1小时前更新HISTORY近 30 天历史柱高表示当天去重热搜数量
570360045305550657075308590955106011571259135914611556165817561855195720592153225323592467255826512777286129593062315801
08/03—09/01 有历史数据
- 01Revisiting Dynamic Graphs from the Perspective of Time Series期刊：Transactions on Machine Learning Research · 摘要：Numerous studies have investigated temporal modeling in dynamic graphs. Existing approaches predominantly fall into two categories: discrete-time dynamic graph (DTDG) methods and continuous-time dynamic graph (CTDG) methods. While both paradigms have shown effectiveness in capturing temporal dependencies, they suffer from several inherent limitations. Specifically, DTDG approaches often lose fine-grained temporal information due to snapshot-based discretizations, whereas CTDG methods preserve precise timestamps but may struggle to capture long-range temporal dependencies because of computatio… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：rq697OS6FUWeili Shi, Dongliang Guo, Sheng Li
- 02On Next-Token Prediction in LLMs: How End Goals Determine the Consistency of Decoding Algorithms期刊：Transactions on Machine Learning Research · 摘要：Probabilistic next-token prediction trained using cross-entropy loss is the basis of most large language models. Given a sequence of previous values, next-token prediction assigns a probability to each possible next value in the vocabulary. From this, there are many ways to turn these next-token predictions into token sequences. This paper examines a few of these algorithms (greedy/lookahead decoding, random sampling, and temperature-scaled random sampling) and studies their consistency with respect to the user end goals of information retrieval and creative generation through encoding these… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：96S5nE51L8Jacob Trauger, Ambuj Tewari
- 03Verification and Training of Neural Networks for Robustness Against Neuron Pruning期刊：Transactions on Machine Learning Research · 摘要：Structured neuron pruning removes entire hidden units to reduce model size and computation but often leads to unpredictable accuracy degradation. Existing pruning methods typically rely on heuristic importance scores and provide no formal guarantees on the behavior of pruned models. In this work, we propose a certifiable approach for structured neuron pruning in fully connected layers of feedforward neural networks that guarantees robustness against all pruning masks satisfying a given layer-wise sparsity budget. We further develop a computable upper bound on the worst-case change in pairwise… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/danghue18/Robustness-against-pruning · OpenReview ID：6m3LtUZwvFHue Dang, Nguyen Duc Anh Pham, Hung Manh Pham et al.
- 04Simulation-free Structure Learning for Stochastic Population Dynamics期刊：Transactions on Machine Learning Research · 摘要：Modeling dynamical systems and unraveling their underlying structural dependencies is central to many domains in the natural sciences. Various physical systems, such as those arising in cell biology, are inherently high-dimensional and stochastic in nature, and admit only partial, noisy state measurements. Our primary motivating setting is single-cell biology, where destructive measurements yield unpaired population snapshots rather than longitudinal trajectories of the same cells. This poses a significant challenge for addressing the problems of modeling the underlying dynamics and inferring… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/NoahElRimawiFine/StructureFlow · OpenReview ID：Kj9cibMkV3Noah El Rimawi-Fine, Adam Stecklov, Lucas Nelson et al.
- 05A Few Large Shifts: Layer-Inconsistency Based Minimal Overhead Adversarial Example Detection期刊：Transactions on Machine Learning Research · 摘要：Deep neural networks (DNNs) are highly susceptible to adversarial examples—small, malicious perturbations that can cause incorrect predictions. We introduce a lightweight, plug-in detector that uses internal layer-wise inconsistencies within the target model and requires only benign data for fitting and calibration. The approach is motivated by the **A Few Large Shifts Assumption**, an empirical hypothesis that adversarial perturbations often produce large, localized growth in representation changes across a small number of consecutive layers, connecting adversarial behavior to layer-wise Lip… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/c0510gy/AFLS-AED · OpenReview ID：0xXYBxDNHASanggeon Yun, Ryozo Masukawa, Hyunwoo Oh et al.
- 06FocusAgent: Simple Yet Effective Ways of Trimming the Large Context of Web Agents期刊：Transactions on Machine Learning Research · 摘要：Web agents powered by large language models (LLMs) must process lengthy web page observations to complete user goals; these pages often exceed tens of thousands of tokens. This saturates context limits and increases computational cost processing; moreover, processing full pages exposes agents to security risks such as prompt injection. Existing pruning strategies either discard relevant content or retain irrelevant context, leading to suboptimal action prediction. We introduce FocusAgent, a simple yet effective approach that leverages a lightweight LLM retriever to extract the most relevant l… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/imenelydiaker/focus_agent · OpenReview ID：mINaJKSy7AImene Kerboua, Sahar Omidi Shayegan, Xing Han Lù et al.
