OpenReview · 实时热榜

期刊：Transactions on Machine Learning Research · 摘要：Foundation model developers are among the world’s most important companies. As these companies become increasingly consequential, how do their transparency practices evolve? The 2025 Foundation Model Transparency Index is the third edition of an annual effort to characterize and quantify the transparency of foundation model developers. The 2025 FMTI introduces new indicators related to data acquisition, usage data, and monitoring and evaluates companies like Alibaba, DeepSeek, and xAI for the first time. The 2024 FMTI reported that transparency was improving, but the 2025 FMTI finds this prog… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：1jT253Xtyf

期刊：Transactions on Machine Learning Research · 摘要：Designing a reinforcement learning from human feedback (RLHF) algorithm to approximate a human's unobservable reward function requires assuming, implicitly or explicitly, a model of human preferences. In sequential decision making tasks, a preference model that poorly describes how humans generate preferences risks learning a poor approximation of the human’s reward function. In this paper, we conduct human studies to assess whether one can influence the expression of real human preferences to more closely conform to a desired preference model. Importantly, our approach does not seek to alter… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：7YPlw1nUmW

期刊：Transactions on Machine Learning Research · 摘要：Recently there has been a surge of interest in optimal decision tree (ODT) methods that globally optimize accuracy directly, in contrast to traditional approaches that locally optimize an impurity or information metric. However, the literature shows conflicting evidence on the value of ODTs, with some demonstrating superior out-of-sample performance of ODTs over greedy approaches, while others show the opposite. The value and performance of ODTs therefore remains one of several open question regarding ODTs, most of which could not be answered before due to lack of scalability. With our experi… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/ConSol-Lab/opt-vs-greedy-dts · OpenReview ID：DvDOAtskXl

期刊：Transactions on Machine Learning Research · 摘要：Large Language Models (LLMs) are increasingly applied to tasks involving structured inputs such as semantic graphs, yet adapting them to such inputs remains non-trivial. Common approaches either linearize graphs, discarding structural information, or rely on specialized architectures that are not directly compatible with standard pretrained LLMs. We present SAFT, a structure-aware fine-tuning method that augments LLMs with graph positional encodings derived from the magnetic Laplacian of the input graph. These encodings are projected into the LLM embedding space, introducing relational induct… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/guerrantif/saft · OpenReview ID：QZoUMyzYDB

期刊：Transactions on Machine Learning Research · 摘要：While large language models (LLMs) have shown promise in automating data science, existing agents often struggle with the complexity of real-world workflows that require exploring multiple sources and synthesizing open-ended insights. In this paper, we introduce DS-STAR, a specialized agent to bridge this gap. Unlike prior approaches, DS-STAR is designed to (1) seamlessly process and integrate data across diverse, heterogeneous formats, and (2) move beyond simple QA to generate comprehensive research reports for open-ended queries. Extensive evaluation shows that DS-STAR achieves state-of-the… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/google-research/ds-star · OpenReview ID：Yz3ZPLzYaU

期刊：Transactions on Machine Learning Research · 摘要：We study classification problems using binary estimators where the decision boundary is described by horizon functions and where the data distribution satisfies a geometric margin condition. A key novelty of our work is the derivation of lower bounds for the worst-case learning rates over broad classes of functions, under a geometric margin condition---a setting that remains theoretically challenging. Moreover, we work in the noiseless setting, where lower bounds are particularly hard to establish. Our general results cover, in particular, classification problems with decision boundaries belo… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：ZIshsqojB6

期刊：Transactions on Machine Learning Research · 摘要：We study the maximization of nonnegative, non-monotone $\gamma$-weakly diminishing-returns (DR) submodular functions over down-closed convex bodies. The weakly DR model relaxes classical diminishing returns by allowing marginal gains to decay up to a multiplicative factor $\gamma \in (0,1]$, capturing a broad class of objectives that interpolate between monotone and fully non-monotone DR submodularity. Existing methods in this regime achieve guarantees that deteriorate rapidly as $\gamma$ decreases and fail to recover the best known bounds in the fully DR case. We develop a $\gamma$-aware alg… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：yS78Cb1CnX

