AI conference papers

AI conference paper library

Search papers from 30+ top conferences, including NeurIPS, ICLR, ICML, CVPR, ACL and AAAI. Filter by keyword, then open any paper for its abstract, authors and links.

arXiv

Is "Knowing It's Malicious Enough?" Evaluating LLMs for Fine-Grained Malware Behavior Auditing

Xinran Zheng, Xingzhi Qian, Yiling He, Shuo Yang, Lorenzo Cavallaro

Automated malware classifiers achieve strong detection performance, but auditing requires more than flagging a sample: analysts must explain malicious behaviors and justify them with code evidence. Traditional signature-based methods and learning-based XAI often fail to provide such support in a human-interpretable form. Large Language Models (LLMs) appear promising, yet their reliability for malware auditing remains unclear. Evaluation faces three challenges: (1) the lack of human-written behavioral ground truth; (2) real-world codebases that exceed current context limits; and (3) the lack of reliable mechanisms to verify whether generated claims are grounded in code evidence. These obstacles make benchmarking difficult and leave model capabilities and failure modes opaque. We introduce MalEval, a diagnostic framework for measuring the capability boundaries of LLMs in malware auditing. MalEval pairs real-world application codebases with expert-written audit reports to provide fine-grained behavior-level ground truth. It compresses large codebases into behavior-relevant program contexts through a context-driven intermediate representation that preserves call relations. Expert reports and model outputs are mapped, via constrained reasoning, into structured evidence chains linking code-level facts to high-level behaviors in a shared space. MalEval decomposes auditing into four stage-wise tasks, enabling each intermediate judgment to be verified under limited context windows. We evaluate seven LLMs and find that they rely on surface cues rather than verifiable evidence, struggle to compose dispersed facts into coherent attack chains, and are highly sensitive to context formulation. These findings shift attention from isolated outputs to reliable LLM and agentic workflows for malware auditing. MalEval is publicly available at https://github.com/ZhengXR930/MalEval.git

arXiv

Supervised Reward Inference

Will Schwarzer, Jordan Schneider, Philip S. Thomas, Scott Niekum

Existing approaches to reward inference typically assume that humans provide demonstrations according to specific behavior models. However, humans often indicate their goals through a wide range of behaviors, from actions that are suboptimal due to poor planning or execution to behaviors intended to communicate goals rather than achieve them. One existing solution for inferring rewards from such behavior $\unicode{x2013}$ provided it is drawn from the same distribution at training and deployment $\unicode{x2013}$ is to construct a dataset of behavior paired with known rewards, and to learn the mapping from behavior to rewards; however, prior methods in this family face notable limitations, such as restrictions to tabular settings. Given such a dataset, we propose instead that supervised learning offers a parsimonious yet powerful solution, which we term Supervised Reward Inference (SRI). Theoretically, we prove that SRI is asymptotically Bayes-optimal under standard assumptions. Empirically, SRI achieves near-ceiling performance on a prior benchmark for reward inference from suboptimal behavior, while on Meta-World robotics tasks, it infers rewards from even arbitrarily suboptimal demonstrations as accurately as those demonstrations allow. Finally, we demonstrate our framework's universality with straightforward generalizations to action- and goal-prediction.

arXiv

BADTV: Unveiling Backdoor Threats in Third-Party Task Vectors

Chia-Yi Hsu, Yu-Lin Tsai, Yu Zhe, Yan-Lun Chen, Chih-Hsun Lin, Chia-Mu Yu, Yang Zhang, Chun-Ying Huang, Jun Sakuma

Task arithmetic in large-scale pre-trained models enables agile adaptation to diverse downstream tasks without extensive retraining. By leveraging task vectors (TVs), users can perform modular updates through simple arithmetic operations like addition and subtraction. Yet, this flexibility presents new security challenges. In this paper, we investigate how TVs are vulnerable to backdoor attacks, revealing how malicious actors can exploit them to compromise model integrity. By creating composite backdoors that are designed asymmetrically, we introduce BadTV, a backdoor attack specifically crafted to remain effective simultaneously under task learning, forgetting, and analogy operations. Extensive experiments show that BadTV achieves near-perfect attack success rates across diverse scenarios, posing a serious threat to models relying on task arithmetic. We also evaluate current defenses, finding they fail to detect or mitigate BadTV. Our results highlight the urgent need for robust countermeasures to secure TVs in real-world deployments.