- 07CURATE: Automatic Curriculum Learning for Reinforcement Learning Agents through Competence-Based Curriculum Policy Search in Structured Task Spaces期刊：Transactions on Machine Learning Research · 摘要：Due to fundamental exploration challenges without informed priors or specialized algorithms, agents may be unable to consistently receive informative rewards, leading to inefficient or intractable learning. To address these challenges, we introduce CURATE, an automatic curriculum learning algorithm for reinforcement learning agents in structured task spaces of monotonic difficulty. Through "exploration by exploitation," CURATE dynamically scales the task difficulty to match the agent's current competence. By exploiting its current capabilities that were learned in easier tasks, the agent impr… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/montrealrobotics/curate-code · OpenReview ID：DlnvWfoIgvTabitha Edith Lee, Nan Rosemary Ke, Sarvesh Patil et al.
- 08Break the Brake, Not the Wheel: Untargeted Jailbreak via Entropy Maximization期刊：Transactions on Machine Learning Research · 摘要：Recent studies show that gradient-based universal image jailbreaks on vision-language models (VLMs) exhibit little or no cross-model transferability, casting doubt on the feasibility of transferable multimodal jailbreaks. We revisit this conclusion under a strictly untargeted threat model without enforcing a fixed prefix or response pattern. Our preliminary experiment reveals that refusal behavior concentrates at high-entropy tokens during autoregressive decoding, and non-refusal tokens already carry substantial probability mass among the top-ranked candidates before attack. Motivated by this… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：dfX4zpagYFMengqi He, Xinyu Tian, Xin Shen et al.
- 09Scalable Learning from Probability Measures with Mean Measure Quantization期刊：Transactions on Machine Learning Research · 摘要：We consider statistical learning problems in which data are observed as a set of probability measures. Optimal transport (OT) is a popular tool to compare and manipulate such objects, but its computational cost becomes prohibitive when the measures have large support. We study a quantization-based approach in which all input measures are approximated by $K$-point discrete measures sharing a common support. We establish consistency of the resulting quantized measures. We further derive convergence guarantees for several OT-based downstream tasks computed from the quantized measures. Numerical… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：eA4cQkY7UgErell Gachon, Elsa Cazelles, Jérémie Bigot
- 10FORGE: Forward-Only Test-Time Adaptation for Integer-Only Vision Models on Microcontrollers期刊：Transactions on Machine Learning Research · 摘要：Vision models deployed on microcontrollers (MCUs) are quantized to integer-only arithmetic and run in inference-only runtimes that do not carry the machinery backpropagation needs: the standard tool for adapting a model to the distribution shift (sensor noise, blur, lighting) it meets in the field. Existing forward-only test-time adaptation (TTA) methods either run only on server- or edge-GPU-class models (not true microcontroller integer execution), or require the batch-normalization (BN) layers that integer deployment fuses away. We present a forward-only TTA method that operates on deploye… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/Rehan000/forge-tta · OpenReview ID：A45I5p25ddMuhammad Rehan, Haider Ali, Muhammad Ali Munir et al.
- 11MLLM4TS: Leveraging Vision and Multimodal Language Models for General Time-Series Analysis期刊：Transactions on Machine Learning Research · 摘要：Effective analysis of time series data presents significant challenges due to the complex temporal dependencies and cross-channel interactions in multivariate data. Inspired by the way human analysts visually inspect time series to uncover hidden patterns, we ask: can incorporating visual representations enhance automated time-series analysis? Recent advances in multimodal large language models have demonstrated impressive generalization and visual understanding capability, yet their application to time series remains constrained by the modality gap between continuous numerical data and discr… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：bhd6naKDoLQinghua Liu, Sam Heshmati, Zheda Mai et al.
- 12FlashBind: Towards Accurate and Efficient Structure-based Virtual Screening期刊：Transactions on Machine Learning Research · 摘要：Accurate prediction of protein-ligand interactions is central to computational drug discovery. Recent foundation models such as Boltz-2 have achieved remarkable accuracy in binding affinity prediction, yet their prohibitive computational cost remains a major barrier to large-scale virtual screening. Here we introduce FlashBind, a lightweight structure-based model that achieves a 50× speedup over Boltz-2 at inference time by replacing expensive structure prediction with a fast docking model and substituting costly PairFormer modules with a streamlined EGNN architecture. FlashBind attains early… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/AIDD-Lab/FlashBind · OpenReview ID：2D91AcVcMiSonglin Jiang, Yifan Chen, Aarti Krishnan et al.