期刊：Transactions on Machine Learning Research · 摘要：Natural gradient methods are appealing in policy optimization due to their invariance to smooth reparameterization and their ability to account for the local geometry of the policy manifold. These properties often lead to improved conditioning of the optimization problem compared to Euclidean policy gradients. However, their reliance on Monte Carlo estimation introduces high variance and sensitivity to hyperparameters. In this paper, we address these limitations by integrating Randomized Quasi-Monte Carlo (RQMC) sampling into the natural actor-critic (NAC) framework. We revisit the NAC linear… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：kOSx9v6dfb

期刊：Transactions on Machine Learning Research · 摘要：Recent progress in large language models demonstrates that hybrid architectures--combining self-attention mechanisms with state-space layers--can achieve a compelling balance between modeling quality and computational efficiency, particularly for long-context tasks. While these Transformer–Mamba-2 hybrid models show promising performance, systematic comparisons of hybridization strategies and analyses on the key factors behind their effectiveness have not been clearly shared with the community. In this work, we present a holistic evaluation of hybrid architectures based on inter-layer (sequen… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：x7qyXl8ecT

期刊：Transactions on Machine Learning Research · 摘要：Recent symmetry-based approaches in Variational Autoencoders (VAEs) have advanced disentanglement learning and compositional generalization. However, existing methods can encode identical semantic transformations differently depending on the specific sample pairs, which reduce the representational consistency of identical transformations. In this paper, we analyze how three commonly used symmetry parameterization families in prior work, namely (1) matrix-exponential parameterizations over the general linear group GL(n), (2) vector-additive actions in latent space, and (3) surjective mappings… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/GIST-IRR/RCIT · OpenReview ID：VjbBxj4aWb

期刊：Transactions on Machine Learning Research · 摘要：The use of Vision-Language Models (VLMs) in automated driving applications is becoming increasingly common, with the aim of leveraging their reasoning and generalisation capabilities to handle long-tail scenarios. However, these models often fail on simple visual questions that are highly relevant to automated driving, and the reasons behind these failures remain poorly understood. In this work, we examine the intermediate activations of VLMs and assess the extent to which specific visual concepts are linearly encoded, with the goal of identifying bottlenecks in the flow of visual information… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：HlBBy19ojC

期刊：Transactions on Machine Learning Research · 摘要：Instance-level recognition (ILR) focuses on identifying individual objects rather than broad categories, offering the highest granularity in image classification. However, this fine-grained nature makes creating large-scale annotated datasets challenging, limiting ILR’s real-world applicability across domains. To overcome this, we introduce a novel approach that synthetically generates diverse object instances from multiple domains under varied conditions and backgrounds, forming a large-scale training set. Unlike prior work on automatic data synthesis, our method is the first to address ILR-… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/yankungou/ILGen · OpenReview ID：T3JgJXH3ZK

期刊：Transactions on Machine Learning Research · 摘要：Learning utility functions over sets of elements is central to many machine learning and decision-making tasks such as feature selection, sensor placement, and content recommendation, where the goal is to evaluate and select an optimal subset of elements that provide the largest utility. These utility functions often exhibit desirable properties like monotonicity and submodularity over sets, but are typically expensive to evaluate and may lack an explicit analytical form. Moreover, the utility of a set can vary depending on certain contextual variables, further complicating the learning task.… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：SZ8mOziJBx

期刊：Transactions on Machine Learning Research · 摘要：Research in artificial intelligence is undergoing a paradigm shift from prioritizing model innovations and benchmark scores towards emphasizing problem definition and rigorous real-world evaluation. As the field enters the "second half," the central challenge becomes real utility in long-horizon, dynamic, and user-dependent settings such as agentic coding, deep research, and computer use, where LLM-based agents face context explosion beyond fixed context windows and must continuously accumulate, manage, and selectively reuse large volumes of information across extended interactions. Memory, w… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/AgentMemoryWorld/Awesome-Agent-Memory · OpenReview ID：XycbogUAeJ

期刊：Transactions on Machine Learning Research · 摘要：Recent work in neural network interpretability has suggested that hidden activations of some deep models can be viewed as linear projections of much higher-dimensional vectors of sparse latent ``features.'' In general, this kind of representation is known as a superposition code. This work presents an information-theoretic account of superposition codes in a setting applicable to interpretability. We show that when the number $k$ of active features is very small compared to the number $N$ of total features, simple inference methods currently used by sparse autoencoders can reliably decode a $… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：Nby4pCPIZI

