OpenReview · 实时热榜

期刊：Transactions on Machine Learning Research · 摘要：Causal discovery through experimentation and intervention is fundamental to robust problem solving. It requires not just updating beliefs within a fixed framework but revising the hypothesis space itself, a capacity current AI agents lack when evidence demands representations they have not previously constructed. We extend the blicket detector paradigm from developmental science to test this capacity in AI agents equipped with architectural scaffolding that targets hypothesis-space restructuring. Our compositional architecture has two discrete components: context graphs, which structure explo… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/concurrence-hq/extended-blicket-benchmark · OpenReview ID：05VkYfgXm6

期刊：Transactions on Machine Learning Research · 摘要：While experimental reproduction remains a pillar of the scientific method, we observe that the software best practices supporting the reproduction of Machine Learning (ML) research are often undervalued or overlooked. We quantify these concerns by surveying the usage of software best practices in software repositories associated with publications at major ML conferences and journals such as NeurIPS, ICML, ICLR, TMLR, and MLOSS within the last decade. We report the results of this survey that identify areas where software best practices are lacking and areas with potential for growth in the ML… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/BonnBytes/ml-swe-analysis · OpenReview ID：13ThuknGwY

期刊：Transactions on Machine Learning Research · 摘要：Medical Agents are an emerging class of agentic systems deployed in clinical settings that operate over multimodal, longitudinal data, maintain internal state, plan and adapt sequences of actions, and interact with clinical information systems under governance constraints. They extend traditional medical artificial intelligence (MedAI) beyond narrow diagnostic and predictive models toward workflow-centric architectures that address persistent challenges such as administrative burden, fragmented workflows, and workforce strain. In this paper, we (i) propose a functional definition and three-le… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：4pwRl2G6je

期刊：Transactions on Machine Learning Research · 摘要：Clinical prognosis requires integrating heterogeneous evidence from structured measurements, physiological waveforms, longitudinal history, and population context. Multimodal language models provide a natural framework for this integration, but standard approaches often serialize structured measurements as text, treat encounters independently, and lack a mechanism for incorporating representations of clinically similar patients. We introduce PRISM (Population-Retrieved Inference for Structured Multimodal Prognosis), which represents heterogeneous evidence through per-feature numerical tokeniz… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：5ChJTFLi5r

期刊：Transactions on Machine Learning Research · 摘要：Diffusion models have achieved remarkable progress in the field of video generation. However, their iterative denoising nature requires a large number of inference steps to generate a video, which is slow and computationally expensive. In this paper, we begin with a detailed analysis of the challenges present in existing diffusion distillation methods and propose a novel efficient method, namely AccVideo, to reduce the inference steps for accelerating video diffusion models with synthetic dataset. We leverage the pretrained video diffusion model to generate multiple valid denoising trajectori… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/aejion/AccVideo · OpenReview ID：5ntdEzTTa2

期刊：Transactions on Machine Learning Research · 摘要：Structure-based drug design has been accelerated by pocket-aware 3D generative models, yet most methods primarily fit the training distribution and may fall short of satisfying multiple properties required in real-world therapeutic drug discovery. Recently, increasing attention has focused on structure-based molecule optimization (SBMO), which targets fine-grained control over multiple specified molecular properties. In this paper, we present DEPPA, a novel SBMO approach building upon Denoising Diffusion Policy Optimization for fine-tuning a pre-trained pocket-aware diffusion model via reinfo… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/xy9485/DePPA · OpenReview ID：5sjQt9I2mt

期刊：Transactions on Machine Learning Research · 摘要：Standard supervised classification trains models to imitate the labels in their training data. Since this typically happens in a single forward pass, models are locked into a fixed compute budget whether an input is simple or complex. Moreover, the imitation objective forces the model to express absolute certainty on its training data, which often carries over into evaluation, leading to overconfident predictions. Rather than forcing a rigid, single-step guess, early learning systems often framed learning as a trial-and-error process of gradual improvement. Reclaiming this perspective for cla… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：7xXijNfm8j

