Recently, the personalization of Large Language Models (LLMs) to generate content that aligns with individual user preferences has garnered widespread attention. Personalized Retrieval-Augmented Generation (RAG), which retrieves relevant documents from the user's history to reflect their preferences and enhance LLM generation, is one commonly used approach for personalization. However, existing personalized RAG methods do not consider that the histories of similar users can also assist in personalized generation for the current user, meaning that collaborative information between users can also benefit personalized generation. Inspired by the application of collaborative filtering in recommender systems, we propose a method called CFRAG, which adapts Collaborative Filtering to RAG for personalized text generation. However, this presents two challenges: (1)~how to incorporate collaborative information without explicit user similarity labels? (2)~how to retrieve documents that support personalized LLM generation? For Challenge 1, we use contrastive learning to train user embeddings to retrieve similar users and introduce collaborative information. For Challenge 2, we design a personalized retriever and reranker to retrieve the top-$k$ documents from these users' histories. We take into account the user's preference during retrieval and reranking. Then we leverage feedback from the LLM to fine-tune the personalized retriever and reranker, enabling them to retrieve documents that meet the personalized generation needs of the LLM. Experimental results on the Language Model Personalization (LaMP) benchmark validate the effectiveness of CFRAG. Further analysis confirms the importance of incorporating collaborative information.
Charts are ubiquitous, as people often use them to analyze data, answer questions, and discover critical insights. However, performing complex analytical tasks with charts requires significant perceptual and cognitive effort. Chart Question Answering (CQA) systems automate this process by enabling models to interpret and reason with visual representations of data. However, existing benchmarks like ChartQA lack real-world diversity and have recently shown performance saturation with modern large vision-language models (LVLMs). To address these limitations, we introduce ChartQAPro, a new benchmark that includes 1,341 charts from 157 diverse sources, spanning various chart types, including infographics and dashboards, and featuring 1,948 questions in various types, such as multiple-choice, conversational, hypothetical, and unanswerable questions, to better reflect real-world challenges. Our evaluations with 21 models show a substantial performance drop for LVLMs on ChartQAPro; e.g., Claude Sonnet 3.5 scores 90.5% on ChartQA but only 55.81% on ChartQAPro, underscoring the complexity of chart reasoning. We complement our findings with detailed error analyses and ablation studies, identifying key challenges and opportunities for advancing LVLMs in chart understanding and reasoning. We release ChartQAPro at https://github.com/vis-nlp/ChartQAPro.
Despite significant advancements in large language models (LLMs), a major drawback of reasoning models is their enormous token usage, which increases computational cost, resource requirements, and response time. In this work, we revisit the core principles of reinforcement learning (RL) and, through mathematical analysis, demonstrate that the tendency to generate lengthy responses arises inherently from RL-based optimization during training. This finding questions the prevailing assumption that longer responses inherently improve reasoning accuracy. Instead, we uncover a natural correlation between conciseness and accuracy that has been largely overlooked. We show that introducing a secondary phase of RL training, using a very small set of problems, can significantly reduce chains of thought while maintaining or even enhancing accuracy. Additionally, we demonstrate that, while GRPO shares some interesting properties of PPO, it suffers from collapse modes, which limit its reliability for concise reasoning. Finally, we validate our conclusions through extensive experimental results.
Image restoration has witnessed significant advancements with the development of deep learning models. Although Transformer architectures have progressed considerably in recent years, challenges remain, particularly the limited receptive field in window-based self-attention. In this work, we propose DSwinIR, a Deformable Sliding window Transformer for Image Restoration. DSwinIR introduces a novel deformable sliding window self-attention that adaptively adjusts receptive fields based on image content, enabling the attention mechanism to focus on important regions and enhance feature extraction aligned with salient features. Additionally, we introduce a central ensemble pattern to reduce the inclusion of irrelevant content within attention windows. In this way, the proposed DSwinIR model integrates the deformable sliding window Transformer and central ensemble pattern to amplify the strengths of both CNNs and Transformers while mitigating their limitations. Extensive experiments on various image restoration tasks demonstrate that DSwinIR achieves state-of-the-art performance. For example, in image deraining, compared to DRSformer on the SPA dataset, DSwinIR achieves a 0.66 dB PSNR improvement. In all-in-one image restoration, compared to PromptIR, DSwinIR achieves over a 0.66 dB and 1.04 dB improvement on three-task and five-task settings, respectively. Pretrained models and code are available at our project https://github.com/Aitical/DSwinIR.
