id
stringlengths
9
16
title
stringlengths
4
278
categories
stringlengths
5
104
abstract
stringlengths
6
4.09k
2502.11312
AI Generations: From AI 1.0 to AI 4.0
cs.AI
This paper proposes that Artificial Intelligence (AI) progresses through several overlapping generations: AI 1.0 (Information AI), AI 2.0 (Agentic AI), AI 3.0 (Physical AI), and now a speculative AI 4.0 (Conscious AI). Each of these AI generations is driven by shifting priorities among algorithms, computing power, and data. AI 1.0 ushered in breakthroughs in pattern recognition and information processing, fueling advances in computer vision, natural language processing, and recommendation systems. AI 2.0 built on these foundations through real-time decision-making in digital environments, leveraging reinforcement learning and adaptive planning for agentic AI applications. AI 3.0 extended intelligence into physical contexts, integrating robotics, autonomous vehicles, and sensor-fused control systems to act in uncertain real-world settings. Building on these developments, AI 4.0 puts forward the bold vision of self-directed AI capable of setting its own goals, orchestrating complex training regimens, and possibly exhibiting elements of machine consciousness. This paper traces the historical foundations of AI across roughly seventy years, mapping how changes in technological bottlenecks from algorithmic innovation to high-performance computing to specialized data, have spurred each generational leap. It further highlights the ongoing synergies among AI 1.0, 2.0, 3.0, and 4.0, and explores the profound ethical, regulatory, and philosophical challenges that arise when artificial systems approach (or aspire to) human-like autonomy. Ultimately, understanding these evolutions and their interdependencies is pivotal for guiding future research, crafting responsible governance, and ensuring that AI transformative potential benefits society as a whole.
2502.11323
A statistical theory of overfitting for imbalanced classification
math.ST cs.LG stat.ML stat.TH
Classification with imbalanced data is a common challenge in data analysis, where certain classes (minority classes) account for a small fraction of the training data compared with other classes (majority classes). Classical statistical theory based on large-sample asymptotics and finite-sample corrections is often ineffective for high-dimensional data, leaving many overfitting phenomena in empirical machine learning unexplained. In this paper, we develop a statistical theory for high-dimensional imbalanced classification by investigating support vector machines and logistic regression. We find that dimensionality induces truncation or skewing effects on the logit distribution, which we characterize via a variational problem under high-dimensional asymptotics. In particular, for linearly separable data generated from a two-component Gaussian mixture model, the logits from each class follow a normal distribution $\mathsf{N}(0,1)$ on the testing set, but asymptotically follow a rectified normal distribution $\max\{\kappa, \mathsf{N}(0,1)\}$ on the training set -- which is a pervasive phenomenon we verified on tabular data, image data, and text data. This phenomenon explains why the minority class is more severely affected by overfitting. Further, we show that margin rebalancing, which incorporates class sizes into the loss function, is crucial for mitigating the accuracy drop for the minority class. Our theory also provides insights into the effects of overfitting on calibration and other uncertain quantification measures.
2502.11324
Robust High-Dimensional Mean Estimation With Low Data Size, an Empirical Study
stat.ML cs.LG
Robust statistics aims to compute quantities to represent data where a fraction of it may be arbitrarily corrupted. The most essential statistic is the mean, and in recent years, there has been a flurry of theoretical advancement for efficiently estimating the mean in high dimensions on corrupted data. While several algorithms have been proposed that achieve near-optimal error, they all rely on large data size requirements as a function of dimension. In this paper, we perform an extensive experimentation over various mean estimation techniques where data size might not meet this requirement due to the high-dimensional setting.
2502.11329
Differentially private fine-tuned NF-Net to predict GI cancer type
cs.CV
Based on global genomic status, the cancer tumor is classified as Microsatellite Instable (MSI) and Microsatellite Stable (MSS). Immunotherapy is used to diagnose MSI, whereas radiation and chemotherapy are used for MSS. Therefore, it is significant to classify a gastro-intestinal (GI) cancer tumor into MSI vs. MSS to provide appropriate treatment. The existing literature showed that deep learning could directly predict the class of GI cancer tumors from histological images. However, deep learning (DL) models are susceptible to various threats, including membership inference attacks, model extraction attacks, etc. These attacks render the use of DL models impractical in real-world scenarios. To make the DL models useful and maintain privacy, we integrate differential privacy (DP) with DL. In particular, this paper aims to predict the state of GI cancer while preserving the privacy of sensitive data. We fine-tuned the Normalizer Free Net (NF-Net) model. We obtained an accuracy of 88.98\% without DP to predict (GI) cancer status. When we fine-tuned the NF-Net using DP-AdamW and adaptive DP-AdamW, we got accuracies of 74.58% and 76.48%, respectively. Moreover, we investigate the Weighted Random Sampler (WRS) and Class weighting (CW) to solve the data imbalance. We also evaluated and analyzed the DP algorithms in different settings.
2502.11330
System Message Generation for User Preferences using Open-Source Models
cs.CL cs.AI
System messages play a crucial role in interactions with large language models (LLMs), often serving as prompts to initiate conversations. Through system messages, users can assign specific roles, perform intended tasks, incorporate background information, specify various output formats and communication styles. Despite such versatility, publicly available data are often lack system messages and subject to strict license constraints in the industry field. Manual labeling of publicly available data with system messages that align with user instructions demands significant resources. In view of such challenges, our work introduces SysGen, a pipeline for generating system messages with better aligned assistant responses from the supervised fine-tuning dataset without system messages. Training on SysGen data has demonstrated substantial improvements in the alignment of model responses with system messages and user instructions, as demonstrated across various open-source models on the Multifacet benchmark, while maintaining minimal impact on other unseen benchmarks such as Open LLM Leaderboard 2. Our qualitative analysis highlights the importance of diverse system messages to ensure better adaptability across different contexts.
2502.11331
Transfer Learning of CATE with Kernel Ridge Regression
stat.ME cs.LG stat.ML
The proliferation of data has sparked significant interest in leveraging findings from one study to estimate treatment effects in a different target population without direct outcome observations. However, the transfer learning process is frequently hindered by substantial covariate shift and limited overlap between (i) the source and target populations, as well as (ii) the treatment and control groups within the source. We propose a novel method for overlap-adaptive transfer learning of conditional average treatment effect (CATE) using kernel ridge regression (KRR). Our approach involves partitioning the labeled source data into two subsets. The first one is used to train candidate CATE models based on regression adjustment and pseudo-outcomes. An optimal model is then selected using the second subset and unlabeled target data, employing another pseudo-outcome-based strategy. We provide a theoretical justification for our method through sharp non-asymptotic MSE bounds, highlighting its adaptivity to both weak overlaps and the complexity of CATE function. Extensive numerical studies confirm that our method achieves superior finite-sample efficiency and adaptability. We conclude by demonstrating the effectiveness of our approach using a 401(k) eligibility dataset.
2502.11333
Inverse Flow and Consistency Models
cs.LG cs.AI
Inverse generation problems, such as denoising without ground truth observations, is a critical challenge in many scientific inquiries and real-world applications. While recent advances in generative models like diffusion models, conditional flow matching, and consistency models achieved impressive results by casting generation as denoising problems, they cannot be directly used for inverse generation without access to clean data. Here we introduce Inverse Flow (IF), a novel framework that enables using these generative models for inverse generation problems including denoising without ground truth. Inverse Flow can be flexibly applied to nearly any continuous noise distribution and allows complex dependencies. We propose two algorithms for learning Inverse Flows, Inverse Flow Matching (IFM) and Inverse Consistency Model (ICM). Notably, to derive the computationally efficient, simulation-free inverse consistency model objective, we generalized consistency training to any forward diffusion processes or conditional flows, which have applications beyond denoising. We demonstrate the effectiveness of IF on synthetic and real datasets, outperforming prior approaches while enabling noise distributions that previous methods cannot support. Finally, we showcase applications of our techniques to fluorescence microscopy and single-cell genomics data, highlighting IF's utility in scientific problems. Overall, this work expands the applications of powerful generative models to inversion generation problems.
2502.11335
Personalized Ranking on Cascading Behavior Graphs for Accurate Multi-Behavior Recommendation
cs.IR
Multi-behavior recommendation predicts items a user may purchase by analyzing diverse behaviors like viewing, adding to a cart, and purchasing. Existing methods fall into two categories: representation learning and graph ranking. Representation learning generates user and item embeddings to capture latent interaction patterns, leveraging multi-behavior properties for better generalization. However, these methods often suffer from over-smoothing and bias toward frequent interactions, limiting their expressiveness. Graph ranking methods, on the other hand, directly compute personalized ranking scores, capturing user preferences more effectively. Despite their potential, graph ranking approaches have been primarily explored in single-behavior settings and remain underutilized for multi-behavior recommendation. In this paper, we propose CascadingRank, a novel graph ranking method for multi-behavior recommendation. It models the natural sequence of user behaviors (e.g., viewing, adding to cart, and purchasing) through a cascading behavior graph. An iterative algorithm computes ranking scores, ensuring smoothness, query fitting, and cascading alignment. Experiments on three real-world datasets demonstrate that CascadingRank outperforms state-of-the-art methods, with up to 9.56% and 7.16% improvements in HR@10 and NDCG@10, respectively. Furthermore, we provide theoretical analysis highlighting its effectiveness, convergence, and scalability, showcasing the advantages of graph ranking in multi-behavior recommendation.
2502.11336
ExaGPT: Example-Based Machine-Generated Text Detection for Human Interpretability
cs.CL
Detecting texts generated by Large Language Models (LLMs) could cause grave mistakes due to incorrect decisions, such as undermining student's academic dignity. LLM text detection thus needs to ensure the interpretability of the decision, which can help users judge how reliably correct its prediction is. When humans verify whether a text is human-written or LLM-generated, they intuitively investigate with which of them it shares more similar spans. However, existing interpretable detectors are not aligned with the human decision-making process and fail to offer evidence that users easily understand. To bridge this gap, we introduce ExaGPT, an interpretable detection approach grounded in the human decision-making process for verifying the origin of a text. ExaGPT identifies a text by checking whether it shares more similar spans with human-written vs. with LLM-generated texts from a datastore. This approach can provide similar span examples that contribute to the decision for each span in the text as evidence. Our human evaluation demonstrates that providing similar span examples contributes more effectively to judging the correctness of the decision than existing interpretable methods. Moreover, extensive experiments in four domains and three generators show that ExaGPT massively outperforms prior powerful detectors by up to +40.9 points of accuracy at a false positive rate of 1%.
2502.11337
A Comparison of Human and Machine Learning Errors in Face Recognition
cs.HC cs.CV cs.CY
Machine learning applications in high-stakes scenarios should always operate under human oversight. Developing an optimal combination of human and machine intelligence requires an understanding of their complementarities, particularly regarding the similarities and differences in the way they make mistakes. We perform extensive experiments in the area of face recognition and compare two automated face recognition systems against human annotators through a demographically balanced user study. Our research uncovers important ways in which machine learning errors and human errors differ from each other, and suggests potential strategies in which human-machine collaboration can improve accuracy in face recognition.
2502.11338
WRT-SAM: Foundation Model-Driven Segmentation for Generalized Weld Radiographic Testing
cs.CV
Radiographic testing is a fundamental non-destructive evaluation technique for identifying weld defects and assessing quality in industrial applications due to its high-resolution imaging capabilities. Over the past decade, deep learning techniques have significantly advanced weld defect identification in radiographic images. However, conventional approaches, which rely on training small-scale, task-specific models on single-scenario datasets, exhibit poor cross-scenario generalization. Recently, the Segment Anything Model (SAM), a pre-trained visual foundation model trained on large-scale datasets, has demonstrated exceptional zero-shot generalization capabilities. Fine-tuning SAM with limited domain-specific data has yielded promising results in fields such as medical image segmentation and anomaly detection. To the best of our knowledge, this work is the first to introduce SAM-based segmentation for general weld radiographic testing images. We propose WRT-SAM, a novel weld radiographic defect segmentation model that leverages SAM through an adapter-based integration with a specialized prompt generator architecture. To improve adaptability to grayscale weld radiographic images, we introduce a frequency prompt generator module, which enhances the model's sensitivity to frequency-domain information. Furthermore, to address the multi-scale nature of weld defects, we incorporate a multi-scale prompt generator module, enabling the model to effectively extract and encode defect information across varying scales. Extensive experimental evaluations demonstrate that WRT-SAM achieves a recall of 78.87%, a precision of 84.04%, and an AUC of 0.9746, setting a new state-of-the-art (SOTA) benchmark. Moreover, the model exhibits superior zero-shot generalization performance, highlighting its potential for practical deployment in diverse radiographic testing scenarios.
2502.11340
S2TX: Cross-Attention Multi-Scale State-Space Transformer for Time Series Forecasting
cs.LG
Time series forecasting has recently achieved significant progress with multi-scale models to address the heterogeneity between long and short range patterns. Despite their state-of-the-art performance, we identify two potential areas for improvement. First, the variates of the multivariate time series are processed independently. Moreover, the multi-scale (long and short range) representations are learned separately by two independent models without communication. In light of these concerns, we propose State Space Transformer with cross-attention (S2TX). S2TX employs a cross-attention mechanism to integrate a Mamba model for extracting long-range cross-variate context and a Transformer model with local window attention to capture short-range representations. By cross-attending to the global context, the Transformer model further facilitates variate-level interactions as well as local/global communications. Comprehensive experiments on seven classic long-short range time-series forecasting benchmark datasets demonstrate that S2TX can achieve highly robust SOTA results while maintaining a low memory footprint.