- 13Graph Contrastive Learning via Weisfeiler-Leman Dual-View Sampling期刊：Transactions on Machine Learning Research · 摘要：Graph contrastive learning (GCL) approaches have gained momentum over the past few years. By augmenting the original graph data, the common GCL pipeline learns from such multiple contrastive graph views in a self-supervised manner, tackling critical issues in the literature, such as node label scarcity. To obtain contrastive views, most GCL techniques heavily rely on feature-space similarity measures. We consider this as a limiting factor in GCL, since it implies that node features are (in general) informative and closely aligned with the graph topology, an assumption that does not hold, for… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/172698691/WLGCL · OpenReview ID：uuk14WVKyjZiyu Wu, Daniele Malitesta, Giannis Nikolentzos et al.
- 14Trust Region Policy Optimization for Functional Linear Policies期刊：Transactions on Machine Learning Research · 摘要：Reinforcement Learning (RL) tasks where the states are given by spatial or temporal measurements often lead to high-dimensional state spaces, making function approximation difficult and unstable. We adapt the classic RL framework to allow the direct use of the inherent functional state, which can be estimated from the discrete measurements. We propose a suitable family of policies based on functional linear models, allowing us to take actions conditionally on functional states. Moreover, we extend Trust Region Policy Optimization (TRPO) to improve such policies and address the challenge of op… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：mRNOs9Y8t1Simon Weinberger, Jairo Cugliari
- 15Dynamic Model Routing and Cascading for Efficient LLM Inference: A Survey期刊：Transactions on Machine Learning Research · 摘要：The rapid growth of large language models (LLMs) with diverse capabilities, costs, and domains has created a critical need for intelligent model selection at inference time. While smaller models suffice for routine queries, complex tasks demand more capable models. However, static model deployment does not account for the complexity and domain of incoming queries, leading to suboptimal performance and increased costs. Dynamic routing systems that adaptively select models based on query characteristics have emerged as a solution to this challenge. This survey provides a systematic analysis of… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/ymoslem/awesome-llm-routing-cascading · OpenReview ID：ypRg1TvQaMYasmin Moslem, John Kelleher
- 16Scaffold-Conditioned Preference Triplets for Controllable Molecular Optimization with Large Language Models期刊：Transactions on Machine Learning Research · 摘要：Molecular property optimization is central to drug discovery, yet many deep learning methods rely on black-box scoring and offer limited control over scaffold preservation, often producing unstable or biologically implausible edits. While large language models (LLMs) are promising molecular generators, optimization remains constrained by the lack of chemistry-grounded preference supervision and principled data curation. We introduce \textbf{Scaffold-Conditioned Preference Triplets (SCPT)}, a pipeline that constructs similarity-constrained triplets $\langle\text{scaffold}, \text{better}, \text… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：KwszuDv9owYi Xiong, Xiaohong Ji, Liang Xiong et al.
- 17Protein structural superfamily classification using hand-crafted and language model features: A performance vs interpretability trade-off期刊：Transactions on Machine Learning Research · 摘要：The newfound rise of protein language models (PLMs) that leverage data and compute has introduced an interesting conflict: a trade-off between the high predictive performance of non-interpretable features and the scientific insight that can be gained from interpretable, hand-crafted ones. In this work, we highlight and study this conflict via the task of classifying protein domains into their CATH superfamilies. We train one-vs-all (OvA) linear SVM classifiers for 45 diverse CATH superfamilies, each characterised by significant class imbalance. Our analysis compares nine feature vector types,… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/Pranav-Machingal/CATH_OvA_classification · OpenReview ID：huTeyYU0yDPranav Machingal, Rakesh Busi, Nandyala Hemachandra et al.
