论文检索

Deep Learning · Foundation Models

Yiran Huang, Karsten Roth, Quentin Bouniot, Wenjia Xu, Zeynep Akata

Transformer-based multimodal large language models often exhibit in-context learning (ICL) capabilities. Motivated by this phenomenon, we ask: how do transformers learn to associate information across modalities from in-context examples? We investigate this through controlled experiments on small transformers trained on synthetic classification tasks, enabling precise manipulation of data statistics and model architecture. We begin by revisiting core principles of unimodal ICL in modern transformers. While several prior findings replicate, we find that Rotary Position Embeddings (RoPE) can delay the onset of ICL circuits. Extending to the multimodal setting reveals a fundamental learning asymmetry: when pretrained on high-diversity data from a primary modality, surprisingly low data complexity in the secondary modality suffices for multimodal ICL to emerge. Mechanistic analysis shows that both settings rely on an induction-style mechanism that copies labels from matching in-context exemplars; multimodal training refines and extends these circuits across modalities. Our findings provide a mechanistic foundation for understanding multimodal ICL in modern transformers and introduce a controlled testbed for future investigation.

Deep Learning · Theory

Yang Yang, Zhengmin Kong, Yuan Liu, Tao Huang, Wei Xiang

Modern neural networks derive much of their effectiveness from rich connectivity patterns. Yet, existing architectures often fix the topology at either the sparse or dense extremes, thereby limiting structural flexibility and analysis. We propose Kronecker Generative Networks (KGNs), an algebraic framework that constructs neural network topologies via recursive generation rules, treating topology as a first-class design object. KGNs generate families of directed acyclic graphs with controllable connectivity complexity, enabling systematic interpolation between sparse and dense aggregation regimes. Under this formulation, architectures such as FractalNet and DenseNet arise as specific instantiations corresponding to different generation rules. We provide theoretical analysis of acyclicity, connectivity scaling, and expressiveness, and demonstrate experimentally that KGN instantiations achieve favorable accuracy-efficiency trade-offs across multiple domains.

Probabilistic Methods · Gaussian Processes

Donghyun Lee, Young Myoung Ko

Modeling the learning curve is critical for cost-effective data collection in deep learning systems. Most prior approaches assume a specific parametric learning curve, but these can be inappropriate when no reliable parametric form can be assumed for the learning curve. While Gaussian processes offer flexible nonparametric modeling, existing GP approaches that enforce monotonicity typically introduce intractable factors or require derivative observations. To address this, we propose a Monotonic Variational Gaussian Process for Efficient Data Collection (MOVE), which (i) introduces a novel monotonic variational GP formulation with virtual-derivative factors to enable tractable posterior inference, and (ii) develops an expected shortfall based objective for target-driven data collection. Furthermore, our theoretical analysis shows that expected shortfall provides non-vanishing gradient signals that enable reliable gradient-based optimization. Extensive experiments on classification, segmentation, and detection benchmarks demonstrate consistent improvements over the prior method.

General Machine Learning · Everything Else

Yaqian Zhang, Bernhard Pfahringer, Eibe Frank, Albert Bifet

Memory is a critical component in replay-based continual learning (CL). Prior research has largely treated CL memory as a monolithic store of past data, focusing on how to select and store representative past examples. However, this perspective overlooks the higher-level memory architecture that governs the interaction between old and new data. In this work, we identify and characterize a dual-memory system that is inherently present in both online and offline CL settings. This system comprises: a short-term memory, which temporarily buffers recent data for immediate model updates, and a long-term memory, which maintains a carefully curated subset of past experiences for future replay and consolidation. We propose \textit{memory capacity ratio} (MCR), the ratio between short-term memory and long-term memory capacities, to characterize online and offline CL. Based on this framework, we systematically investigate how MCR influences generalization, stability, and plasticity. Across diverse CL settings—class-incremental, task-incremental, and domain-incremental—and multiple data modalities (e.g., image and text classification), we observe that a smaller MCR, characteristic of \textit{online CL}, can yield comparable or even superior performance relative to a larger one, characteristic of \textit{offline CL}, when both are evaluated under equivalent computational and data storage budgets. This advantage holds consistently across several state-of-the-art replay strategies, such as ER, DER, and SCR. Theoretical analysis further reveals that a reduced MCR yields a better trade-off between stability and plasticity by lowering a bound on generalization error when learning from non-stationary data streams with limited memory. These findings offer new insights into the role of memory allocation in continual learning and underscore the underexplored potential of online CL approaches.