2502.11345
Hierarchical Graph Topic Modeling with Topic Tree-based Transformer
cs.CL
Textual documents are commonly connected in a hierarchical graph structure where a central document links to others with an exponentially growing connectivity. Though Hyperbolic Graph Neural Networks (HGNNs) excel at capturing such graph hierarchy, they cannot model the rich textual semantics within documents. Moreover, text contents in documents usually discuss topics of different specificity. Hierarchical Topic Models (HTMs) discover such latent topic hierarchy within text corpora. However, most of them focus on the textual content within documents, and ignore the graph adjacency across interlinked documents. We thus propose a Hierarchical Graph Topic Modeling Transformer to integrate both topic hierarchy within documents and graph hierarchy across documents into a unified Transformer. Specifically, to incorporate topic hierarchy within documents, we design a topic tree and infer a hierarchical tree embedding for hierarchical topic modeling. To preserve both topic and graph hierarchies, we design our model in hyperbolic space and propose Hyperbolic Doubly Recurrent Neural Network, which models ancestral and fraternal tree structure. Both hierarchies are inserted into each Transformer layer to learn unified representations. Both supervised and unsupervised experiments verify the effectiveness of our model.
2502.11346
Power-Measurement-Based Channel Autocorrelation Estimation for IRS-Assisted Wideband Communications
cs.IT math.IT
Channel state information (CSI) is essential to the performance optimization of intelligent reflecting surface (IRS)-aided wireless communication systems. However, the passive and frequency-flat reflection of IRS, as well as the high-dimensional IRS-reflected channels, have posed practical challenges for efficient IRS channel estimation, especially in wideband communication systems with significant multi-path channel delay spread. To tackle the above challenge, we propose a novel neural network (NN)-empowered IRS channel estimation and passive reflection design framework for the wideband orthogonal frequency division multiplexing (OFDM) communication system based only on the user's reference signal received power (RSRP) measurements with time-varying random IRS training reflections. In particular, we show that the average received signal power over all OFDM subcarriers at the user terminal can be represented as the prediction of a single-layer NN composed of multiple subnetworks with the same structure, such that the autocorrelation matrix of the wideband IRS channel can be recovered as their weights via supervised learning. To exploit the potential sparsity of the channel autocorrelation matrix, a progressive training method is proposed by gradually increasing the number of subnetworks until a desired accuracy is achieved, thus reducing the training complexity. Based on the estimates of IRS channel autocorrelation matrix, the IRS passive reflection is then optimized to maximize the average channel power gain over all subcarriers. Numerical results indicate the effectiveness of the proposed IRS channel autocorrelation matrix estimation and passive reflection design under wideband channels, which can achieve significant performance improvement compared to the existing IRS reflection designs based on user power measurements.
2502.11349
Biases in Edge Language Models: Detection, Analysis, and Mitigation
cs.LG cs.PF stat.ML
The integration of large language models (LLMs) on low-power edge devices such as Raspberry Pi, known as edge language models (ELMs), has introduced opportunities for more personalized, secure, and low-latency language intelligence that is accessible to all. However, the resource constraints inherent in edge devices and the lack of robust ethical safeguards in language models raise significant concerns about fairness, accountability, and transparency in model output generation. This paper conducts a comparative analysis of text-based bias across language model deployments on edge, cloud, and desktop environments, aiming to evaluate how deployment settings influence model fairness. Specifically, we examined an optimized Llama-2 model running on a Raspberry Pi 4; GPT 4o-mini, Gemini-1.5-flash, and Grok-beta models running on cloud servers; and Gemma2 and Mistral models running on a MacOS desktop machine. Our results demonstrate that Llama-2 running on Raspberry Pi 4 is 43.23% and 21.89% more prone to showing bias over time compared to models running on the desktop and cloud-based environments. We also propose the implementation of a feedback loop, a mechanism that iteratively adjusts model behavior based on previous outputs, where predefined constraint weights are applied layer-by-layer during inference, allowing the model to correct bias patterns, resulting in 79.28% reduction in model bias.
2502.11352
A Framework for Learning Scoring Rules in Autonomous Driving Planning Systems
cs.RO cs.LG
In autonomous driving systems, motion planning is commonly implemented as a two-stage process: first, a trajectory proposer generates multiple candidate trajectories, then a scoring mechanism selects the most suitable trajectory for execution. For this critical selection stage, rule-based scoring mechanisms are particularly appealing as they can explicitly encode driving preferences, safety constraints, and traffic regulations in a formalized, human-understandable format. However, manually crafting these scoring rules presents significant challenges: the rules often contain complex interdependencies, require careful parameter tuning, and may not fully capture the nuances present in real-world driving data. This work introduces FLoRA, a novel framework that bridges this gap by learning interpretable scoring rules represented in temporal logic. Our method features a learnable logic structure that captures nuanced relationships across diverse driving scenarios, optimizing both rules and parameters directly from real-world driving demonstrations collected in NuPlan. Our approach effectively learns to evaluate driving behavior even though the training data only contains positive examples (successful driving demonstrations). Evaluations in closed-loop planning simulations demonstrate that our learned scoring rules outperform existing techniques, including expert-designed rules and neural network scoring models, while maintaining interpretability. This work introduces a data-driven approach to enhance the scoring mechanism in autonomous driving systems, designed as a plug-in module to seamlessly integrate with various trajectory proposers. Our video and code are available on xiong.zikang.me/FLoRA.
2502.11355
"Nuclear Deployed!": Analyzing Catastrophic Risks in Decision-making of Autonomous LLM Agents
cs.CL cs.AI cs.CR cs.CY
Large language models (LLMs) are evolving into autonomous decision-makers, raising concerns about catastrophic risks in high-stakes scenarios, particularly in Chemical, Biological, Radiological and Nuclear (CBRN) domains. Based on the insight that such risks can originate from trade-offs between the agent's Helpful, Harmlessness and Honest (HHH) goals, we build a novel three-stage evaluation framework, which is carefully constructed to effectively and naturally expose such risks. We conduct 14,400 agentic simulations across 12 advanced LLMs, with extensive experiments and analysis. Results reveal that LLM agents can autonomously engage in catastrophic behaviors and deception, without being deliberately induced. Furthermore, stronger reasoning abilities often increase, rather than mitigate, these risks. We also show that these agents can violate instructions and superior commands. On the whole, we empirically prove the existence of catastrophic risks in autonomous LLM agents. We will release our code upon request.
2502.11356
SAIF: A Sparse Autoencoder Framework for Interpreting and Steering Instruction Following of Language Models
cs.LG cs.AI cs.CL
The ability of large language models (LLMs) to follow instructions is crucial for their practical applications, yet the underlying mechanisms remain poorly understood. This paper presents a novel framework that leverages sparse autoencoders (SAE) to interpret how instruction following works in these models. We demonstrate how the features we identify can effectively steer model outputs to align with given instructions. Through analysis of SAE latent activations, we identify specific latents responsible for instruction following behavior. Our findings reveal that instruction following capabilities are encoded by a distinct set of instruction-relevant SAE latents. These latents both show semantic proximity to relevant instructions and demonstrate causal effects on model behavior. Our research highlights several crucial factors for achieving effective steering performance: precise feature identification, the role of final layer, and optimal instruction positioning. Additionally, we demonstrate that our methodology scales effectively across SAEs and LLMs of varying sizes.
2502.11357
Explorer: Scaling Exploration-driven Web Trajectory Synthesis for Multimodal Web Agents
cs.AI cs.HC
Recent success in large multimodal models (LMMs) has sparked promising applications of agents capable of autonomously completing complex web tasks. While open-source LMM agents have made significant advances in offline evaluation benchmarks, their performance still falls substantially short of human-level capabilities in more realistic online settings. A key bottleneck is the lack of diverse and large-scale trajectory-level datasets across various domains, which are expensive to collect. In this paper, we address this challenge by developing a scalable recipe to synthesize the largest and most diverse trajectory-level dataset to date, containing over 94K successful multimodal web trajectories, spanning 49K unique URLs, 720K screenshots, and 33M web elements. In particular, we leverage extensive web exploration and refinement to obtain diverse task intents. The average cost is 28 cents per successful trajectory, making it affordable to a wide range of users in the community. Leveraging this dataset, we train Explorer, a multimodal web agent, and demonstrate strong performance on both offline and online web agent benchmarks such as Mind2Web-Live, Multimodal-Mind2Web, and MiniWob++. Additionally, our experiments highlight data scaling as a key driver for improving web agent capabilities. We hope this study makes state-of-the-art LMM-based agent research at a larger scale more accessible.
2502.11358
Mimicking the Familiar: Dynamic Command Generation for Information Theft Attacks in LLM Tool-Learning System
cs.AI cs.CR
Information theft attacks pose a significant risk to Large Language Model (LLM) tool-learning systems. Adversaries can inject malicious commands through compromised tools, manipulating LLMs to send sensitive information to these tools, which leads to potential privacy breaches. However, existing attack approaches are black-box oriented and rely on static commands that cannot adapt flexibly to the changes in user queries and the invocation chain of tools. It makes malicious commands more likely to be detected by LLM and leads to attack failure. In this paper, we propose AutoCMD, a dynamic attack comment generation approach for information theft attacks in LLM tool-learning systems. Inspired by the concept of mimicking the familiar, AutoCMD is capable of inferring the information utilized by upstream tools in the toolchain through learning on open-source systems and reinforcement with target system examples, thereby generating more targeted commands for information theft. The evaluation results show that AutoCMD outperforms the baselines with +13.2% $ASR_{Theft}$, and can be generalized to new tool-learning systems to expose their information leakage risks. We also design four defense methods to effectively protect tool-learning systems from the attack.
2502.11360
GeoDANO: Geometric VLM with Domain Agnostic Vision Encoder
cs.CV cs.CL
We introduce GeoDANO, a geometric vision-language model (VLM) with a domain-agnostic vision encoder, for solving plane geometry problems. Although VLMs have been employed for solving geometry problems, their ability to recognize geometric features remains insufficiently analyzed. To address this gap, we propose a benchmark that evaluates the recognition of visual geometric features, including primitives such as dots and lines, and relations such as orthogonality. Our preliminary study shows that vision encoders often used in general-purpose VLMs, e.g., OpenCLIP, fail to detect these features and struggle to generalize across domains. We develop GeoCLIP, a CLIP based model trained on synthetic geometric diagram-caption pairs to overcome the limitation. Benchmark results show that GeoCLIP outperforms existing vision encoders in recognizing geometric features. We then propose our VLM, GeoDANO, which augments GeoCLIP with a domain adaptation strategy for unseen diagram styles. GeoDANO outperforms specialized methods for plane geometry problems and GPT-4o on MathVerse.
2502.11361
VLDBench: Vision Language Models Disinformation Detection Benchmark
cs.CL
The rapid rise of AI-generated content has made detecting disinformation increasingly challenging. In particular, multimodal disinformation, i.e., online posts-articles that contain images and texts with fabricated information are specially designed to deceive. While existing AI safety benchmarks primarily address bias and toxicity, multimodal disinformation detection remains largely underexplored. To address this challenge, we present the Vision-Language Disinformation Detection Benchmark VLDBench, the first comprehensive benchmark for detecting disinformation across both unimodal (text-only) and multimodal (text and image) content, comprising 31,000} news article-image pairs, spanning 13 distinct categories, for robust evaluation. VLDBench features a rigorous semi-automated data curation pipeline, with 22 domain experts dedicating 300 plus hours} to annotation, achieving a strong inter-annotator agreement (Cohen kappa = 0.78). We extensively evaluate state-of-the-art Large Language Models (LLMs) and Vision-Language Models (VLMs), demonstrating that integrating textual and visual cues in multimodal news posts improves disinformation detection accuracy by 5 - 35 % compared to unimodal models. Developed in alignment with AI governance frameworks such as the EU AI Act, NIST guidelines, and the MIT AI Risk Repository 2024, VLDBench is expected to become a benchmark for detecting disinformation in online multi-modal contents. Our code and data will be publicly available.
2502.11362
Teleportation With Null Space Gradient Projection for Optimization Acceleration
cs.LG
Optimization techniques have become increasingly critical due to the ever-growing model complexity and data scale. In particular, teleportation has emerged as a promising approach, which accelerates convergence of gradient descent-based methods by navigating within the loss invariant level set to identify parameters with advantageous geometric properties. Existing teleportation algorithms have primarily demonstrated their effectiveness in optimizing Multi-Layer Perceptrons (MLPs), but their extension to more advanced architectures, such as Convolutional Neural Networks (CNNs) and Transformers, remains challenging. Moreover, they often impose significant computational demands, limiting their applicability to complex architectures. To this end, we introduce an algorithm that projects the gradient of the teleportation objective function onto the input null space, effectively preserving the teleportation within the loss invariant level set and reducing computational cost. Our approach is readily generalizable from MLPs to CNNs, transformers, and potentially other advanced architectures. We validate the effectiveness of our algorithm across various benchmark datasets and optimizers, demonstrating its broad applicability.
2502.11364
Blessing of Multilinguality: A Systematic Analysis of Multilingual In-Context Learning
cs.CL
While multilingual large language models generally perform adequately, and sometimes even rival English performance on high-resource languages (HRLs), they often significantly underperform on low-resource languages (LRLs). Among several prompting strategies aiming at bridging the gap, multilingual in-context learning (ICL) has been particularly effective when demonstration in target languages is unavailable. However, there lacks a systematic understanding of when and why it works well. In this work, we systematically analyze multilingual ICL, using demonstrations in HRLs to enhance cross-lingual transfer. We show that demonstrations in mixed HRLs consistently outperform English-only ones across the board, particularly for tasks written in LRLs. Surprisingly, our ablation study shows that the presence of irrelevant non-English sentences in the prompt yields measurable gains, suggesting the effectiveness of multilingual exposure itself. Our results highlight the potential of strategically leveraging multilingual resources to bridge the performance gap for underrepresented languages.