- 18Entropy-Controlled Time-Discretization Bounds and Loss-Adaptive Schedules for Diffusion Models期刊：Transactions on Machine Learning Research · 摘要：Diffusion generative models synthesize samples by discretizing reverse-time dynamics driven by a learned score or denoiser. Existing analyses of reverse-diffusion discretization often exhibit explicit dependence on the ambient dimension, while guarantees without explicit dimension dependence typically require structural or geometric assumptions on the target distribution. We develop an information-theoretic approach to reverse-diffusion discretization that avoids such assumptions. We decompose the pathwise KL error into initialization, denoiser approximation, and time-discretization terms, an… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：5CNcyOBV3jAhmad Aghapour, Erhan Bayraktar, Ziqing Zhang
- 19GATTA: Graph Active Learning with Test-Time Augmentation期刊：Transactions on Machine Learning Research · 摘要：Test-time augmentation (TTA) has proven effective for improving model robustness and uncertainty estimation in computer vision, yet its application to graph-structured data remains largely unexplored. We introduce GATTA (Graph Active Learning with Test-Time Augmentation), a framework for enhancing active learning by aggregating predictions across multiple augmented views to produce more reliable uncertainty estimates. To address the challenge of label-preserving graph augmentations, GATTA incorporates a consistency-based filtering mechanism that discards augmented views yielding unreliable pr… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：gQjXARc6ttZsombor Bánfi, András Gézsi, András Formanek
- 20Length-MAX Tokenizer for Language Models期刊：Transactions on Machine Learning Research · 摘要：Length-MAX is a tokenizer built to maximize average token length, and therefore encodes text in fewer tokens, by removing pre-tokenization and admitting multi-word tokens under explicit boundary discipline. Candidates are scored by $f(t)\,(n(t)-1)$, where $n(t)-1$ is the exact token saving at each realized occurrence of a candidate spanning $n(t)$ segmentation units, with the BPE merge step as the $n{=}2$ special case. Termhood filters govern admission, and a boundary-enforced decoder reproduces the construction-time segmentation at deployment, a property we test rather than prove, at zero mi… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/dongwinterdong-upenn/length-max-tokenizer · OpenReview ID：CRjrpwGD9ADONG DONG, Weijie J Su
- 21FreeFuse: Multi-Subject LoRA Fusion via Adaptive Token-Level Routing at Test Time期刊：Transactions on Machine Learning Research · 摘要：This paper proposes FreeFuse, a training-free framework for multi-subject text-to-image generation through automatic fusion of multiple subject LoRAs. \rev{In contrast to prior studies that focus on retraining LoRAs to alleviate feature conflicts, our analysis shows that spatially routing LoRA residuals to their intended semantic regions provides an effective mechanism for suppressing direct cross-region LoRA interference while preserving the base model's global contextual reasoning.} Accordingly, we implement Adaptive Token-Level Routing during the inference phase. However, obtaining reliabl… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/yaoliliu/FreeFuse · OpenReview ID：Bvela8OAMcYaoli Liu, Yao-Xiang Ding, Kun Zhou
- 22Bridging Formal Language with Chain-of-Thought Reasoning to Geometry Problem Solving期刊：Transactions on Machine Learning Research · 摘要：Large vision language models exhibit notable limitations on Geometry Problem Solving (GPS) because of their unreliable diagram interpretation and pure natural-language reasoning. A recent line of work mitigates this by using symbolic solvers: the model directly generates a formal program that a solver can execute. However, this direct program generation lacks intermediate reasoning, making the program prone to errors. In this work, we explore integrating Chain-of-Thought (CoT) with formal language for visually grounded geometry problem solving. The model interleaves natural language reasoning… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：6eaWuhhdi8Tianyun Yang, Yunwen Li, Ziniu Li et al.
- 23FuseLIP: Multimodal Embeddings via Early Fusion of Discrete Tokens期刊：Transactions on Machine Learning Research · 摘要：Contrastive language-image pre-training aligns features of text-image pairs in a common latent space via distinct encoders for each modality. While this approach achieves impressive performance in several zero-shot tasks, it cannot natively handle multimodal inputs, i.e., encoding image and text into a single feature vector. As a remedy, it is common practice to use additional modules to merge the features extracted by unimodal encoders. In this work, we present FuseLIP, a new architecture for multimodal embedding. Leveraging recent progress in discrete image tokenizers, we propose to use a s… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/chs20/fuselip · OpenReview ID：yq9je6kLC6Christian Schlarmann, Francesco Croce, Nicolas Flammarion et al.
- 24Dynamically Scaled Activation Steering期刊：Transactions on Machine Learning Research · 摘要：Activation steering has emerged as a powerful method for guiding the behavior of generative models towards desired outcomes such as toxicity mitigation. However, most existing methods apply interventions uniformly across all inputs, degrading model performance when steering is unnecessary. We introduce Dynamically Scaled Activation Steering (DSAS), a method-agnostic steering framework that decouples when to steer from how to steer. DSAS adaptively modulates the strength of existing steering transformations across layers and inputs, intervening strongly only when undesired behavior is detected… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：EXKv1KMNtRAlex Ferrando, Xavier Suau, Jordi Gonzàlez et al.
- 25An Efficient Framework for Length Extension via Dynamically Growing Positional Embedding and Routing Attention期刊：Transactions on Machine Learning Research · 摘要：Modeling long sequences is critical for numerous large-scale models. However, extending existing architectures to handle significantly longer sequences poses substantial technical and computational challenges. One inevitable issue is the overfitting of large models to positional encodings during pretraining, which limits their ability to generalize to unseen positional encoding scales. Additionally, extending sequence lengths requires extensive computational resources and time. Existing positional encoding methods often rely on carefully designed scaling factors but typically yield suboptimal… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/fisalt/TMLR_CRG-NTK_and_MS · OpenReview ID：qLNYDuNYKZwang ning, Zekun Li, Tongxin Bai et al.