期刊：Transactions on Machine Learning Research · 摘要：One-shot federated learning (OSFL) addresses the communication overhead of federated learning by limiting training to a single round, but doing so without sacrificing model quality is non-trivial, particularly when client data distributions diverge. Recent work has addressed this challenge by aggregating client knowledge on the server through the construction of transferable synthetic datasets or distillates. However, most of these methods lack formal privacy guarantees, leaving a gap in jointly achieving low communication, robustness to heterogeneity, and rigorous privacy. We propose FedKT-C… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：9H3b5MX87m

期刊：Transactions on Machine Learning Research · 摘要：Differentiable Neural Architecture Search (NAS) provides efficient, gradient-based methods for automatically designing neural networks, yet its adoption remains limited in practice. We present MIDAS, an approach that modernizes DARTS by replacing static architecture parameters with dynamic, input-specific parameters computed via self-attention. To improve robustness, MIDAS (i) localizes the architecture selection by computing it separately for each spatial patch of the activation map, and (ii) introduces a parameter-free, topology-aware search space that models the two-edge node topology duri… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：9R6EBFTzuy

期刊：Transactions on Machine Learning Research · 摘要：Graph-structured datasets often suffer from class imbalance, which complicates node classification tasks. In this work, we address this issue by first providing an upper bound on population risk for imbalanced transductive node classification. We then propose a simple and novel algorithm, Uncertainty-aware Pseudo-labeling (UPL). Our approach leverages pseudo-labels assigned to unlabeled nodes to mitigate the adverse effects of imbalance on classification accuracy. Furthermore, the UPL algorithm enhances the accuracy of pseudo-labeling by reducing training noise of pseudo-labels through a nove… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/teimuri/UPL · OpenReview ID：9ovThQD3LQ

期刊：Transactions on Machine Learning Research · 摘要：In contemporary cloud-based services, protecting users’ sensitive data and ensuring the confidentiality of the server’s model are critical. Fully homomorphic encryption (FHE) enables inference directly on encrypted inputs, but its practicality is hindered by expensive bootstrapping and inefficient approximations of non-linear activations. We introduce Safhire, a hybrid inference framework that executes linear layers under encryption on the server while offloading non-linearities to the client in plaintext. This design eliminates bootstrapping, supports exact activations, and significantly red… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/sacs-epfl/SAFHIRE · OpenReview ID：AAneiBMYgS

期刊：Transactions on Machine Learning Research · 摘要：Human-move predictors are typically evaluated by whether their highest-ranked move matches the move played. This evaluation, however, does not distinguish between the observed move falling outside the model's leading candidates and being ranked highly but not first. We study this distinction on 884{,}049 positions from the Allie Lichess 2022 blitz test set using two strong, independently developed population-level human-move predictors, MAIA3 and Allie policy-only, evaluated under the same legal-move protocol. MAIA3 and Allie policy-only rank the observed human move in Top5 on 91.843\% and 90… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/SarveshVGharat/argmax-gap · OpenReview ID：AKQ2b7VLA4

期刊：Transactions on Machine Learning Research · 摘要：Fine-tuning adapts pretrained models to new tasks, but identifying which weight changes support the resulting behavior remains difficult. Weight-space interpretability assigns semantic descriptions to these changes, called fine-tuning deltas. Such descriptions leave open how much task performance depends on the identified updates, particularly when other updates can support the same predictions. We address this gap by pairing reversion, which restores selected parameters to their pretrained values, with retention, which keeps only selected fine-tuning updates. We evaluate Contrastive Language… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：AQ7a5vGI0v

期刊：Transactions on Machine Learning Research · 摘要：Known Lie symmetries specify how a partial differential equation (PDE) solution operator should transform under group actions, but finite supervised training data constrain this relationship only at observed input--output pairs. To encourage equivariance without changing the architecture, we introduce \textbf{LOCO} (\textbf{L}ocal \textbf{O}rbit \textbf{Co}nsistency), a lightweight and flexible training objective for neural operators. LOCO samples a small Lie-group transformation during training and explicitly penalizes the mismatch between the operator's prediction on the transformed input a… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/michalmariuszlewandowski/orbit-tanget-no · OpenReview ID：AQIyD6jSSl

期刊：Transactions on Machine Learning Research · 摘要：Videos are inherently temporal sequences by their very nature. In this work, we explore the potential of modeling videos in a chronological and scalable manner with autoregressive (AR) language models, inspired by their success in natural language processing. We introduce \textbf{DiCoDe}, a novel approach that leverages \textbf{Di}ffusion-\textbf{Co}mpressed \textbf{De}ep Tokens to generate videos with a language model in an autoregressive manner. Unlike existing methods that employ low-level representations with limited compression rates, DiCoDe utilizes deep tokens with a considerable compr… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：B7a0IAIEXZ

