This paper studies policy transfer, one of the well-known transfer learning techniques adopted in large language models, for continuous-time reinforcement learning problems. In the case of continuous-time linear-quadratic systems with Shannon's entropy regularization, we fully exploit the Gaussian structure of their optimal policy and the stability of their associated Riccati equations. In the general case where the system has possibly non-linear and bounded dynamics, the key technical component is the stability of diffusion SDEs which is established by invoking the rough path theory. Our work provides the first theoretical proof of policy transfer for continuous-time RL: an optimal policy learned for one RL problem can be used to initialize to search for a near-optimal policy for another closely related RL problem, while achieving (at least) the same rate of convergence for the original algorithm. As a byproduct of our analysis, we derive the stability of a concrete class of continuous-time score-based diffusion models via their connection with LQRs. To illustrate the benefit of policy transfer for RL, we propose a novel policy learning algorithm for continuous-time LQRs, which achieves global linear convergence and local super-linear convergence.

Space exploration increasingly relies on Virtual Reality for several tasks, such as mission planning, multidisciplinary scientific analysis, and astronaut training. A key factor for the reliability of the simulations is having accurate 3D representations of planetary terrains. Extraterrestrial heightmaps derived from satellite imagery often contain missing values due to acquisition and transmission constraints. Mars is among the most studied planets beyond Earth, and its extensive terrain datasets make the Martian surface reconstruction a valuable task, although many areas remain unmapped. Deep learning algorithms can support void-filling tasks; however, whereas Earth's comprehensive datasets enables the use of conditional methods, such approaches cannot be applied to Mars. Current approaches rely on simpler interpolation techniques which, however, often fail to preserve geometric coherence. In this work, we propose a method for reconstructing the surface of Mars based on an unconditional diffusion model. Training was conducted on an augmented dataset of 12000 Martian heightmaps derived from NASA's HiRISE survey. A non-homogeneous rescaling strategy captures terrain features across multiple scales before resizing to a fixed 128x128 model resolution. We compared our method against established void-filling and inpainting techniques, including Inverse Distance Weighting, kriging, and Navier-Stokes algorithm, on an evaluation set of 1000 samples. Results show that our approach consistently outperforms these methods in terms of reconstruction accuracy (4-15% on RMSE) and perceptual similarity (29-81% on LPIPS) with the original data.

Large Vision-Language Models (LVLMs) have demonstrated remarkable multimodal capabilities, but they inherit the tendency to hallucinate from their underlying language models. While visual contrastive decoding has been proposed to mitigate this issue, existing methods often apply generic visual augmentations that disregard the specific context provided by the text query, limiting their effectiveness. This study introduces a novel training-free decoding strategy that addresses these limitations, featuring two key contributions. First, a self-augmentation prompting strategy that leverages the intrinsic knowledge of the model to dynamically align semantics between the query and the visual augmentation. Second, an adaptive thresholding algorithm that adaptively adjusts next token candidate size based on the output sparsity, utilizing full information from the logit distribution. Extensive experiments across four LVLMs and seven benchmarks demonstrate that the proposed decoding significantly enhances factual consistency compared to state-of-the-art decoding methods. This work highlights the importance of integrating query-dependent augmentation and entropy-aware decoding for improving effective generation of LVLMs.

Reasoning-based image quality assessment (IQA) models trained through reinforcement learning (RL) exhibit exceptional generalization, yet the underlying mechanisms and critical factors driving this capability remain underexplored in current research. Moreover, despite their superior performance, these models incur inference energy usage and latency orders of magnitude higher than their earlier counterparts, restricting their deployment in specific scenarios. Through extensive experiments, this paper verifies and elaborates that through RL training, MLLMs leverage their reasoning capability to convert redundant visual representations into compact, cross-domain aligned text representations. This conversion is precisely the source of the generalization exhibited by these reasoning-based IQA models. Building on this fundamental insight, we propose a novel algorithm, RALI, which employs contrastive learning to directly align images with these generalizable text representations learned by RL. This approach eliminates the reliance on reasoning processes and even obviates the need to load an LLM. For the quality scoring task, this framework achieves generalization performance comparable to reasoning-based models while requiring less than 5% of their model parameters and inference time.