2502.11367
Sparse Autoencoder Features for Classifications and Transferability
cs.LG cs.AI cs.CL
Sparse Autoencoders (SAEs) provide potentials for uncovering structured, human-interpretable representations in Large Language Models (LLMs), making them a crucial tool for transparent and controllable AI systems. We systematically analyze SAE for interpretable feature extraction from LLMs in safety-critical classification tasks. Our framework evaluates (1) model-layer selection and scaling properties, (2) SAE architectural configurations, including width and pooling strategies, and (3) the effect of binarizing continuous SAE activations. SAE-derived features achieve macro F1 > 0.8, outperforming hidden-state and BoW baselines while demonstrating cross-model transfer from Gemma 2 2B to 9B-IT models. These features generalize in a zero-shot manner to cross-lingual toxicity detection and visual classification tasks. Our analysis highlights the significant impact of pooling strategies and binarization thresholds, showing that binarization offers an efficient alternative to traditional feature selection while maintaining or improving performance. These findings establish new best practices for SAE-based interpretability and enable scalable, transparent deployment of LLMs in real-world applications. Full repo: https://github.com/shan23chen/MOSAIC.
2502.11368
LLMs can Perform Multi-Dimensional Analytic Writing Assessments: A Case Study of L2 Graduate-Level Academic English Writing
cs.CL cs.AI
The paper explores the performance of LLMs in the context of multi-dimensional analytic writing assessments, i.e. their ability to provide both scores and comments based on multiple assessment criteria. Using a corpus of literature reviews written by L2 graduate students and assessed by human experts against 9 analytic criteria, we prompt several popular LLMs to perform the same task under various conditions. To evaluate the quality of feedback comments, we apply a novel feedback comment quality evaluation framework. This framework is interpretable, cost-efficient, scalable, and reproducible, compared to existing methods that rely on manual judgments. We find that LLMs can generate reasonably good and generally reliable multi-dimensional analytic assessments. We release our corpus for reproducibility.
2502.11369
Physics-Informed Gaussian Process Classification for Constraint-Aware Alloy Design
cond-mat.mtrl-sci cs.LG
Alloy design can be framed as a constraint-satisfaction problem. Building on previous methodologies, we propose equipping Gaussian Process Classifiers (GPCs) with physics-informed prior mean functions to model the boundaries of feasible design spaces. Through three case studies, we highlight the utility of informative priors for handling constraints on continuous and categorical properties. (1) Phase Stability: By incorporating CALPHAD predictions as priors for solid-solution phase stability, we enhance model validation using a publicly available XRD dataset. (2) Phase Stability Prediction Refinement: We demonstrate an in silico active learning approach to efficiently correct phase diagrams. (3) Continuous Property Thresholds: By embedding priors into continuous property models, we accelerate the discovery of alloys meeting specific property thresholds via active learning. In each case, integrating physics-based insights into the classification framework substantially improved model performance, demonstrating an efficient strategy for constraint-aware alloy design.
2502.11370
HI-GVF: Shared Control based on Human-Influenced Guiding Vector Fields for Human-multi-robot Cooperation
cs.RO
Human-multi-robot shared control leverages human decision-making and robotic autonomy to enhance human-robot collaboration. While widely studied, existing systems often adopt a leader-follower model, limiting robot autonomy to some extent. Besides, a human is required to directly participate in the motion control of robots through teleoperation, which significantly burdens the operator. To alleviate these two issues, we propose a layered shared control computing framework using human-influenced guiding vector fields (HI-GVF) for human-robot collaboration. HI-GVF guides the multi-robot system along a desired path specified by the human. Then, an intention field is designed to merge the human and robot intentions, accelerating the propagation of the human intention within the multi-robot system. Moreover, we give the stability analysis of the proposed model and use collision avoidance based on safety barrier certificates to fine-tune the velocity. Eventually, considering the firefighting task as an example scenario, we conduct simulations and experiments using multiple human-robot interfaces (brain-computer interface, myoelectric wristband, eye-tracking), and the results demonstrate that our proposed approach boosts the effectiveness and performance of the task.
2502.11371
RAG vs. GraphRAG: A Systematic Evaluation and Key Insights
cs.IR
Retrieval-Augmented Generation (RAG) enhances the performance of LLMs across various tasks by retrieving relevant information from external sources, particularly on text-based data. For structured data, such as knowledge graphs, GraphRAG has been widely used to retrieve relevant information. However, recent studies have revealed that structuring implicit knowledge from text into graphs can benefit certain tasks, extending the application of GraphRAG from graph data to general text-based data. Despite their successful extensions, most applications of GraphRAG for text data have been designed for specific tasks and datasets, lacking a systematic evaluation and comparison between RAG and GraphRAG on widely used text-based benchmarks. In this paper, we systematically evaluate RAG and GraphRAG on well-established benchmark tasks, such as Question Answering and Query-based Summarization. Our results highlight the distinct strengths of RAG and GraphRAG across different tasks and evaluation perspectives. Inspired by these observations, we investigate strategies to integrate their strengths to improve downstream tasks. Additionally, we provide an in-depth discussion of the shortcomings of current GraphRAG approaches and outline directions for future research.
2502.11372
Weibull Processes in Network Degree Distributions
cs.SI physics.soc-ph
This study examines degree distributions in two large collaboration networks: the Microsoft Academic Graph (1800-2020) and Internet Movie Database (1900-2020), comprising $2.72 \times 10^8$ and $1.88 \times 10^6$ nodes respectively. Statistical comparison using $\chi^2$ measures showed that Weibull distributions fit the degree distributions better than power-law or log-normal models, especially at later stages in the network evolution. The Weibull shape parameters exhibit notable stability ($k \approx 0.8$-$1.0$ for academic, $k \approx 0.9$-$1.1$ for entertainment collaborations) despite orders of magnitude growth in network size. While early-stage networks display approximate power-law scaling, mature networks develop characteristic flattening in the low-degree region that Weibull distributions appear to capture better. In the academic network, the cutoff between the flattened region and power-law tail shows a gradual increase from $5$ to $9$ edges over time, while the entertainment network maintains a distinctive degree structure that may reflect storytelling and cast-size constraints. These patterns suggest the possibility that collaboration network evolution might be influenced more by constraint-based growth than by pure preferential attachment or multiplicative processes.
2502.11374
Leave No One Behind: Enhancing Diversity While Maintaining Accuracy in Social Recommendation
cs.IR
Social recommendation, a branch of algorithms that utilizes social connection information to construct recommender systems, has demonstrated its effectiveness in enhancing recommendation accuracy. However, apart from accuracy, the diversity of recommendations also plays a critical role in user engagement. Unfortunately, the impact of social recommendation models on recommendation diversity remains largely unexplored. In this study, we investigate the dual performance of existing social recommendation algorithms in terms of accuracy and diversity. Our empirical findings highlight a concerning trend: social recommendation models tend to decrease diversity, despite their accuracy improvements. To address this issue, we propose a novel approach called Diversified Social Recommendation (DivSR), which leverages relational knowledge distillation techniques to transfer high-diversity structured knowledge from non-social recommendation models to social recommendation models. DivSR is designed as a simple, model-agnostic framework that integrates seamlessly with existing social recommendation architectures. Experimental results on three benchmark datasets demonstrate that DivSR significantly increases diversity without markedly compromising accuracy across various social recommendation backbones, achieving a better accuracy-diversity trade-off. Our code and data are publicly available at: https://github.com/ll0ruc/DivSR
2502.11375
Robot Deformable Object Manipulation via NMPC-generated Demonstrations in Deep Reinforcement Learning
cs.RO cs.LG
In this work, we conducted research on deformable object manipulation by robots based on demonstration-enhanced reinforcement learning (RL). To improve the learning efficiency of RL, we enhanced the utilization of demonstration data from multiple aspects and proposed the HGCR-DDPG algorithm. It uses a novel high-dimensional fuzzy approach for grasping-point selection, a refined behavior-cloning method to enhance data-driven learning in Rainbow-DDPG, and a sequential policy-learning strategy. Compared to the baseline algorithm (Rainbow-DDPG), our proposed HGCR-DDPG achieved 2.01 times the global average reward and reduced the global average standard deviation to 45% of that of the baseline algorithm. To reduce the human labor cost of demonstration collection, we proposed a low-cost demonstration collection method based on Nonlinear Model Predictive Control (NMPC). Simulation experiment results show that demonstrations collected through NMPC can be used to train HGCR-DDPG, achieving comparable results to those obtained with human demonstrations. To validate the feasibility of our proposed methods in real-world environments, we conducted physical experiments involving deformable object manipulation. We manipulated fabric to perform three tasks: diagonal folding, central axis folding, and flattening. The experimental results demonstrate that our proposed method achieved success rates of 83.3%, 80%, and 100% for these three tasks, respectively, validating the effectiveness of our approach. Compared to current large-model approaches for robot manipulation, the proposed algorithm is lightweight, requires fewer computational resources, and offers task-specific customization and efficient adaptability for specific tasks.
2502.11377
PrivilegedDreamer: Explicit Imagination of Privileged Information for Rapid Adaptation of Learned Policies
cs.RO cs.LG
Numerous real-world control problems involve dynamics and objectives affected by unobservable hidden parameters, ranging from autonomous driving to robotic manipulation, which cause performance degradation during sim-to-real transfer. To represent these kinds of domains, we adopt hidden-parameter Markov decision processes (HIP-MDPs), which model sequential decision problems where hidden variables parameterize transition and reward functions. Existing approaches, such as domain randomization, domain adaptation, and meta-learning, simply treat the effect of hidden parameters as additional variance and often struggle to effectively handle HIP-MDP problems, especially when the rewards are parameterized by hidden variables. We introduce Privileged-Dreamer, a model-based reinforcement learning framework that extends the existing model-based approach by incorporating an explicit parameter estimation module. PrivilegedDreamer features its novel dual recurrent architecture that explicitly estimates hidden parameters from limited historical data and enables us to condition the model, actor, and critic networks on these estimated parameters. Our empirical analysis on five diverse HIP-MDP tasks demonstrates that PrivilegedDreamer outperforms state-of-the-art model-based, model-free, and domain adaptation learning algorithms. Additionally, we conduct ablation studies to justify the inclusion of each component in the proposed architecture.
2502.11379
CCJA: Context-Coherent Jailbreak Attack for Aligned Large Language Models
cs.CR cs.AI cs.CL
Despite explicit alignment efforts for large language models (LLMs), they can still be exploited to trigger unintended behaviors, a phenomenon known as "jailbreaking." Current jailbreak attack methods mainly focus on discrete prompt manipulations targeting closed-source LLMs, relying on manually crafted prompt templates and persuasion rules. However, as the capabilities of open-source LLMs improve, ensuring their safety becomes increasingly crucial. In such an environment, the accessibility of model parameters and gradient information by potential attackers exacerbates the severity of jailbreak threats. To address this research gap, we propose a novel \underline{C}ontext-\underline{C}oherent \underline{J}ailbreak \underline{A}ttack (CCJA). We define jailbreak attacks as an optimization problem within the embedding space of masked language models. Through combinatorial optimization, we effectively balance the jailbreak attack success rate with semantic coherence. Extensive evaluations show that our method not only maintains semantic consistency but also surpasses state-of-the-art baselines in attack effectiveness. Additionally, by integrating semantically coherent jailbreak prompts generated by our method into widely used black-box methodologies, we observe a notable enhancement in their success rates when targeting closed-source commercial LLMs. This highlights the security threat posed by open-source LLMs to commercial counterparts. We will open-source our code if the paper is accepted.
2502.11380
Exploring the Small World of Word Embeddings: A Comparative Study on Conceptual Spaces from LLMs of Different Scales
cs.CL
A conceptual space represents concepts as nodes and semantic relatedness as edges. Word embeddings, combined with a similarity metric, provide an effective approach to constructing such a space. Typically, embeddings are derived from traditional distributed models or encoder-only pretrained models, whose objectives directly capture the meaning of the current token. In contrast, decoder-only models, including large language models (LLMs), predict the next token, making their embeddings less directly tied to the current token's semantics. Moreover, comparative studies on LLMs of different scales remain underexplored. In this paper, we construct a conceptual space using word embeddings from LLMs of varying scales and comparatively analyze their properties. We establish a network based on a linguistic typology-inspired connectivity hypothesis, examine global statistical properties, and compare LLMs of varying scales. Locally, we analyze conceptual pairs, WordNet relations, and a cross-lingual semantic network for qualitative words. Our results indicate that the constructed space exhibits small-world properties, characterized by a high clustering coefficient and short path lengths. Larger LLMs generate more intricate spaces, with longer paths reflecting richer relational structures and connections. Furthermore, the network serves as an efficient bridge for cross-lingual semantic mapping.
2502.11381
Without Paired Labeled Data: An End-to-End Self-Supervised Paradigm for UAV-View Geo-Localization
cs.CV cs.AI
UAV-View Geo-Localization (UVGL) aims to ascertain the precise location of a UAV by retrieving the most similar GPS-tagged satellite image. However, existing methods predominantly rely on supervised learning paradigms that necessitate annotated paired data for training, which incurs substantial annotation costs and impedes large-scale deployment. To overcome this limitation, we propose the Dynamic Memory-Driven and Neighborhood Information Learning (DMNIL) network, a lightweight end-to-end self-supervised framework for UAV-view geo-localization. The DMNIL framework utilizes a dual-path clustering-based contrastive learning architecture as its baseline to model intra-view structural relationships, enhancing feature consistency and discriminability. Additionally, a dynamic memory-driven hierarchical learning module is proposed to progressively mine local and global information, reinforcing multi-level feature associations to improve model robustness. To bridge the domain gap between UAV and satellite views, we design an information-consistent evolutionary learning mechanism that systematically explores latent correlations within intra-view neighborhoods and across cross-view domains, ultimately constructing a unified cross-view feature representation space. Extensive experiments on three benchmarks (University-1652, SUES-200, and DenseUAV) demonstrate that DMNIL achieves competitive performance against state-of-the-art supervised methods while maintaining computational efficiency. Notably, this superiority is attained without relying on paired training data, underscoring the framework's practicality for real-world deployment. Codes will be released soon.
