PyTorchKR Papers

최신 논문

총 139,336편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2504.12996공식 코드 · Papers with Code

SHA256 at SemEval-2025 Task 4: Selective Amnesia -- Constrained Unlearning for Large Language Models via Knowledge Isolation

Large language models (LLMs) frequently memorize sensitive information during training, posing risks when deploying publicly accessible models. Current machine unlearning methods struggle to selectively remove specific data associations without degrading overall model capabilities. This paper presents our solution to SemEval-2025 Task 4 on targeted unlearning, which introduces a two-stage methodology that combines causal mediation analysis with layer-specific optimization. Through systematic causal tracing experiments on OLMo architectures (1B and 7B parameters), we identify the critical role of the first few transformer layers (layers 0-5) in storing subject-attribute associations within MLP modules. Building on this insight, we develop a constrained optimization approach that freezes upper layers while applying a novel joint loss function to lower layers-simultaneously maximizing forget set loss via output token cross-entropy penalties and minimizing retain set deviation through adaptive regularization. Our method achieves 2nd place in the 1B model track, demonstrating strong task performance while maintaining 88% of baseline MMLU accuracy. These results establish causal-informed layer optimization as a promising paradigm for efficient, precise unlearning in LLMs, offering a significant step forward in addressing data privacy concerns in AI systems.

arXiv:2504.12915공식 코드 · Papers with Code

ConExion: Concept Extraction with Large Language Models

In this paper, an approach for concept extraction from documents using pre-trained large language models (LLMs) is presented. Compared with conventional methods that extract keyphrases summarizing the important information discussed in a document, our approach tackles a more challenging task of extracting all present concepts related to the specific domain, not just the important ones. Through comprehensive evaluations of two widely used benchmark datasets, we demonstrate that our method improves the F1 score compared to state-of-the-art techniques. Additionally, we explore the potential of using prompts within these models for unsupervised concept extraction. The extracted concepts are intended to support domain coverage evaluation of ontologies and facilitate ontology learning, highlighting the effectiveness of LLMs in concept extraction tasks. Our source code and datasets are publicly available at https://github.com/ISE-FIZKarlsruhe/concept_extraction.

arXiv:2504.12879공식 코드 · Papers with Code

Building Russian Benchmark for Evaluation of Information Retrieval Models

We introduce RusBEIR, a comprehensive benchmark designed for zero-shot evaluation of information retrieval (IR) models in the Russian language. Comprising 17 datasets from various domains, it integrates adapted, translated, and newly created datasets, enabling systematic comparison of lexical and neural models. Our study highlights the importance of preprocessing for lexical models in morphologically rich languages and confirms BM25 as a strong baseline for full-document retrieval. Neural models, such as mE5-large and BGE-M3, demonstrate superior performance on most datasets, but face challenges with long-document retrieval due to input size constraints. RusBEIR offers a unified, open-source framework that promotes research in Russian-language information retrieval.

arXiv:2504.12833공식 코드 · Papers with Code

Image-Editing Specialists: An RLAIF Approach for Diffusion Models

We present a novel approach to training specialized instruction-based image-editing diffusion models, addressing key challenges in structural preservation with input images and semantic alignment with user prompts. We introduce an online reinforcement learning framework that aligns the diffusion model with human preferences without relying on extensive human annotations or curating a large dataset. Our method significantly improves the realism and alignment with instructions in two ways. First, the proposed models achieve precise and structurally coherent modifications in complex scenes while maintaining high fidelity in instruction-irrelevant areas. Second, they capture fine nuances in the desired edit by leveraging a visual prompt, enabling detailed control over visual edits without lengthy textual prompts. This approach simplifies users' efforts to achieve highly specific edits, requiring only 5 reference images depicting a certain concept for training. Experimental results demonstrate that our models can perform intricate edits in complex scenes, after just 10 training steps. Finally, we showcase the versatility of our method by applying it to robotics, where enhancing the visual realism of simulated environments through targeted sim-to-real image edits improves their utility as proxies for real-world settings.