arXiv

From Scene-Centric to Observer-Centric: Modeling Observer-Aware Relations for 3D Scene Graph Generation

Jingjun Sun, Chaowei Wang, Zhirui Liu, Jiaxu Tian, Ming Yang, Yaoxing Wang, Yan Di, Shan Gao

3D Scene Graph Generation (3DSGG) represents 3D scenes as structured object--relation--object graphs for spatial understanding. In observer-centric spatial perception, the same scene may be expressed under different local observer frames while its structure remains unchanged. However, existing models typically assume a fixed scene-aligned reference frame and may produce semantically inconsistent predictions when the scene is re-expressed in another observer frame. We attribute this failure to the heterogeneous frame dependency of relational predicates. Directional predicates such as $\textit{left}$, $\textit{front}$, $\textit{right}$, and $\textit{behind}$ are $\textbf{Observer-Dependent Relations}$, whereas most contact, support, and semantic predicates, such as $\textit{standing on}$ and $\textit{attached to}$, are approximately $\textbf{Observer-Independent Relations}$. Conventional models do not distinguish these frame responses, leading to degraded relation prediction under observer-frame reorientation. We introduce $\textbf{Observer-Aware Relations (OAR)}$, which combines observer-aware geometric encoding and relation specialization, supported by frame-stable object encoding, for unified multi-label predicate prediction. Experiments on 3DSSG show that OAR consistently outperforms baselines across controlled observer-frame reorientations without training-time frame-reorientation augmentation, while remaining competitive on the standard benchmark. The project page is available at https://oar-predicate.github.io/.

arXiv

Symmetric Behavior Regularized Policy Optimization

Lingwei Zhu, Haseeb Shah, Zheng Chen, Martha White

Behavior Regularized Policy Optimization (BRPO) leverages asymmetric divergence regularization to mitigate distribution shift in offline reinforcement learning. This paper is the first to study the open question of symmetric BRPO. Using didactic examples, we show that symmetric regularization can outperform asymmetric regularization in addressing one-sided bias, near-boundary policy updates, and projection geometry consistency. However, symmetric divergences do not fit BRPO naturally: they do not permit a closed-form solution when used as regularizers, and can lead to numerical instability when used as optimization objectives. We first introduce a universal BRPO framework using an infinite series of Pearson-Vajda divergences to represent any $f$-divergence, which includes both symmetric and asymmetric divergences. We use a finite-series approximation to obtain the following results for symmetric BRPO: (1) a closed-form optimal policy expression; (2) a numerically stable optimization surrogate; and (3) a tight upper bound on the approximation quality. On the D4RL benchmark and in didactic examples, we show that the proposed method achieves consistently strong results and is robust to the number of terms in the approximation.

arXiv

MMGraphRAG: Bridging Vision and Language with Interpretable Multimodal Knowledge Graphs

Xueyao Wan, Hang Yu

Large Language Models (LLMs) suffer from hallucinations due to their static parametric knowledge. Retrieval-Augmented Generation (RAG) and GraphRAG mitigate this issue by incorporating external knowledge and structured reasoning over knowledge graphs (KGs). However, existing approaches remain largely text-centric, as constructing fine-grained multimodal knowledge graphs (MMKGs) with explicit cross-modal semantics remains challenging. In this paper, we propose MMGraphRAG, a framework for building interpretable MMKGs that unify textual and visual knowledge. Our approach represents visual content as structured scene graphs and integrates them with textual KGs through a novel cross-modal entity linking method, SpecLink, which leverages spectral clustering to jointly model semantic similarity and graph structure. This design preserves explicit entities, relations, and reasoning paths across modalities, enabling structure-aware retrieval and generation. To support evaluation, we introduce the CMEL dataset, a benchmark for fine-grained cross-modal entity alignment. Experimental results on CMEL demonstrate improved entity linking accuracy, while evaluations on DocBench and MMLongBench show that MMGraphRAG achieves superior performance and stronger robustness, particularly in complex multimodal reasoning scenarios.