Disclosing information via the publication of a machine learning model poses significant privacy risks. However, auditing this disclosure across every datapoint during the training of Large Language Models (LLMs) is computationally prohibitive. In this paper, we present Gradient Uniqueness (GNQ), a principled, attack-agnostic metric derived from an information-theoretic upper bound on the amount of information embedded in a model about individual training points via gradient descent. While naively computing GNQ requires forming and inverting an $P \times P$ matrix for every datapoint (for a model with $P$ parameters), we introduce Batch-Space Ghost GNQ (BS-Ghost GNQ). This efficient algorithm performs all computations in a much smaller batch-space and leverages ghost kernels to compute GNQ ``in-run'' with minimal computational overhead. We empirically validate that GNQ successfully accounts for prior/common knowledge. Our evaluation demonstrates that GNQ strongly predicts sequence extractability in targeted attacks and reveals how disclosure risk concentrates heterogeneously on specific examples over the course of LLM training.

As AI increasingly shapes daily life, energy consumption and data privacy have become pressing concerns. On-device learning trains models directly on edge devices, cutting energy consumption and safeguarding data privacy. However, the expanding scale of modern neural networks creates a major obstacle for on-device training. Although prior work has concentrated on compact convolutional architectures, we instead apply subspace-based training to transformer models. Motivated by the idea that a model's essential information lies in a fixed subspace, we introduce Weight-Activation Subspace Iteration (WASI), a method that mitigates the memory bottleneck of backpropagation and boosts inference efficiency in transformer models by restricting training to this subspace. Our results demonstrate that WASI maintains accuracy comparable to vanilla training while reducing memory usage by up to $62\times$ and computational cost (FLOPs) by up to $2\times$. On a Raspberry Pi 5, WASI achieves roughly $1.4\times$ faster training and inference than vanilla training. The code is available at https://github.com/Le-TrungNguyen/ICLR2026-WASI.git.

Reasoning LLMs are trained to verbalize their reasoning process, yielding strong gains on complex tasks. This transparency also opens a promising direction: multiple reasoners can directly collaborate on each other's thinking within a shared trajectory, yielding better inference efficiency and exploration. A key prerequisite, however, is the ability to assess the usefulness and build on another model's partial thinking -- we call this off-trajectory reasoning. Our paper investigates a critical question: can standard solo-reasoning training pipelines deliver desired off-trajectory behaviors? We propose twin tests that capture the two extremes of the off-trajectory spectrum, namely Recoverability, which tests whether LLMs can backtrack from "distractions" induced by misleading reasoning traces, and Guidability, which tests their ability to build upon correct reasoning from stronger collaborators. Our study evaluates 15 open-weight LLMs (1.5B-32B) and reveals a counterintuitive finding -- "stronger" LLMs on benchmarks are often more fragile under distraction. Moreover, all models tested fail to effectively leverage guiding steps from collaborators on problems beyond their inherent capabilities with solve rates remaining under 9.2%. Finally, we conduct control studies to isolate the effects of three factors in post-training on these behaviors: the choice of distillation teacher, the use of RL, and data selection strategy. Our results provide actionable insights for training natively strong reasoning collaborators; e.g., we find that suboptimal recoverability behaviors of teacher models are transferred to distilled students even if the distillation trajectories are correct. Taken together, this work lays the groundwork for evaluating multi-model collaborations in shared reasoning trajectories and highlights the limitations of off-the-shelf reasoning LLMs.

Multi-LLM systems harness the complementary strengths of diverse Large Language Models, achieving performance and efficiency gains that are not attainable by a single model. In existing designs, LLMs communicate through text, forcing internal representations to be transformed into output token sequences. This process both loses rich semantic information and incurs token-by-token generation latency. Motivated by these limitations, we ask: Can LLMs communicate beyond text? Oracle experiments show that enriching the KV-Cache semantics can improve response quality without increasing cache size, supporting KV-Cache as an effective medium for inter-model communication. Thus, we propose Cache-to-Cache (C2C), a new paradigm for direct semantic communication between LLMs. C2C uses a neural network to project and fuse the source model's KV-cache with that of the target model to enable direct semantic transfer. A learnable gating mechanism selects the target layers that benefit from cache communication. Compared with text communication, C2C utilizes the deep, specialized semantics from both models, while avoiding explicit intermediate text generation. Experiments show that C2C achieves 6.4-14.2% higher average accuracy than individual models. It further outperforms the text communication paradigm by approximately 3.1-5.4%, while delivering an average 2.5x speedup in latency. Our code is available at https://github.com/thu-nics/C2C.