期刊：Transactions on Machine Learning Research · 摘要：New multi-modal large language models (MLLMs) are continuously being trained and deployed, following rapid development cycles. This generative AI frenzy is driving steady increases in energy consumption, greenhouse gas emissions, and a plethora of other environmental impacts linked to datacenter construction and hardware manufacturing. Mitigating the environmental consequences of GenAI remains challenging due to an overall lack of transparency by the main actors in the field. Even when the environmental impacts of specific models are mentioned, they are typically restricted to the carbon foot… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/marta-lopez-rauhut/gen-ai-footprint · OpenReview ID：uurX0xsr8G

期刊：Transactions on Machine Learning Research · 摘要：We examine the reasoning and planning capabilities of large language models (LLMs) in solving complex tasks. Recent advances in inference-time techniques demonstrate the potential to enhance LLM reasoning without additional training by exploring intermediate steps during inference. Here, we explore how scaling inference-time techniques can improve reasoning and planning, focusing on understanding the tradeoff between computational cost and performance. To this end, we construct a comprehensive benchmark, known as *Sys2Bench*, and perform extensive experiments evaluating existing inference-tim… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/divelab/sys2bench · OpenReview ID：budZJyCK8G

期刊：Transactions on Machine Learning Research · 摘要：Large Language Models (LLMs) have benefited enormously from scaling, yet these gains are bounded by five fundamental limitations: (1) hallucination, (2) context compression, (3) reasoning degradation, (4) retrieval fragility, and (5) multimodal misalignment. While existing surveys describe these phenomena empirically, they lack a rigorous theoretical synthesis connecting them to the foundational limits of computation, information, and learning. This work closes that gap by presenting a unified, proof-informed framework that formalizes the innate theoretical ceilings of LLM scaling. First, com… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：BIRDGVrom8

期刊：Transactions on Machine Learning Research · 摘要：Representation Learning (RL) is fundamental for image matching, retrieval, classification, and other applications, enabling task-specific feature learning. RL algorithms aim to learn compact embeddings that preserve the neighbourhood structure of the input data. A general approach to this is contrastive learning, which pulls similar images (positives) closer together and pushes dissimilar images (negatives) farther apart in the embedding space. In Visual Place Recognition (VPR), positive images of a query share specific geographical and visual attributes with the query and can form a cluster.… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/Anuradha-Uggi/ANU-RL · OpenReview ID：mXE4OP55il

期刊：Transactions on Machine Learning Research · 摘要：Physics-informed neural networks (PINNs) have demonstrated effectiveness in solving partial differential equations (PDEs), yet they often struggle in high-dimensional regimes and lack interpretable representations and in scenarios involving sharp solution structures. Moreover, existing approaches typically rely on manually specified domain partitions. We propose a unified Dimension–Domain Co-Decomposition (3D) framework that jointly integrates dimension-wise decomposition with mixture-of-experts (MoE)–based domain decomposition. At the dimension level, we introduce an interpretable decomposit… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/TIML-Group/3DPINN · OpenReview ID：kuzkynVyRq

期刊：Transactions on Machine Learning Research · 摘要：Federated Learning (FL) aims at unburdening the training of deep models by distributing computation across multiple devices (clients) while safeguarding data privacy. On top of that, Federated Continual Learning (FCL) also accounts for data distribution evolving over time, mirroring the dynamic nature of real-world environments. While previous studies have identified Catastrophic Forgetting and Client Drift as primary causes of performance degradation in FCL, we shed light on the importance of Incremental Bias and Federated Bias, which cause models to prioritize classes that are recently intr… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/aimagelab/fed-mammoth · OpenReview ID：k2TT42Ei8W

