论文检索

Applications · Everything Else

Yinfang Chen, Jiaqi Pan, Jackson Clark, Yiming Su, Noah Zheutlin, Bhavya Bhavya, Rohan R. Arora, Yu Deng, Saurabh Jha, Tianyin Xu

In cloud-scale systems, failures are the norm. A distributed computing cluster exhibits hundreds of machine failures and thousands of disk failures; software bugs and misconfigurations are reported to be more frequent. The demand for autonomous, AI-driven reliability engineering continues to grow, as existing human-in-the-loop practices can hardly keep up with the scale of modern clouds. This paper presents STRATUS, an LLM-based multi-agent system for realizing autonomous Site Reliability Engineering (SRE) of cloud services. STRATUS consists of multiple specialized agents (e.g., for failure detection, diagnosis, mitigation), organized in a state machine to assist system-level safety reasoning and enforcement. We formalize a key safety specification of agentic SRE systems like STRATUS, termed Transactional No-Regression (TNR), which enables safe exploration and iteration. We show that TNR can effectively improve autonomous failure mitigation. STRATUS significantly outperforms state-of-the-art SRE agents in terms of success rate of failure mitigation problems in AIOpsLab and ITBench (two SRE benchmark suites), by at least 1.5 times across various models. STRATUS shows a promising path toward practical deployment of agentic systems for cloud reliability.

Deep Learning · Generative Models and Autoencoders

Emanuele La Malfa, Gabriele La Malfa, Samuele Marro, Jie Zhang, Elizabeth Black, Michael Luck, Philip Torr, Michael Wooldridge

Recent interest in Multi-Agent Systems of Large Language Models (MAS LLMs) has led to an increase in frameworks leveraging multiple LLMs to tackle complex tasks. However, much of this literature appropriates the terminology of MAS without engaging with its foundational principles. In this position paper, we highlight critical discrepancies between MAS theory and current MAS LLMs implementations, focusing on four key areas: the social aspect of agency, environment design, coordination and communication protocols, and measuring emergent behaviours. Our position is that many MAS LLMs lack multi-agent characteristics such as autonomy, social interaction, and structured environments, and often rely on oversimplified, LLM-centric architectures. The field may slow down and lose traction by revisiting problems the MAS literature has already addressed. Therefore, we systematically analyse this issue and outline associated research opportunities; we advocate for better integrating established MAS concepts and more precise terminology to avoid mischaracterisation and missed opportunities.

Deep Learning · Robustness

Daniel Wesego, Pedram Rooshenas

Deep neural networks remain highly susceptible to adversarial attacks, where small, subtle perturbations to input images may induce misclassification. We propose a novel optimization-based purification framework that directly removes these perturbations by maximizing a Bayesian-inspired objective combining a pretrained diffusion prior with a likelihood term tailored to the adversarial perturbation space. Our method iteratively refines a given input through gradient-based updates of a combined score-based loss to guide the purification process. Unlike existing optimization-based defenses that treat adversarial noise as generic corruption, our approach explicitly integrates the adversarial landscape into the objective. Experiments performed on CIFAR-10 and CIFAR-100 demonstrate strong robust accuracy against a range of common adversarial attacks. Our work offers a principled test-time defense grounded in probabilistic inference using score-based generative models. Our code can be found at \url{https://github.com/rooshenasgroup/aaopt}.

Social Aspects · Fairness, Equity, Justice and Safety

Shuai Yuan, Xingshuo Han, Hongwei Li, Guowen Xu, Wenbo Jiang, Tao Ni, Qingchuan Zhao, Yuguang Fang

Recently, traffic sign recognition (TSR) systems have become a prominent target for physical adversarial attacks. These attacks typically rely on conspicuous stickers and projections, or using invisible light and acoustic signals that can be easily blocked. In this paper, we introduce a novel attack medium, i.e., fluorescent ink, to design a stealthy and effective physical adversarial patch, namely FIPatch, to advance the state-of-the-art. Specifically, we first model the fluorescence effect in the digital domain to identify the optimal attack settings, which guide the real-world fluorescence parameters. By applying a carefully designed fluorescence perturbation to the target sign, the attacker can later trigger a fluorescent effect using invisible ultraviolet light, causing the TSR system to misclassify the sign and potentially leading to traffic accidents. We conducted a comprehensive evaluation to investigate the effectiveness of FIPatch, which shows a success rate of 98.31% in low-light conditions. Furthermore, our attack successfully bypasses five popular defenses and achieves a success rate of 96.72%.