期刊：Transactions on Machine Learning Research · 摘要：Unified decoder-only transformers have shown promise for multimodal generation, yet how sequence design affects the synchronization of modalities with heterogeneous sampling rates remains under explored. We investigate this through video-text-to-speech (VTTS) synthesis---a controlled task requiring fine-grained temporal alignment between sparse text, video, and continuous speech. Using a unified decoder-only transformer, dubbed Visatronic, trained on VoxCeleb2, we study: (i) how modalities contribute complementary information, (ii) how positional encoding strategies enable synchronization acr… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：D5QpMy6yUH

期刊：Transactions on Machine Learning Research · 摘要：Understanding the agent's learning process, particularly the factors that contribute to its success or failure post-training, is crucial for comprehending the rationale behind the agent's decision-making process. Prior methods clarify the learning process by creating a structural causal model (SCM) or visually representing the distribution of value functions. Nevertheless, these approaches have constraints as they exclusively function in 2D environments or with simple transition dynamics. Understanding the agent's learning process in complex environments or tasks is more challenging. In this… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：Dmu7agmmTQ

期刊：Transactions on Machine Learning Research · 摘要：Recent advances in vision-language model (VLM) reasoning have been largely attributed to the rise of reinforcement learning (RL), which has shifted the community’s focus away from the supervised fine-tuning (SFT) paradigm. Many studies suggest that introducing the SFT stage not only fails to improve reasoning ability but may also negatively impact model training. In this study, we revisit this RL-centric belief through a systematic and controlled comparison of SFT and RL on VLM reasoning. Using identical data sources, we find that the relative effectiveness of SFT and RL is conditional and st… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：EWk0VxDpax

期刊：Transactions on Machine Learning Research · 摘要：In a controlled fixed-budget study of linear attention decay mechanisms, we find that the delta rule is the largest observed factor under our protocol. We evaluate seven linear-attention variants plus a standard softmax baseline on two datasets (TinyStories, WikiText-103) with 3 random seeds each at both 18M and 125M parameters, and 5 variants at 42M on TinyStories. The data-independent subset, where decay gates are learned parameters rather than token-conditioned functions, forms a complete 2×2 comparison over granularity (scalar vs. channel-wise) and delta rule (absent vs. present), while t… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/emailtovamos/delta-rule-linear-attention · OpenReview ID：FdXB4u27LD

期刊：Transactions on Machine Learning Research · 摘要：Federated Learning (FL) faces significant challenges related to communication efficiency and performance reduction when scaling to many clients. To address these issues, we explore the potential of using low-rank updates and provide one of the first theoretical studies comparing client and server Hessian rank structures in FL based on random-matrix theory. Our theoretical analysis shows that a client’s loss exhibits a higher-rank structure (i.e., gradients span higher-rank subspaces of the Hessian) compared to the server’s loss, and that low-rank approximations of the clients’ gradients have… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：MXSAMrmktX

期刊：Transactions on Machine Learning Research · 摘要：One of the most striking features of Large Language Models (LLMs) is their ability to learn in-context. Namely at inference time an LLM is able to learn new patterns without any additional weight update when these patterns are presented in the form of examples in the prompt, even if these patterns were not seen during training. The mechanisms through which this can happen are still largely unknown. In this work, we show that the stacking of a self-attention layer with an MLP allows the transformer block to implicitly modify the weights of the MLP layer according to the context. We argue throu… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：Mx0ZcLznPS

期刊：Transactions on Machine Learning Research · 摘要：Zero-shot reinforcement learning (RL) has emerged as a setting for developing general agents, capable of solving downstream tasks without additional training or planning at test-time. While conventional RL optimizes policies for fixed rewards, zero-shot RL requires learning representations that enable immediate adaptation to arbitrary reward functions. As the field matures, the growing diversity of approaches demands a foundational framework reconciling different perspectives under a common structure. In this work, we introduce a formal, unified framework for zero-shot RL, allowing for rigoro… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：Nv26zx1Ywa