arXiv:2504.12778공식 코드 · Papers with Code

Towards Lossless Token Pruning in Late-Interaction Retrieval Models

Late interaction neural IR models like ColBERT offer a competitive effectiveness-efficiency trade-off across many benchmarks. However, they require a huge memory space to store the contextual representation for all the document tokens. Some works have proposed using either heuristics or statistical-based techniques to prune tokens from each document. This however doesn't guarantee that the removed tokens have no impact on the retrieval score. Our work uses a principled approach to define how to prune tokens without impacting the score between a document and a query. We introduce three regularization losses, that induce a solution with high pruning ratios, as well as two pruning strategies. We study them experimentally (in and out-domain), showing that we can preserve ColBERT's performance while using only 30\% of the tokens.

arXiv:2504.12717공식 코드 · Papers with Code

Post-pre-training for Modality Alignment in Vision-Language Foundation Models

Contrastive language image pre-training (CLIP) is an essential component of building modern vision-language foundation models. While CLIP demonstrates remarkable zero-shot performance on downstream tasks, the multi-modal feature spaces still suffer from a modality gap, which is a gap between image and text feature clusters and limits downstream task performance. Although existing works attempt to address the modality gap by modifying pre-training or fine-tuning, they struggle with heavy training costs with large datasets or degradations of zero-shot performance. This paper presents CLIP-Refine, a post-pre-training method for CLIP models at a phase between pre-training and fine-tuning. CLIP-Refine aims to align the feature space with 1 epoch training on small image-text datasets without zero-shot performance degradations. To this end, we introduce two techniques: random feature alignment (RaFA) and hybrid contrastive-distillation (HyCD). RaFA aligns the image and text features to follow a shared prior distribution by minimizing the distance to random reference vectors sampled from the prior. HyCD updates the model with hybrid soft labels generated by combining ground-truth image-text pair labels and outputs from the pre-trained CLIP model. This contributes to achieving both maintaining the past knowledge and learning new knowledge to align features. Our extensive experiments with multiple classification and retrieval tasks show that CLIP-Refine succeeds in mitigating the modality gap and improving the zero-shot performance.

arXiv:2504.12680공식 코드 · Papers with Code

Embodied-R: Collaborative Framework for Activating Embodied Spatial Reasoning in Foundation Models via Reinforcement Learning

Humans can perceive and reason about spatial relationships from sequential visual observations, such as egocentric video streams. However, how pretrained models acquire such abilities, especially high-level reasoning, remains unclear. This paper introduces Embodied-R, a collaborative framework combining large-scale Vision-Language Models (VLMs) for perception and small-scale Language Models (LMs) for reasoning. Using Reinforcement Learning (RL) with a novel reward system considering think-answer logical consistency, the model achieves slow-thinking capabilities with limited computational resources. After training on only 5k embodied video samples, Embodied-R with a 3B LM matches state-of-the-art multimodal reasoning models (OpenAI-o1, Gemini-2.5-pro) on both in-distribution and out-of-distribution embodied spatial reasoning tasks. Embodied-R also exhibits emergent thinking patterns such as systematic analysis and contextual integration. We further explore research questions including response length, training on VLM, strategies for reward design, and differences in model generalization after SFT (Supervised Fine-Tuning) and RL training.

arXiv:2504.12576공식 코드 · Papers with Code

CM3AE: A Unified RGB Frame and Event-Voxel/-Frame Pre-training Framework

Event cameras have attracted increasing attention in recent years due to their advantages in high dynamic range, high temporal resolution, low power consumption, and low latency. Some researchers have begun exploring pre-training directly on event data. Nevertheless, these efforts often fail to establish strong connections with RGB frames, limiting their applicability in multi-modal fusion scenarios. To address these issues, we propose a novel CM3AE pre-training framework for the RGB-Event perception. This framework accepts multi-modalities/views of data as input, including RGB images, event images, and event voxels, providing robust support for both event-based and RGB-event fusion based downstream tasks. Specifically, we design a multi-modal fusion reconstruction module that reconstructs the original image from fused multi-modal features, explicitly enhancing the model's ability to aggregate cross-modal complementary information. Additionally, we employ a multi-modal contrastive learning strategy to align cross-modal feature representations in a shared latent space, which effectively enhances the model's capability for multi-modal understanding and capturing global dependencies. We construct a large-scale dataset containing 2,535,759 RGB-Event data pairs for the pre-training. Extensive experiments on five downstream tasks fully demonstrated the effectiveness of CM3AE. Source code and pre-trained models will be released on https://github.com/Event-AHU/CM3AE.