To utilize pre-trained neural networks on edge and mobile devices, we often require efficient adaptation to user-specific runtime data distributions while operating under limited compute and memory resources. On-device retraining with a target dataset can facilitate such adaptations; however, it remains impractical due to the increasing depth of modern neural nets, as well as the computational overhead associated with gradient-based optimization across all layers. Current approaches reduce training cost by selecting a subset of layers for retraining, however, they rely on labeled data, at least one full-model backpropagation, or server-side meta-training; limiting their suitability for constrained devices. We introduce AdaBet, a gradient-free layer selection approach to rank important layers by analyzing topological features of their activation spaces through Betti Numbers and using forward passes alone. AdaBet allows selecting layers with high learning capacity, which are important for retraining and adaptation, without requiring labels or gradients. Evaluating AdaBet on sixteen pairs of benchmark models and datasets, shows AdaBet achieves an average gain of 2.5% more classification accuracy over gradient-based baselines while reducing average peak memory consumption by 40%.

Samples of brain signals collected by EEG sensors have inherent anti-correlations that are well modeled by negative edges in a finite graph. To differentiate epilepsy patients from healthy subjects using collected EEG signals, we build lightweight and interpretable transformer-like neural nets by unrolling a spectral denoising algorithm for signals on a balanced signed graph -- graph with no cycles of odd number of negative edges. A balanced signed graph has well-defined frequencies that map to a corresponding positive graph via similarity transform of the graph Laplacian matrices. We implement an ideal low-pass filter efficiently on the mapped positive graph via Lanczos approximation, where the optimal cutoff frequency is learned from data. Given that two balanced signed graph denoisers learn posterior probabilities of two different signal classes during training, we evaluate their reconstruction errors for binary classification of EEG signals. Experiments show that our method achieves classification performance comparable to representative deep learning schemes, while employing dramatically fewer parameters.

Diffusion models are widely used for generative tasks across domains. Given a pre-trained diffusion model, it is often desirable to fine-tune it further either to correct for errors in learning or to align with downstream applications. Towards this, we examine the effect of shaping the distribution at intermediate noise levels induced by diffusion models. First, we show that existing variants of Rejection sAmpling based Fine-Tuning (RAFT), which we unify as GRAFT, can implicitly perform KL regularized reward maximization with reshaped rewards. Motivated by this observation, we introduce P-GRAFT to shape distributions at intermediate noise levels and demonstrate empirically that this can lead to more effective fine-tuning. We mathematically explain this via a bias-variance tradeoff. Next, we look at correcting learning errors in pre-trained flow models based on the developed mathematical framework. In particular, we propose inverse noise correction, a novel algorithm to improve the quality of pre-trained flow models without explicit rewards. We empirically evaluate our methods on text-to-image(T2I) generation, layout generation, molecule generation and unconditional image generation. Notably, our framework, applied to Stable Diffusion v2, improves over policy gradient methods on popular T2I benchmarks in terms of VQAScore and shows an $8.81\%$ relative improvement over the base model. For unconditional image generation, inverse noise correction improves FID of generated images at lower FLOPs/image.

Video frame interpolation (VFI), which generates intermediate frames from given start and end frames, has become a fundamental function in video generation applications. However, existing generative VFI methods are constrained to synthesize a fixed number of intermediate frames, lacking the flexibility to adjust generated frame rates or total sequence duration. In this work, we present ArbInterp, a novel generative VFI framework that enables efficient interpolation at any timestamp and of any length. Specifically, to support interpolation at any timestamp, we propose the Timestamp-aware Rotary Position Embedding (TaRoPE), which modulates positions in temporal RoPE to align generated frames with target normalized timestamps. This design enables fine-grained control over frame timestamps, addressing the inflexibility of fixed-position paradigms in prior work. For any-length interpolation, we decompose long-sequence generation into segment-wise frame synthesis. We further design a novel appearance-motion decoupled conditioning strategy: it leverages prior segment endpoints to enforce appearance consistency and temporal semantics to maintain motion coherence, ensuring seamless spatiotemporal transitions across segments. Experimentally, we build comprehensive benchmarks for multi-scale frame interpolation (2x to 32x) to assess generalizability across arbitrary interpolation factors. Results show that ArbInterp outperforms prior methods across all scenarios with higher fidelity and more seamless spatiotemporal continuity. Project website: https://mcg-nju.github.io/ArbInterp-Web/.