期刊：Transactions on Machine Learning Research · 摘要：Proper citation of relevant literature is essential for contextualising and validating scientific contributions. While current citation recommendation systems leverage local and global textual information, they often overlook the nuances of the human citation behaviour. Recent methods that incorporate such patterns improve performance but incur high computational costs and introduce systematic biases into downstream rerankers. To address this, we propose $\textit{Profiler}$, a lightweight, non-learnable module that captures human citation patterns efficiently with mitigated bias, significantl… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/goyalkaraniit/PPM · OpenReview ID：R1g1Z4yX57

期刊：Transactions on Machine Learning Research · 摘要：Conformal regression provides finite-sample marginal coverage, but interval efficiency depends on how uncertainty is allocated across inputs. We propose \emph{Conformal Laplace-Aware Predictive Scaling} (CLAPS), which combines a learned heteroscedastic scale with last-layer Laplace uncertainty. The learned variance also weights the training features used to construct the posterior, yielding a noise-weighted support correction in the learned representation. The correction contracts as weighted feature information increases and becomes most influential under localized support imbalance. Experim… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/AndrewKim1997/CLAPS · OpenReview ID：RvoPfNBjwp

期刊：Transactions on Machine Learning Research · 摘要：Vision-Language Models (VLMs) have demonstrated strong capabilities in aligning visual and textual modalities, enabling a wide range of applications in multimodal understanding and generation. While they excel in zero-shot and transfer learning scenarios, VLMs remain susceptible to misclassification, often yielding confident yet incorrect predictions. This limitation poses a significant risk in safety-critical domains, where erroneous predictions can lead to severe consequences. In this work, we introduce TrustVLM, a training-free framework designed to address the critical challenge of estima… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/EPFL-IMOS/TrustVLM · OpenReview ID：S2SdvYaPKm

期刊：Transactions on Machine Learning Research · 摘要：Reinforcement Learning (RL) and Imitation Learning (IL) offer complementary capabilities: RL can learn high-performing policies but is data-intensive, whereas IL enables rapid learning from demonstrations but is limited by the demonstrator’s quality. Combining them offers the potential for improved sample efficiency in learning high-performing policies, yet naïve integrations often suffer from two fundamental issues: (1) negative transfer, where optimizing the IL loss hinders effective RL fine-tuning, and (2) gradient conflict, where a negative inner product between IL and RL gradients might… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/ParanoidAndroid96/CAIR · OpenReview ID：S98skK3FfD

期刊：Transactions on Machine Learning Research · 摘要：Evaluating thousands of candidate formulas against the same data guarantees that some will fit purely by chance, yet current feature-construction tools offer no formal error control. We treat feature construction as a selective-inference problem. Candidates are generated via beam search over an expression Directed Acyclic Graph (DAG) using one half of the training rows. To control the false discovery rate (FDR), these candidates are rigorously screened on the remaining data using an exact permutation test of marginal association under Benjamini–Hochberg, Benjamini–Yekutieli, or knockoff contr… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/LD-Shell/beamfeat · OpenReview ID：UibXoy4h3u

期刊：Transactions on Machine Learning Research · 摘要：Bayesian neural networks (BNNs) offer a principled framework for quantifying epistemic uncertainty. However, standard weight-space priors yield complex, high-dimensional neural network posteriors that are computationally prohibitive to sample from, and popular approximate inference schemes often fail to capture epistemic uncertainty effectively---a failure mode particularly acute in reinforcement learning (RL) \citep{osband2018randomized}. To resolve this challenge, we introduce a new class of BNNs that quantify uncertainty by placing Bayesian nonparametric Dirichlet Process (DP) priors direc… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/sumilations/Dirichlet-Process-Bayesian-Neural-networks · OpenReview ID：VrWIwB8g3Z

期刊：Transactions on Machine Learning Research · 摘要：Association Rule Mining (ARM) is a fundamental data mining technique for knowledge discovery in tabular data and is widely used in high-stakes decision-making. Existing approaches suffer from two limitations: reliance on frequent itemset mining leading to rule explosion in classical methods, and failure in low-data regimes in recent neural approaches. We reframe ARM as a conditional probability estimation problem and introduce a model-agnostic rule learning framework that extracts association rules from any conditional probabilistic model, decoupling rule extraction from frequent itemset mini… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/DiTEC-project/tabprobe · OpenReview ID：Y70lnNnA6T