2502.11382
A Physics-Informed Blur Learning Framework for Imaging Systems
cs.CV
Accurate blur estimation is essential for high-performance imaging across various applications. Blur is typically represented by the point spread function (PSF). In this paper, we propose a physics-informed PSF learning framework for imaging systems, consisting of a simple calibration followed by a learning process. Our framework could achieve both high accuracy and universal applicability. Inspired by the Seidel PSF model for representing spatially varying PSF, we identify its limitations in optimization and introduce a novel wavefront-based PSF model accompanied by an optimization strategy, both reducing optimization complexity and improving estimation accuracy. Moreover, our wavefront-based PSF model is independent of lens parameters, eliminate the need for prior knowledge of the lens. To validate our approach, we compare it with recent PSF estimation methods (Degradation Transfer and Fast Two-step) through a deblurring task, where all the estimated PSFs are used to train state-of-the-art deblurring algorithms. Our approach demonstrates improvements in image quality in simulation and also showcases noticeable visual quality improvements on real captured images.
2502.11386
Intelligent Mobile AI-Generated Content Services via Interactive Prompt Engineering and Dynamic Service Provisioning
cs.NI cs.LG
Due to massive computational demands of large generative models, AI-Generated Content (AIGC) can organize collaborative Mobile AIGC Service Providers (MASPs) at network edges to provide ubiquitous and customized content generation for resource-constrained users. However, such a paradigm faces two significant challenges: 1) raw prompts (i.e., the task description from users) often lead to poor generation quality due to users' lack of experience with specific AIGC models, and 2) static service provisioning fails to efficiently utilize computational and communication resources given the heterogeneity of AIGC tasks. To address these challenges, we propose an intelligent mobile AIGC service scheme. Firstly, we develop an interactive prompt engineering mechanism that leverages a Large Language Model (LLM) to generate customized prompt corpora and employs Inverse Reinforcement Learning (IRL) for policy imitation through small-scale expert demonstrations. Secondly, we formulate a dynamic mobile AIGC service provisioning problem that jointly optimizes the number of inference trials and transmission power allocation. Then, we propose the Diffusion-Enhanced Deep Deterministic Policy Gradient (D3PG) algorithm to solve the problem. By incorporating the diffusion process into Deep Reinforcement Learning (DRL) architecture, the environment exploration capability can be improved, thus adapting to varying mobile AIGC scenarios. Extensive experimental results demonstrate that our prompt engineering approach improves single-round generation success probability by 6.3 times, while D3PG increases the user service experience by 67.8% compared to baseline DRL approaches.
2502.11387
RoleMRC: A Fine-Grained Composite Benchmark for Role-Playing and Instruction-Following
cs.CL
Role-playing is important for Large Language Models (LLMs) to follow diverse instructions while maintaining role identity and the role's pre-defined ability limits. Existing role-playing datasets mostly contribute to controlling role style and knowledge boundaries, but overlook role-playing in instruction-following scenarios. We introduce a fine-grained role-playing and instruction-following composite benchmark, named RoleMRC, including: (1) Multi-turn dialogues between ideal roles and humans, including free chats or discussions upon given passages; (2) Role-playing machine reading comprehension, involving response, refusal, and attempts according to passage answerability and role ability; (3) More complex scenarios with nested, multi-turn and prioritized instructions. The final RoleMRC features a 10.2k role profile meta-pool, 37.9k well-synthesized role-playing instructions, and 1.4k testing samples. We develop a pipeline to quantitatively evaluate the fine-grained role-playing and instruction-following capabilities of several mainstream LLMs, as well as models that are fine-tuned on our data. Moreover, cross-evaluation on external role-playing datasets confirms that models fine-tuned on RoleMRC enhances instruction-following without compromising general role-playing and reasoning capabilities. We also probe the neural-level activation maps of different capabilities over post-tuned LLMs. Access to our RoleMRC, RoleMRC-mix and Codes: https://github.com/LuJunru/RoleMRC.
2502.11390
MARS: Mesh AutoRegressive Model for 3D Shape Detailization
cs.CV
State-of-the-art methods for mesh detailization predominantly utilize Generative Adversarial Networks (GANs) to generate detailed meshes from coarse ones. These methods typically learn a specific style code for each category or similar categories without enforcing geometry supervision across different Levels of Detail (LODs). Consequently, such methods often fail to generalize across a broader range of categories and cannot ensure shape consistency throughout the detailization process. In this paper, we introduce MARS, a novel approach for 3D shape detailization. Our method capitalizes on a novel multi-LOD, multi-category mesh representation to learn shape-consistent mesh representations in latent space across different LODs. We further propose a mesh autoregressive model capable of generating such latent representations through next-LOD token prediction. This approach significantly enhances the realism of the generated shapes. Extensive experiments conducted on the challenging 3D Shape Detailization benchmark demonstrate that our proposed MARS model achieves state-of-the-art performance, surpassing existing methods in both qualitative and quantitative assessments. Notably, the model's capability to generate fine-grained details while preserving the overall shape integrity is particularly commendable.
2502.11393
HellaSwag-Pro: A Large-Scale Bilingual Benchmark for Evaluating the Robustness of LLMs in Commonsense Reasoning
cs.CL
Large language models (LLMs) have shown remarkable capabilities in commonsense reasoning; however, some variations in questions can trigger incorrect responses. Do these models truly understand commonsense knowledge, or just memorize expression patterns? To investigate this question, we present the first extensive robustness evaluation of LLMs in commonsense reasoning. We introduce HellaSwag-Pro, a large-scale bilingual benchmark consisting of 11,200 cases, by designing and compiling seven types of question variants. To construct this benchmark, we propose a two-stage method to develop Chinese HellaSwag, a finely annotated dataset comprising 12,000 instances across 56 categories. We conduct extensive experiments on 41 representative LLMs, revealing that these LLMs are far from robust in commonsense reasoning. Furthermore, this robustness varies depending on the language in which the LLM is tested. This work establishes a high-quality evaluation benchmark, with extensive experiments offering valuable insights to the community in commonsense reasoning for LLMs.
2502.11394
Oversmoothing as Loss of Sign: Towards Structural Balance in Graph Neural Networks
cs.LG
Oversmoothing is a common issue in graph neural networks (GNNs), where node representations become excessively homogeneous as the number of layers increases, resulting in degraded performance. Various strategies have been proposed to combat oversmoothing in practice, yet they are based on different heuristics and lack a unified understanding of their inherent mechanisms. In this paper, we show that three major classes of anti-oversmoothing techniques can be mathematically interpreted as message passing over signed graphs comprising both positive and negative edges. By analyzing the asymptotic behavior of signed graph propagation, we demonstrate that negative edges can repel nodes to a certain extent, providing deeper insights into how these methods mitigate oversmoothing. Furthermore, our results suggest that the structural balance of a signed graph-where positive edges exist only within clusters and negative edges appear only between clusters-is crucial for clustering node representations in the long term through signed graph propagation. Motivated by these observations, we propose a solution to mitigate oversmoothing with theoretical guarantees-Structural Balance Propagation (SBP), by incorporating label and feature information to create a structurally balanced graph for message-passing. Experiments on nine datasets against twelve baselines demonstrate the effectiveness of our method, highlighting the value of our signed graph perspective.
2502.11396
Maintenance of Structural Hole Spanners in Dynamic Networks
cs.SI
Structural Hole (SH) spanners are the set of users who bridge different groups of users and are vital in numerous applications. Despite their importance, existing work for identifying SH spanners focuses only on static networks. However, real-world networks are highly dynamic where the underlying structure of the network evolves continuously. Consequently, we study SH spanner problem for dynamic networks. We propose an efficient solution for updating SH spanners in dynamic networks. Our solution reuses the information obtained during the initial runs of the static algorithm and avoids the recomputations for the nodes unaffected by the updates. Experimental results show that the proposed solution achieves a minimum speedup of 3.24 over recomputation. To the best of our knowledge, this is the first attempt to address the problem of maintaining SH spanners in dynamic networks.
2502.11400
Revisiting Robust RAG: Do We Still Need Complex Robust Training in the Era of Powerful LLMs?
cs.CL
Retrieval-augmented generation (RAG) systems often suffer from performance degradation when encountering noisy or irrelevant documents, driving researchers to develop sophisticated training strategies to enhance their robustness against such retrieval noise. However, as large language models (LLMs) continue to advance, the necessity of these complex training methods is increasingly questioned. In this paper, we systematically investigate whether complex robust training strategies remain necessary as model capacity grows. Through comprehensive experiments spanning multiple model architectures and parameter scales, we evaluate various document selection methods and adversarial training techniques across diverse datasets. Our extensive experiments consistently demonstrate that as models become more powerful, the performance gains brought by complex robust training methods drop off dramatically. We delve into the rationale and find that more powerful models inherently exhibit superior confidence calibration, better generalization across datasets (even when trained with randomly selected documents), and optimal attention mechanisms learned with simpler strategies. Our findings suggest that RAG systems can benefit from simpler architectures and training strategies as models become more powerful, enabling more scalable applications with minimal complexity.
2502.11401
Following the Autoregressive Nature of LLM Embeddings via Compression and Alignment
cs.CL
A new trend uses LLMs as dense text encoders via contrastive learning. However, since LLM embeddings predict the probability distribution of the next token, they are inherently generative and distributive, conflicting with contrastive learning, which requires embeddings to capture full-text semantics and align via cosine similarity. This discrepancy hinders the full utilization of LLMs' pre-training capabilities, resulting in inefficient learning. In response to this issue, we propose AutoRegEmbed, a new contrastive learning method built on embedding conditional probability distributions, which integrates two core tasks: information compression and conditional distribution alignment. The information compression task encodes text into the embedding space, ensuring that the embedding vectors capture global semantics. The conditional distribution alignment task focuses on aligning text embeddings with positive samples embeddings by leveraging the conditional distribution of embeddings while simultaneously reducing the likelihood of generating negative samples from text embeddings, thereby achieving embedding alignment and uniformity. Experimental results demonstrate that our method significantly outperforms traditional contrastive learning approaches and achieves performance comparable to state-of-the-art models when using the same amount of data.
2502.11404
ToolCoder: A Systematic Code-Empowered Tool Learning Framework for Large Language Models
cs.CL
Tool learning has emerged as a crucial capability for large language models (LLMs) to solve complex real-world tasks through interaction with external tools. Existing approaches face significant challenges, including reliance on hand-crafted prompts, difficulty in multi-step planning, and lack of precise error diagnosis and reflection mechanisms. We propose ToolCoder, a novel framework that reformulates tool learning as a code generation task. Inspired by software engineering principles, ToolCoder transforms natural language queries into structured Python function scaffold and systematically breaks down tasks with descriptive comments, enabling LLMs to leverage coding paradigms for complex reasoning and planning. It then generates and executes function implementations to obtain final responses. Additionally, ToolCoder stores successfully executed functions in a repository to promote code reuse, while leveraging error traceback mechanisms for systematic debugging, optimizing both execution efficiency and robustness. Experiments demonstrate that ToolCoder achieves superior performance in task completion accuracy and execution reliability compared to existing approaches, establishing the effectiveness of code-centric approaches in tool learning.
2502.11405
LayAlign: Enhancing Multilingual Reasoning in Large Language Models via Layer-Wise Adaptive Fusion and Alignment Strategy
cs.CL
Despite being pretrained on multilingual corpora, large language models (LLMs) exhibit suboptimal performance on low-resource languages. Recent approaches have leveraged multilingual encoders alongside LLMs by introducing trainable parameters connecting the two models. However, these methods typically focus on the encoder's output, overlooking valuable information from other layers. We propose \aname (\mname), a framework that integrates representations from all encoder layers, coupled with the \attaname mechanism to enable layer-wise interaction between the LLM and the multilingual encoder. Extensive experiments on multilingual reasoning tasks, along with analyses of learned representations, show that our approach consistently outperforms existing baselines.
2502.11408
Precise GPS-Denied UAV Self-Positioning via Context-Enhanced Cross-View Geo-Localization
cs.CV
Image retrieval has been employed as a robust complementary technique to address the challenge of Unmanned Aerial Vehicles (UAVs) self-positioning. However, most existing methods primarily focus on localizing objects captured by UAVs through complex part-based representations, often overlooking the unique challenges associated with UAV self-positioning, such as fine-grained spatial discrimination requirements and dynamic scene variations. To address the above issues, we propose the Context-Enhanced method for precise UAV Self-Positioning (CEUSP), specifically designed for UAV self-positioning tasks. CEUSP integrates a Dynamic Sampling Strategy (DSS) to efficiently select optimal negative samples, while the Rubik's Cube Attention (RCA) module, combined with the Context-Aware Channel Integration (CACI) module, enhances feature representation and discrimination by exploiting interdimensional interactions, inspired by the rotational mechanics of a Rubik's Cube. Extensive experimental validate the effectiveness of the proposed method, demonstrating notable improvements in feature representation and UAV self-positioning accuracy within complex urban environments. Our approach achieves state-of-the-art performance on the DenseUAV dataset, which is specifically designed for dense urban contexts, and also delivers competitive results on the widely recognized University-1652 benchmark.
2502.11410
Structure based SAT dataset for analysing GNN generalisation
cs.LG
Satisfiability (SAT) solvers based on techniques such as conflict driven clause learning (CDCL) have produced excellent performance on both synthetic and real world industrial problems. While these CDCL solvers only operate on a per-problem basis, graph neural network (GNN) based solvers bring new benefits to the field by allowing practitioners to exploit knowledge gained from solved problems to expedite solving of new SAT problems. However, one specific area that is often studied in the context of CDCL solvers, but largely overlooked in GNN solvers, is the relationship between graph theoretic measure of structure in SAT problems and the generalisation ability of GNN solvers. To bridge the gap between structural graph properties (e.g., modularity, self-similarity) and the generalisability (or lack thereof) of GNN based SAT solvers, we present StructureSAT: a curated dataset, along with code to further generate novel examples, containing a diverse set of SAT problems from well known problem domains. Furthermore, we utilise a novel splitting method that focuses on deconstructing the families into more detailed hierarchies based on their structural properties. With the new dataset, we aim to help explain problematic generalisation in existing GNN SAT solvers by exploiting knowledge of structural graph properties. We conclude with multiple future directions that can help researchers in GNN based SAT solving develop more effective and generalisable SAT solvers.