Diffusion Transformer has shown remarkable abilities in generating high-fidelity videos, delivering visually coherent frames and rich details over extended durations. However, existing video generation models still fall short in subject-consistent video generation due to an inherent difficulty in parsing prompts that specify complex spatial relationships, temporal logic, and interactions among multiple subjects. To address this issue, we propose BindWeave, a unified framework that handles a broad range of subject-to-video scenarios from single-subject cases to complex multi-subject scenes with heterogeneous entities. To bind complex prompt semantics to concrete visual subjects, we introduce an MLLM-DiT framework in which a pretrained multimodal large language model performs deep cross-modal reasoning to ground entities and disentangle roles, attributes, and interactions, yielding subject-aware hidden states that condition the diffusion transformer for high-fidelity subject-consistent video generation. Experiments on the OpenS2V benchmark demonstrate that our method achieves superior performance across subject consistency, naturalness, and text relevance in generated videos, outperforming existing open-source and commercial models.

3D Gaussian Splatting (3DGS) has emerged as an efficient approach for achieving photorealistic rendering. Recent MLP-based variants further improve visual fidelity but introduce substantial decoding overhead during rendering. To alleviate computation cost, several pruning strategies and level-of-detail (LOD) techniques have been introduced, aiming to effectively reduce the number of Gaussian primitives in large-scale scenes. However, our analysis reveals that significant redundancy still remains due to the lack of occlusion awareness. In this work, we propose Proxy-GS, a novel pipeline that exploits a proxy to introduce Gaussian occlusion awareness from any view. At the core of our approach is a fast proxy system capable of producing precise occlusion depth maps at a resolution of 1000x1000 under 1ms. This proxy serves two roles: first, it guides the culling of anchors and Gaussians to accelerate rendering speed. Second, it guides the densification towards surfaces during training, avoiding inconsistencies in occluded regions, and improving the rendering quality. In heavily occluded scenarios, such as the MatrixCity Streets dataset, Proxy-GS not only equips MLP-based Gaussian splatting with stronger rendering capability but also achieves faster rendering speed. Specifically, it achieves more than 2.5x speedup over Octree-GS, and consistently delivers substantially higher rendering quality. Code will be public upon acceptance.

Answering complex medical questions requires not only domain expertise and patient-specific information, but also structured and multi-perspective reasoning. Existing multi-agent approaches often rely on fixed roles or shallow interaction prompts, limiting their ability to detect and resolve fine-grained logical inconsistencies. To address this, we propose \textsc{MedLA}, a logic-driven multi-agent framework built on large language models. Each agent organizes its reasoning process into an explicit logical tree based on syllogistic triads (major premise, minor premise, and conclusion), enabling transparent inference and premise-level alignment. Agents engage in a multi-round, graph-guided discussion to compare and iteratively refine their logic trees, achieving consensus through error correction and contradiction resolution. We demonstrate that \textsc{MedLA} consistently outperforms both static role-based systems and single-agent baselines on challenging benchmarks such as MedDDx and standard medical QA tasks. Furthermore, \textsc{MedLA} scales effectively across both open-source and commercial LLM backbones, achieving state-of-the-art performance and offering a generalizable paradigm for trustworthy medical reasoning.

The Entropic Optimal Transport (EOT) problem and its dynamic counterpart, the Schrödinger bridge (SB) problem, play an important role in modern machine learning, linking generative modeling with optimal transport theory. While recent advances in discrete diffusion and flow models have sparked growing interest in applying SB methods to discrete domains, there remains no reliable way to assess how well these methods actually solve the underlying problem. We address this challenge by introducing a benchmark for SB on discrete spaces. Our construction yields pairs of probability distributions with analytically known SB solutions, enabling rigorous evaluation. As a byproduct of building this benchmark, we obtain two new SB algorithms, DLightSB and DLightSB-M, and additionally extend prior related work to construct the $α$-CSBM algorithm. We demonstrate the utility of our benchmark by evaluating both existing and new solvers in high-dimensional discrete settings. This work provides the first step toward proper evaluation of SB methods on discrete spaces, paving the way for more reproducible future studies. The code for the benchmark and all associated experiments is available at https://github.com/gregkseno/catsbench.