arXiv:2504.12562공식 코드 · Papers with Code

ZeroSumEval: Scaling LLM Evaluation with Inter-Model Competition

Evaluating the capabilities of Large Language Models (LLMs) has traditionally relied on static benchmark datasets, human assessments, or model-based evaluations - methods that often suffer from overfitting, high costs, and biases. ZeroSumEval is a novel competition-based evaluation protocol that leverages zero-sum games to assess LLMs with dynamic benchmarks that resist saturation. ZeroSumEval encompasses a diverse suite of games, including security challenges (PyJail), classic games (Chess, Liar's Dice, Poker), knowledge tests (MathQuiz), and persuasion challenges (Gandalf, Debate). These games are designed to evaluate a range of AI capabilities such as strategic reasoning, planning, knowledge application, and creativity. Building upon recent studies that highlight the effectiveness of game-based evaluations for LLMs, ZeroSumEval enhances these approaches by providing a standardized and extensible framework. To demonstrate this, we conduct extensive experiments with >7000 simulations across 7 games and 13 models. Our results show that while frontier models from the GPT and Claude families can play common games and answer questions, they struggle to play games that require creating novel and challenging questions. We also observe that models cannot reliably jailbreak each other and fail generally at tasks requiring creativity. We release our code at https://github.com/facebookresearch/ZeroSumEval.

arXiv:2504.12558공식 코드 · Papers with Code

Benchmarking LLM-based Relevance Judgment Methods

Large Language Models (LLMs) are increasingly deployed in both academic and industry settings to automate the evaluation of information seeking systems, particularly by generating graded relevance judgments. Previous work on LLM-based relevance assessment has primarily focused on replicating graded human relevance judgments through various prompting strategies. However, there has been limited exploration of alternative assessment methods or comprehensive comparative studies. In this paper, we systematically compare multiple LLM-based relevance assessment methods, including binary relevance judgments, graded relevance assessments, pairwise preference-based methods, and two nugget-based evaluation methods~--~document-agnostic and document-dependent. In addition to a traditional comparison based on system rankings using Kendall correlations, we also examine how well LLM judgments align with human preferences, as inferred from relevance grades. We conduct extensive experiments on datasets from three TREC Deep Learning tracks 2019, 2020 and 2021 as well as the ANTIQUE dataset, which focuses on non-factoid open-domain question answering. As part of our data release, we include relevance judgments generated by both an open-source (Llama3.2b) and a commercial (gpt-4o) model. Our goal is to \textit{reproduce} various LLM-based relevance judgment methods to provide a comprehensive comparison. All code, data, and resources are publicly available in our GitHub Repository at https://github.com/Narabzad/llm-relevance-judgement-comparison.