Deep Learning · Graph Neural Networks

Danial Saber, Amirali Salehi-Abari

Graph Neural Networks (GNNs) suffer from overfitting and over-squashing of long-range information. Stochastic graph augmentations (e.g., edge deletion) regularize training against overfitting but can introduce train-inference misalignment and do not improve over-squashing. In contrast, rewiring methods improve connectivity to mitigate over-squashing, but are not designed to regularize training. We propose Random Add-Drop Edge (RADE), a stochastic graph augmentation method that jointly drops and adds edges to address both overfitting and over-squashing simultaneously. RADE is provably designed to align training and inference so that random augmentations regularize training without distribution shift, while supporting long-range communication at inference. We further propose and study a mini-batch gradient-norm balancing algorithm that adapts deletion and addition rates during training, rendering RADE hyperparameter-free in practice. Experiments on node- and graph-classification benchmarks show that RADE is a strong regularizer and mitigates over-squashing. Ablations support the roles of train-inference alignment, adaptive rate selection, and the complementary effects of random edge deletion and edge addition.

Applications · Chemistry, Physics, and Earth Sciences

Xiaohan Qin, Wenjie Du, Yang Wang

Identifying molecular structures from spectral data is essential for early-stage chemical analysis, yet it remains a difficult task due to the imbalance in functional group distributions. Current methods often overfit to prevalent groups while neglecting underrepresented ones, failing to capture key dependencies between functional groups. This highlights the need for a unified approach that addresses both data imbalance and structural constraints. In this work, we present SymSpectra, a Symmetric Conditional Information Bottleneck (SCIB) framework designed to seamlessly integrate multi-modal Spectra features. Our model employs the SCIB framework to fuse multi-modal spectroscopic data into a unified representation, effectively preserving discriminative signals while mitigating redundancy. To enhance robustness against data imbalance, we incorporate conditional mutual information into the training objective, increasing the model’s sensitivity to rare functional groups and challenging molecular cases. Additionally, a specialized module captures the dependencies among functional groups, improving both prediction accuracy and chemically meaningful interpretability. Experiments on multimodal spectral datasets demonstrate that SymSpectra significantly outperforms state-of-the-art methods, achieving an F1-score of 0.970 in substructure classification. More importantly, SymSpectra consistently outperforms baselines under various imbalanced scenarios, exhibiting superior robustness and generalizability, which may help advance the automation of chemical discovery. Our code can be found at https://anonymous.4open.science/r/SymSpectra-0017.

Jitian Zhao, Changho Shin, Tzu-Heng Huang, Satya Sai Srinath Namburi GNVV, Frederic Sala

LLM-as-a-judge ensembles are the standard paradigm for scalable evaluation, but their aggregation mechanisms suffer from a fundamental flaw: they implicitly assume that judges provide independent estimates of true quality. However, in practice, LLM judges exhibit correlated errors caused by shared latent confounders---such as verbosity, stylistic preferences, or training artifacts---causing standard aggregation rules like majority vote or averaging to provide little gain or even amplify systematic mistakes. To address this, we introduce CARE, a confounder-aware aggregation framework that explicitly models LLM judge scores as arising from both a latent true-quality signal and shared confounding factors. Rather than heuristically re-weighting judges, CARE separates quality from confounders without access to ground-truth labels. We provide theoretical guarantees for identifiability and finite-sample recovery under shared confounders, and we quantify the systematic bias incurred when aggregation models omit confounding latent factors. Across 12 public benchmarks spanning continuous scoring, binary classification, and pairwise preference settings, CARE improves aggregation accuracy, reducing error by up to 26.8%.