Inference-time control of diffusion models aims to steer model outputs to satisfy new constraints without retraining. Previous approaches have mostly relied on heuristic guidance or have been coupled with Sequential Monte Carlo (SMC) for bias correction. In this paper, we propose a flexible alternative based on replica exchange, an algorithm designed initially for sampling problems. We refer to this method as CREPE (Controlling with REPlica Exchange). Unlike SMC, CREPE: (1) generates particles sequentially, (2) maintains high diversity in the generated samples after a burn-in period, and (3) enables online refinement or early termination. We demonstrate its versatility across various tasks, including temperature annealing, reward-tilting, model composition and classifier-free guidance debiasing, with competitive performance compared to prior SMC methods.

Earth observation (EO) is essential for understanding the evolving states of the Earth system. Although recent MLLMs have advanced EO research, they still lack the capability to tackle complex tasks that require multi-step reasoning and the use of domain-specific tools. Agent-based methods offer a promising direction, but current attempts remain in their infancy, confined to RGB perception, shallow reasoning, and lacking systematic evaluation protocols. To overcome these limitations, we introduce Earth-Agent, the first agentic framework that unifies RGB and spectral EO data within an MCP-based tool ecosystem, enabling cross-modal, multi-step, and quantitative spatiotemporal reasoning beyond pretrained MLLMs. Earth-Agent supports complex scientific tasks such as geophysical parameter retrieval and quantitative spatiotemporal analysis by dynamically invoking expert tools and models across modalities. To support comprehensive evaluation, we further propose Earth-Bench, a benchmark of 248 expert-curated tasks with 13,729 images, spanning spectrum, products and RGB modalities, and equipped with a dual-level evaluation protocol that assesses both reasoning trajectories and final outcomes. We conduct comprehensive experiments varying different LLM backbones, comparisons with general agent frameworks, and comparisons with MLLMs on remote sensing benchmarks, demonstrating both the effectiveness and potential of Earth-Agent. Earth-Agent establishes a new paradigm for EO analysis, moving the field toward scientifically grounded, next-generation applications of LLMs in Earth observation. More information about Earth-Agent can be found at https://github.com/opendatalab/Earth-Agent

N-gram novelty is widely used to evaluate language models' ability to generate text outside of their training data. More recently, it has also been adopted as a metric for measuring textual creativity. However, theoretical work on creativity suggests that this approach may be inadequate, as it does not account for creativity's dual nature: novelty (how original the text is) and appropriateness (how sensical and pragmatic it is). We investigate the relationship between this notion of creativity and n-gram novelty through 8,618 expert writer annotations of novelty, pragmaticality, and sensicality via close reading of human- and AI-generated text. We find that while n-gram novelty is positively associated with expert writer-judged creativity, approximately 91% of top-quartile n-gram novel expressions are not judged as creative, cautioning against relying on n-gram novelty alone. Furthermore, unlike in human-written text, higher n-gram novelty in open-source LLMs correlates with lower pragmaticality. In an exploratory study with frontier closed-source models, we additionally confirm that they are less likely to produce creative expressions than humans. Using our dataset, we test whether zero-shot, few-shot, and finetuned models are able to identify expressions perceived as novel by experts (a positive aspect of writing) or non-pragmatic (a negative aspect). Overall, frontier LLMs exhibit performance much higher than random but leave room for improvement, especially struggling to identify non-pragmatic expressions. We further find that LLM-as-a-Judge novelty ratings align with expert writer preferences in an out-of-distribution dataset, more so than an n-gram based metric.