2502.11411
Detecting and Filtering Unsafe Training Data via Data Attribution
cs.LG
Large language models (LLMs) are vulnerable to unsafe training data that even small amounts of unsafe data can lead to harmful model behaviors. Detecting and filtering such unsafe training data is essential for trustworthy model development. Current state-of-the-art (SOTA) approaches typically rely on training moderation classifiers which requires significant computational overhead and are limited to predefined taxonomies, making them less adaptable to evolving safety concerns. Moreover, these classifiers lack insight into the training process, limiting their effectiveness in filtering unsafe data. To address these limitations, we propose DABUF, leveraging data attribution to detect and filter unsafe training data by attributing harmful model outputs to influential training data points. DABUF enables flexible identification of various unsafe data types without predefined taxonomies. However, in practice, model outputs can be complex with combined safe linguistic features and unsafe content, leading to reduced attribution accuracy. In such cases, DABUF will integrate moderation classifiers to identify a minimal subset of unsafe training data for targeted attribution (such as jailbreak). When model outputs are relatively straightforward, DABUF uses model outputs directly as the attribution targets. We evaluate the performance on two different tasks: in filtering jailbreaking training data and in identifying and mitigating gender bias. DABUF outperforms SOTA approaches by up to 7.5\% in detection AUPRC in jailbreaking scenarios, and 44.1\% in detecting gender bias. Moreover, retraining on DABUF-filtered data leads to higher model safety across experiments, underscoring its versatility in addressing a broad spectrum of unsafe data issues.
2502.11413
Statistical Query Hardness of Multiclass Linear Classification with Random Classification Noise
cs.LG stat.ML
We study the task of Multiclass Linear Classification (MLC) in the distribution-free PAC model with Random Classification Noise (RCN). Specifically, the learner is given a set of labeled examples $(x, y)$, where $x$ is drawn from an unknown distribution on $R^d$ and the labels are generated by a multiclass linear classifier corrupted with RCN. That is, the label $y$ is flipped from $i$ to $j$ with probability $H_{ij}$ according to a known noise matrix $H$ with non-negative separation $\sigma: = \min_{i \neq j} H_{ii}-H_{ij}$. The goal is to compute a hypothesis with small 0-1 error. For the special case of two labels, prior work has given polynomial-time algorithms achieving the optimal error. Surprisingly, little is known about the complexity of this task even for three labels. As our main contribution, we show that the complexity of MLC with RCN becomes drastically different in the presence of three or more labels. Specifically, we prove super-polynomial Statistical Query (SQ) lower bounds for this problem. In more detail, even for three labels and constant separation, we give a super-polynomial lower bound on the complexity of any SQ algorithm achieving optimal error. For a larger number of labels and smaller separation, we show a super-polynomial SQ lower bound even for the weaker goal of achieving any constant factor approximation to the optimal loss or even beating the trivial hypothesis.
2502.11414
Unbiased Learning to Rank with Query-Level Click Propensity Estimation: Beyond Pointwise Observation and Relevance
cs.IR
Most existing unbiased learning-to-rank (ULTR) approaches are based on the user examination hypothesis, which assumes that users will click a result only if it is both relevant and observed (typically modeled by position). However, in real-world scenarios, users often click only one or two results after examining multiple relevant options, due to limited patience or because their information needs have already been satisfied. Motivated by this, we propose a query-level click propensity model to capture the probability that users will click on different result lists, allowing for non-zero probabilities that users may not click on an observed relevant result. We hypothesize that this propensity increases when more potentially relevant results are present, and refer to this user behavior as relevance saturation bias. Our method introduces a Dual Inverse Propensity Weighting (DualIPW) mechanism -- combining query-level and position-level IPW -- to address both relevance saturation and position bias. Through theoretical derivation, we prove that DualIPW can learn an unbiased ranking model. Experiments on the real-world Baidu-ULTR dataset demonstrate that our approach significantly outperforms state-of-the-art ULTR baselines. The code and dataset information can be found at https://github.com/Trustworthy-Information-Access/DualIPW.
2502.11417
DiSCo: Device-Server Collaborative LLM-Based Text Streaming Services
cs.LG cs.DC
The rapid rise of large language models (LLMs) in text streaming services has introduced significant cost and Quality of Experience (QoE) challenges in serving millions of daily requests, especially in meeting Time-To-First-Token (TTFT) and Time-Between-Token (TBT) requirements for real-time interactions. Our real-world measurements show that both server-based and on-device deployments struggle to meet diverse QoE demands: server deployments face high costs and last-hop issues (e.g., Internet latency and dynamics), while on-device LLM inference is constrained by resources. We introduce DiSCo, a device-server cooperative scheduler designed to optimize users' QoE by adaptively routing requests and migrating response generation between endpoints while maintaining cost constraints. DiSCo employs cost-aware scheduling, leveraging the predictable speed of on-device LLM inference with the flexible capacity of server-based inference to dispatch requests on the fly, while introducing a token-level migration mechanism to ensure consistent token delivery during migration. Evaluations on real-world workloads -- including commercial services like OpenAI GPT and DeepSeek, and open-source deployments such as LLaMA3 -- show that DiSCo can improve users' QoE by reducing tail TTFT (11-52\%) and mean TTFT (6-78\%) across different model-device configurations, while dramatically reducing serving costs by up to 84\% through its migration mechanism while maintaining comparable QoE levels.
2502.11418
TimeCAP: Learning to Contextualize, Augment, and Predict Time Series Events with Large Language Model Agents
cs.AI cs.LG
Time series data is essential in various applications, including climate modeling, healthcare monitoring, and financial analytics. Understanding the contextual information associated with real-world time series data is often essential for accurate and reliable event predictions. In this paper, we introduce TimeCAP, a time-series processing framework that creatively employs Large Language Models (LLMs) as contextualizers of time series data, extending their typical usage as predictors. TimeCAP incorporates two independent LLM agents: one generates a textual summary capturing the context of the time series, while the other uses this enriched summary to make more informed predictions. In addition, TimeCAP employs a multi-modal encoder that synergizes with the LLM agents, enhancing predictive performance through mutual augmentation of inputs with in-context examples. Experimental results on real-world datasets demonstrate that TimeCAP outperforms state-of-the-art methods for time series event prediction, including those utilizing LLMs as predictors, achieving an average improvement of 28.75% in F1 score.
2502.11419
InsBank: Evolving Instruction Subset for Ongoing Alignment
cs.CL
Large language models (LLMs) typically undergo instruction tuning to enhance alignment. Recent studies emphasize that quality and diversity of instruction data are more crucial than quantity, highlighting the need to select diverse, high-quality subsets to reduce training costs. However, how to evolve these selected subsets alongside the development of new instruction data remains insufficiently explored. To achieve LLMs' ongoing alignment, we introduce Instruction Bank (InsBank), a continuously updated repository that integrates the latest valuable instruction data. We further propose Progressive Instruction Bank Evolution (PIBE), a novel framework designed to evolve InsBank effectively and efficiently over time. PIBE employs a gradual data selection strategy to maintain long-term efficiency, leveraging a representation-based diversity score to capture relationships between data points and retain historical information for comprehensive diversity evaluation. This also allows for flexible combination of diversity and quality scores during data selection and ranking. Extensive experiments demonstrate that PIBE significantly outperforms baselines in InsBank evolution and is able to extract budget-specific subsets, demonstrating its effectiveness and adaptability.
2502.11420
Training-Free Guidance Beyond Differentiability: Scalable Path Steering with Tree Search in Diffusion and Flow Models
cs.LG
Training-free guidance enables controlled generation in diffusion and flow models, but most existing methods assume differentiable objectives and rely on gradients. This work focuses on training-free guidance addressing challenges from non-differentiable objectives and discrete data distributions. We propose an algorithmic framework TreeG: Tree Search-Based Path Steering Guidance, applicable to both continuous and discrete settings in diffusion and flow models. TreeG offers a unified perspective on training-free guidance: proposing candidates for the next step, evaluating candidates, and selecting the best to move forward, enhanced by a tree search mechanism over active paths or parallelizing exploration. We comprehensively investigate the design space of TreeG over the candidate proposal module and the evaluation function, instantiating TreeG into three novel algorithms. Our experiments show that TreeG consistently outperforms the top guidance baselines in symbolic music generation, small molecule generation, and enhancer DNA design, all of which involve non-differentiable challenges. Additionally, we identify an inference-time scaling law showing TreeG's scalability in inference-time computation.
2502.11422
Planning of Heuristics: Strategic Planning on Large Language Models with Monte Carlo Tree Search for Automating Heuristic Optimization
cs.AI
Heuristics have achieved great success in solving combinatorial optimization problems (COPs). However, heuristics designed by humans require too much domain knowledge and testing time. Given the fact that Large Language Models (LLMs) possess strong capabilities to understand and generate content, and a knowledge base that covers various domains, which offer a novel way to automatically optimize heuristics. Therefore, we propose Planning of Heuristics (PoH), an optimization method that integrates the self-reflection of LLMs with the Monte Carlo Tree Search (MCTS), a well-known planning algorithm. PoH iteratively refines generated heuristics by evaluating their performance and providing improvement suggestions. Our method enables to iteratively evaluate the generated heuristics (states) and improve them based on the improvement suggestions (actions) and evaluation results (rewards), by effectively simulating future states to search for paths with higher rewards. In this paper, we apply PoH to solve the Traveling Salesman Problem (TSP) and the Flow Shop Scheduling Problem (FSSP). The experimental results show that PoH outperforms other hand-crafted heuristics and Automatic Heuristic Design (AHD) by other LLMs-based methods, and achieves the significant improvements and the state-of-the-art performance of our proposed method in automating heuristic optimization with LLMs to solve COPs.
2502.11423
Exploring Persona Sentiment Sensitivity in Personalized Dialogue Generation
cs.CL
Personalized dialogue systems have advanced considerably with the integration of user-specific personas into large language models (LLMs). However, while LLMs can effectively generate personalized responses, the influence of persona sentiment on dialogue quality remains underexplored. In this work, we conduct a large-scale analysis of dialogues generated using a range of polarized user profiles. Our experiments reveal that dialogues involving negatively polarized users tend to overemphasize persona attributes, leading to increased entailment and contradiction instances and lower overall coherence. In contrast, positively polarized profiles yield dialogues that selectively incorporate persona information, resulting in smoother and more coherent interactions. Furthermore, we find that personas with weak or neutral sentiment generally produce lower-quality dialogues. Motivated by these findings, we propose a dialogue generation approach that explicitly accounts for persona polarity by combining a turn-based generation strategy with a profile ordering mechanism. Our study provides new insights into the sensitivity of LLMs to persona sentiment and offers guidance for developing more robust and nuanced personalized dialogue systems.
2502.11425
Counterfactual-Consistency Prompting for Relative Temporal Understanding in Large Language Models
cs.CL cs.AI
Despite the advanced capabilities of large language models (LLMs), their temporal reasoning ability remains underdeveloped. Prior works have highlighted this limitation, particularly in maintaining temporal consistency when understanding events. For example, models often confuse mutually exclusive temporal relations like ``before'' and ``after'' between events and make inconsistent predictions. In this work, we tackle the issue of temporal inconsistency in LLMs by proposing a novel counterfactual prompting approach. Our method generates counterfactual questions and enforces collective constraints, enhancing the model's consistency. We evaluate our method on multiple datasets, demonstrating significant improvements in event ordering for explicit and implicit events and temporal commonsense understanding by effectively addressing temporal inconsistencies.
2502.11426
Verti-Bench: A General and Scalable Off-Road Mobility Benchmark for Vertically Challenging Terrain
cs.RO
Recent advancement in off-road autonomy has shown promises in deploying autonomous mobile robots in outdoor off-road environments. Encouraging results have been reported from both simulated and real-world experiments. However, unlike evaluating off-road perception tasks on static datasets, benchmarking off-road mobility still faces significant challenges due to a variety of factors, including variations in vehicle platforms and terrain properties. Furthermore, different vehicle-terrain interactions need to be unfolded during mobility evaluation, which requires the mobility systems to interact with the environments instead of comparing against a pre-collected dataset. In this paper, we present Verti-Bench, a mobility benchmark that focuses on extremely rugged, vertically challenging off-road environments. 100 unique off-road environments and 1000 distinct navigation tasks with millions of off-road terrain properties, including a variety of geometry and semantics, rigid and deformable surfaces, and large natural obstacles, provide standardized and objective evaluation in high-fidelity multi-physics simulation. Verti-Bench is also scalable to various vehicle platforms with different scales and actuation mechanisms. We also provide datasets from expert demonstration, random exploration, failure cases (rolling over and getting stuck), as well as a gym-like interface for reinforcement learning. We use Verti-Bench to benchmark ten off-road mobility systems, present our findings, and identify future off-road mobility research directions.
2502.11427
Do we Really Need Visual Instructions? Towards Visual Instruction-Free Fine-tuning for Large Vision-Language Models
cs.CL cs.CV
Visual instruction tuning has become the predominant technology in eliciting the multimodal task-solving capabilities of large vision-language models (LVLMs). Despite the success, as visual instructions require images as the input, it would leave the gap in inheriting the task-solving capabilities from the backbone LLMs, and make it costly to collect a large-scale dataset. To address it, we propose ViFT, a visual instruction-free fine-tuning framework for LVLMs. In ViFT, we only require the text-only instructions and image caption data during training, to separately learn the task-solving and visual perception abilities. During inference, we extract and combine the representations of the text and image inputs, for fusing the two abilities to fulfill multimodal tasks. Experimental results demonstrate that ViFT can achieve state-of-the-art performance on several visual reasoning and visual instruction following benchmarks, with rather less training data. Our code and data will be publicly released.