arXiv

Gradient Span Algorithms Make Predictable Progress in High Dimension

Felix Benning, Leif Döring

We prove that all 'gradient span algorithms' have asymptotically deterministic behavior on scaled Gaussian random functions as the dimension tends to infinity. This is a functional generalization of similar results for random quadratic functions and spin glasses. They explain the counterintuitive phenomenon that different training runs of many large machine learning models result in approximately equal cost curves despite random initialization on a complicated non-convex landscape. This 'predictable progress' phenomenon is exploited by the AutoML community: Since the optimization progress of a single run is already representative, multiple retries with the same hyperparameters are not necessary.

arXiv

Online-Score-Aided Federated Learning for Resource-Constrained Wireless Clients with Continual Data Arrival

Ferdous Pervej, Minseok Choi, Andreas F. Molisch

Heterogeneous system configurations of distributed clients connected to the central server (CS) via a time-varying wireless network pose significant challenges for popular distributed machine learning (ML) algorithms such as federated learning (FL). Although the limited (radio and computational) resources are widely acknowledged, two critical yet often ignored aspects are (a) client devices can only dedicate a small chunk of their limited storage for the FL task and (b) new training samples may arrive continually in many practical wireless applications. Therefore, we propose a new FL algorithm, online-score-aided federated learning (OSAFL), specifically designed for tasks with continual data arrival in resource-constrained environments. We first theoretically show how the convergence bound is affected by continual data distribution shifts, uncertain client participation, gradient quantization errors, and noise from stochastic gradients and statistical data heterogeneity across clients. We then show how to (sub-optimally) minimize these errors by choosing appropriate aggregation weights at the CS during global update. Our extensive simulation results across three popular image classification datasets and three ML models with different numbers of trainable parameters validate the effectiveness of the proposed OSAFL algorithm compared to (modified) state-of-the art FL baselines.

arXiv

Curvature-Adaptive Consistency Flow Matching: Autonomous Trajectory Optimization via Reinforcement Learning

Songtao Tian, Guhan Chen, Bohan Li, Jingyi Ma, Zixiong Yu

Consistency distillation has significantly accelerated diffusion-model inference, but its sampling dynamics remain underexplored. We reveal an asymmetry: although Logit-Normal sampling priors work well for standard iterative generation, consistency distillation exhibits a different difficulty profile (e.g., U-shaped), with optimization bottlenecks concentrated at the boundary stages rather than intermediate steps. To address the limitations of static sampling under evolving learning demands, we propose Curvature-Adaptive Consistency Flow Matching (CACFM). By formulating distillation as a dynamic decision process, CACFM uses a lightweight reinforcement learning agent to probe Probability Flow ODE trajectories and construct an efficiency-oriented curriculum that prioritizes critical regions without manual scheduling. Combined with Flow-adapted DMD and adversarial consistency objectives, our RL-based scheduler achieves state-of-the-art results on large-scale models such as FLUX and SDXL, mitigating structural deformities and preserving high-frequency details in extreme few-step regimes.

arXiv

How Much Is a Dataset Worth? Scaling Laws, the Vendi Score, and Matrix Spectral Functions