The Minimum Description Length (MDL) principle offers a formal framework for applying Occam's razor in machine learning. However, its application to neural networks such as Transformers is challenging due to the lack of a principled, universal measure for model complexity. This paper introduces the theoretical notion of asymptotically optimal description length objectives, grounded in the theory of Kolmogorov complexity. We establish that a minimizer of such an objective achieves optimal compression, for any dataset, up to an additive constant, in the limit as model resource bounds increase. We prove that asymptotically optimal objectives exist for Transformers, building on a new demonstration of their computational universality. We further show that such objectives can be tractable and differentiable by constructing and analyzing a variational objective based on an adaptive Gaussian mixture prior. Our empirical analysis shows that this variational objective selects for a low-complexity solution with strong generalization on an algorithmic task, but standard optimizers fail to find such solutions from a random initialization, highlighting key optimization challenges. More broadly, by providing a theoretical framework for identifying description length objectives with strong asymptotic guarantees, we outline a potential path towards training neural networks that achieve greater compression and generalization.

Vision Transformers are widely adopted as the backbone of vision foundation models, but they are known to produce high-norm artifacts that degrade representation quality. When knowledge distillation transfers these features to students, high-norm artifacts dominate the objective, so students overfit to artifacts and underweight informative signals, diminishing the gains from larger models. Prior work attempted to remove artifacts but encountered an inherent trade-off between artifact suppression and preserving informative signals from teachers. To address this, we introduce Singular Nullspace-Guided Energy Reallocation (SiNGER), a novel distillation framework that suppresses artifacts while preserving informative signals. The key idea is principled teacher feature refinement: during refinement, we leverage the nullspace-guided perturbation to preserve information while suppressing artifacts. Then, the refined teacher's features are distilled to a student. We implement this perturbation efficiently with a LoRA-based adapter that requires minimal structural modification. Extensive experiments show that \oursname consistently improves student models, achieving state-of-the-art performance in multiple downstream tasks and producing clearer and more interpretable representations.

HRI research increasingly demands robots that go beyond task execution to respond meaningfully to user emotions. This is especially needed when supporting students with learning difficulties in game-based learning scenarios. Here, the objective of these robots is to train users with game-playing skills, and this requires robots to get input about users' interests and engagement. In this paper, we present a system for an adaptive social game-playing robot. However, creating such an agent through online RL requires extensive real-world training data and potentially be uncomfortable for users. To address this, we investigate offline RL as a safe and efficient alternative. We introduce a system architecture that integrates multimodal emotion recognition and adaptive robotic responses. We also evaluate the performance of various offline RL algorithms using a dataset collected from a real-world human-robot game-playing scenario. Our results indicate that BCQ and DDQN offer the greatest robustness to hyperparameter variations, whereas CQL is the most effective at mitigating overestimation bias. Through this research, we aim to inform the selection and design of reliable offline RL policies for real-world social robotics. Ultimately, this work provides a foundational step toward creating socially intelligent agents that can learn complex and emotion-adaptive behaviors entirely from offline datasets, ensuring both human comfort and practical scalability.

We study the gradient-based training of large-depth residual networks (ResNets) from standard random initializations. We show that infinite-depth ResNets behave as if they were infinitely wide, regardless of their actual width. More precisely, we obtain that with a fixed embedding dimension $D$, the training dynamics converges to a unique Neural Mean ODE training dynamics as the depth $L$ diverges, regardless of the scaling of the hidden width $M$. For a residual scale $Θ_D\big(\fracα{LM}\big)$ with $α=Θ_D(1)$, we obtain the error bound $O_D\big(\frac{1}{L}+ \frac{1}{\sqrt{LM}}\big)$ between the model's output and its limit after a fixed number gradient of steps. In this regime, the limit exhibits maximal local feature updates, i.e. the Mean ODE is genuinely non-linearly parameterized. In contrast, we show that $α\to \infty$ yields a lazy ODE regime where the Mean ODE is linearly parameterized, and we derive a convergence rate in this case as well. We then focus on the particular case of ResNets with two-layer perceptron blocks, for which we study how these scalings depend on the embedding dimension $D$. We identify the residual scale $O\big(\frac{\sqrt{D}}{LM}\big)$ as necessary and sufficient for maximal local feature updates. In this regime, we prove a high-probability error bound $O\big(\frac{1}{L}+ \frac{\sqrt{D}}{\sqrt{LM}}\big)$ between the ResNet and its limit after a fixed number of gradient steps. Our convergence results rely on a novel mathematical perspective on ResNets : (i) due to the randomness of the initialization, the forward and backward pass through the ResNet behave as the stochastic approximation of certain mean ODEs, and (ii) by propagation of chaos (that is, asymptotic independence of the units) this behavior is preserved through the training dynamics. We verify empirically that all our rates are tight.