- 26Linear Model Merging Unlocks Simple and Scalable Multimodal Data Mixture Optimization期刊：Transactions on Machine Learning Research · 摘要：Selecting the best data mixture is critical for successful Supervised Fine-Tuning (SFT) of Multimodal Large Language Models. However, determining the optimal mixture weights across multiple domain-specific datasets remains a significant bottleneck due to the combinatorial search space and the high cost associated with even a single training run. This is the so-called Data Mixture Optimization (DMO) problem. On the other hand, model merging unifies domain-specific experts through parameter interpolation. This strategy is efficient, as it requires only a single training run per domain, yet it o… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/BerasiDavide/mLLMs_merging_4_DMO · OpenReview ID：rxaQbSybeTDavide Berasi, Matteo Farina, Massimiliano Mancini et al.
- 27EquiReg: Equivariance Regularized Diffusion for Inverse Problems期刊：Transactions on Machine Learning Research · 摘要：Diffusion models represent the state-of-the-art for solving inverse problems such as image restoration tasks. Diffusion-based inverse solvers incorporate a likelihood term to guide prior sampling, generating data consistent with the posterior distribution. However, due to the intractability of the likelihood, most methods rely on isotropic Gaussian approximations, which can push estimates off the data manifold and produce inconsistent, poor reconstructions. We propose Equivariance Regularized (EquiReg) diffusion, a general plug-in framework that improves posterior sampling by penalizing traje… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/Anima-Lab/Equireg · OpenReview ID：3iaYKJLcyGBahareh Tolooshams, Aditi Chandrashekar, Rayhan Zirvi et al.
- 28PAVO: Pipeline-Aware Voice Orchestration with Demand-Conditioned Inference Routing期刊：Transactions on Machine Learning Research · 摘要：Voice agents built on ASR-LLM-TTS pipelines allocate compute statically. It wastes resources on simple queries but does not cater to complex ones. So, we created PAVO (Pipeline-Aware Voice Orchestrator), which routes each turn through a three-stage pipeline. We orchestrate these routes based on demand signals extracted before the transcription even begins. We noticed that ASR errors propagate to downstream LLMs in two distinct regimes. One of them was a sharp factual accuracy cliff and the other one was gradual semantic degradation. This resulted in creating inter stage coupling constraints t… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：zrneoIxlFxNarasingaMoorthy VeiluKanthaPerumal, Mohammed Imthathullah
- 29Unified Deployment-Aware Evaluation of Open Reasoning Language Models期刊：Transactions on Machine Learning Research · 摘要：Open reasoning language models are often compared using mixed sample sizes, partially standardized prompts, and accuracy-centered summaries, making practical model selection difficult. We present a unified evaluation of seven open reasoning language model configurations across ARC-Challenge, GSM8K, MATH levels 1--3, and TruthfulQA MC1 under zero-shot, chain-of-thought (CoT), and few-shot CoT prompting. The study uses a complete matched-core matrix with 238 examples per model--dataset--strategy condition and a larger-sample robustness matrix with 500 examples for ARC-Challenge, GSM8K, and Trut… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/mkboch/UDAE · OpenReview ID：FBmgI8UMt8Md Motaleb Hossen Manik, Ge Wang
- 30Alert: Learning Trigger Functions for Early Classification of Time Series using Deep-RL期刊：Transactions on Machine Learning Research · 摘要：Early Classification of Time Series (ECTS) is vital in fields like industrial monitoring and medical triage, where quick and accurate predictions are essential. One of the core challenges lies in the trigger function, which decides when to make a prediction, independently of the classifier itself. Most existing methods rely on handcrafted rules, but can data-driven approaches outperform them? This paper introduces \textsc{Alert}, a Deep-RL framework that learns trigger functions from any state representation. Systematic comparisons on 30 datasets show that the design of the state space signif… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：J7EIFs4sjxAurélien Renault, Alexis Bondu, Antoine Cornuéjols et al.
