking behavior and that it is not straightforward to build a reliable process verifier for perceptual tasks. Thus, we propose a novel three-stage data synthesis framework that first synthesizes verifiable multiple-choice questions from dense image descriptions, then extracts simple CoTs from VLMs for those verifiable problems, and finally expands those simple thoughts to elaborate long thoughts via frontier reasoning models. In controlled experiments with a strong instruction-tuned 7B model, we demonstrate notable improvements over existing visual reasoning data-generation methods. Our model, trained on the generated dataset, achieves an average +3.4 points improvement over 5 vision-centric benchmarks, including +11.8 points on V$^*$ Bench. Notably, despite being tuned for vision tasks, it also improves performance on the text reasoning benchmark, MMLU-Pro, by +2 points.

arXiv:2505.10937

Reasoning with OmniThought: A Large CoT Dataset with Verbosity and Cognitive Difficulty Annotations

The emergence of large reasoning models (LRMs) has transformed Natural Language Processing by excelling in complex tasks such as mathematical problem-solving and code generation. These models leverage chain-of-thought (CoT) processes, enabling them to emulate human-like reasoning strategies. However, the advancement of LRMs is hindered by the lack of comprehensive CoT datasets. Current resources often fail to provide extensive reasoning problems with coherent CoT processes distilled from multiple teacher models and do not account for multifaceted properties describing the internal characteristics of CoTs. To address these challenges, we introduce OmniThought, a large-scale dataset featuring 2 million CoT processes generated and validated by two powerful LRMs as teacher models. Each CoT process in OmniThought is annotated with novel Reasoning Verbosity (RV) and Cognitive Difficulty (CD) scores, which describe the appropriateness of CoT verbosity and cognitive difficulty level for models to comprehend these reasoning processes. We further establish a self-reliant pipeline to curate this dataset. Extensive experiments using Qwen2.5 models of various sizes demonstrate the positive impact of our proposed scores on LRM training effectiveness. Based on the proposed OmniThought dataset, we further train and release a series of high-performing LRMs, specifically equipped with stronger reasoning abilities and optimal CoT output length and difficulty level. Our contributions significantly enhance the development and training of LRMs for solving complex tasks.