2502.11429
What's in a Query: Polarity-Aware Distribution-Based Fair Ranking
cs.LG cs.CY
Machine learning-driven rankings, where individuals (or items) are ranked in response to a query, mediate search exposure or attention in a variety of safety-critical settings. Thus, it is important to ensure that such rankings are fair. Under the goal of equal opportunity, attention allocated to an individual on a ranking interface should be proportional to their relevance across search queries. In this work, we examine amortized fair ranking -- where relevance and attention are cumulated over a sequence of user queries to make fair ranking more feasible in practice. Unlike prior methods that operate on expected amortized attention for each individual, we define new divergence-based measures for attention distribution-based fairness in ranking (DistFaiR), characterizing unfairness as the divergence between the distribution of attention and relevance corresponding to an individual over time. This allows us to propose new definitions of unfairness, which are more reliable at test time. Second, we prove that group fairness is upper-bounded by individual fairness under this definition for a useful class of divergence measures, and experimentally show that maximizing individual fairness through an integer linear programming-based optimization is often beneficial to group fairness. Lastly, we find that prior research in amortized fair ranking ignores critical information about queries, potentially leading to a fairwashing risk in practice by making rankings appear more fair than they actually are.
2502.11431
Any Information Is Just Worth One Single Screenshot: Unifying Search With Visualized Information Retrieval
cs.CL
With the popularity of multimodal techniques, it receives growing interests to acquire useful information in visual forms. In this work, we formally define an emerging IR paradigm called \textit{Visualized Information Retrieval}, or \textbf{Vis-IR}, where multimodal information, such as texts, images, tables and charts, is jointly represented by a unified visual format called \textbf{Screenshots}, for various retrieval applications. We further make three key contributions for Vis-IR. First, we create \textbf{VIRA} (Vis-IR Aggregation), a large-scale dataset comprising a vast collection of screenshots from diverse sources, carefully curated into captioned and question-answer formats. Second, we develop \textbf{UniSE} (Universal Screenshot Embeddings), a family of retrieval models that enable screenshots to query or be queried across arbitrary data modalities. Finally, we construct \textbf{MVRB} (Massive Visualized IR Benchmark), a comprehensive benchmark covering a variety of task forms and application scenarios. Through extensive evaluations on MVRB, we highlight the deficiency from existing multimodal retrievers and the substantial improvements made by UniSE. Our work will be shared with the community, laying a solid foundation for this emerging field.
2502.11433
FLAG-Trader: Fusion LLM-Agent with Gradient-based Reinforcement Learning for Financial Trading
cs.AI cs.CE q-fin.TR
Large language models (LLMs) fine-tuned on multimodal financial data have demonstrated impressive reasoning capabilities in various financial tasks. However, they often struggle with multi-step, goal-oriented scenarios in interactive financial markets, such as trading, where complex agentic approaches are required to improve decision-making. To address this, we propose \textsc{FLAG-Trader}, a unified architecture integrating linguistic processing (via LLMs) with gradient-driven reinforcement learning (RL) policy optimization, in which a partially fine-tuned LLM acts as the policy network, leveraging pre-trained knowledge while adapting to the financial domain through parameter-efficient fine-tuning. Through policy gradient optimization driven by trading rewards, our framework not only enhances LLM performance in trading but also improves results on other financial-domain tasks. We present extensive empirical evidence to validate these enhancements.
2502.11435
SMART: Self-Aware Agent for Tool Overuse Mitigation
cs.AI cs.CL cs.LG
Current Large Language Model (LLM) agents demonstrate strong reasoning and tool use capabilities, but often lack self-awareness, failing to balance these approaches effectively. This imbalance leads to Tool Overuse, where models unnecessarily rely on external tools for tasks solvable with parametric knowledge, increasing computational overhead. Inspired by human metacognition, we introduce SMART (Strategic Model-Aware Reasoning with Tools), a paradigm that enhances an agent's self-awareness to optimize task handling and reduce tool overuse. To support this paradigm, we introduce SMART-ER, a dataset spanning three domains, where reasoning alternates between parametric knowledge and tool-dependent steps, with each step enriched by rationales explaining when tools are necessary. Through supervised training, we develop SMARTAgent, a family of models that dynamically balance parametric knowledge and tool use. Evaluations show that SMARTAgent reduces tool use by 24% while improving performance by over 37%, enabling 7B-scale models to match its 70B counterpart and GPT-4o. Additionally, SMARTAgent generalizes to out-of-distribution test data like GSM8K and MINTQA, maintaining accuracy with just one-fifth the tool calls. These highlight the potential of strategic tool use to enhance reasoning, mitigate overuse, and bridge the gap between model size and performance, advancing intelligent and resource-efficient agent designs.
2502.11436
ADO: Automatic Data Optimization for Inputs in LLM Prompts
cs.LG
This study explores a novel approach to enhance the performance of Large Language Models (LLMs) through the optimization of input data within prompts. While previous research has primarily focused on refining instruction components and augmenting input data with in-context examples, our work investigates the potential benefits of optimizing the input data itself. We introduce a two-pronged strategy for input data optimization: content engineering and structural reformulation. Content engineering involves imputing missing values, removing irrelevant attributes, and enriching profiles by generating additional information inferred from existing attributes. Subsequent to content engineering, structural reformulation is applied to optimize the presentation of the modified content to LLMs, given their sensitivity to input format. Our findings suggest that these optimizations can significantly improve the performance of LLMs in various tasks, offering a promising avenue for future research in prompt engineering. The source code is available at https://anonymous.4open.science/r/ADO-6BC5/
2502.11437
Learning Dexterous Bimanual Catch Skills through Adversarial-Cooperative Heterogeneous-Agent Reinforcement Learning
cs.RO cs.AI
Robotic catching has traditionally focused on single-handed systems, which are limited in their ability to handle larger or more complex objects. In contrast, bimanual catching offers significant potential for improved dexterity and object handling but introduces new challenges in coordination and control. In this paper, we propose a novel framework for learning dexterous bimanual catching skills using Heterogeneous-Agent Reinforcement Learning (HARL). Our approach introduces an adversarial reward scheme, where a throw agent increases the difficulty of throws-adjusting speed-while a catch agent learns to coordinate both hands to catch objects under these evolving conditions. We evaluate the framework in simulated environments using 15 different objects, demonstrating robustness and versatility in handling diverse objects. Our method achieved approximately a 2x increase in catching reward compared to single-agent baselines across 15 diverse objects.
2502.11438
SAFE-SQL: Self-Augmented In-Context Learning with Fine-grained Example Selection for Text-to-SQL
cs.CL
Text-to-SQL aims to convert natural language questions into executable SQL queries. While previous approaches, such as skeleton-masked selection, have demonstrated strong performance by retrieving similar training examples to guide large language models (LLMs), they struggle in real-world scenarios where such examples are unavailable. To overcome this limitation, we propose Self-Augmentation in-context learning with Fine-grained Example selection for Text-to-SQL (SAFE-SQL), a novel framework that improves SQL generation by generating and filtering self-augmented examples. SAFE-SQL first prompts an LLM to generate multiple Text-to-SQL examples relevant to the test input. Then SAFE-SQL filters these examples through three relevance assessments, constructing high-quality in-context learning examples. Using self-generated examples, SAFE-SQL surpasses the previous zero-shot, and few-shot Text-to-SQL frameworks, achieving higher execution accuracy. Notably, our approach provides additional performance gains in extra hard and unseen scenarios, where conventional methods often fail.
2502.11439
An Efficient Row-Based Sparse Fine-Tuning
cs.CL cs.AI cs.LG
Fine-tuning is an important step in adapting foundation models such as large language models to downstream tasks. To make this step more accessible to users with limited computational budgets, it is crucial to develop fine-tuning methods that are memory and computationally efficient. Sparse Fine-tuning (SFT) and Low-rank adaptation (LoRA) are two frameworks that have emerged for addressing this problem and have been adopted widely in practice. In this work, we develop a new SFT framework, based on ideas from neural network pruning. At a high level, we first identify "important" neurons/nodes using feature importance metrics from network pruning (specifically, we use the structural pruning method), and then perform fine-tuning by restricting to weights involving these neurons. Using experiments on common language tasks, we demonstrate that our method significantly improves the memory efficiency of SFT without increasing training time complexity and implementation complexity, while achieving accuracy comparable to state-of-the-art methods such as LoRA and its variants.
2502.11440
Medical Image Registration Meets Vision Foundation Model: Prototype Learning and Contour Awareness
cs.CV
Medical image registration is a fundamental task in medical image analysis, aiming to establish spatial correspondences between paired images. However, existing unsupervised deformable registration methods rely solely on intensity-based similarity metrics, lacking explicit anatomical knowledge, which limits their accuracy and robustness. Vision foundation models, such as the Segment Anything Model (SAM), can generate high-quality segmentation masks that provide explicit anatomical structure knowledge, addressing the limitations of traditional methods that depend only on intensity similarity. Based on this, we propose a novel SAM-assisted registration framework incorporating prototype learning and contour awareness. The framework includes: (1) Explicit anatomical information injection, where SAM-generated segmentation masks are used as auxiliary inputs throughout training and testing to ensure the consistency of anatomical information; (2) Prototype learning, which leverages segmentation masks to extract prototype features and aligns prototypes to optimize semantic correspondences between images; and (3) Contour-aware loss, a contour-aware loss is designed that leverages the edges of segmentation masks to improve the model's performance in fine-grained deformation fields. Extensive experiments demonstrate that the proposed framework significantly outperforms existing methods across multiple datasets, particularly in challenging scenarios with complex anatomical structures and ambiguous boundaries. Our code is available at https://github.com/HaoXu0507/IPMI25-SAM-Assisted-Registration.
2502.11441
Which Retain Set Matters for LLM Unlearning? A Case Study on Entity Unlearning
cs.CL
Large language models (LLMs) risk retaining unauthorized or sensitive information from their training data, which raises privacy concerns. LLM unlearning seeks to mitigate these risks by selectively removing specified data while maintaining overall model performance. However, most existing work focus on methods to achieve effective forgetting and does not provide a detailed analysis of the retain set, the portion of training data that is not targeted for removal. In this paper, we investigate the effects of unlearning on various subsets of the retain set through a case study on entity unlearning. We introduce the Syntactically Similar Neighbor Set, a group of queries that share similar syntactic structures with the data targeted for removal, and show that this subset suffers the greatest performance drop during unlearning. Moreover, when used for regularization, this set not only preserves performance on syntactically similar queries but also delivers comparable or improved results across other data subsets. Our results highlight that syntactic similarity is a critical factor, potentially more so than domain or entity relationships, in achieving effective and practical LLM unlearning.
2502.11442
Multi-Turn Multi-Modal Question Clarification for Enhanced Conversational Understanding
cs.IR cs.AI cs.CL cs.LG
Conversational query clarification enables users to refine their search queries through interactive dialogue, improving search effectiveness. Traditional approaches rely on text-based clarifying questions, which often fail to capture complex user preferences, particularly those involving visual attributes. While recent work has explored single-turn multi-modal clarification with images alongside text, such methods do not fully support the progressive nature of user intent refinement over multiple turns. Motivated by this, we introduce the Multi-turn Multi-modal Clarifying Questions (MMCQ) task, which combines text and visual modalities to refine user queries in a multi-turn conversation. To facilitate this task, we create a large-scale dataset named ClariMM comprising over 13k multi-turn interactions and 33k question-answer pairs containing multi-modal clarifying questions. We propose Mario, a retrieval framework that employs a two-phase ranking strategy: initial retrieval with BM25, followed by a multi-modal generative re-ranking model that integrates textual and visual information from conversational history. Our experiments show that multi-turn multi-modal clarification outperforms uni-modal and single-turn approaches, improving MRR by 12.88%. The gains are most significant in longer interactions, demonstrating the value of progressive refinement for complex queries.
2502.11444
Does RAG Really Perform Bad For Long-Context Processing?
cs.CL
The efficient processing of long context poses a serious challenge for large language models (LLMs). Recently, retrieval-augmented generation (RAG) has emerged as a promising strategy for this problem, as it enables LLMs to make selective use of the long context for efficient computation. However, existing RAG approaches lag behind other long-context processing methods due to inherent limitations on inaccurate retrieval and fragmented contexts. To address these challenges, we introduce RetroLM, a novel RAG framework for long-context processing. Unlike traditional methods, RetroLM employs KV-level retrieval augmentation, where it partitions the LLM's KV cache into contiguous pages and retrieves the most crucial ones for efficient computation. This approach enhances robustness to retrieval inaccuracy, facilitates effective utilization of fragmented contexts, and saves the cost from repeated computation. Building on this framework, we further develop a specialized retriever for precise retrieval of critical pages and conduct unsupervised post-training to optimize the model's ability to leverage retrieved information. We conduct comprehensive evaluations with a variety of benchmarks, including LongBench, InfiniteBench, and RULER, where RetroLM significantly outperforms existing long-context LLMs and efficient long-context processing methods, particularly in tasks requiring intensive reasoning or extremely long-context comprehension.
2502.11447
Does Editing Provide Evidence for Localization?
cs.LG cs.AI
A basic aspiration for interpretability research in large language models is to "localize" semantically meaningful behaviors to particular components within the LLM. There are various heuristics for finding candidate locations within the LLM. Once a candidate localization is found, it can be assessed by editing the internal representations at the corresponding localization and checking whether this induces model behavior that is consistent with the semantic interpretation of the localization. The question we address here is: how strong is the evidence provided by such edits? To evaluate the localization claim, we want to assess the effect of the optimal intervention at a particular location. The key new technical tool is a way of adapting LLM alignment techniques to find such optimal localized edits. With this tool in hand, we give an example where the edit-based evidence for localization appears strong, but where localization clearly fails. Indeed, we find that optimal edits at random localizations can be as effective as aligning the full model. In aggregate, our results suggest that merely observing that localized edits induce targeted changes in behavior provides little to no evidence that these locations actually encode the target behavior.