- 31MEGA: Message Passing Neural Networks for Multigraphs with EdGe Attributes期刊：Transactions on Machine Learning Research · 摘要：Edge-attributed multigraphs, in which multiple edges with distinct attributes connect the same pair of nodes, arise naturally in many real-world systems. In these graphs, effective learning requires preserving information from repeated interactions while distinguishing contributions from different neighbors. Existing neural network solutions for edge-attributed multigraphs remain limited: some lose information from repeated interactions, while others break permutation equivariance. To address this, we introduce \emph{neighbor-aware aggregation}, an operator that first combines multi-edge feat… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/hcagri/MEGA-GNN-TMLR · OpenReview ID：Jo4KBvnMIPH. Çağrı Bilgi, Kubilay Atasu
- 32PrIntMesh: Precise Intersection Surfaces for 3D Organ Mesh Reconstruction期刊：Transactions on Machine Learning Research · 摘要：Human organs are composed of interconnected substructures whose geometry and spatial relationships constrain one another. Yet, most deep-learning approaches treat these parts independently, producing anatomically implausible reconstructions. We introduce PrIntMesh, a template-based, topology-preserving framework that reconstructs organs as unified systems. Starting from a connected template, PrIntMesh jointly deforms all substructures to match patient-specific anatomy, while explicitly preserving internal boundaries and enforcing smooth, artifact-free surfaces. We demonstrate its effectivenes… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/cvlab-epfl/printmesh · OpenReview ID：httuGMEO9QDeniz Sayin Mercadier, Hieu Le, Yihong Chen et al.
- 33Higher Resolution, Better Generalization: Unlocking Visual Scaling in Deep Reinforcement Learning期刊：Transactions on Machine Learning Research · 摘要：Pixel-based deep reinforcement learning agents are typically trained on heavily downsampled visual observations, a convention inherited from early benchmarks rather than grounded in principled design. In this work, we show that observation resolution is a critical yet overlooked variable for policy learning: higher-resolution inputs can substantially improve both performance and generalization, provided the network architecture can process them effectively. We find that the widely used Impala encoder, which flattens spatial features into a vector, suffers from quadratic parameter growth as re… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/raphajaner/visual-scaling-drl · OpenReview ID：xNm5W5WidpRaphael Trumpp, Ömer Veysel Çağatan, Baris Akgun et al.
- 34Learning Embeddings for Discrete Tree-Structured Data via Structural Prediction期刊：Transactions on Machine Learning Research · 摘要：Tree-structured data in natural language syntax, program analysis, and other symbolic domains are typically discrete, rooted, and ordered combinatorial objects. Despite their ubiquity, reusable and learnable representations for comparing such discrete structural trees remain limited. Classical methods such as tree edit distance (TED) and tree kernels provide principled structural measures but require pairwise computation, while previous neural encoders often produce latent representations without defining a consistent or interpretable space. We introduce a framework for learning embeddings fo… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：W1bzDTSwzAChen haoyang, Kumiko Tanaka-Ishii
- 35Wasserstein-type Gaussian Process Regressions for Input Measurement Uncertainty期刊：Transactions on Machine Learning Research · 摘要：Gaussian process (GP) regression is widely used for uncertainty quantification, yet the standard formulation assumes noise-free covariates. When inputs are measured with error, this errors-in-variables (EIV) setting can lead to optimistically narrow posterior intervals and biased decisions. We study GP regression under input measurement uncertainty by representing each noisy input as a probability measure and defining covariance through Wasserstein distances between these measures. Building on this perspective, we instantiate a deterministic projected Wasserstein ARD (PWA) kernel whose one-di… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：yEZjju9CUwHengrui Luo, Sherry Li, Yang Liu et al.
- 36Attention Trajectories as a Diagnostic Axis for Deep Reinforcement Learning期刊：Transactions on Machine Learning Research · 摘要：The emergence and evolution of feature reliance in deep reinforcement learning agents remain poorly understood. Here, we introduce a methodological framework for analyzing the learning process through quantitative analysis of saliency maps. This approach aggregates saliency information at the object and modality level into hierarchical attention profiles, quantifying how agents allocate attention over time, thereby forming attention trajectories throughout training. These profiles are then compared across controlled conditions, connected to behavioral measurements and reproduced with differen… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：0aa9zthk7kCharlotte Beylier, Hannah Selder, Arthur Fleig et al.
- 37DeBLAS: Accelerate LLM Pretraining by Length-based Sequence Scheduling期刊：Transactions on Machine Learning Research · 摘要：Pretraining large language models (LLMs) is computationally intensive, typically requiring massive datasets and training iterations. Although recent advances in data selection have shown improvement in training efficiency, their gains often diminish under scaling laws. In this work, we dive into the impact of sequence length on language model pretraining and propose a length-based online data scheduling method for acceleration. Specifically, we design a dense-balanced sequence scheduling framework for LLM pretraining: 1) at the first stage, the model is exposed to uniform-length dense token b… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/warriors-30/DeBLAS · OpenReview ID：s0MTAeYvwwGeng Yu, Jianing Zhu, Jiangchao Yao et al.