期刊：Transactions on Machine Learning Research · 摘要：Individual fairness in machine learning requires that similar individuals receive similar predictions. Most existing definitions are pairwise in nature, imposing a Lipschitz constraint that must hold either globally across all pairs of data instances or locally within specified neighborhoods. In practice, however, such pairwise constraints may be overly restrictive and may not directly identify which individuals receive locally inconsistent predictions. We study local predictive consistency as a node-level relaxation of pairwise individual fairness, where each prediction is compared with a si… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/noorazhaoz/local-predictive-consistency · OpenReview ID：Yd1cLycbmR

期刊：Transactions on Machine Learning Research · 摘要：Trajectory-User Linking (TUL) is a critical task in mobility applications that links unlabeled spatial trajectories to the users or entities that generated them. In these applications, data often arrives as a continuous stream and may experience distributional shifts over time. While adapting TUL models via online learning could address these challenges, this approach remains unexplored in current research. Our work bridges this gap by conducting comprehensive evaluations of common TUL techniques in an online learning context. To improve the performance of existing TUL techniques in this sett… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/lucasczz/MiLES · OpenReview ID：YfyOCduclR

期刊：Transactions on Machine Learning Research · 摘要：Ensemble classifiers are predictive models that combine the results of simpler base models, often by majority vote. A classic example is random forests, which combine the predictions of decision trees. Ensembles that use more base models can be more accurate but also more costly to train and run. In this paper, we consider strategies for reducing the computational cost of binary classification using an approach from the field of sequential testing. Rather than evaluating all the base models and taking a majority vote, we evaluate the base models sequentially and stop execution when a clear ma… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/joseph-e-k/random-forest-early-stopping · OpenReview ID：aUQFMWmW0v

期刊：Transactions on Machine Learning Research · 摘要：Parameter estimation in stochastic differential equations is a classical statistical problem of much importance in many scientific fields. Recent work by Tapia-Costa et al. (2026) introduced a novel technique for estimating the drift when the diffusion parameter is known, using discrete samples from multiple trajectories. Their method treats drift estimation as a denoising problem, and leverages tools from (conditional) score-matching diffusion models. Although their experiments showed promising results across different drift classes, the question of theoretical guarantees for their estimator… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：cyHG8JSTGe

期刊：Transactions on Machine Learning Research · 摘要：Accurate uncertainty quantification is critical for many time series forecasting applications, and recent studies have increasingly adopted generative models to capture predictive uncertainty. Existing approaches often combine deterministic forecasting with generative modeling to improve predictive performance. However, their joint training is typically unstable because the deterministic model is embedded within the generative process, resulting in deep computational graphs and changes to the probability transport path. In this work, we propose DV-Flow, a unified flow-matching formulation tha… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：dDEDwFnnoA

期刊：Transactions on Machine Learning Research · 摘要：During autoregressive generation, transformer-based language models rely on the KV cache as their record of prior context. We study information recall locality in LLMs. Specifically, we ask \textit{whether transformer-based language models can recall the information that token(s) carried after their original KV cache entries are removed}. We study this question on a suite of association recall tasks in which we measure the recall of factual and sentiment information (e.g., ``doctor Alice'' vs. ``dingy bowl''), under various KV cache eviction scenarios. Our evaluation of eight instruction-tune… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：eAMLpW74rV

期刊：Transactions on Machine Learning Research · 摘要：Training deep neural networks requires extensive experimentation to estimate how long it will take, and how much it will cost, for a model to reach convergence. Chronos is a unified framework for pre-training prediction of both the convergence horizon and the resulting end-to-end training time and cost for a target configuration. Here, convergence is defined as the epoch at which early stopping is triggered after a specified number of epochs without sufficient improvement in validation loss. Unlike approaches that estimate these quantities from observations collected during an executing train… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：gDLPOEjQTH