Deep Learning · Large Language Models

Zheng Qi, Chao Shang, Evangelia Spiliopoulou, Nikolaos Pappas

Vision language models (VLMs) often generate hallucination, i.e., content that cannot be substantiated by either textual or visual inputs. Prior work primarily attributes this to over-reliance on linguistic prior knowledge rather than visual inputs. Some methods attempt to mitigate hallucination by amplifying visual token attention proportionally to their attention scores. However, these methods overlook the visual attention sink problem, where attention is frequently misallocated to task-irrelevant visual regions, and neglect cross-modal fusion balance by enhancing only visual attention without adjusting attention to the user query. This can result in amplifying incorrect areas while failing to properly interpret the user query. To address these challenges, we propose a simple yet effective method called \textbf{G}aze Sh\textbf{i}ft-Guided Cross-modal \textbf{F}usion Enhancemen\textbf{t} (\textbf{GIFT}). GIFT pre-computes a holistic visual saliency map by tracking positive changes in visual attention, or \textit{"gaze shifts"}, during user query comprehension, and leverages this map to amplify attention to both salient visual information and the user query at each decoding step. This reduces the impact of visual attention sink, as irrelevant tokens exhibit minimal shifts, while ensuring balanced cross-modal fusion for well-integrated representation. Extensive experiments show that GIFT effectively mitigates hallucination in VLMs across both generative and classification tasks, achieving up to 20.7% improvement over greedy decoding, while maintaining general vision-language performance with low computational overhead.

Applications · Time Series

Lucas Zoroddu, Pierre Humbert, Laurent Oudre

Expert-annotated time series datasets often suffer from low agreement, especially in medical applications where decisions rely on subjective criteria and inconsistent thresholds. Such variability degrades annotation quality and thus limits the reliability of supervised classification models. To address this, we propose to rely on a pairwise comparison-based approach, which provides a more robust alternative to individual annotation, since relative judgments are typically easier and yield higher consistency. The problem is thus transformed into a ranking problem and we introduce an ideal point model adapted to time series data using elastic similarity measures such as Dynamic Time Warping (DTW) and Time Warp Edit Distance (TWED). We prove Lipschitz continuity of these distances and demonstrate several convergence guarantees for this model. To facilitate gradient-based optimization, we also introduce a differentiable version of the TWED. Finally, we show through multiple experiments that our approach produces accurate and robust rankings under noisy annotation conditions.

Social Aspects · Accountability, Transparency, and Interpretability

Adam Karvonen, James Chua, Clément Dumas, Kit Fraser-Taliente, Subhash Kantamneni, Julian Minder, Euan Ong, Arnab Sen Sharma, Daniel Wen, Owain Evans 等

Large language model (LLM) activations are notoriously difficult to understand, with most existing techniques using complex, specialized methods for interpreting them. Recent work has proposed a simpler approach known as LatentQA: training LLMs to directly accept LLM activations as inputs and answer arbitrary questions about them in natural language. However, prior work has focused on narrow task settings for both training and evaluation. In this paper, we instead take a generalist perspective. We evaluate LatentQA-trained models, which we call Activation Oracles (AOs), in far out-of-distribution settings and examine how performance scales with training data diversity. We find that AOs can recover information fine-tuned into a model (e.g., biographical knowledge or malign propensities) that does not appear in the input text, despite never being trained with activations from a fine-tuned model. Our main evaluations are four downstream tasks where we can compare to prior white- and black-box techniques. We find that even narrowly-trained LatentQA models can generalize well, and that adding additional training datasets (such as classification tasks and a self-supervised context prediction task) yields consistent further improvements. Our best AOs match or exceed white-box baselines on all four tasks and the best overall baseline on 3 of 4. These results suggest that diversified training to answer natural-language queries imparts a general capability to verbalize information about LLM activations.