2502.11448
AGrail: A Lifelong Agent Guardrail with Effective and Adaptive Safety Detection
cs.AI
The rapid advancements in Large Language Models (LLMs) have enabled their deployment as autonomous agents for handling complex tasks in dynamic environments. These LLMs demonstrate strong problem-solving capabilities and adaptability to multifaceted scenarios. However, their use as agents also introduces significant risks, including task-specific risks, which are identified by the agent administrator based on the specific task requirements and constraints, and systemic risks, which stem from vulnerabilities in their design or interactions, potentially compromising confidentiality, integrity, or availability (CIA) of information and triggering security risks. Existing defense agencies fail to adaptively and effectively mitigate these risks. In this paper, we propose AGrail, a lifelong agent guardrail to enhance LLM agent safety, which features adaptive safety check generation, effective safety check optimization, and tool compatibility and flexibility. Extensive experiments demonstrate that AGrail not only achieves strong performance against task-specific and system risks but also exhibits transferability across different LLM agents' tasks.
2502.11449
Tractable General Equilibrium
cs.GT cs.CE econ.TH
We study Walrasian economies (or general equilibrium models) and their solution concept, the Walrasian equilibrium. A key challenge in this domain is identifying price-adjustment processes that converge to equilibrium. One such process, t\^atonnement, is an auction-like algorithm first proposed in 1874 by L\'eon Walras. While continuous-time variants of t\^atonnement are known to converge to equilibrium in economies satisfying the Weak Axiom of Revealed Preferences (WARP), the process fails to converge in a pathological Walrasian economy known as the Scarf economy. To address these issues, we analyze Walrasian economies using variational inequalities (VIs), an optimization framework. We introduce the class of mirror extragradient algorithms, which, under suitable Lipschitz-continuity-like assumptions, converge to a solution of any VI satisfying the Minty condition in polynomial time. We show that the set of Walrasian equilibria of any balanced economy-which includes among others Arrow-Debreu economies-corresponds to the solution set of an associated VI that satisfies the Minty condition but is generally discontinuous. Applying the mirror extragradient algorithm to this VI we obtain a class of t\^atonnement-like processes, which we call the mirror extrat\^atonnement process. While our VI formulation is generally discontinuous, it is Lipschitz-continuous in variationally stable Walrasian economies with bounded elasticity-including those satisfying WARP and the Scarf economy-thus establishing the polynomial-time convergence of mirror extrat\^atonnement in these economies. We validate our approach through experiments on large Arrow-Debreu economies with Cobb-Douglas, Leontief, and CES consumers, as well as the Scarf economy, demonstrating fast convergence in all cases without failure.
2502.11450
Fishing For Cheap And Efficient Pruners At Initialization
cs.LG cs.AI
Pruning offers a promising solution to mitigate the associated costs and environmental impact of deploying large deep neural networks (DNNs). Traditional approaches rely on computationally expensive trained models or time-consuming iterative prune-retrain cycles, undermining their utility in resource-constrained settings. To address this issue, we build upon the established principles of saliency (LeCun et al., 1989) and connection sensitivity (Lee et al., 2018) to tackle the challenging problem of one-shot pruning neural networks (NNs) before training (PBT) at initialization. We introduce Fisher-Taylor Sensitivity (FTS), a computationally cheap and efficient pruning criterion based on the empirical Fisher Information Matrix (FIM) diagonal, offering a viable alternative for integrating first- and second-order information to identify a model's structurally important parameters. Although the FIM-Hessian equivalency only holds for convergent models that maximize the likelihood, recent studies (Karakida et al., 2019) suggest that, even at initialization, the FIM captures essential geometric information of parameters in overparameterized NNs, providing the basis for our method. Finally, we demonstrate empirically that layer collapse, a critical limitation of data-dependent pruning methodologies, is easily overcome by pruning within a single training epoch after initialization. We perform experiments on ResNet18 and VGG19 with CIFAR-10 and CIFAR-100, widely used benchmarks in pruning research. Our method achieves competitive performance against state-of-the-art techniques for one-shot PBT, even under extreme sparsity conditions. Our code is made available to the public.
2502.11451
From Personas to Talks: Revisiting the Impact of Personas on LLM-Synthesized Emotional Support Conversations
cs.CL
The rapid advancement of Large Language Models (LLMs) has revolutionized the generation of emotional support conversations (ESC), offering scalable solutions with reduced costs and enhanced data privacy. This paper explores the role of personas in the creation of ESC by LLMs. Our research utilizes established psychological frameworks to measure and infuse persona traits into LLMs, which then generate dialogues in the emotional support scenario. We conduct extensive evaluations to understand the stability of persona traits in dialogues, examining shifts in traits post-generation and their impact on dialogue quality and strategy distribution. Experimental results reveal several notable findings: 1) LLMs can infer core persona traits, 2) subtle shifts in emotionality and extraversion occur, influencing the dialogue dynamics, and 3) the application of persona traits modifies the distribution of emotional support strategies, enhancing the relevance and empathetic quality of the responses. These findings highlight the potential of persona-driven LLMs in crafting more personalized, empathetic, and effective emotional support dialogues, which has significant implications for the future design of AI-driven emotional support systems.
2502.11453
Connector-S: A Survey of Connectors in Multi-modal Large Language Models
cs.LG cs.AI
With the rapid advancements in multi-modal large language models (MLLMs), connectors play a pivotal role in bridging diverse modalities and enhancing model performance. However, the design and evolution of connectors have not been comprehensively analyzed, leaving gaps in understanding how these components function and hindering the development of more powerful connectors. In this survey, we systematically review the current progress of connectors in MLLMs and present a structured taxonomy that categorizes connectors into atomic operations (mapping, compression, mixture of experts) and holistic designs (multi-layer, multi-encoder, multi-modal scenarios), highlighting their technical contributions and advancements. Furthermore, we discuss several promising research frontiers and challenges, including high-resolution input, dynamic compression, guide information selection, combination strategy, and interpretability. This survey is intended to serve as a foundational reference and a clear roadmap for researchers, providing valuable insights into the design and optimization of next-generation connectors to enhance the performance and adaptability of MLLMs.
2502.11454
UniCBE: An Uniformity-driven Comparing Based Evaluation Framework with Unified Multi-Objective Optimization
cs.CL
Human preference plays a significant role in measuring large language models and guiding them to align with human values. Unfortunately, current comparing-based evaluation (CBE) methods typically focus on a single optimization objective, failing to effectively utilize scarce yet valuable preference signals. To address this, we delve into key factors that can enhance the accuracy, convergence, and scalability of CBE: suppressing sampling bias, balancing descending process of uncertainty, and mitigating updating uncertainty. Following the derived guidelines, we propose UniCBE, a unified uniformity-driven CBE framework which simultaneously optimize these core objectives by constructing and integrating three decoupled sampling probability matrices, each designed to ensure uniformity in specific aspects. We further ablate the optimal tuple sampling and preference aggregation strategies to achieve efficient CBE. On the AlpacaEval benchmark, UniCBE saves over 17% of evaluation budgets while achieving a Pearson correlation with ground truth exceeding 0.995, demonstrating excellent accuracy and convergence. In scenarios where new models are continuously introduced, UniCBE can even save over 50% of evaluation costs, highlighting its improved scalability.
2502.11456
Leveraging Labelled Data Knowledge: A Cooperative Rectification Learning Network for Semi-supervised 3D Medical Image Segmentation
cs.CV cs.AI
Semi-supervised 3D medical image segmentation aims to achieve accurate segmentation using few labelled data and numerous unlabelled data. The main challenge in the design of semi-supervised learning methods consists in the effective use of the unlabelled data for training. A promising solution consists of ensuring consistent predictions across different views of the data, where the efficacy of this strategy depends on the accuracy of the pseudo-labels generated by the model for this consistency learning strategy. In this paper, we introduce a new methodology to produce high-quality pseudo-labels for a consistency learning strategy to address semi-supervised 3D medical image segmentation. The methodology has three important contributions. The first contribution is the Cooperative Rectification Learning Network (CRLN) that learns multiple prototypes per class to be used as external knowledge priors to adaptively rectify pseudo-labels at the voxel level. The second contribution consists of the Dynamic Interaction Module (DIM) to facilitate pairwise and cross-class interactions between prototypes and multi-resolution image features, enabling the production of accurate voxel-level clues for pseudo-label rectification. The third contribution is the Cooperative Positive Supervision (CPS), which optimises uncertain representations to align with unassertive representations of their class distributions, improving the model's accuracy in classifying uncertain regions. Extensive experiments on three public 3D medical segmentation datasets demonstrate the effectiveness and superiority of our semi-supervised learning method.
2502.11457
Aligning Sentence Simplification with ESL Learner's Proficiency for Language Acquisition
cs.CL cs.AI
Text simplification is crucial for improving accessibility and comprehension for English as a Second Language (ESL) learners. This study goes a step further and aims to facilitate ESL learners' language acquisition by simplification. Specifically, we propose simplifying complex sentences to appropriate levels for learners while also increasing vocabulary coverage of the target level in the simplifications. We achieve this without a parallel corpus by conducting reinforcement learning on a large language model. Our method employs token-level and sentence-level rewards, and iteratively trains the model on its self-generated outputs to guide the model to search for simplification hypotheses that satisfy the target attributes. Experiment results on CEFR-SP and TurkCorpus datasets show that the proposed method can effectively increase the frequency and diversity of vocabulary of the target level by more than $20\%$ compared to baseline models, while maintaining high simplification quality.
2502.11458
Towards Efficient Pre-training: Exploring FP4 Precision in Large Language Models
cs.LG cs.AI
The burgeoning computational demands for training large language models (LLMs) necessitate efficient methods, including quantized training, which leverages low-bit arithmetic operations to reduce costs. While FP8 precision has shown potential, leveraging FP4 remains challenging due to inherent quantization errors and limited representation capability. Based on the Transformer architecture, we present an FP4 training scheme for LLMs, overcoming these obstacles through mixed-precision quantization strategies tailed for different modules and training stages. This allows us to apply the precision level suitable to distinct components within the model, ensuring that multi-head attention and linear layers are handled appropriately. Our pretraining recipe ensures stability in backpropagation by incorporating fine-grained quantization methods with a target precision training schedule. Experimental results demonstrate that our FP4 training scheme achieves accuracy comparable to BF16 and FP8, with smaller theoretical computational cost. With the advent of next-generation hardware supporting FP4, our method sets the foundation for efficient ultra-low precision training.
2502.11459
Towards Responsible and Fair Data Science: Resource Allocation for Inclusive and Sustainable Analytics
cs.DB
This project addresses the challenges of responsible and fair resource allocation in data science (DS), focusing on DS queries evaluation. Current DS practices often overlook the broader socio-economic, environmental, and ethical implications, including data sovereignty, fairness, and inclusivity. By integrating a decolonial perspective, the project aims to establish innovative fairness metrics that respect cultural and contextual diversity, optimise computational and energy efficiency, and ensure equitable participation of underrepresented communities. The research includes developing algorithms to align resource allocation with fairness constraints, incorporating ethical and sustainability considerations, and fostering interdisciplinary collaborations to bridge technical advancements and societal impact gaps. This work aims to reshape into an equitable, transparent, and community-empowering practice challenging the technological power developed by the Big Tech.
2502.11460
UnitCoder: Scalable Iterative Code Synthesis with Unit Test Guidance
cs.CL cs.SE
Large Language Models (LLMs) have demonstrated remarkable capabilities in various tasks, yet code generation remains a major challenge. Current approaches for obtaining high-quality code data primarily focus on (i) collecting large-scale pre-training data and (ii) synthesizing instruction data through prompt engineering with powerful models. While pre-training data faces quality consistency issues, instruction-based synthesis suffers from limited instruction diversity and inherent biases of LLMs. To address this gap, we introduce UnitCoder, a systematic pipeline leveraging model-generated unit tests to both guide and validate the code generation process. Combined with large-scale package-based retrieval from pre-training corpus, we generate a dataset of 500K+ verifiable programs containing diverse API calls. Evaluations on multiple Python benchmarks (BigCodeBench, HumanEval, MBPP) demonstrate that models fine-tuned on our synthetic data exhibit consistent performance improvements. Notably, Llama3.1-8B and InternLM2.5-7B improve from 31\% and 28\% to 40\% and 39\% success rates on BigCodeBench, respectively. Our work presents a scalable approach that leverages model-generated unit tests to guide the synthesis of high-quality code data from pre-training corpora, demonstrating the potential for producing diverse and high-quality post-training data at scale. All code and data will be released (https://github.com).
2502.11461
Doppler Correspondence: Non-Iterative Scan Matching With Doppler Velocity-Based Correspondence
cs.RO
Achieving successful scan matching is essential for LiDAR odometry. However, in challenging environments with adverse weather conditions or repetitive geometric patterns, LiDAR odometry performance is degraded due to incorrect scan matching. Recently, the emergence of frequency-modulated continuous wave 4D LiDAR and 4D radar technologies has provided the potential to address these unfavorable conditions. The term 4D refers to point cloud data characterized by range, azimuth, and elevation along with Doppler velocity. Although 4D data is available, most scan matching methods for 4D LiDAR and 4D radar still establish correspondence by repeatedly identifying the closest points between consecutive scans, overlooking the Doppler information. This paper introduces, for the first time, a simple Doppler velocity-based correspondence -- Doppler Correspondence -- that is invariant to translation and small rotation of the sensor, with its geometric and kinematic foundations. Extensive experiments demonstrate that the proposed method enables the direct matching of consecutive point clouds without an iterative process, making it computationally efficient. Additionally, it provides a more robust correspondence estimation in environments with repetitive geometric patterns.
2502.11462
LMFCA-Net: A Lightweight Model for Multi-Channel Speech Enhancement with Efficient Narrow-Band and Cross-Band Attention
eess.AS cs.LG cs.SD
Deep learning based end-to-end multi-channel speech enhancement methods have achieved impressive performance by leveraging sub-band, cross-band, and spatial information. However, these methods often demand substantial computational resources, limiting their practicality on terminal devices. This paper presents a lightweight multi-channel speech enhancement network with decoupled fully connected attention (LMFCA-Net). The proposed LMFCA-Net introduces time-axis decoupled fully-connected attention (T-FCA) and frequency-axis decoupled fully-connected attention (F-FCA) mechanisms to effectively capture long-range narrow-band and cross-band information without recurrent units. Experimental results show that LMFCA-Net performs comparably to state-of-the-art methods while significantly reducing computational complexity and latency, making it a promising solution for practical applications.