期刊：Transactions on Machine Learning Research · 摘要：We propose GENERIC-FNO, a neural operator that places the metriplectic (GENERIC) degeneracy structure of non-equilibrium thermodynamics directly in function space: reversible, energy-conserving dynamics are coupled to irreversible, entropy-producing dynamics via the degeneracy conditions. Structure-preserving neural operators developed previously enforce no more than one conservation law, or else a purely Hamiltonian form, and thermodynamically consistent learning has been restricted to finite-dimensional, graph, or particle systems. In GENERIC-FNO, the energy and entropy functionals are lear… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/jaysulk/GENERIC-FNO · OpenReview ID：gbpEYDdfDB

期刊：Transactions on Machine Learning Research · 摘要：Discrete Diffusion Language Models have progressed rapidly as an alternative to Autoregressive (AR) models, motivated by their ability to generate tokens in parallel. However, for tractability, discrete diffusion models sample from a factorized distribution, which is less expressive than the AR factorization. Recent Flow Language Models (FLMs) apply continuous flows to language, transporting noise to data with a deterministic ODE that avoids factorized sampling. FLMs operate on one-hot vectors whose dimension scales with the vocabulary size, making FLMs costly to train. Moreover, since all di… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/jdeschena/s-flm · OpenReview ID：jBh2WNssnL

期刊：Transactions on Machine Learning Research · 摘要：Sparse Autoencoders (SAEs) have become essential for decomposing model activations into interpretable concepts. However, despite their effectiveness, SAEs lack a natural ordering of features, making it difficult to prioritize important concepts under compute constraints. The Matryoshka Sparse Autoencoder (MSAE) was introduced as a means of learning nested subspaces, theoretically forcing high-level features into earlier dimensions to enable adaptive granularity. In this paper, we reproduce and analyze the MSAE framework. The reproduction of the study suffers from certain challenges, but the m… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/nkhllg/MSAE-sandwich-rep · OpenReview ID：l0qZkSk2aH

期刊：Transactions on Machine Learning Research · 摘要：Evaluating spoken role-play requires judging not only linguistic content but also prosody, emotional expression, character embodiment, and contextual fit. Existing automatic evaluation methods often collapse these aspects into coarse scores and are evaluated primarily on synthesized speech. We introduce Speech-DRAME, a framework comprising Speech-DRAME-EvalBench, a bilingual, human-annotated benchmark with contrastive role-, context-, and delivery-mismatch conditions, and DRAME-Eval, a fine-tuned audio-language evaluator. The benchmark supports two complementary protocols: top-down archetype… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/Anuttacon/speech_drame · OpenReview ID：m9IP9ECrk2

期刊：Transactions on Machine Learning Research · 摘要：Although Multimodal Large Language Models (MLLMs) have shown remarkable potential in Visual Document Retrieval (VDR) through generating high-quality multi-vector embeddings, the substantial storage overhead caused by representing a page with thousands of visual tokens limits their practicality in real-world applications. To address this challenge, we propose an auto-regressive generation approach, CausalEmbed, for constructing multi-vector embeddings. By incorporating iterative margin loss during contrastive training, CausalEmbed encourages the embedding models to learn compact and well-struc… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/Z1zs/Causal-Embed · OpenReview ID：mZXoOgBxX2

期刊：Transactions on Machine Learning Research · 摘要：Many real-world tasks require optimizing expensive black-box functions accessible only through noisy evaluations, a setting commonly addressed with Bayesian optimization. While Gaussian processes (GPs) and Bayesian neural networks have driven significant progress, GP-based methods scale cubically with observations, and BNN-based approaches remain burdened by expensive posterior sampling such as Hamiltonian Monte Carlo, and inner-loop acquisition optimization in high-dimensional setting. In this work, we propose VBO-MI (Variational Black-Box Optimization with Mutual Information), a fully gradi… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：nkj5zYqJ2p

期刊：Transactions on Machine Learning Research · 摘要：Training agents directly on the task they will be deployed on often leads to the best policy performance, but training on physical robots can cause damage through falls, so minimizing training falls is critical. A standard mitigation hands control to a separate recovery policy whenever the agent leaves a designer-specified safe region. The resulting mixed-policy rollouts bias the on-policy update, and importance sampling cannot remove this bias when the recovery policy is deterministic. We propose SafeExplorer, a drop-in modification of proximal policy optimization (PPO) built on an unbiased… · 篇幅：Regular submission (no more than 12 pages of main content) · OpenReview ID：ppOrbSsMM0