Social Aspects · Safety

Jinjie Shen, Jing Wu, Yaxiong Wang, Lechao Cheng, Shengeng Tang, Tianrui Hui, Nan Pu, Zhun Zhong

Existing forgery detection methods are often limited to uni-modal or bi-modal settings, failing to handle the interleaved text, images, and videos prevalent in real-world misinformation. To bridge this gap, we propose **OmniVL-Guard**, a unified framework for omni vision-language forgery detection and grounding. In this unified setting, the interplay between diverse modalities and the dual requirements of simultaneous detection and localization pose significant optimization challenges. Through extensive investigations, we identify a critical difficulty bias in this multi-task optimization: the simpler veracity classification task tends to dominate the gradients, leading to suboptimal performance in fine-grained grounding. To address this imbalance, we first develop a Self-Evolving CoT Generation pipeline to synthesize high-quality reasoning paths, effectively overcoming the cold-start challenge. Building upon this, we propose **A**daptive **R**eward **S**caling **P**olicy **O**ptimization (**ARSPO**). By dynamically modulating reward scales and task weights, ARSPO ensures a balanced joint optimization that prioritizes challenging grounding objectives. Extensive experiments demonstrate that OmniVL-Guard significantly outperforms state-of-the-art methods and exhibits robust zero-shot generalization across out-of-domain scenarios.

Deep Learning · Large Language Models

Amrut Nadgir, Vijay Balasubramanian, Pratik Chaudhari

Many language tasks can be modeled as classification problems where a large language model (LLM) is given a prompt and selects one among many possible answers. We show that the classification error in such problems scales as a power law in the number of classes. This has a dramatic consequence: the prediction error can be reduced substantially by splitting the overall task into a sequence of smaller classification problems, each with the same number of classes ("degree"). This tree-structured decomposition models chain-of-thought (CoT). It has been observed that CoT-based predictors perform better when they "think", i.e., when they develop a deeper tree, thus decomposing the problem into a larger number of steps. We identify a critical threshold for the degree, below which thinking is detrimental, and above which there exists an optimal depth that minimizes the error. It is impossible to surpass this minimal error by increasing the depth of thinking.

General Machine Learning · Evaluation

Yizheng Huang, Wenjun Zeng, Aditi Kumaresan, Zi Wang

Evaluating generative AI models is increasingly resource-intensive due to slow generation, expensive raters, and a rapidly growing landscape of models and benchmarks. We propose ProEval, a proactive evaluation framework that leverages transfer learning to efficiently estimate performance and identify failure cases. ProEval employs pre-trained Gaussian Processes (GPs) as surrogates for the performance score function, mapping model inputs to metrics such as error likelihood or safety violations. By framing performance estimation as Bayesian quadrature (BQ) and failure discovery as superlevel set sampling, we develop strategies that proactively select or synthesize the most informative inputs for testing. Theoretically, we prove that our pre-trained GP-based BQ estimator is unbiased and bounded. Empirically, extensive experiments on reasoning, safety alignment, and classification benchmarks demonstrate that ProEval is significantly more efficient than competitive baselines. It requires 10–100x fewer samples to achieve estimates within $\pm1\%$ of ground truth, while simultaneously revealing more diverse failure cases under a stricter evaluation budget.