arXiv:2504.12501cs.LG공식 코드 · Papers with Code

Reinforcement Learning from Human Feedback

인간 피드백을 통한 강화학습(Reinforcement Learning from Human Feedback, RLHF)은 최신 머신러닝 시스템의 배포에 있어 중요한 기술적 도구로 자리 잡고 있다. 이 연구는 RLHF의 기원과 이론적 배경을 탐구하며, 경제학, 철학, 최적 제어 등 다양한 분야의 융합을 통해 발전해온 과정을 설명한다. RLHF의 정의와 문제 정의를 명확히 하여, 에이전트가 인간의 피드백을 활용하여 더 나은 결정을 내릴 수 있도록 하는 방법론을 제시한다. 데이터 수집은 RLHF의 성능을 결정짓는 중요한 요소로, 적절한 데이터 수집 방법론의 필요성을 강조한다. 이어서 RLHF의 핵심 최적화 단계인 지시 조정(instruction tuning), 보상 모델 훈련(reward model training), 거부 샘플링(rejection sampling), 강화학습(reinforcement learning), 직접 정렬 알고리즘(direct alignment algorithms)을 상세히 설명한다. 각 단계는 서로 연결되어 있으며, 에이전트가 특정 작업을 수행하도록 지시하고, 보상 신호를 통해 학습하며, 최종적으로 인간의 피드백을 반영하여 성능을 극대화하는 과정을 포함한다. 또한, 합성 데이터(synthetic data)와 평가(evaluation)와 같은 고급 주제에 대한 논의도 포함되어 있으며, 이는 RLHF의 성능 향상과 객관적인 평가 방법론을 제시하는 데 기여한다. 이 연구는 RLHF의 기초부터 고급 주제까지 포괄적으로 다루어, 독자들이 이 분야에 대한 깊은 이해를 높일 수 있도록 돕는다. 향후 연구 방향에 대한 제안도 포함되어 있어, RLHF가 머신러닝의 중요한 발전을 이끌어갈 수 있는 가능성을 제시한다.

arXiv:2504.12463공식 코드 · Papers with Code

Dense Backpropagation Improves Training for Sparse Mixture-of-Experts

Mixture of Experts (MoE) pretraining is more scalable than dense Transformer pretraining, because MoEs learn to route inputs to a sparse set of their feedforward parameters. However, this means that MoEs only receive a sparse backward update, leading to training instability and suboptimal performance. We present a lightweight approximation method that gives the MoE router a dense gradient update while continuing to sparsely activate its parameters. Our method, which we refer to as Default MoE, substitutes missing expert activations with default outputs consisting of an exponential moving average of expert outputs previously seen over the course of training. This allows the router to receive signals from every expert for each token, leading to significant improvements in training performance. Our Default MoE outperforms standard TopK routing in a variety of settings without requiring significant computational overhead. Code: https://github.com/vatsal0/default-moe.

arXiv:2504.12458공식 코드 · Papers with Code

M$^2$FGB: A Min-Max Gradient Boosting Framework for Subgroup Fairness

In recent years, fairness in machine learning has emerged as a critical concern to ensure that developed and deployed predictive models do not have disadvantageous predictions for marginalized groups. It is essential to mitigate discrimination against individuals based on protected attributes such as gender and race. In this work, we consider applying subgroup justice concepts to gradient-boosting machines designed for supervised learning problems. Our approach expanded gradient-boosting methodologies to explore a broader range of objective functions, which combines conventional losses such as the ones from classification and regression and a min-max fairness term. We study relevant theoretical properties of the solution of the min-max optimization problem. The optimization process explored the primal-dual problems at each boosting round. This generic framework can be adapted to diverse fairness concepts. The proposed min-max primal-dual gradient boosting algorithm was theoretically shown to converge under mild conditions and empirically shown to be a powerful and flexible approach to address binary and subgroup fairness.

arXiv:2504.12436공식 코드 · Papers with Code

Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation

Adapting Vision-Language Models (VLMs) to new domains with few labeled samples remains a significant challenge due to severe overfitting and computational constraints. State-of-the-art solutions, such as low-rank reparameterization, mitigate these issues but often struggle with generalization and require extensive hyperparameter tuning. In this paper, a novel Sparse Optimization (SO) framework is proposed. Unlike low-rank approaches that typically constrain updates to a fixed subspace, our SO method leverages high sparsity to dynamically adjust very few parameters. We introduce two key paradigms. First, we advocate for \textit{local sparsity and global density}, which updates a minimal subset of parameters per iteration while maintaining overall model expressiveness. As a second paradigm, we advocate for \textit{local randomness and global importance}, which sparsifies the gradient using random selection while pruning the first moment based on importance. This combination significantly mitigates overfitting and ensures stable adaptation in low-data regimes. Extensive experiments on 11 diverse datasets show that SO achieves state-of-the-art few-shot adaptation performance while reducing memory overhead.

arXiv:2504.12408공식 코드 · Papers with Code

A Human-AI Comparative Analysis of Prompt Sensitivity in LLM-Based Relevance Judgment