Theory · Reinforcement Learning and Planning

Steven Cao, Gregory Valiant, Percy Liang

We study the problem of entropy calibration, which asks whether a language model's entropy over generations matches its log loss on human text. Past work found that models are miscalibrated, with entropy per step increasing (and text quality decreasing) as generations grow longer. This error accumulation is a fundamental problem in autoregressive models, and the standard solution is to truncate the distribution, which improves text quality at the cost of diversity. In this paper, we ask: is miscalibration likely to improve with scale, and is it theoretically possible to calibrate without tradeoffs? To build intuition, we first study a simplified theoretical setting to characterize the scaling behavior of miscalibration with respect to dataset size. We find that the scaling behavior depends on the power law exponent of the data distribution --- in particular, for a power law exponent close to 1, the scaling exponent is close to 0, meaning that miscalibration improves very slowly with scale. Next, we measure miscalibration empirically in language models ranging from 0.5B to 70B parameters. We find that the observed scaling behavior is similar to what is predicted by the simplified setting: our fitted scaling exponents for text are close to 0, meaning that larger models accumulate error at a similar rate as smaller ones. This scaling (or, lack thereof) provides one explanation for why we sample from larger models with similar amounts of truncation as smaller models, even though the larger models are of higher quality. However, truncation is not a satisfying solution because it comes at the cost of increased log loss. In theory, is it even possible to reduce entropy while preserving log loss? We prove that it is possible, if we assume access to a black box which can fit models to predict the future entropy of text.

Social Aspects · Accountability, Transparency and Interpretability

Thomas Decker, Volker Tresp, Florian Buettner

Perturbation-based explanations are widely utilized to enhance the transparency of machine-learning models in practice. However, their reliability is often compromised by the unknown model behavior under the specific perturbations used. This paper investigates the relationship between uncertainty calibration - the alignment of model confidence with actual accuracy - and perturbation-based explanations. We show that models systematically produce unreliable probability estimates when subjected to explainability-specific perturbations and theoretically prove that this directly undermines global and local explanation quality. To address this, we introduce ReCalX, a novel approach to recalibrate models for improved explanations while preserving their original predictions. Empirical evaluations across diverse models and datasets demonstrate that ReCalX consistently reduces perturbation-specific miscalibration most effectively while enhancing explanation robustness and the identification of globally important input features.

Optimization · Large Scale, Parallel and Distributed

Gang Yan, Sikai Yang, Wan Du

Integrating large pre-trained models into federated learning (FL) can significantly improve generalization and convergence efficiency. A widely adopted strategy freezes the pre-trained backbone and fine-tunes a lightweight task head, thereby reducing computational and communication costs. However, this partial fine-tuning paradigm introduces new security risks, making the system vulnerable to poisoned updates and backdoor attacks. To address these challenges, we propose FedRACE, a unified framework for robust FL with partially frozen models. FedRACE comprises two core components: HStat-Net, a hierarchical network that refines frozen features into compact, linearly separable representations; and DevGuard, a server-side mechanism that detects malicious clients by evaluating statistical deviance in class-level predictions modeling generalized linear models (GLMs). DevGuard further incorporates adaptive thresholding based on theoretical misclassification bounds and employs randomized majority voting to enhance detection reliability. We implement FEDRACE on the FedScale platform and evaluate it on CIFAR-100, Food-101, and Tiny ImageNet under diverse attack scenarios. FedRACE achieves a true positive rate of up to 99.3% with a false positive rate below 1.2%, while preserving model accuracy and improving generalization.

Theory · Everything Else

Etienne Gauthier, Francis Bach, Michael Jordan

We introduce *Backward Conformal Prediction*, a method that guarantees conformal coverage while providing flexible control over the size of prediction sets. Unlike standard conformal prediction, which fixes the coverage level and allows the conformal set size to vary, our approach defines a rule that constrains how prediction set sizes behave based on the observed data, and adapts the coverage level accordingly. Our method builds on two key foundations: (i) recent results by Gauthier et al. [2025] on post-hoc validity using e-values, which ensure marginal coverage of the form $\mathbb{P}(Y_{\rm test} \in \hat C_n^{\tilde{\alpha}}(X_{\rm test})) \ge 1 - \mathbb{E}[\tilde{\alpha}]$ up to a first-order Taylor approximation for any data-dependent miscoverage $\tilde{\alpha}$, and (ii) a novel leave-one-out estimator $\hat{\alpha}^{\rm LOO}$ of the marginal miscoverage $\mathbb{E}[\tilde{\alpha}]$ based on the calibration set, ensuring that the theoretical guarantees remain computable in practice. This approach is particularly useful in applications where large prediction sets are impractical such as medical diagnosis. We provide theoretical results and empirical evidence supporting the validity of our method, demonstrating that it maintains computable coverage guarantees while ensuring interpretable, well-controlled prediction set sizes.