Deep Learning · Large Language Models

Aymane El Gadarri, Ali Aouad, Vivek Farias

Traditional large language model (LLM) alignment methods are based on Reinforcement Learning From Human Feedback (RLHF), which learns a single reward model (implicitly or explicitly) from pairwise comparison data. This approach implicitly assumes homogeneous preferences across human labelers---an assumption that is violated in practice. As a result, the learned reward model in RLHF is generally misspecified: Prior work shows that it is inconsistent with the population-average utility, incurring large distortion, and that recovering this utilitarian objective is provably impossible in the worst case. In this work, we show that the average utility is recoverable under a mild assumption. Our method, the Sign Estimator, simply replaces the standard cross-entropy loss function with a notion of binary classification loss and yields a reward model that is ordinally consistent with the population-average utility. We further establish a fast finite-sample convergence rate of $O({n^{-{1}/{3}}})$ , which provides, to our knowledge, the first consistent estimator for heterogeneous preferences that does not suffer from the curse of dimensionality.

Deep Learning · Generative Models and Autoencoders

Xiao Li, Yixuan Jia, Zekai Zhang, Xiang Li, Lianghe Shi, Jinxin Zhou, Zhihui Zhu, Liyue Shen, Qing Qu

Diffusion models are effective generative frameworks with strong representation learning capabilities, yet the intrinsic properties that govern their semantic structure and generalization remain poorly understood. Drawing inspiration from self-supervised representation learning (SSL), we introduce an evaluation framework that decomposes diffusion features into a perturbation invariant component and a residual component induced by noise and augmentations. From this decomposition we derive the Invariant Contamination Ratio (ICR), a Fisher-based metric that measures how residual, augmentation-sensitive energy contaminates invariant signal in feature space. We use this framework to analyze both discriminative and generative behavior. On the representation side, we find invariance peaks at intermediate noise levels, which also yield the best downstream classification performance. On the generative side, we study how training transitions from genuine generalization to memorization in data-limited regimes, and find that $\mathrm{ICR}$ serves as a sensitive training time indicator of the early learning phenomenon: rising residual energy along Fisher directions marks the onset of memorization, detectable from training features alone without external evaluators or held-out test sets. Overall, our results show diffusion models can be monitored from a self-supervised perspective via the geometry of their learned representations.

Theory · Deep Learning

Peng Wang, Xiao Li, Can Yaras, Zhihui Zhu, Laura Balzano, Wei Hu, Qing Qu

Over the past decade, deep learning has proven to be a highly effective tool for learning meaningful features from raw data. However, it remains an open question how deep networks perform hierarchical feature learning across layers. In this work, we attempt to unveil this mystery by investigating the structures of intermediate features. Motivated by our empirical findings that linear layers mimic the roles of deep layers in nonlinear networks for feature learning, we explore how deep linear networks transform input data into output by investigating the output (i.e., features) of each layer after training in the context of multi-class classification problems. Toward this goal, we first define metrics to measure within-class compression and between-class discrimination of intermediate features, respectively. Through theoretical analysis of these two metrics, we show that the evolution of features follows a simple and quantitative pattern from shallow to deep layers when the input data is nearly orthogonal and the network weights are minimum-norm, balanced, and approximately low-rank: each layer of the linear network progressively compresses within-class features at a geometric rate and discriminates between-class features at a linear rate with respect to the number of layers that data have passed through. To the best of our knowledge, this is the first quantitative characterization of feature evolution in hierarchical representations of deep linear networks. Moreover, our extensive experiments not only validate our theoretical results but also reveal a similar pattern in deep nonlinear networks, which aligns well with recent empirical studies. Finally, we demonstrate the practical value of our results in transfer learning.

Ziming Liu, Surya Ganguli, Andreas Tolias

Vafa et al. recently showed that a transformer fails to acquire an internal Newtonian world model when trained on synthetic planetary-motion data. How can we fix this problem? We find that inductive biases are key to learning the veridical world model: (1) **Spatial smoothness** is required for any world model to be learned. However, naive tokenization may disrupt smoothness since two close points in physical space may be far apart in token embedding space without sufficient training or data. We fix this by formulating the prediction problem as regression instead of classification. (2) **Spatial stability** makes the prediction robust to noise, which is not guaranteed by default, but can be taught via correcting in-context noise perturbations. (3) With both spatial smoothness and stability built in, further imposing **temporal locality** induces a Newtonian world model, while the lack of this knowledge induces a Keplerian world model -- fitting elliptical parameters instead of computing gravitational forces. Our results suggest that even simple general inductive biases are powerful enough to induce correct and specific world models. The inductive biases do not need to know that much about the underlying law to be learned, but without them, it is impossible to learn.