- 38Graph-based Subset Selection for Efficient Training of Gene Perturbation Models期刊：Transactions on Machine Learning Research · 摘要：Genomic studies face a vast hypothesis space, while interventions such as gene perturbations remain costly and time-consuming. To accelerate such experiments, gene perturbation models predict the transcriptional outcome of interventions. Since constructing the training set is challenging, active learning is often employed in a “lab-in-the-loop” process. While this strategy makes training more targeted, it is substantially slower, as it fails to exploit the inherent parallelizability of Perturb-seq experiments. Here, we focus on graph neural network–based gene perturbation models and propose a… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/geopanag/graph-based-subset-selection · OpenReview ID：S5YDiO3OoxGeorge Panagopoulos, Johannes F. Lutzeyer, Sofiane ENNADIR et al.
- 39Guided Diffusion by Optimized Loss Functions on Relaxed Parameters for Inverse Material Design期刊：Transactions on Machine Learning Research · 摘要：Inverse design problems are common in engineering and materials science. The forward direction, i.e., computing output quantities from design parameters, typically requires running a numerical simulation, such as a FEM, as an intermediate step, which is an optimization problem by itself. In many scenarios, several design parameters can lead to the same or similar output values. For such cases, multi-modal probabilistic approaches are advantageous to obtain diverse solutions. A major difficulty in inverse design stems from the structure of the design space, since discrete parameters or further… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/EmbodiedVision/diffoptloss-invmatdesign · OpenReview ID：JVUk7fT0LlJens U. Kreber, Christian Weißenfels, Joerg Stueckler
- 40Use Bayesian Paired Tests with a ROPE to Improve the Comparison of Machine Learning Models期刊：Transactions on Machine Learning Research · 摘要：This tutorial paper argues that model comparison in machine learning can be much improved by using \emph{paired testing}, i.e.\ comparing the predictions of methods A and B on each (common) test example. Due to the limitations of null hypothesis significance testing, a Bayesian approach is recommended, including the use of the region of practical equivalence (ROPE; Kruschke 2015a; Kruschke and Liddell 2018; Benavoli, Corani, Dem\v{s}ar, and Zaffalon 2017). We discuss a Bayesian $t$-test and a Bayesian McNemar test for comparisons on a single task, and Bayesian hierarchical models for comparis… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/ckiwgithub/compareMLModels · OpenReview ID：t6xhcnLlElChris Williams
- 41Quantitative LLM Judges Using Post-Hoc Score Calibration期刊：Transactions on Machine Learning Research · 摘要：LLM-as-a-judge is a large language model (LLM) that evaluates the output of another LLM. While LLMs excel at qualitative evaluation, they often struggle to predict human preferences and numeric scores. We propose quantitative LLM judges, which align LLM judges in a given domain using post-hoc calibration. Specifically, we train regression models to improve the score of the original judge using its rationale and score. We present four quantitative judges for different types of absolute and relative feedback, which shows the generality and versatility of our framework. Our framework does not re… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：isUKzVxmOBAishwarya Sahoo, Jeevana Kruthi Karnuthala, Tushar Parmanand Budhwani et al.
- 42Path-Integrated Loss-Gradient Kernels: Auditing and Simi- larity for Trained Neural Networks期刊：Transactions on Machine Learning Research · 摘要：Despite their success, deep neural networks remain opaque: it is often unclear why a model fails on a particular input, and classical generalization theory offers limited guidance in the overparameterized regime. Gradient-descent training naturally gives rise to path-dependent inner products between data points, but the resulting kernel matrices are asymmetric and can have negative eigenvalues, precluding their use as proper kernels or similarity measures. We show that a simple modification -- replacing output gradients with loss gradients in these inner products -- restores symmetry and posi… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/RyanPyle1/PLGK-Kernel · OpenReview ID：tE0RWpe5z2Ryan Pyle, Ankit Patel, Justin Sahs
- 43Attributional Safety Failures in Large Language Models under Code-Mixed Perturbations期刊：Transactions on Machine Learning Research · 摘要：LLMs that appear safety-aligned in English can fail under code-mixed prompts, where two languages are blended within a single utterance. Across ten languages and twelve systems, semantically equivalent code-mixed prompts consistently elicit more harmful responses than their English counterparts, with attack success rates rising by up to +58 percentage points in smaller open-weight models and by +8 to +12 points even in frontier API models; the effect replicates on naturally occurring code-mixed social-media data across four language pairs. To explain these failures, we introduce Saliency Drif… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：5q6S8E7BrxSomnath Banerjee, Pratyush Chatterjee, Shanu Kumar et al.