Computer Vision · Classification and Understanding

Aaron Sun, Subhransu Maji, Grant Van Horn

In the single-positive multi-label (SPML) setting, each image in a dataset is labeled with the presence of a single class, while the true presence of other classes remains unknown. The challenge is to narrow the performance gap between this partially-labeled setting and fully-supervised learning, which often requires a significant annotation budget. Prior SPML methods were developed and benchmarked on synthetic datasets created by randomly sampling single positive labels from fully-annotated datasets like Pascal VOC, COCO, NUS-WIDE, and CUB200. However, this synthetic approach does not reflect real-world scenarios and fails to capture the fine-grained complexities that can lead to difficult misclassifications. In this work, we introduce the L48 dataset, a fine-grained, real-world multi-label dataset derived from recordings of bird sounds. L48 provides a natural SPML setting with single-positive annotations on a challenging, fine-grained domain, as well as two extended settings in which domain priors give access to additional negative labels. We benchmark existing SPML methods on L48 and observe significant performance differences compared to synthetic datasets and analyze method weaknesses, underscoring the need for more realistic and difficult benchmarks.

Social Aspects · Fairness, Equity, Justice and Safety

Jiangtong Li, Dongyi Liu, Kun Zhu, Dawei Cheng, changjun jiang

Graph Neural Networks (GNNs) have achieved significant success in various real-world applications, including social networks, finance systems, and traffic management. Recent researches highlight their vulnerability to backdoor attacks in node classification, where GNNs trained on a poisoned graph misclassify a test node only when specific triggers are attached. These studies typically focus on single attack categories and use adaptive trigger generators to create node-specific triggers. However, adaptive trigger generators typically have a simple structure, limited parameters, and lack category-aware graph knowledge, which makes them struggle to handle backdoor attacks across multiple categories as the number of target categories increases. We address this gap by proposing a novel approach for Effective and Unnoticeable Multi-Category (EUMC) graph backdoor attacks, leveraging subgraph from the attacked graph as category-aware triggers to precisely control the target category. To ensure the effectiveness of our method, we construct a Multi-Category Subgraph Triggers Pool (MC-STP) using the subgraphs of the attacked graph as triggers. We then exploit the attachment probability shifts of each subgraph trigger as category-aware priors for target category determination. Moreover, we develop a ``select then attach'' strategy that connects suitable category-aware trigger to attacked nodes for unnoticeability. Extensive experiments across different real-world datasets confirm the efficacy of our method in conducting multi-category graph backdoor attacks on various GNN models and defense strategies.

Social Aspects · Fairness, Equity, Justice and Safety

Annette Zimmermann, Andrew Zeppa, Srijan Pandey, Kenneth Diao

The claim that the AI community, or society at large, should ‘democratize AI’ has attracted considerable critical attention and controversy. Two core problems have arisen and remain unsolved: conceptual disagreement persists about what democratizing AI means; normative disagreement persists over whether democratizing AI is ethically and politically desirable. We identify eight common AI democratization traps: democratization-skeptical arguments that seem plausible at first glance, but turn out to be misconceptions. We develop arguments about how to resist each trap. We conclude that, while AI democratization may well have drawbacks, we should be cautious about dismissing AI democratization prematurely and for the wrong reasons. We offer a constructive roadmap for developing alternative conceptual and normative approaches to democratizing AI that successfully avoid the traps.

Xinyi Yang, Runzhe Zhan, Shu Yang, Junchao Wu, Lidia S. Chao, Derek F. Wong

Investigating bias in large language models (LLMs) is crucial for developing trustworthy AI. While prompt-based through prompt engineering is common, its effectiveness relies on the assumption that models inherently understand biases. Our study systematically analyzed this assumption using the BBQ and StereoSet benchmarks on both open-source models as well as commercial GPT model. Experimental results indicate that prompt-based is often superficial; for instance, the Llama2-7B-Chat model misclassified over 90% of unbiased content as biased, despite achieving high accuracy in identifying bias issues on the BBQ dataset. Additionally, specific evaluation and question settings in bias benchmarks often lead LLMs to choose “evasive answers”, disregarding the core of the question and the relevance of the response to the context. Moreover, the apparent success of previous methods may stem from flawed evaluation metrics. Our research highlights a potential “false prosperity” in prompt-base efforts and emphasizes the need to rethink bias evaluation metrics to ensure truly trustworthy AI. We will release our data and code upon acceptance.