Data augmentation has become a promising method of mitigating data sparsity in sequential recommendation. Existing methods generate new yet effective data during model training to improve performance. However, deploying them requires retraining, architecture modification, or introducing additional learnable parameters. The above steps are time-consuming and costly for well-trained models, especially when the model scale becomes large. In this work, we explore the test-time augmentation (TTA) for sequential recommendation, which augments the inputs during the model inference and then aggregates the model's predictions for augmented data to improve final accuracy. It avoids significant time and cost overhead from loss calculation and backward propagation. We first experimentally disclose the potential of existing augmentation operators for TTA and find that the Mask and Substitute consistently achieve better performance. Further analysis reveals that these two operators are effective because they retain the original sequential pattern while adding appropriate perturbations. Meanwhile, we argue that these two operators still face time-consuming item selection or interference information from mask tokens. Based on the analysis and limitations, we present TNoise and TMask. The former injects uniform noise into the original representation, avoiding the computational overhead of item selection. The latter blocks mask token from participating in model calculations or directly removes interactions that should have been replaced with mask tokens. Comprehensive experiments demonstrate the effectiveness, efficiency, and generalizability of our method. We provide an anonymous implementation at https://github.com/KingGugu/TTA4SR.
Graphical User Interface (GUI) grounding models are crucial for enabling intelligent agents to understand and interact with complex visual interfaces. However, these models face significant robustness challenges in real-world scenarios due to natural noise and adversarial perturbations, and their robustness remains underexplored. In this study, we systematically evaluate the robustness of state-of-the-art GUI grounding models, such as UGround, under three conditions: natural noise, untargeted adversarial attacks, and targeted adversarial attacks. Our experiments, which were conducted across a wide range of GUI environments, including mobile, desktop, and web interfaces, have clearly demonstrated that GUI grounding models exhibit a high degree of sensitivity to adversarial perturbations and low-resolution conditions. These findings provide valuable insights into the vulnerabilities of GUI grounding models and establish a strong benchmark for future research aimed at enhancing their robustness in practical applications. Our code is available at https://github.com/ZZZhr-1/Robust_GUI_Grounding.
The increasing size of the Key-Value (KV) cache during the Large Language Models long-context inference is the main obstacle for its balance between the deployment cost and task accuracy. To reduce the KV cache size in such scenarios, most previous efforts leveraged on the attention weight to evict non-critical cache tokens. But there is a trade-off in those methods, they usually require major modifiation of the inference infrastructure and significant computation overhead. Base on the fact that the Large Lanuage models are autoregresssive models, we propose {\it LagKV}, a KV allocation strategy only relying on straight forward comparison among KV themself. It is a totally attention free method which offers easy integration to the main stream inference platform and comparable performance comparing to other complicated KV compression methods. Results on LongBench and PasskeyRetrieval show that, our approach achieves nearly zero loss when the ratio is $2\times$ and $\approx 90\%$ of the original model performance for $8\times$. Especially in the 64-digit passkey retrieval task, our mehod outperforms the attention weight based method $H_2O$ over $60\%$ with same compression ratios. Our code is available at \url{https://github.com/AI-Lab-China-Merchants-Bank/LagKV}.
Contrastive Learning (CL) has recently emerged as a powerful technique in recommendation systems, particularly for its capability to harness self-supervised signals from perturbed views to mitigate the persistent challenge of data sparsity. The process of constructing perturbed views of the user-item bipartite graph and performing contrastive learning between perturbed views in a graph convolutional network (GCN) is called graph contrastive learning (GCL), which aims to enhance the robustness of representation learning. Although existing GCL-based models are effective, the weight assignment method for perturbed views has not been fully explored. A critical problem in existing GCL-based models is the irrational allocation of feature attention. This problem limits the model's ability to effectively leverage crucial features, resulting in suboptimal performance. To address this, we propose a Weighted Graph Contrastive Learning framework (WeightedGCL). Specifically, WeightedGCL applies a robust perturbation strategy, which perturbs only the view of the final GCN layer. In addition, WeightedGCL incorporates a squeeze and excitation network (SENet) to dynamically weight the features of the perturbed views. Our WeightedGCL strengthens the model's focus on crucial features and reduces the impact of less relevant information. Extensive experiments on widely used datasets demonstrate that our WeightedGCL achieves significant accuracy improvements compared to competitive baselines.