期刊：Transactions on Machine Learning Research · 摘要：Peer incentivization (PI) is a popular multi-agent reinforcement learning approach where all agents can reward or penalize each other to achieve cooperation in social dilemmas. Despite their potential for scalable cooperation, current PI methods heavily depend on fixed incentive values that need to be appropriately chosen with respect to the environmental rewards and thus are highly sensitive to their changes. Therefore, they fail to maintain cooperation under changing rewards in the environment, e.g., caused by modified specifications, varying supply and demand, or sensory flaws — even when… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/philippaltmann/DRIVE · OpenReview ID：9Ltu1HV2YI

期刊：Transactions on Machine Learning Research · 摘要：Identifying linear time-invariant (LTI) dynamical systems from data is especially challenging when trajectories are short, noisy, or high-dimensional. Traditional system identification methods typically treat each system in isolation and therefore fail to exploit shared structure across related systems. We propose a PAC-Bayesian meta-learning framework for few-shot LTI system identification (PBML-LTI), which learns a transferable prior over task-specific dynamics while preserving task-level heterogeneity. Each task corresponds to an unknown LTI system, and a meta-learner uses a collection of… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/chenfeng-huang/PBML-LTI-TMLR-2026 · OpenReview ID：CiGFpSLzFv

期刊：Transactions on Machine Learning Research · 摘要：Recent advances in reasoning with large language models (LLMs) have demonstrated strong performance on complex mathematical tasks. Techniques such as Chain-of-Thought and In-Context Learning have further enhanced this capability, making LLMs both powerful and accessible tools for a wide range of users, including non-experts. However, their application to problems arising in operations research, particularly those at the intersection of combinatorial optimization and game theory that require domain expertise, remains underexplored. To address this gap, we introduce a benchmark of 369 instances… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/maryloufauchard/CAP_Benchmark · OpenReview ID：2dpt2Ughzt

期刊：Transactions on Machine Learning Research · 摘要：Learning general-purpose reasoning capabilities has long been a challenging problem in AI. Recent research in LLMs, such as DeepSeek-R1, has shown that reinforcement learning techniques like GRPO enable pre-trained LLMs to develop reasoning capabilities using simple question-answer pairs. In this paper, we aim to train visual language models (VLMs) to perform reasoning on image data through reinforcement learning and visual question-answer pairs, without explicitly using any chain-of-thought (CoT) supervision. Our key finding indicates that simply applying GRPO to a VLM---by prompting the mod… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/maifoundations/Visionary-R1 · OpenReview ID：JWkZXBgh5a

期刊：Transactions on Machine Learning Research · 摘要：The proliferation of open-platform text-to-image generative models has made prompt-wise model selection critical to maximize generation quality and semantic alignment. However, current strategies, such as contextual bandits, often converge slowly and fail to exploit the semantic relationships across prompts. To bridge this gap, we propose NeMoS, a non-parametric bandit framework that couples nearest neighbor reward estimation with a budget-constrained active learning strategy. Specifically, our approach operates in the prompt embedding space and estimates the reward of incoming prompts based… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/julesdamidaux/nemos-tmlr · OpenReview ID：CSjewjplO1

期刊：Transactions on Machine Learning Research · 摘要：Despite its foundational role in human discovery and sense-making, abductive reasoning—the inference of the most plausible explanation for an observation—has been relatively underexplored in Large Language Models (LLMs). Although LLMs have advanced rapidly, research on abductive reasoning and its diverse facets has remained disjointed rather than cohesive. To the best of our knowledge, this paper presents the first survey dedicated specifically to abductive reasoning in LLMs, tracing its trajectory from philosophical foundations to contemporary LLM-based approaches. To address the widespread… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：oeVkugH0WB

期刊：Transactions on Machine Learning Research · 摘要：Flow-based generative models, such as diffusion models and flow matching models, have achieved remarkable success in learning complex data distributions. However, a critical gap remains for their deployment in safety-critical domains: the lack of formal guarantees that generated samples will satisfy hard constraints. We propose a safety filtering framework that acts as an online shield for any pre-trained generative model. Our key insight is to cooperate with the generative process rather than override it. We define a constricting safety tube that is relaxed at the initial noise distribution… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/darshangm/constricted-diffusion · OpenReview ID：iZi471b4Pf