Narjis Asad, Nihar Ranjan Sahoo, Rudra Murthy, Swaprava Nath, Pushpak Bhattacharyya

While a few high-quality bias benchmark datasets exist to address stereotypes in Language Models (LMs), a notable lack of focus remains on body image stereotypes. To bridge this gap, we propose \textbf{BIStereo}, a suite to uncover LMs’ biases towards people of certain physical appearance characteristics, namely, \textit{skin complexion, body shape, height, attire,} and a \textit{miscellaneous category} including \textit{hair texture, eye color, and more}. Our dataset comprises 40k sentence pairs designed to assess LMs’ biased preference for certain body types. We further include 60k premise-hypothesis pairs designed to comprehensively assess LMs’ preference for fair skin tone. Additionally, we curate 553 tuples consisting of a \textit{body image descriptor, gender, and a stereotypical attribute}, validated by a diverse pool of annotators for physical appearance stereotypes.We propose a metric, \textbf{TriSentBias}, that captures the biased preferences of LMs towards a certain body type over others. Using \textbf{BIStereo}, we assess the presence of body image biases in ten different language models, revealing significant biases in models Muril, XLMR, Llama3, and Gemma. We further evaluate the LMs through downstream NLI and Analogy tasks.Our NLI experiments highlight notable patterns in the LMs that align with the well-documented cognitive bias in humans known as \textbf{\textit{the Halo Effect}}.

Anna Mosolova, Marie Candito, Carlos Ramisch

In the absence of sense-annotated data, word sense induction (WSI) is a compelling alternative to word sense disambiguation, particularly in low-resource or domain-specific settings. In this paper, we emphasize methodological problems in current WSI evaluation. We propose an evaluation on a SemCor-derived dataset, respecting the original corpus polysemy and frequency distributions. We assess pre-trained embeddings and clustering algorithms across parts of speech, and propose and evaluate an LLM-based WSI method for English. We evaluate data augmentation sources (LLM-generated, corpus and lexicon), and semi-supervised scenarios using Wiktionary for data augmentation, must-link constraints, number of clusters per lemma.We find that no unsupervised method (whether ours or previous) surpasses the strong “one cluster per lemma” heuristic (1cpl). We also show that (i) results and best systems may vary across POS, (ii) LLMs have troubles performing this task, (iii) data augmentation is beneficial and (iv) capitalizing on Wiktionary does help. It surpasses previous SOTA system on our test set by 3.3%. WSI is not solved, and calls for a better articulation of lexicons and LLMs’ lexical semantics capabilities.

Ioana Ivan, Carlos Ramisch, Alexis Nasr

The scaling of causal language models in size and training data enabled them to tackle increasingly complex tasks. Despite the development of sophisticated tests to reveal their new capabilities, the underlying basis of these complex skills remains unclear. We argue that complex skills might be explained using simpler ones, represented by linguistic concepts. As an initial step in exploring this hypothesis, we focus on the lexical-semantic concept of synonymy, laying the groundwork for research into its relationship with more complex skills. We develop a comprehensive test suite to assess various aspects of synonymy under different conditions, and evaluate causal open-source models ranging up to 10 billion parameters. We find that these models effectively recognize synonymy but struggle to generate synonyms when prompted with relevant context.

Chengzu Li, Han Zhou, Goran Glavaš, Anna Korhonen, Ivan Vulić

When adapting ICL with or without fine-tuning, we are curious about whether the instruction-tuned language model is able to achieve well-calibrated results without suffering from the problem of overconfidence (i.e., miscalibration) considering its strong instruction following ability, especially in such limited data setups. In this work, we deliver an in-depth analysis of the behavior across different choices of learning methods from the perspective of both performance and calibration. Through extensive controlled experiments, we observe that the miscalibration problem exists across all learning methods in low-resource setups. To achieve simultaneous gain for both in-task performance and calibration, we then study the potential of self-ensembling applied at different modeling stages (e.g., variations of in-context examples or variations in prompts or different ensembling strategies) to make the predictions more calibrated and have comparable or even better performance. We find that self-ensembling with max probability produces robust and calibrated predictions. Our work reveals the potential calibration problem of using ICL despite the improvements in task performance and sheds light on which learning paradigm to choose. We also provide practical guidelines for choosing learning paradigms depending on whether the data has been seen by the model before and a worthwhile solution via self-ensembling on how to enhance both task performance and calibration of LMs, which we hope could encourage further study.