期刊：Transactions on Machine Learning Research · 摘要：Performative prediction (PP) provides a framework for learning in settings where model deployment influences the data distribution on which the model is trained. Unlike conventional learning under a fixed data distribution, PP involves model-induced distribution shifts that make it challenging to design algorithms converging to stable points, known as stationary performatively stable (SPS) solutions. Existing algorithms for finding SPS solutions, including repeated gradient descent (RGD) and greedy stochastic gradient descent (SGD-GD), have largely focused on strongly convex losses. Although… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/osu-srml/SPRINT · OpenReview ID：q0L3a1zSro

期刊：Transactions on Machine Learning Research · 摘要：The geometry of generative models serves as the basis for interpolation, model inspection, and more. Although certain generative models admit an implicit geometric structure, there is no broadly applicable framework that captures a principled notion of geometry across generative models without imposing restrictive assumptions on the model class or data dimensionality. In this paper, we show how to equip generative models with a general geometry compatible with different metrics and probability distributions to analyze generative models. Our method does not require additional training. We cons… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/FrederikMR/likely_geometry · OpenReview ID：qKM72YmSSw

期刊：Transactions on Machine Learning Research · 摘要：This paper studies continual test-time adaptation (CTTA), the task of adapting a model to constantly changing unseen domains in testing while preserving previously learned knowledge. Existing CTTA methods mostly focus on adaptation to the current test domain only, overlooking generalization to arbitrary test domains a model may face in the future. To tackle this limitation, we present a novel online test-time domain generalization framework for CTTA, dubbed TestDG. TestDG aims to learn features invariant to both current and previous test domains on the fly during testing, improving the potent… · 篇幅：Regular submission (no more than 12 pages of main content) · 代码：https://github.com/sohyun-l/TestDG · OpenReview ID：qZ1HuY07Ud

期刊：Transactions on Machine Learning Research · 摘要：In many machine learning problems, understanding variable importance is a central concern. Two common approaches are Permute-and-Predict (PaP), which randomly permutes a feature in a validation set, and Permute-and-Relearn (PaRL), which retrains models after permuting a training feature. Both methods deem a variable important if predictions with the original data substantially outperform those with permutations. In linear regression, empirical and theoretical studies have linked PaP to regression coefficients and PaRL to $t$-statistics. We here employ a latent variable framework to offer a ne… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/KelvynBladen/theoryVarImportance · OpenReview ID：rdCPIjY3Dr

期刊：Transactions on Machine Learning Research · 摘要：Integer programming (IP) is central to many combinatorial optimization tasks but remains challenging due to its NP-hard nature. A practical way to improve IP solvers is to manually design inequality families that exploit problem structure. However, this creative process requires deep expertise and has therefore been difficult to automate. Our proposed framework, Symbolic Acceleration via Generative Evolution (SAGE), automates symbolic discovery of acceleration inequality families at the modeling level. It reasons over a symbolic IP formulation and a natural language description of the problem… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/milad1378yz/sage · OpenReview ID：rxJvINT074

期刊：Transactions on Machine Learning Research · 摘要：Deep neural networks (DNNs) and Kolmogorov-Arnold networks (KANs) provide flexible function approximators, but their accuracy can require substantial network depth, width, and trainable parameter counts. Beyond making the resulting networks less transparent, large parameter spaces can increase training cost and produce solutions that are sensitive to optimization and parameter initialization, leading to variability in generalization error. For these reasons, we propose Shallow Universal Polynomial Networks (SUPNs), which retain a single hidden layer while increasing the expressivity of each u… · 篇幅：Long submission (more than 12 pages of main content) · OpenReview ID：srHGkCrSTM

期刊：Transactions on Machine Learning Research · 摘要：We introduce `dualGNN', an autoregressive message-passing GNN for sampling fine, regular triangulations of lattice polytopes. dualGNN operates on a generalization of the dual graph of a triangulation, with edges labeled by `signed circuits' --- combinatorial invariants from the theory of oriented matroids. We show that these circuits are necessary and sufficient to determine a triangulation's regularity from the graph, provided certain magnitude information is retained. The model is independent of the polytope's point count and invariant under its orientation-preserving symmetries ($\mathrm{S… · 篇幅：Long submission (more than 12 pages of main content) · 代码：https://github.com/natemacfadden/dualGNN · OpenReview ID：tlscqOYjmZ