期刊：Transactions on Machine Learning Research · 摘要：Language models continue to hallucinate despite scaling parameters, compute, and data. We propose neural diversity — decorrelated parallel representations — as a provable mechanism to reduce hallucination rates at fixed parameter and data budgets. While existing mitigation strategies largely target accuracy, we reframe it as a second-moment reliability problem governed by representational covariance and provide the first formal tail bounds for hallucination probability in ensembled language models, explaining 94.3% of reliability variation across configurations in our setting (Qwen2.5-0.5B, 2… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/kushalc/nd-lora · OpenReview ID：5l9ZflyApA

期刊：Transactions on Machine Learning Research · 摘要：Humans possess a remarkable ability to filter out irrelevant sensory clutter, extracting only the information needed to anticipate and act within dynamic environments. Prior attempts to mitigate this through augmentation and masking strategies have improved robustness, but remain limited by computational overhead, weak semantic grounding, or instability in actor-critic training. Inspired by how language guides human perception, we introduce Task Relevant Language-conditioned Segmentation (TaLaS), a framework that leverages language-conditioned segmentation to impose semantic structure on visu… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：MRHXB6eooE

期刊：Transactions on Machine Learning Research · 摘要：Tool-using agents make intermediate decisions that alter persistent state, shape later observations, and create failures that are not equally easy to recover from. When verification is costly, the central question is not whether checking helps in general, but which decisions are worth checking. Policies driven only by local uncertainty capture whether a step may be wrong, but not how much that error would matter if left uncorrected. We formulate budgeted verification for tool-using agents as an intervention-allocation problem in which the value of checking a step depends on verifier efficacy,… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/tang03130313/verify-what-matters · OpenReview ID：nv1jzr0FaZ

期刊：Transactions on Machine Learning Research · 摘要：Heavy-tailed data make covariance estimation sensitive to the clipping level: stronger clipping reduces variance but increases bias. We study how to choose this clipping level from the data within a radial clipped covariance family. We propose the quantile tail-energy surrogate (QTES), a fully data-driven rule that combines a cross-fitted variance certificate with a held-out estimate of the tail energy removed by clipping. QTES requires no distributional prior parameters. For Euclidean clipping, the operator-norm bias is bounded by this scalar tail-energy quantity. Under a finite $L_4$ moment… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：MyNXLdRFJ3

期刊：Transactions on Machine Learning Research · 摘要：Zero-Shot Anomaly Detection (ZSAD) aims to detect anomalies in a target dataset without any training samples, leveraging models trained on auxiliary data. While CLIP offers strong cross-modal representations for ZSAD, its pretraining objective inherently emphasizes global foreground semantics over fine-grained local defects. Consequently, its anomaly localization remains highly sensitive to prompt wording, limiting the effectiveness of existing methods that rely on explicit category labels. To overcome this limitation, we introduce ViP$^{2}$-CLIP, a lightweight CLIP-based ZSAD framework featu… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：KCRRuiQSIm

期刊：Transactions on Machine Learning Research · 摘要：The rapid advances in generative AI models have empowered the creation of highly realistic images with arbitrary content, raising concerns about potential misuse and harm, such as Deepfakes. Current research focuses on training detectors using large datasets of generated images. However, these training-based solutions are often computationally expensive and show limited generalization to unseen generated images. In this paper, we propose a training-free method to distinguish between real and AI-generated images. We first observe that real images are more robust to tiny noise perturbations tha… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/IBM/RIGID · OpenReview ID：NBkBI2Zjlm

期刊：Transactions on Machine Learning Research · 摘要：Spatio-temporal sensor data in real-world systems is often sparse, noisy, and irregular, making it difficult to infer global structure from limited observations. Under extreme sparsity, we run into the limits of identifiability of latent system states, making latent field reconstruction fundamentally underconstrained. In such scenarios, multiple physically plausible fields may remain consistent with the same observations, requiring reconstruction models to rely heavily on inductive biases regarding locality, transport structure, and spatial regularity. Under such sparsity regimes, reliable re… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/ankitbha/fieldformer · OpenReview ID：we4FYGOE2y