2502.11465
All Models Are Miscalibrated, But Some Less So: Comparing Calibration with Conditional Mean Operators
stat.ML cs.LG
When working in a high-risk setting, having well calibrated probabilistic predictive models is a crucial requirement. However, estimators for calibration error are not always able to correctly distinguish which model is better calibrated. We propose the \emph{conditional kernel calibration error} (CKCE) which is based on the Hilbert-Schmidt norm of the difference between conditional mean operators. By working directly with the definition of strong calibration as the distance between conditional distributions, which we represent by their embeddings in reproducing kernel Hilbert spaces, the CKCE is less sensitive to the marginal distribution of predictive models. This makes it more effective for relative comparisons than previously proposed calibration metrics. Our experiments, using both synthetic and real data, show that CKCE provides a more consistent ranking of models by their calibration error and is more robust against distribution shift.
2502.11466
GiFT: Gibbs Fine-Tuning for Code Generation
cs.LG cs.CL cs.SE
Training Large Language Models (LLMs) with synthetic data is a prevalent practice in code generation. A key approach is self-training, where LLMs are iteratively trained on self-generated correct code snippets. In this case, the self-generated codes are drawn from a conditional distribution, conditioned on a specific seed description. However, the seed description is not the only valid representation that aligns with its intended meaning. With all valid descriptions and codes forming a joint space, codes drawn from the conditional distribution would lead to an underrepresentation of the full description-code space. As such, we propose Gibbs Fine-Tuning (GiFT), a novel self-training method inspired by Gibbs sampling. GiFT allows self-generated data to be drawn from the marginal distribution of the joint space, thereby mitigating the biases inherent in conditional sampling. We provide a theoretical analysis demonstrating the potential benefits of fine-tuning LLMs with code derived from the marginal distribution. Furthermore, we propose a perplexity-based code selection method to mitigate the imbalanced long-tail distribution of the self-generated codes. Empirical evaluation of two LLMs across four datasets demonstrates that GiFT achieves superior performance, particularly on more challenging benchmarks.
2502.11467
Approximation of Permutation Invariant Polynomials by Transformers: Efficient Construction in Column-Size
cs.LG math.FA
Transformers are a type of neural network that have demonstrated remarkable performance across various domains, particularly in natural language processing tasks. Motivated by this success, research on the theoretical understanding of transformers has garnered significant attention. A notable example is the mathematical analysis of their approximation power, which validates the empirical expressive capability of transformers. In this study, we investigate the ability of transformers to approximate column-symmetric polynomials, an extension of symmetric polynomials that take matrices as input. Consequently, we establish an explicit relationship between the size of the transformer network and its approximation capability, leveraging the parameter efficiency of transformers and their compatibility with symmetry by focusing on the algebraic properties of symmetric polynomials.
2502.11468
Semantically Robust Unsupervised Image Translation for Paired Remote Sensing Images
cs.CV
Image translation for change detection or classification in bi-temporal remote sensing images is unique. Although it can acquire paired images, it is still unsupervised. Moreover, strict semantic preservation in translation is always needed instead of multimodal outputs. In response to these problems, this paper proposes a new method, SRUIT (Semantically Robust Unsupervised Image-to-image Translation), which ensures semantically robust translation and produces deterministic output. Inspired by previous works, the method explores the underlying characteristics of bi-temporal Remote Sensing images and designs the corresponding networks. Firstly, we assume that bi-temporal Remote Sensing images share the same latent space, for they are always acquired from the same land location. So SRUIT makes the generators share their high-level layers, and this constraint will compel two domain mapping to fall into the same latent space. Secondly, considering land covers of bi-temporal images could evolve into each other, SRUIT exploits the cross-cycle-consistent adversarial networks to translate from one to the other and recover them. Experimental results show that constraints of sharing weights and cross-cycle consistency enable translated images with both good perceptual image quality and semantic preservation for significant differences.
2502.11469
If Attention Serves as a Cognitive Model of Human Memory Retrieval, What is the Plausible Memory Representation?
cs.CL
Recent work in computational psycholinguistics has revealed intriguing parallels between attention mechanisms and human memory retrieval, focusing primarily on Transformer architectures that operate on token-level representations. However, computational psycholinguistic research has also established that syntactic structures provide compelling explanations for human sentence processing that word-level factors alone cannot fully account for. In this study, we investigate whether the attention mechanism of Transformer Grammar (TG), which uniquely operates on syntactic structures as representational units, can serve as a cognitive model of human memory retrieval, using Normalized Attention Entropy (NAE) as a linking hypothesis between model behavior and human processing difficulty. Our experiments demonstrate that TG's attention achieves superior predictive power for self-paced reading times compared to vanilla Transformer's, with further analyses revealing independent contributions from both models. These findings suggest that human sentence processing involves dual memory representations -- one based on syntactic structures and another on token sequences -- with attention serving as the general retrieval algorithm, while highlighting the importance of incorporating syntactic structures as representational units.
2502.11470
Optimized detection of cyber-attacks on IoT networks via hybrid deep learning models
cs.CR cs.AI
The rapid expansion of Internet of Things (IoT) devices has increased the risk of cyber-attacks, making effective detection essential for securing IoT networks. This work introduces a novel approach combining Self-Organizing Maps (SOMs), Deep Belief Networks (DBNs), and Autoencoders to detect known and previously unseen attack patterns. A comprehensive evaluation using simulated and real-world traffic data is conducted, with models optimized via Particle Swarm Optimization (PSO). The system achieves an accuracy of up to 99.99% and Matthews Correlation Coefficient (MCC) values exceeding 99.50%. Experiments on NSL-KDD, UNSW-NB15, and CICIoT2023 confirm the model's strong performance across diverse attack types. These findings suggest that the proposed method enhances IoT security by identifying emerging threats and adapting to evolving attack strategies.
2502.11471
GLTW: Joint Improved Graph Transformer and LLM via Three-Word Language for Knowledge Graph Completion
cs.CL cs.IR
Knowledge Graph Completion (KGC), which aims to infer missing or incomplete facts, is a crucial task for KGs. However, integrating the vital structural information of KGs into Large Language Models (LLMs) and outputting predictions deterministically remains challenging. To address this, we propose a new method called GLTW, which encodes the structural information of KGs and merges it with LLMs to enhance KGC performance. Specifically, we introduce an improved Graph Transformer (iGT) that effectively encodes subgraphs with both local and global structural information and inherits the characteristics of language model, bypassing training from scratch. Also, we develop a subgraph-based multi-classification training objective, using all entities within KG as classification objects, to boost learning efficiency.Importantly, we combine iGT with an LLM that takes KG language prompts as input.Our extensive experiments on various KG datasets show that GLTW achieves significant performance gains compared to SOTA baselines.
2502.11476
FastMCTS: A Simple Sampling Strategy for Data Synthesis
cs.CL
Synthetic high-quality multi-step reasoning data can significantly enhance the performance of large language models on various tasks. However, most existing methods rely on rejection sampling, which generates trajectories independently and suffers from inefficiency and imbalanced sampling across problems of varying difficulty. In this work, we introduce FastMCTS, an innovative data synthesis strategy inspired by Monte Carlo Tree Search. FastMCTS provides a more efficient sampling method for multi-step reasoning data, offering step-level evaluation signals and promoting balanced sampling across problems of different difficulty levels. Experiments on both English and Chinese reasoning datasets demonstrate that FastMCTS generates over 30\% more correct reasoning paths compared to rejection sampling as the number of generated tokens scales up. Furthermore, under comparable synthetic data budgets, models trained on FastMCTS-generated data outperform those trained on rejection sampling data by 3.9\% across multiple benchmarks. As a lightweight sampling strategy, FastMCTS offers a practical and efficient alternative for synthesizing high-quality reasoning data. Our code will be released soon.
2502.11477
Learning to Sample Effective and Diverse Prompts for Text-to-Image Generation
cs.CV
Recent advances in text-to-image diffusion models have achieved impressive image generation capabilities. However, it remains challenging to control the generation process with desired properties (e.g., aesthetic quality, user intention), which can be expressed as black-box reward functions. In this paper, we focus on prompt adaptation, which refines the original prompt into model-preferred prompts to generate desired images. While prior work uses reinforcement learning (RL) to optimize prompts, we observe that applying RL often results in generating similar postfixes and deterministic behaviors. To this end, we introduce \textbf{P}rompt \textbf{A}daptation with \textbf{G}FlowNets (\textbf{PAG}), a novel approach that frames prompt adaptation as a probabilistic inference problem. Our key insight is that leveraging Generative Flow Networks (GFlowNets) allows us to shift from reward maximization to sampling from an unnormalized density function, enabling both high-quality and diverse prompt generation. However, we identify that a naive application of GFlowNets suffers from mode collapse and uncovers a previously overlooked phenomenon: the progressive loss of neural plasticity in the model, which is compounded by inefficient credit assignment in sequential prompt generation. To address this critical challenge, we develop a systematic approach in PAG with flow reactivation, reward-prioritized sampling, and reward decomposition for prompt adaptation. Extensive experiments validate that PAG successfully learns to sample effective and diverse prompts for text-to-image generation. We also show that PAG exhibits strong robustness across various reward functions and transferability to different text-to-image models.
2502.11478
TAPS: Throat and Acoustic Paired Speech Dataset for Deep Learning-Based Speech Enhancement
cs.SD cs.LG eess.AS
In high-noise environments such as factories, subways, and busy streets, capturing clear speech is challenging due to background noise. Throat microphones provide a solution with their noise-suppressing properties, reducing the noise while recording speech. However, a significant limitation remains: high-frequency information is attenuated as sound waves pass through skin and tissue, reducing speech clarity. Recent deep learning approaches have shown promise in enhancing throat microphone recordings, but further progress is constrained by the absence of standardized dataset. We introduce a throat and acoustic paired speech dataset (TAPS), a collection of paired utterances recorded from 60 native Korean speakers using throat and acoustic microphones. To demonstrate the TAPS's utility, we tested three baseline deep learning models and identified the mapping-based approach as superior in improving speech quality and restoring content. Additionally, we propose an optimal method to mitigate the signal mismatch between throat and acoustic microphones, ensuring model performance. These results highlight the potential of TAPS to serve as a standardized dataset and advance research in throat microphone-based speech enhancement.
2502.11480
Enhancing Offline Model-Based RL via Active Model Selection: A Bayesian Optimization Perspective
cs.LG stat.ML
Offline model-based reinforcement learning (MBRL) serves as a competitive framework that can learn well-performing policies solely from pre-collected data with the help of learned dynamics models. To fully unleash the power of offline MBRL, model selection plays a pivotal role in determining the dynamics model utilized for downstream policy learning. However, offline MBRL conventionally relies on validation or off-policy evaluation, which are rather inaccurate due to the inherent distribution shift in offline RL. To tackle this, we propose BOMS, an active model selection framework that enhances model selection in offline MBRL with only a small online interaction budget, through the lens of Bayesian optimization (BO). Specifically, we recast model selection as BO and enable probabilistic inference in BOMS by proposing a novel model-induced kernel, which is theoretically grounded and computationally efficient. Through extensive experiments, we show that BOMS improves over the baseline methods with a small amount of online interaction comparable to only $1\%$-$2.5\%$ of offline training data on various RL tasks.
2502.11481
Variable-frame CNNLSTM for Breast Nodule Classification using Ultrasound Videos
cs.CV cs.AI
The intersection of medical imaging and artificial intelligence has become an important research direction in intelligent medical treatment, particularly in the analysis of medical images using deep learning for clinical diagnosis. Despite the advances, existing keyframe classification methods lack extraction of time series features, while ultrasonic video classification based on three-dimensional convolution requires uniform frame numbers across patients, resulting in poor feature extraction efficiency and model classification performance. This study proposes a novel video classification method based on CNN and LSTM, introducing NLP's long and short sentence processing scheme into video classification for the first time. The method reduces CNN-extracted image features to 1x512 dimension, followed by sorting and compressing feature vectors for LSTM training. Specifically, feature vectors are sorted by patient video frame numbers and populated with padding value 0 to form variable batches, with invalid padding values compressed before LSTM training to conserve computing resources. Experimental results demonstrate that our variable-frame CNNLSTM method outperforms other approaches across all metrics, showing improvements of 3-6% in F1 score and 1.5% in specificity compared to keyframe methods. The variable-frame CNNLSTM also achieves better accuracy and precision than equal-frame CNNLSTM. These findings validate the effectiveness of our approach in classifying variable-frame ultrasound videos and suggest potential applications in other medical imaging modalities.
2502.11482
DATA: Decomposed Attention-based Task Adaptation for Rehearsal-Free Continual Learning
cs.LG cs.AI cs.CL
Continual learning (CL) is essential for Large Language Models (LLMs) to adapt to evolving real-world demands, yet they are susceptible to catastrophic forgetting (CF). While traditional CF solutions rely on expensive data rehearsal, recent rehearsal-free methods employ model-based and regularization-based strategies to address this issue. However, these approaches often neglect the model's plasticity, which is crucial to achieving optimal performance on newly learned tasks. Consequently, a key challenge in CL is striking a balance between preserving plasticity and mitigating CF. To tackle this challenge, we propose the $\textbf{D}$ecomposed $\textbf{A}$ttention-based $\textbf{T}$ask $\textbf{A}$daptation (DATA), which explicitly decouples and learns both task-specific and task-shared knowledge using high-rank and low-rank task adapters (e.g., LoRAs). For new tasks, DATA dynamically adjusts the weights of adapters of different ranks based on their relevance and distinction from previous tasks, allowing the model to acquire new task-specific skills while effectively retaining previously learned knowledge. Specifically, we implement a decomposed component weighting strategy comprising learnable components that collectively generate attention-based weights, allowing the model to integrate and utilize diverse knowledge from each DATA. Extensive experiments on three widely used benchmarks demonstrate that our proposed method achieves state-of-the-art performance. Notably, our approach significantly enhances model plasticity and mitigates CF by extending learnable components and employing stochastic restoration during training iterations.