Jeff A. Bilmes, Gantavya Bhatt, Arnav M. Das

Neural scaling laws appraise data through dataset size, while the Vendi Score uses quantum entropy to measure dataset value. We show both that common neural-scaling-law objectives and the Vendi Score are submodular. We further show that the Vendi Score is a special case of a broader class of submodular objectives that we call matrix spectral functions. This also includes determinantal (DPP) objectives, as well as many others. We also introduce weakly matrix monotone functions and show how they lead to weakly submodular matrix spectral functions, yielding a broad family of practical objectives for data appraisal. We develop secular-equation-based updates that avoid repeated eigendecompositions during greedy optimization, reducing marginal-gain evaluation for $m$-dimensional embeddings by an $O(m)$ factor relative to oracle queries. This yields an average empirical speedup of about 35,000x, making direct optimization of the Vendi Score feasible on ImageNet-1K-scale datasets. Thus enabled, we compare how well several objectives predict the value of training subsets for held-out test performance under fixed-size, class-balanced, and fixed training-budget regimes, including the Vendi Score, DPPs, facility location, and three new matrix spectral variants. Across multiple datasets, facility location performs the best. Direct optimization also reveals that, while the Vendi Score is predictive over moderate score ranges, pushing the objective to higher values can make it a poor downstream performance proxy. We also find that uniformly at random fixed-size subsets, both unconstrained and class-balanced, are remarkably concentrated in both appraisal scores and held-out performance. Finally, we show that size, class balance, and training budget do not alone determine data value: even when controlling for these factors, performance ranges smoothly from good to bad.

arXiv

Tetris: Tile-level Sampling for Efficient and High-Fidelity Video Object Tracking

Chanwut Kittivorawong, Alena Chao, Charlie Si, Alvin Cheung

Track materialization converts raw videos into reusable object tracks that downstream queries can run against without rerunning tracking, but extracting those tracks efficiently and with high fidelity remains expensive. Prior systems reduce track materialization cost through temporal frame sampling, but aggressive sampling spaces each track's detection points too far apart to faithfully capture the object's actual trajectory. In stationary video, however, large portions of each frame contain no objects of interest, and different sampling rates can be used to extract tracks from the remaining regions. Leveraging this idea, we present Tetris, a track-extraction system that decomposes videos into a tile-based polyomino data model, enabling fine-grained spatiotemporal pruning that reduces detector calls with minimal fidelity loss. Tetris implements track materialization in three steps: first, a classifier identifies relevant tiles and groups them into polyominoes. Then, we use an integer linear program (ILP) to prune redundant polyominoes under a user-specified accuracy constraint, before packing the remaining polyominoes into canvases to minimize detector calls. Across 7 stationary-video datasets, Tetris stays within a 5% tracking accuracy loss as compared to a reference pipeline that processes every frame in its entirety, while prior systems exceed this bound on 3 of the 7 datasets. Moreover, with this 5% bound, Tetris achieves up to 17.4x higher throughput than prior systems, and up to 68.8x higher than the reference pipeline. Conversely, Tetris delivers up to 0.42 higher HOTA tracking accuracy than the best prior system at matched throughput. The project page is at https://tetris-db.github.io .

arXiv

The challenge of hidden gifts in multi-agent reinforcement learning

Dane Malenfant, Blake A. Richards

Sometimes we benefit from actions that others have taken even when we are unaware that they took those actions. For example, if your neighbor chooses not to take a parking spot in front of your house when you are not there, you can benefit, even without being aware that they took this action. These ``hidden gifts'' represent an interesting challenge for multi-agent reinforcement learning (MARL), since assigning credit when the beneficial actions of others are hidden is non-trivial. Here, we study the impact of hidden gifts with a simple MARL task. In this task, agents in a grid-world environment have individual doors to unlock in order to obtain individual rewards. As well, if all the agents unlock their door the group receives a larger collective reward. However, there is only one key for all of the doors, such that the collective reward can only be obtained when the agents drop the key for others after they use it. Notably, there is nothing to indicate to an agent that the other agents have dropped the key, thus this act for others is a ``hidden gift''. We show that several different state-of-the-art MARL algorithms, including MARL specific architectures, fail to learn how to obtain the collective reward in this simple task. Interestingly, we find that decentralized actor-critic policy gradient agents can succeed when we provide them with information about their own action history, but MARL agents still cannot solve the task with action history. Finally, we derive a correction term for policy gradient agents, inspired by learning aware approaches, which reduces the variance in learning and helps them to converge to collective success more reliably. These results show that credit assignment in multi-agent settings can be particularly challenging in the presence of ``hidden gifts'', and demonstrate that self learning-awareness in decentralized agents can benefit these settings.