Applications · Computer Vision

Xinyue Zhang, Xu Zou, Wanjia Luo, Yanjie Wang, Jiahuan Zhou, Sheng Zhong, Luxin Yan

Class-incremental semantic segmentation learns new classes while retaining old ones without access to past data. Although existing methods alleviate catastrophic forgetting on old classes, new-class performance remains limited. We identify the key bottleneck arises from low-margin regions, where the logit of the ground-truth class is close to that of the most competitive non-ground-truth class. Our theoretical analysis suggests that optimization in these regions is characterized by high curvature and a small stability radius, making learning prone to class confusion. Based on the above analysis, we propose Learnability-Driven Knowledge Assimilation (LDKA), which targets low-margin learning via three complementary optimization strategies: (i) Progressive Margin Learning continuously reallocates pixel-wise optimization budget in a threshold-free manner, shifting emphasis from high-margin to low-margin regions; (ii) Smooth Knowledge Distillation applies curvature damping and perturbation stabilization to suppress high-frequency updates and increase stability radius; (iii) Misclassification-Aware Decoupling measures inter-class confusion with a competition matrix and decouples highly competitive class representations. Experiments show that LDKA improves mIoU on new classes while preserving performance on old classes across 9 incremental protocols.

General Machine Learning · Transfer, Multitask and Meta-learning

Gang Liu, Xiaoxuan Zhang, Yuhong Feng, Mingyang Zhou, Xiaoqun Wu, Hao Liao, Rui Mao

Few-shot classification aims to adapt a pretrained model to novel classes with limited examples. While current methods often heuristically combine pretrained knowledge and few-shot evidence, we seek a more principled understanding of their relationship. In this paper, we propose a Bayesian-inspired optimal integration framework(BOIF) that interprets pretrained models as priors and few-shot evidence as likelihoods. Under conditional independence approximation, we show that the optimal log-posterior decomposes into the sum of prior logits and likelihood logits. This leads to a simple yet effective design principle: decouple the prior and likelihood pathways and combine their logits additively. Guided by this principle, we implement BOIF using CLIP with two novel enhancements: (1) a multi-level feature adapter to enrich visual representations, and (2) a simplified cache module for likelihood estimation. Extensive experiments on 11 benchmarks show BOIF achieves state-of-the-art performance (e.g., 80.61\% average accuracy at 16-shot) and strong out-of-distribution robustness. Our work provides both a principled perspective and an effective instantiation for few-shot adaptation.

General Machine Learning · Transfer, Multitask and Meta-learning

Li-Jun Zhao, Zhen-Duo Chen, Xin Luo, Xin-Shun Xu

Few-shot class-incremental learning (FSCIL) aims at recognizing novel classes continually with limited novel class samples. A mainstream baseline for FSCIL is first to train the whole model in the base session, then freeze the feature extractor in the incremental sessions. Despite achieving high overall accuracy, most methods exhibit notably low accuracy on incremental classes. While some recent methods have recognized this issue, their strategies remain constrained by a unified classification objective across all samples, making it difficult to simultaneously satisfy the performance requirements of both base and incremental classes. In this paper, considering that base and incremental classes play different yet both critical roles in FSCIL, we approach FSCIL from a more structured perspective by decomposing the overall classification objective into three sub-objectives. Building on this insight, we propose a novel classification framework called Hierarchical Filtering and Refinement Classification (HFRC) to hierarchically decompose and address the classification task. Extensive experiments demonstrate that our method effectively balances the classification accuracy between base and incremental classes, and achieves superior performance compared to state-of-the-art methods.