Multi-modal Large Language Models (MLLMs) integrate visual and linguistic reasoning to address complex tasks such as image captioning and visual question answering. While MLLMs demonstrate remarkable versatility, MLLMs appears limited performance on special applications. But tuning MLLMs for downstream tasks encounters two key challenges: Task-Expert Specialization, where distribution shifts between pre-training and target datasets constrain target performance, and Open-World Stabilization, where catastrophic forgetting erases the model general knowledge. In this work, we systematically review recent advancements in MLLM tuning methodologies, classifying them into three paradigms: (I) Selective Tuning, (II) Additive Tuning, and (III) Reparameterization Tuning. Furthermore, we benchmark these tuning strategies across popular MLLM architectures and diverse downstream tasks to establish standardized evaluation analysis and systematic tuning principles. Finally, we highlight several open challenges in this domain and propose future research directions. To facilitate ongoing progress in this rapidly evolving field, we provide a public repository that continuously tracks developments: https://github.com/WenkeHuang/Awesome-MLLM-Tuning.
In this paper, we systematically evaluate the robustness of multi-exit language models against adversarial slowdown. To audit their robustness, we design a slowdown attack that generates natural adversarial text bypassing early-exit points. We use the resulting WAFFLE attack as a vehicle to conduct a comprehensive evaluation of three multi-exit mechanisms with the GLUE benchmark against adversarial slowdown. We then show our attack significantly reduces the computational savings provided by the three methods in both white-box and black-box settings. The more complex a mechanism is, the more vulnerable it is to adversarial slowdown. We also perform a linguistic analysis of the perturbed text inputs, identifying common perturbation patterns that our attack generates, and comparing them with standard adversarial text attacks. Moreover, we show that adversarial training is ineffective in defeating our slowdown attack, but input sanitization with a conversational model, e.g., ChatGPT, can remove perturbations effectively. This result suggests that future work is needed for developing efficient yet robust multi-exit models. Our code is available at: https://github.com/ztcoalson/WAFFLE
Mixture-of-Experts (MoE) Large Language Models (LLMs) suffer from severely sub-optimal expert pathways-our study reveals that naive expert selection learned from pretraining leaves a surprising 10-20% accuracy gap for improvement. Motivated by this observation, we develop a novel class of test-time optimization methods to re-weight or "re-mixing" the experts in different layers jointly for each test sample. Since the test sample's ground truth is unknown, we propose to optimize a surrogate objective defined by the sample's "successful neighbors" from a reference set of samples. We introduce three surrogates and algorithms based on mode-finding, kernel regression, and the average loss of similar reference samples/tasks. To reduce the cost of optimizing whole pathways, we apply our algorithms merely to the core experts' mixing weights in critical layers, which enjoy similar performance but save significant computation. This leads to "Critical-Layer, Core-Expert, Collaborative Pathway Optimization (C3PO)". We apply C3PO to two recent MoE LLMs and examine it on six widely-used benchmarks. It consistently improves the base model by 7-15% in accuracy and outperforms widely used test-time learning baselines, e.g., in-context learning and prompt/prefix tuning, by a large margin. Moreover, C3PO enables MoE LLMs with 1-3B active parameters to outperform LLMs of 7-9B parameters, hence improving MoE's advantages on efficiency. Our thorough ablation study further sheds novel insights on achieving test-time improvement on MoE.
Despite their impressive performance on complex tasks, current language models (LMs) typically operate in a vacuum: Each input query is processed separately, without retaining insights from previous attempts. Here, we present Dynamic Cheatsheet (DC), a lightweight framework that endows a black-box LM with a persistent, evolving memory. Rather than repeatedly re-discovering or re-committing the same solutions and mistakes, DC enables models to store and reuse accumulated strategies, code snippets, and general problem-solving insights at inference time. This test-time learning enhances performance substantially across a range of tasks without needing explicit ground-truth labels or human feedback. Leveraging DC, Claude 3.5 Sonnet's accuracy more than doubled on AIME math exams once it began retaining algebraic insights across questions. Similarly, GPT-4o's success rate on Game of 24 increased from 10% to 99% after the model discovered and reused a Python-based solution. In tasks prone to arithmetic mistakes, such as balancing equations, DC enabled GPT-4o and Claude to reach near-perfect accuracy by recalling previously validated code, whereas their baselines stagnated around 50%. Beyond arithmetic challenges, DC yields notable accuracy gains on knowledge-demanding tasks. Claude achieved a 9% improvement in GPQA-Diamond and an 8% boost on MMLU-Pro problems. Crucially, DC's memory is self-curated, focusing on concise, transferable snippets rather than entire transcript. Unlike finetuning or static retrieval methods, DC adapts LMs' problem-solving skills on the fly, without modifying their underlying parameters. Overall, our findings present DC as a promising approach for augmenting LMs with persistent memory, bridging the divide between isolated inference events and the cumulative, experience-driven learning characteristic of human cognition.