arXiv

Data Balancing Strategies: A Systematic Survey of Resampling and Augmentation Methods

Behnam Yousefimehr, Mehdi Ghatee, Javad Fazli, Shervin Ghaffari, Zahra Rafei, Mohammad Amin Seifi, Sajed Tavakoli, Abolfazl Nikahd, Mahdi Razi Gandomani, Alireza Orouji, Ramtin Mahmoudi Kashani, Sarina Heshmati, Negin Sadat Mousavi

Imbalanced datasets, where one class significantly outnumbers others, remain a persistent challenge in machine learning, often biasing predictions toward the majority class and degrading classifier performance. This paper provides a comprehensive, systematic review of data balancing methods, extending beyond foundational oversampling techniques such as the Synthetic Minority Oversampling Technique (SMOTE) and its variants (e.g., Borderline SMOTE, K-Means SMOTE, and Safe-Level SMOTE) to encompass advanced adaptive methods (MWMOTE, AMDO), deep generative models (generative adversarial networks, variational autoencoders, and diffusion models), undersampling techniques (NearMiss, Tomek Links), combination/hybrid methods (SMOTE-ENN, SMOTE-Tomek, and SMOTE+OCSVM), ensemble strategies (SMOTEBoost, RUSBoost, Balanced Random Forest, and One-Sided Selection), and specialized approaches for multi-label and clustered data. Beyond descriptive categorization, this review critically examines each method's underlying assumptions, operational mechanisms, and suitability for diverse data characteristics, including high dimensionality, mixed feature types, class overlap, and noise. Key findings demonstrate that no single method universally outperforms others; optimal selection depends critically on dataset characteristics, classifier choice, and evaluation metrics. The paper concludes by identifying emerging research directions, including self-supervised learning for imbalance, diffusion-based generative oversampling, distribution-preserving resampling, knowledge distillation for imbalanced deployment, and the adaptation of foundation models to skewed distributions, offering practical guidelines for practitioners and a roadmap for future methodological development.

arXiv

Training Continuous Chain of Thought Models: A Tale of Two Regimes

Varun Yerram, He He, Eunsol Choi

Continuous Chain-of-Thought methods replace verbose reasoning traces with a short sequence of dense latent representations. Earlier continuous CoT methods indirectly supervise the latent representations such that its final state match that of verbose reasoning traces, requiring autoregressive, slow generation during training. We introduce C-MTP, a simpler, faster direct supervision approach that models each latent as an average of the embeddings in the CoT traces to be compressed. Our approach outperforms a prior direct supervision method that approximates the distribution of compressed tokens, and performs competitively to slower indirect supervision approaches in existing evaluation setup with simplified CoT traces (less than 100 tokens). Lastly, we extend the evaluation of Continuous CoT methods to complex tasks with longer reasoning traces ($\ge$ few hundreds reasoning tokens). We find both direct and indirect supervision training methods perform poorly (roughly 65\% performance drop) in this setting, revealing the limitations of current continuous CoT methods. The code and checkpoints are released at https://github.com/Varun221/cmtp_research

arXiv

A Method for Learning Value Systems in Generative AI

Andrés Holgado-Sánchez, Holger Billhardt, Sascha Ossowski

Value-aware AI systems require explicit computational representations of human values (groundings) and their aggregation into value systems in order to align their decisions with ours. As such representations are difficult to elicit, value learning seeks to infer them by observing human behaviour. This work addresses the lack of grounded value learning methods in generative AI: existing approaches typically replicate human preferences without awareness of the multidimensional structure of value alignment, or lack principled value system elicitation methods. To address these gaps, we adapt a previously validated value system learning method to the generative AI setting, which, based on pairwise prompt-response preference data, simultaneously learns: i) an implementation of a grounding for a set of values given by a multi-objective reward model, and ii) a value system representation in the form of a weighted linear scalarization of the previous grounding model. To ensure that the learned value systems are based on coherent value representations, our algorithm dynamically prioritizes the grounding learning process. We evaluate the method against baselines and a contemporary method on prompt-response preference datasets. Results show competitive performance and minimal trade-offs against the baselines, while improving explainability.