- 44Online Dense Video Captioning with Factorized Action Object Retrieval期刊：Transactions on Machine Learning Research · 摘要：Dense video captioning presents the dual challenge of temporally localizing events and generating descriptive captions within long videos. However, existing methods often struggle to handle evolving contexts in streaming settings or depend on static, global retrieval mechanisms. To address these limitations, we introduce a novel framework that embeds a dynamic, factorized retrieval mechanism directly into a causally-aware video processing backbone. Unlike approaches utilizing static global retrieval, our method dynamically retrieves concise action and object phrases at each timestep as the vi… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：rL7ns0ngAQDahun Kim, AJ Piergiovanni, Anelia Angelova
- 45iLOCO: Distribution-Free Inference for Feature Interactions期刊：Transactions on Machine Learning Research · 摘要：Feature importance measures are widely studied and are essential for understanding model behavior, guiding feature selection, and enhancing interpretability. However, many machine learning fitted models involve complex interactions between features. Existing feature importance metrics fail to capture these pairwise or higher-order effects, while existing interaction metrics often suffer from limited applicability or excessive computation; no methods exist to conduct statistical inference for feature interactions. To bridge this gap, we first propose a new model-agnostic metric, interaction Le… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：pUGr1uA99fCamille Olivia Little, Lili Zheng, Genevera I. Allen
- 46Cog-VADU: A Training-Free Cognitive Reasoning Framework for Video Anomaly Detection and Understanding期刊：Transactions on Machine Learning Research · 摘要：Video Anomaly Detection (VAD) aims to temporally localize abnormal events in videos. Most existing approaches rely on dataset-specific training and curated annotations, limiting generalization in open-set scenarios. Recent zero-shot methods based on Large Vision-Language Models (LVLMs) alleviate this dependency but often lack temporal continuity and structured reasoning. We propose \textbf{Cog-VADU}, a fully training-free framework that reformulates VAD as a sequential cognitive reasoning task. Cog-VADU introduces \emph{Chain-of-Anomaly Detection Thought Prompting} (CoADTP), which unrolls an… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/MohdUbaidwani/Cog-VADU · OpenReview ID：QcuSMNG7J8Mohd Ubaid Wani, Sara Atito, Josef Kittler et al.
- 47Minimally Invasive Machine Unlearning via Posterior Control期刊：Transactions on Machine Learning Research · 摘要：Approximate Machine Unlearning (MU) methods typically forget specific data by modifying model parameters or learning data-dependent augmentations. However, each unlearning request requires a dedicated optimization process, resulting in high computational overhead and cumulative degradation of model performance over time. Recent approaches have proposed manipulating a small subset of neural activations as a more targeted alternative, yet these methods still rely on on-demand searches for relevant parameters and remain computationally expensive. We propose Minimally Invasive Machine Unlearning… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/layer6ai-labs/MIMU · OpenReview ID：bVc04v8dV7Matthew Young, Valentin Villecroze, Sujay Menon et al.
- 48TAH-Quant: Effective Activation Quantization in Pipeline Parallelism over Slow Network期刊：Transactions on Machine Learning Research · 摘要：Decentralized training of large language models offers the opportunity to leverage computational resources across geographically distributed participants, but is often bottlenecked by network communication, particularly in pipeline-parallel settings. While pipeline parallelism partitions model layers across devices to handle large-scale models, it requires frequent communication of intermediate activations, which can be challenging when network bandwidth is limited. To address these issues, we propose TAH-Quant (Tile-wise Adaptive Hadamard Quantization), a novel activation quantization framew… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：6ysPGq2RVDGuangxin He, Yuan Cao, Yutong He et al.
- 49An Interactive Framework for Finding the Preferred Trade-off in Differential Privacy期刊：Transactions on Machine Learning Research · 摘要：Differential privacy (DP) is the gold standard for privacy-preserving analysis but introduces a fundamental trade-off between privacy guarantees and model performance. Selecting the optimal balance is a critical challenge, framed as a multi-objective optimization (MOO) problem of discovering the Pareto front and eliciting a decision-maker's preference. While interactive MOO offers a solution, standard approaches---which model objectives separately and rely on simple pairwise feedback---are suboptimal for DP because they do not utilize the problem structure. In this work, we propose a method,… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/YangYaohong1/PACE · OpenReview ID：ire2TQNxfvYaohong Yang, Aki Rehn, Sammie Katt et al.
- 50Knowing When Not to Answer: Mitigating Social Bias in LLMs via Epistemic Abstention期刊：Transactions on Machine Learning Research · 摘要：The growing application of Large Language Models (LLMs) to social contexts has led to an increase in unjustifiable social group attributions through their own stereotype-based responses; especially when responding to questions where there is little evidence to support a response or ambiguity to context. The lack of sufficient evidence often leads models to hallucinate socially grounded inferences, undermining fairness and trust. In this work, we attempt to mitigate social bias under ambiguity via epistemic uncertainty. We further introduce BHARATBBQ-R, a rationale-augmented extension of BHARA… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：UT5E31pYobChanchal Suman, Kapil Rana, Sanjit Kumar et al.