Prior work has primarily formulated CA-HAR as a multi-label classification problem, where model inputs are time-series sensor data and target labels are binary encodings representing whether a given activity or context occurs. These CA-HAR methods either predicted each label independently or manually imposed relationships using graphs. However, both strategies often neglect an essential aspect: activity labels have rich semantic relationships. For instance, walking, jogging, and running activities share similar movement patterns but differ in pace and intensity, indicating that they are semantically related. Consequently, prior CA-HAR methods often struggled to accurately capture these inherent and nuanced relationships, particularly on datasets with noisy labels typically used for CA-HAR or situations where the ideal sensor type is unavailable (e.g., recognizing speech without audio sensors). To address this limitation, we propose SEAL, which leverage LMs to encode CA-HAR activity labels to capture semantic relationships. LMs generate vector embeddings that preserve rich semantic information from natural language. Our SEAL approach encodes input-time series sensor data from smart devices and their associated activity and context labels (text) as vector embeddings. During training, SEAL aligns the sensor data representations with their corresponding activity/context label embeddings in a shared embedding space. At inference time, SEAL performs a similarity search, returning the CA-HAR label with the embedding representation closest to the input data. Although LMs have been widely explored in other domains, surprisingly, their potential in CA-HAR has been underexplored, making our approach a novel contribution to the field. Our research opens up new possibilities for integrating more advanced LMs into CA-HAR tasks.
Reinforcement learning (RL)-based fine-tuning has become a crucial step in post-training language models for advanced mathematical reasoning and coding. Following the success of frontier reasoning models, recent work has demonstrated that RL fine-tuning consistently improves performance, even in smaller-scale models; however, the underlying mechanisms driving these improvements are not well-understood. Understanding the effects of RL fine-tuning requires disentangling its interaction with pretraining data composition, hyperparameters, and model scale, but such problems are exacerbated by the lack of transparency regarding the training data used in many existing models. In this work, we present a systematic end-to-end study of RL fine-tuning for mathematical reasoning by training models entirely from scratch on different mixtures of fully open datasets. We investigate the effects of various RL fine-tuning algorithms (PPO, GRPO, and Expert Iteration) across models of different scales. Our study reveals that RL algorithms consistently converge towards a dominant output distribution, amplifying patterns in the pretraining data. We also find that models of different scales trained on the same data mixture will converge to distinct output distributions, suggesting that there are scale-dependent biases in model generalization. Moreover, we find that RL post-training on simpler questions can lead to performance gains on harder ones, indicating that certain reasoning capabilities generalize across tasks. Our findings show that small-scale proxies in controlled settings can elicit interesting insights regarding the role of RL in shaping language model behavior.
Recent advances in inference-time compute have significantly improved performance on complex tasks by generating long chains of thought (CoTs) using Large Reasoning Models (LRMs). However, this improved accuracy comes at the cost of high inference latency due to the length of generated reasoning sequences and the autoregressive nature of decoding. Our key insight in tackling these overheads is that LRM inference, and the reasoning that it embeds, is highly tolerant of approximations: complex tasks are typically broken down into simpler steps, each of which brings utility based on the semantic insight it provides for downstream steps rather than the exact tokens it generates. Accordingly, we introduce SpecReason, a system that automatically accelerates LRM inference by using a lightweight model to (speculatively) carry out simpler intermediate reasoning steps and reserving the costly base model only to assess (and potentially correct) the speculated outputs. Importantly, SpecReason's focus on exploiting the semantic flexibility of thinking tokens in preserving final-answer accuracy is complementary to prior speculation techniques, most notably speculative decoding, which demands token-level equivalence at each step. Across a variety of reasoning benchmarks, SpecReason achieves 1.5-2.5$\times$ speedup over vanilla LRM inference while improving accuracy by 1.0-9.9\%. Compared to speculative decoding without SpecReason, their combination yields an additional 19.4-44.2\% latency reduction. We open-source SpecReason at https://github.com/ruipeterpan/specreason.