arXiv

Ghosts in Neural Networks: Existence, Structure and Role of Infinite-Dimensional Null Space

Sho Sonoda, Isao Ishikawa, Masahiro Ikeda

We study parameter nonuniqueness in continuous-width depth-two fully connected neural networks. Our main contribution is a direct method for solving the neural-network equation $S[γ]=f$. Starting from the Fourier expression of the synthesis operator, separation of variables produces a ridgelet particular solution and identifies every homogeneous direction. To isolate the argument, we first prove an abstract reconstruction formula for unitary factorizations, yielding the adjoint, normalized right inverse, and orthogonal solution geometry. We then specialize this formula to neural-network synthesis: for tempered-distribution activations such as ReLU, we equip the activation class $A_{s,t}$ with a Hilbert structure, construct compatible coefficient and parameter Hilbert spaces $H_{s,t}$ and $G_{s,t}$, and prove that $S:G_{s,t}\to L^2(\mathbb R^m)$ is bounded. The resulting ridgelet expansion exhausts the null space and the complete solution set and identifies the unique minimum-norm parameter distribution. Concrete examples give adjoint ridgelet functions for standard activations. Further developments show that finite-measure null elements admit normalized width-$N$ discretizations with $O(N^{-1/2})$ output error and characterize how additive parameter perturbations can reveal information encoded in the null space. A Lean 4 blueprint for the main results is available at https://shosonoda.github.io/lean-ridgelet/ .

arXiv

DeliCIR: Memory-Guided Test-Time Deliberation via Multi-Agent Collaboration for Composed Image Retrieval

Xingtian Pei, Yukun Song, Changwei Wang, Shunpeng Chen, Rongtao Xu, Shengpeng Xu, Shibiao Xu

Composed Image Retrieval (CIR) requires both preserving the visual continuity of the reference image and faithfully executing the semantic variables specified in the modification text, which constitute the core challenge of the task. Existing methods often suffer from Perception Myopia in a single space, or fall into Logic Drift in iterative collaboration due to the perception ceiling of the underlying retriever. To address this issue, we propose a one-stop hierarchical Perception-to-Deliberation Framework (PDF), which, to the best of our knowledge, is the first to introduce Test-Time Scaling (TTS) into CIR. Relying on a hierarchical multi-agent architecture, PDF first utilizes an Intent Routing Manager to dynamically dispatch multi-view Worker perception signals based on modification intents to construct a high-recall candidate pool. Subsequently, the Deliberation Manager combines a training-free reasoning experience distillation mechanism with a tournament-style test-time scaling strategy (T-TTS) to perform experience-guided fine-grained reasoning and produce the final retrieval results. Experimental results demonstrate that PDF achieves SOTA performance on three benchmark datasets: CIRR, CIRCO, and Fashion IQ. This study indicates that experience-driven memory mechanism and TTS represent a highly promising and scalable path for achieving fine-grained multimedia retrieval. The code will be made publicly available upon acceptance.

arXiv

Geometrically Consistent Multi-View Scene Generation from Freehand Sketches

Ahmed Bourouis, Savas Ozkan, Andrea Maracani, Yi-Zhe Song, Mete Ozay

We tackle a new problem: generating geometrically consistent multi-view scenes from a single freehand sketch. Freehand sketches are the most geometrically impoverished input one could offer a multi-view generator. They convey scene intent through abstract strokes while introducing spatial distortions that actively conflict with any consistent 3D interpretation. No prior method attempts this; existing multi-view approaches require photographs or text, while sketch-to-3D methods need multiple views or costly per-scene optimisation. We address three compounding challenges; absent training data, the need for geometric reasoning from distorted 2D input, and cross-view consistency, through three mutually reinforcing contributions: (i) a curated dataset of $\sim$9k sketch-to-multiview samples, constructed via an automated generation and filtering pipeline; (ii) Parallel Camera-Aware Attention Adapters (CA3) that inject geometric inductive biases into the video transformer; and (iii) a Sparse Correspondence Supervision Loss (CSL) derived from Structure-from-Motion reconstructions. Our framework synthesizes all views in a single denoising process without requiring reference images, iterative refinement, or per-scene optimization. Our approach significantly outperforms state-of-the-art two-stage baselines, improving realism (FID) by over 60% and geometric consistency (Corr-Acc) by 23%, while providing up to a 3.7$\times$ inference speedup.