期刊：Transactions on Machine Learning Research · 摘要：Cross-view masked autoencoding has emerged as a powerful pretext task for learning dense correspondences, which are essential for applications such as video label propagation. The cross-view pretext task is modeled with a masked autoencoder, where a masked target view is reconstructed from an anchor view. However, acquiring effective training data remains a challenge - collecting diverse video datasets is costly, while simple image crops lack the necessary pose variations, underperforming video-based methods. This paper introduces CDG-MAE, a novel MAE-based self-supervised method that uses di… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/cvlab-stonybrook/CDG-MAE · OpenReview ID：7XIymKIA0v

期刊：Transactions on Machine Learning Research · 摘要：Post-training quantization (PTQ) is a practical way to reduce the memory footprint of large language models, but low-bit quantization is sensitive to mismatches between the quantization codebook and the empirical weight/activation distributions. We revisit Benford-like leading-digit statistics as a lightweight diagnostic of scale-broad behavior in transformer tensors. Across several model families, we observe a consistent functional dichotomy: transformational nn.Linear weights tend to be Benford-like, whereas LayerNorm and embedding parameters systematically deviate. Motivated by this observ… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/ufopcsilab/benford-quant · OpenReview ID：YiLcQY4Nje

期刊：Transactions on Machine Learning Research · 摘要：The fast scaling speed of histopathology datasets allows researchers to train various foundation models for disease-centered research with applications in classifying disease-state information and predicting gene expression levels. However, it has been shown that current models tend to be overconfident and make classification at a low-calibration level. This case is underexplored for regression-type tasks such as gene expression prediction as well, which could seriously affect the diagnosis and treatment based on the developed models. To resolve this critical issue, we propose a \underline{u}… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：LLlOJs4o2N

期刊：Transactions on Machine Learning Research · 摘要：Single nucleotide polymorphisms (SNPs) are among the most common and informative forms of genetic variation in the human genome and constitute the primary data representation used in genome-wide association studies (GWAS). Due to their extreme dimensionality, strong correlation structure, and the presence of both population-level and familial dependencies, SNP datasets exhibit structural properties that fundamentally distinguish them from standard tabular data. At the same time, genomic data is uniquely sensitive; it is immutable, identifying, and shared across relatives, and has been shown t… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：6BWikkmkOH

期刊：Transactions on Machine Learning Research · 摘要：Existing unlearning approaches typically rely on post hoc weight adaptation or distillation, leading to duplicated memory costs, degraded generalization, and limited scalability. In this work, we introduce ERASE, Erasure via Reconstructive Adversarial Signal Editing, a framework for on-the-go forgetting that suppresses the observable influence of private data without modifying model weights. ERASE leverages structured, class-conditioned input perturbations to induce selective forgetting during inference, eliminating the need for retraining, fine-tuning, or model copies. We rigorously characte… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：PIXVov5LQq

期刊：Transactions on Machine Learning Research · 摘要：Recent progress in large language models enables the creation of realistic machine-generated content. Watermarking is a promising approach to distinguish machine-generated text from human text, embedding statistical signals in the output that are ideally undetectable to humans. We propose a watermarking framework that encodes such signals through an error correcting code. Our method, termed robust binary code (RBC) watermark, introduces no noticeable degradation in quality. We evaluate our watermark on base and instruction fine-tuned models and find that our watermark is robust to edits, dele… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/patrickrchao/watermarking-llms · OpenReview ID：H6oBZxNQk2

期刊：Transactions on Machine Learning Research · 摘要：We propose a family of random feature maps for scalable kernel machines defined over low-dimensional subspaces in high dimensions, \ie over the Grassmannian manifold. This is typically useful in a machine learning context when data classes or clusters are well represented by the span of a few data points. Classical Grassmannian kernels such as the \emph{projection} or \emph{Binet–Cauchy} kernels require constructing full Gram matrices for practical applications, leading to prohibitive computational and memory costs for large subspace datasets in high dimensions. We address this limitation by… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：wq18dZJ2pA

期刊：Transactions on Machine Learning Research · 摘要：We introduce a novel inference-time alignment approach for LLMs that aims to generate safe responses almost surely, i.e., with probability approaching one w.r.t. a given cost model. Our approach models the generation of safe responses as a constrained Markov Decision Process (MDP) within the LLM's latent space. We augment a safety state that tracks the evolution of safety constraints and dynamically penalize unsafe generations to ensure the generation of safe responses. Consequently, we demonstrate formal safety guarantees w.r.t. the given cost model upon solving the MDP in the latent space w… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/rsshyam/inf-guard · OpenReview ID：FlnokjaSEu