Large Language Models (LLMs) are increasingly used to automate relevance judgments for information retrieval (IR) tasks, often demonstrating agreement with human labels that approaches inter-human agreement. To assess the robustness and reliability of LLM-based relevance judgments, we systematically investigate impact of prompt sensitivity on the task. We collected prompts for relevance assessment from 15 human experts and 15 LLMs across three tasks~ -- ~binary, graded, and pairwise~ -- ~yielding 90 prompts in total. After filtering out unusable prompts from three humans and three LLMs, we employed the remaining 72 prompts with three different LLMs as judges to label document/query pairs from two TREC Deep Learning Datasets (2020 and 2021). We compare LLM-generated labels with TREC official human labels using Cohen's $\kappa$ and pairwise agreement measures. In addition to investigating the impact of prompt variations on agreement with human labels, we compare human- and LLM-generated prompts and analyze differences among different LLMs as judges. We also compare human- and LLM-generated prompts with the standard UMBRELA prompt used for relevance assessment by Bing and TREC 2024 Retrieval Augmented Generation (RAG) Track. To support future research in LLM-based evaluation, we release all data and prompts at https://github.com/Narabzad/prompt-sensitivity-relevance-judgements/.

arXiv:2504.12397공식 코드 · Papers with Code

Activated LoRA: Fine-tuned LLMs for Intrinsics

Low-Rank Adaptation (LoRA) has emerged as a highly efficient framework for finetuning the weights of large foundation models, and has become the go-to method for data-driven customization of LLMs. Despite the promise of highly customized behaviors and capabilities, switching between relevant LoRAs in a multiturn setting is highly inefficient, as the key-value (KV) cache of the entire turn history must be recomputed with the LoRA weights before generation can begin. To address this problem, we propose Activated LoRA (aLoRA), which modifies the LoRA framework to only adapt weights for the tokens in the sequence \emph{after} the aLoRA is invoked. This change crucially allows aLoRA to accept the base model's KV cache of the input string, meaning that aLoRA can be instantly activated whenever needed in a chain without recomputing the cache. This enables building what we call \emph{intrinsics}, i.e. highly specialized models invoked to perform well-defined operations on portions of an input chain or conversation that otherwise uses the base model by default. We use aLoRA to train a set of intrinsics models, demonstrating competitive accuracy with standard LoRA while achieving significant inference benefits.

arXiv:2504.11543공식 코드 · Papers with Code

REAL: Benchmarking Autonomous Agents on Deterministic Simulations of Real Websites

We introduce REAL, a benchmark and framework for multi-turn agent evaluations on deterministic simulations of real-world websites. REAL comprises high-fidelity, deterministic replicas of 11 widely-used websites across domains such as e-commerce, travel, communication, and professional networking. We also release a benchmark consisting of 112 practical tasks that mirror everyday complex user interactions requiring both accurate information retrieval and state-changing actions. All interactions occur within this fully controlled setting, eliminating safety risks and enabling robust, reproducible evaluation of agent capability and reliability. Our novel evaluation framework combines programmatic checks of website state for action-based tasks with rubric-guided LLM-based judgments for information retrieval. The framework supports both open-source and proprietary agent systems through a flexible evaluation harness that accommodates black-box commands within browser environments, allowing research labs to test agentic systems without modification. Our empirical results show that frontier language models achieve at most a 41% success rate on REAL, highlighting critical gaps in autonomous web navigation and task completion capabilities. Our framework supports easy integration of new tasks, reproducible evaluation, and scalable post-training data generation, marking a significant step forward in evaluating and advancing agent capabilities.