Practically all large language models have been pre-trained on data that is subject to global uncertainty related to copyright infringement and breach of contract. This creates potential risk for users and developers due to this uncertain legal status. The KL3M Data Project directly confronts this critical issue by introducing the largest comprehensive training data pipeline that minimizes risks related to copyright or breach of contract. The foundation of this project is a corpus of over 132 million documents and trillions of tokens spanning 16 different sources that have been verified to meet the strict copyright and licensing protocol detailed herein. We are releasing the entire pipeline, including 1) the source code to acquire and process these documents, 2) the original document formats with associated provenance and metadata, 3) extracted content in a standardized format, 4) pre-tokenized representations of the documents, and 5) various mid- and post-train resources such as question-answer, summarization, conversion, drafting, classification, prediction, and conversational data. All of these resources are freely available to the public on S3, Hugging Face, and GitHub under CC-BY terms. We are committed to continuing this project in furtherance of a more ethical, legal, and sustainable approach to the development and use of AI models.
Retrieval Augmented Generation (RAG) has emerged as a standard paradigm for enhancing the factual accuracy and contextual relevance of Large Language Models (LLMs) by integrating retrieval mechanisms. However, existing evaluation frameworks fail to provide a holistic black-box approach to assessing RAG systems, especially in real-world deployment scenarios. To address this gap, we introduce SCARF (System for Comprehensive Assessment of RAG Frameworks), a modular and flexible evaluation framework designed to benchmark deployed RAG applications systematically. SCARF provides an end-to-end, black-box evaluation methodology, enabling a limited-effort comparison across diverse RAG frameworks. Our framework supports multiple deployment configurations and facilitates automated testing across vector databases and LLM serving strategies, producing a detailed performance report. Moreover, SCARF integrates practical considerations such as response coherence, providing a scalable and adaptable solution for researchers and industry professionals evaluating RAG applications. Using the REST APIs interface, we demonstrate how SCARF can be applied to real-world scenarios, showcasing its flexibility in assessing different RAG frameworks and configurations. SCARF is available at GitHub repository.
This paper studies the limitations of (retrieval-augmented) large language models (LLMs) in generating diverse and comprehensive responses, and introduces the Plan-and-Refine (P&R) framework based on a two phase system design. In the global exploration phase, P&R generates a diverse set of plans for the given input, where each plan consists of a list of diverse query aspects with corresponding additional descriptions. This phase is followed by a local exploitation phase that generates a response proposal for the input query conditioned on each plan and iteratively refines the proposal for improving the proposal quality. Finally, a reward model is employed to select the proposal with the highest factuality and coverage. We conduct our experiments based on the ICAT evaluation methodology--a recent approach for answer factuality and comprehensiveness evaluation. Experiments on the two diverse information seeking benchmarks adopted from non-factoid question answering and TREC search result diversification tasks demonstrate that P&R significantly outperforms baselines, achieving up to a 13.1% improvement on the ANTIQUE dataset and a 15.41% improvement on the TREC dataset. Furthermore, a smaller scale user study confirms the substantial efficacy of the P&R framework.
This paper introduces NorEval, a new and comprehensive evaluation suite for large-scale standardized benchmarking of Norwegian generative language models (LMs). NorEval consists of 24 high-quality human-created datasets -- of which five are created from scratch. In contrast to existing benchmarks for Norwegian, NorEval covers a broad spectrum of task categories targeting Norwegian language understanding and generation, establishes human baselines, and focuses on both of the official written standards of the Norwegian language: Bokm{\aa}l and Nynorsk. All our datasets and a collection of over 100 human-written prompts are integrated into LM Evaluation Harness, ensuring flexible and reproducible evaluation. We describe the NorEval design and present the results of benchmarking 19 open-source pre-trained and instruction-tuned LMs for Norwegian in various scenarios. Our benchmark, evaluation framework, and annotation materials are publicly available.
Large Language Models (LLMs) demonstrate impressive capabilities in natural language processing but suffer from inaccuracies and logical inconsistencies known as hallucinations. This compromises their reliability, especially in domains requiring factual accuracy. We propose a neuro-symbolic approach integrating symbolic ontological reasoning and machine learning methods to enhance the consistency and reliability of LLM outputs. Our workflow utilizes OWL ontologies, a symbolic reasoner (e.g., HermiT) for consistency checking, and a lightweight machine learning model (logistic regression) for mapping natural language statements into logical forms compatible with the ontology. When inconsistencies between LLM outputs and the ontology are detected, the system generates explanatory feedback to guide the LLM towards a corrected, logically coherent response in an iterative refinement loop. We present a working Python prototype demonstrating this pipeline. Experimental results in a defined domain suggest significant improvements in semantic coherence and factual accuracy of LLM outputs, showcasing the potential of combining LLM fluency with the rigor of formal semantics.