期刊：Transactions on Machine Learning Research · 摘要：We analyze the convergence behavior of stochastic gradient descent with momentum (SGDM) under dynamic learning-rate and batch-size schedules by introducing a novel and simpler Lyapunov function. We extend the existing theoretical framework to cover three practical scheduling strategies commonly used in deep learning: a constant batch size with a decaying learning rate, an increasing batch size with a decaying learning rate, and an increasing batch size with an increasing learning rate. Our results reveal a clear hierarchy in convergence: a constant batch size does not guarantee convergence of… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：s6DTv7Sorj

期刊：Transactions on Machine Learning Research · 摘要：Video motion magnification is a technique to capture and amplify subtle motion in a video that is invisible to the naked eye. The deep learning-based prior work successfully models outstanding quality better than conventional signal processing-based ones. However, it still lags behind real-time performance, which prevents it from being extended to various online systems. In this paper, we revisit the first learning-based model and present experimental analyses, in particular on the identification of redundant components, the insertion of spatial bottlenecks, and the trade-off relationship bet… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/kaist-ami/Fast-MM · OpenReview ID：TAmmPuExE1

期刊：Transactions on Machine Learning Research · 摘要：While large language models (LLMs) fine-tuned with lightweight adapters achieve strong performance across diverse tasks, their performance on individual tasks depends on the fine-tuning strategy. Fusing independently trained models with different strengths has shown promise for multi-task learning through three main strategies: ensembling, which combines outputs from independent models; merging, which fuses model weights via parameter averaging; and routing, which integrates models in an input-dependent fashion. However, many design decisions in these approaches remain understudied, and the r… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：bnRCvRtZv5

期刊：Transactions on Machine Learning Research · 摘要：Efficient sampling from Boltzmann distributions is central to modelling complex physical systems. Markov Chain Monte Carlo (MCMC) methods suffer from critical slowing down, high autocorrelation, and poor mode-mixing, limiting their scalability. Recent advances, like Boltzmann Generators, offer a promising alternative but remain constrained by costly MCMC-based training, inefficient sampling, and poor ergodicity. We introduce an algorithm for learning Boltzmann distributions that does not require any true samples for training. Our approach draws inspiration from flow matching but departs funda… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/madhavlab/2025_fund · OpenReview ID：O05dDDVcyZ

期刊：Transactions on Machine Learning Research · 摘要：Recent techniques such as retrieval-augmented generation or chain-of-thought reasoning have led to longer contexts and increased inference costs. Context compression techniques can reduce these costs, but the most effective approaches require fine-tuning the target model or even modifying its architecture. This can degrade its general abilities when not used for this specific purpose. Here we explore an alternative approach: an encoder that compresses the context into continuous representations which replace token embeddings in decoder LLMs. First, we perform a study of training strategies an… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/kyutai-labs/ARC-Encoder · OpenReview ID：lU1P9dsqfn

期刊：Transactions on Machine Learning Research · 摘要：Real-world forecasting requires models to integrate not only historical data but also relevant contextual information provided in textual form. While large language models (LLMs) show promise for context-aided forecasting, critical challenges remain: we lack diagnostic tools to understand failure modes, performance remains far below their potential, and high computational costs limit practical deployment. We introduce a unified framework of four strategies that address these limitations along three orthogonal dimensions: model diagnostics, accuracy, and efficiency. Through extensive evaluatio… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/ashok-arjun/beyond-naive-prompting · OpenReview ID：dkjHHFJkVI

期刊：Transactions on Machine Learning Research · 摘要：Recurrent Neural Networks (\RNNs) have demonstrated their effectiveness in modeling sequential data and are a key building block of modern deep learning architectures. In this review paper, we study recurrent networks through the lens of automata theory. Given an \RNN, automata learning seeks to model its behavior with an automaton, which enables better interpretability and eases our understanding of its working mechanisms. We begin by examining the theoretical foundations of this approach, demonstrating how it can be applied to learn automata from various types of recurrent architectures, in… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：R52ETbUBVo