arXiv:2504.11536공식 코드 · Papers with Code

ReTool: Reinforcement Learning for Strategic Tool Use in LLMs

While reasoning models (e.g., DeepSeek R1) trained with reinforcement learning (RL), excel in textual reasoning, they struggle in scenarios requiring structured problem-solving, such as geometric reasoning, concise computation, or complex equation solving-areas where computational tools like code interpreters (CI) demonstrate distinct advantages. To bridge this gap, we propose ReTool, which enhances long-form reasoning with tool-integrated learning, including two key features: (1) dynamic interleaving of real-time code execution within natural language reasoning processes, and (2) an automated RL paradigm that allows policy rollouts with multi-turn real-time code execution and teaches the model in learning when and how to invoke tools based on outcome feedback. ReTool employs a systematic training framework, beginning with synthetic cold-start data generation to produce code-augmented long-form reasoning traces for fine-tuning base models. Subsequent RL training leverages task outcomes as rewards to iteratively refine the model's tool use strategy, enabling autonomous discovery of optimal tool invocation patterns without human priors. Experiments on the challenging MATH Olympiad benchmark AIME demonstrate ReTool's superiority: Our 32B model achieves 67% accuracy with 400 training steps, outperforming text-based RL baseline (40% accuracy, 1080 steps) in efficiency and performance. Remarkably, ReTool-32B attains 72.5% accuracy in extended settings, surpassing OpenAI's o1-preview by 27.9%. Further analysis reveals emergent behaviors such as code self-correction, signaling an ''aha moment'' in which the model autonomously masters adaptive tool use. These findings highlight the promise of outcome-driven tool integration for advancing complex mathematical reasoning and offer new insights into hybrid neuro-symbolic systems.

arXiv:2502.14891공식 코드 · Papers with Code

CoDiff: Conditional Diffusion Model for Collaborative 3D Object Detection

Collaborative 3D object detection holds significant importance in the field of autonomous driving, as it greatly enhances the perception capabilities of each individual agent by facilitating information exchange among multiple agents. However, in practice, due to pose estimation errors and time delays, the fusion of information across agents often results in feature representations with spatial and temporal noise, leading to detection errors. Diffusion models naturally have the ability to denoise noisy samples to the ideal data, which motivates us to explore the use of diffusion models to address the noise problem between multi-agent systems. In this work, we propose CoDiff, a novel robust collaborative perception framework that leverages the potential of diffusion models to generate more comprehensive and clearer feature representations. To the best of our knowledge, this is the first work to apply diffusion models to multi-agent collaborative perception. Specifically, we project high-dimensional feature map into the latent space of a powerful pre-trained autoencoder. Within this space, individual agent information serves as a condition to guide the diffusion model's sampling. This process denoises coarse feature maps and progressively refines the fused features. Experimental study on both simulated and real-world datasets demonstrates that the proposed framework CoDiff consistently outperforms existing relevant methods in terms of the collaborative object detection performance, and exhibits highly desired robustness when the pose and delay information of agents is with high-level noise. The code is released at https://github.com/HuangZhe885/CoDiff

arXiv:2504.12845공식 코드 · Papers with Code

Can LLMs reason over extended multilingual contexts? Towards long-context evaluation beyond retrieval and haystacks

Existing multilingual long-context benchmarks, often based on the popular needle-in-a-haystack test, primarily evaluate a model's ability to locate specific information buried within irrelevant texts. However, such a retrieval-centric approach is myopic and inherently limited, as successful recall alone does not indicate a model's capacity to reason over extended contexts. Moreover, these benchmarks are susceptible to data leakage, short-circuiting, and risk making the evaluation a priori identifiable. To address these limitations, we introduce MLRBench, a new synthetic benchmark for multilingual long-context reasoning. Unlike existing benchmarks, MLRBench goes beyond surface-level retrieval by including tasks that assess multi-hop inference, aggregation, and epistemic reasoning. Spanning seven languages, MLRBench is designed to be parallel, resistant to leakage, and scalable to arbitrary context lengths. Our extensive experiments with an open-weight large language model (LLM) reveal a pronounced gap between high- and low-resource languages, particularly for tasks requiring the model to aggregate multiple facts or predict the absence of information. We also find that, in multilingual settings, LLMs effectively utilize less than 30% of their claimed context length. Although off-the-shelf Retrieval Augmented Generation helps alleviate this to a certain extent, it does not solve the long-context problem. We open-source MLRBench to enable future research in improved evaluation and training of multilingual LLMs.