Kaishuai Xu, Tiezheng Yu, Wenjun Hou, Yi Cheng, Chak Tou Leong, Liangyou Li, Xin Jiang, Lifeng Shang, Qun Liu, Wenjie Li

Large Language Models (LLMs) have exhibited strong mathematical reasoning prowess, tackling tasks ranging from basic arithmetic to advanced competition-level problems. However, frequently occurring subtle yet critical errors, such as miscalculations or incorrect substitutions, limit the LLMs’ full potential. Existing studies to improve mathematical ability typically involve applying preference learning to step-wise solution pairs. Although these methods leverage samples of varying granularity to mitigate reasoning errors, they overlook critical subtle errors. In this work, we propose a novel preference learning framework called eRror-Injected Self-Editing (RISE), which injects predefined subtle errors into pivotal tokens in reasoning or computation steps to construct hard pairs for error mitigation. In detail, RISE uses the LLM itself to edit a small number of tokens in the solution, injecting designed subtle errors. Then, pairs composed of self-edited solutions and their corresponding correct ones, along with pairs of correct and incorrect solutions obtained through sampling, are used together for subtle error-aware DPO training. Compared with other preference learning methods, RISE further refines the training objective without requiring fine-grained sampling or preference annotation. Extensive experiments validate the effectiveness of RISE, with preference learning on Qwen2-7B-Instruct yielding notable improvements of 3.0% on GSM8K and 7.9% on MATH with only 4.5K training samples. Moreover, the effect of error mitigation extends from mathematical reasoning to logical reasoning and code generation.

Doyoun Kim, Suin Kim, Yohan Jo

Knowledge tracing (KT) in programming education presents unique challenges due to the complexity of coding tasks and the diverse methods students use to solve problems. Although students’ questions often contain valuable signals about their understanding and misconceptions, traditional KT models often neglect to incorporate these questions as inputs to address these challenges. This paper introduces SQKT (Students’ Question-based Knowledge Tracing), a knowledge tracing model that leverages students’ questions and automatically extracted skill information to enhance the accuracy of predicting students’ performance on subsequent problems in programming education. Our method creates semantically rich embeddings that capture not only the surface-level content of the questions but also the student’s mastery level and conceptual understanding. Experimental results demonstrate SQKT’s superior performance in predicting student completion across various Python programming courses of differing difficulty levels. In in-domain experiments, SQKT achieved a 33.1% absolute improvement in AUC compared to baseline models. The model also exhibited robust generalization capabilities in cross-domain settings, effectively addressing data scarcity issues in advanced programming courses. SQKT can be used to tailor educational content to individual learning needs and design adaptive learning systems in computer science education.

Yooseop Lee, Suin Kim, Yohan Jo

In designing multiple-choice questions (MCQs) in education, creating plausible distractors is crucial for identifying students’ misconceptions and gaps in knowledge and accurately assessing their understanding. However, prior studies on distractor generation have not paid sufficient attention to enhancing the difficulty of distractors, resulting in reduced effectiveness of MCQs. This study presents a pipeline for training a model to generate distractors that are more likely to be selected by students. First, we train a pairwise ranker to reason about students’ misconceptions and assess the relative plausibility of two distractors. Using this model, we create a dataset of pairwise distractor ranks and then train a distractor generator via Direct Preference Optimization (DPO) to generate more plausible distractors. Experiments on computer science subjects (Python, DB, MLDL) demonstrate that our pairwise ranker effectively identifies students’ potential misunderstandings and achieves ranking accuracy comparable to human experts. Furthermore, our distractor generator outperforms several baselines in generating plausible distractors and produces questions with a higher item discrimination index (DI).

Yoo Yeon Sung, Eve Fleisig, Yu Hou, Ishan Upadhyay, Jordan Boyd-Graber

Language models are often miscalibrated, leading to confidently incorrect answers. We introduce GRACE, a benchmark for language model calibration that incorporates comparison with human calibration. GRACE consists of question-answer pairs, in which each question contains a series of clues that gradually become easier, all leading to the same answer; models must answer correctly as early as possible as the clues are revealed. This setting permits granular measurement of model calibration based on how early, accurately, and confidently a model answers. After collecting these questions, we host live human vs. model competitions to gather 1,749 data points on human and model teams’ timing, accuracy, and confidence. We propose a metric, CalScore, that uses GRACE to analyze model calibration errors and identify types of model miscalibration that differ from human behavior. We find that although humans are less accurate than models, humans are generally better calibrated. Since state-of-the-art models struggle on GRACE, it effectively evaluates progress on improving model calibration.