arXiv

Revealing Physical-World Semantic Vulnerabilities: Universal Adversarial Patch for Infrared Vision-Language Models

Chengyin Hu, Yuxian Dong, Yikun Guo, Xiang Chen, Qike Zhang, Junqi Wu, Jiahuan Long, Jiujiang Guo, Yiwei Wei, Tingsong Jiang, Wen Yao

Infrared vision-language models (IR-VLMs) are becoming important for semantic perception in low-visibility environments, yet their robustness to physical semantic attacks remains underexplored. Existing adversarial patch methods are mainly designed for red-green-blue (RGB) images or closed-set infrared detectors and do not directly address open-ended IR-VLM tasks, where one deployable artifact can affect classification, captioning, and visual question answering (VQA) simultaneously. We propose Universal Curved-Grid Patch, abbreviated UCGP, a universal physical adversarial patch framework tailored to IR-VLMs. UCGP represents the patch as a deployable low-frequency curved grid and optimizes it with a representation-driven objective over subspace departure, topology disruption, and local appearance regularization. Meta Differential Evolution (MetaDE) searches for patch parameters with two physical robustness augmentations: Expectation over Transformation (EOT) capture sampling for imaging variations and thin-plate spline (TPS) patch-deformation modeling for non-rigid local shape changes. Rather than manipulating labels or prompts, UCGP disrupts the clean-category manifold in visual representation space, which later appears as degraded cross-modal outputs. Experiments show that a single shared patch degrades classification, captioning, and VQA across diverse IR-VLM architectures while retaining measurable cross-model transfer, cross-dataset generalization, cross-category extensibility, and real-scene physical effectiveness under the evaluated conditions. These results reveal a robustness blind spot in current infrared multimodal systems. Code and reproducibility assets are available at https://github.com/dyx6663/UCGP.

arXiv

Trace-Based On-Policy Distillation for Masked Diffusion Language Models

Haolin Ren, Ziyang Huang, Chenhao Yuan, Jun Zhao, Kang Liu

Diffusion large language models (dLLMs) are a promising alternative to autoregressive generation. However, reasoning-oriented post-training for dLLMs remains challenging. Supervised fine-tuning (SFT) for dLLMs requires dense but often off-policy masked states, while reinforcement learning (RL) relies on sparse rewards or value modeling. This paper proposes \textbf{trace-based on-policy distillation (TOPD)}, a teacher-supervised framework that transfers reasoning ability to a target dLLM without reward estimation. The key idea is to supervise a dLLM on its own denoising trajectory, focusing on the trace-aligned token decisions that form the final response. Specifically, TOPD samples on-policy diffusion trajectories from the target dLLM, obtains teacher token distributions from a teacher model on the corresponding partially denoised states, and updates the target dLLM with a token-level Reverse Kullback-Leibler (Reverse-KL) objective. This design preserves dense teacher supervision while aligning training with the model's own denoising states. On mathematical reasoning benchmarks, TOPD enables SDAR-4B-Chat to match the MATH500 accuracy of its RL-trained counterpart TraDo-4B-Instruct, with gains of +5.7 under static evaluation and +4.5 under dynamic evaluation. Compared with the RL-trained counterpart, TOPD achieves this with 4$\times$ fewer rollout rounds, corresponding to an estimated 96.0$\times$ to-accuracy model-compute speedup.

Keep browsing

More papers from top AI conferences

30+ top conferences, by conference, year, and topic.