A central paradox in fine-tuning Large Language Models (LLMs) with Reinforcement Learning with Verifiable Reward (RLVR) is the frequent degradation of multi-attempt performance (Pass@k) despite improvements in single-attempt accuracy (Pass@1). This is often accompanied by catastrophic forgetting, where models lose previously acquired skills. While various methods have been proposed, the choice and function of the divergence term have been surprisingly unexamined as a proactive solution. We argue that standard RLVR objectives -- both those using the mode-seeking reverse KL-divergence and those forgoing a divergence term entirely -- lack a crucial mechanism for knowledge retention. The reverse-KL actively accelerates this decay by narrowing the policy, while its absence provides no safeguard against the model drifting from its diverse knowledge base. We propose a fundamental shift in perspective: using the divergence term itself as the solution. Our framework, Diversity-Preserving Hybrid RL (DPH-RL), leverages mass-covering f-divergences (like forward-KL and JS-divergence) to function as a rehearsal mechanism. By continuously referencing the initial policy, this approach forces the model to maintain broad solution coverage. Extensive experiments on math and SQL generation demonstrate that DPH-RL not only resolves the Pass@k degradation but improves both Pass@1 and Pass@k in- and out-of-domain. Additionally, DPH-RL is more training-efficient because it computes f-divergence using generator functions, requiring only sampling from the initial policy and no online reference model. Our work highlights a crucial, overlooked axis for improving RLVR, demonstrating that the proper selection of a divergence measure is a powerful tool for building more general and diverse reasoning models.

We show that translation quality can be predicted with surprising accuracy \textit{without ever running the translation system itself}. Using only a handful of features, token fertility ratios, token counts, and basic linguistic metadata (language family, script, and region), we can forecast ChrF scores for GPT-4o translations across 203 languages in the FLORES-200 benchmark. Gradient boosting models achieve favorable performance ($R^{2}=0.66$ for XX$\rightarrow$English and $R^{2}=0.72$ for English$\rightarrow$XX). Feature importance analyses reveal that typological factors dominate predictions into English, while fertility plays a larger role for translations into diverse target languages. These findings suggest that translation quality is shaped by both token-level fertility and broader linguistic typology, offering new insights for multilingual evaluation and quality estimation.

This paper presents a novel application of the Tabular Prior-Data Fitted Network (TabPFN) - a transformer-based foundation model for tabular data - to geotechnical site characterization problems defined in the GEOAI benchmark BM/AirportSoilProperties/2/2025. Two tasks are addressed: (1) predicting the spatial variation of undrained shear strength (su) across borehole depth profiles, and (2) imputing missing mechanical parameters in a dense-site dataset. We apply TabPFN in a zero-training, few-shot, in-context learning setting - without hyper-parameter tuning - and provide it with additional context from the big indirect database (BID). The study demonstrates that TabPFN, as a general-purpose foundation model, achieved superior accuracy and well-calibrated predictive distributions compared to a conventional hierarchical Bayesian model (HBM) baseline, while also offering significant gains in inference efficiency. In Benchmark Problem #1 (spatial su prediction), TabPFN outperformed the HBM in prediction accuracy and delivered an order-of-magnitude faster runtime. In Benchmark Problem #2 (missing mechanical parameter imputation), TabPFN likewise achieved lower RMSE for all target parameters with well-quantified uncertainties, though its cumulative computation cost was higher than HBM's due to its one-variable-at-a-time inference. These results mark the first successful use of a tabular foundation model in geotechnical modeling, suggesting a potential paradigm shift in probabilistic site characterization.

Reinforcement learning (RL) has emerged as an effective post-training paradigm for enhancing the reasoning capabilities of multimodal large language model (MLLM). However, current RL pipelines often suffer from training inefficiencies caused by two underexplored issues: Advantage Collapsing, where most advantages in a batch concentrate near zero, and Rollout Silencing, where the proportion of rollouts contributing non-zero gradients diminishes over time. These issues lead to suboptimal gradient updates and hinder long-term learning efficiency. To address these issues, we propose Shuffle-R1, a simple yet principled framework that improves RL fine-tuning efficiency by dynamically restructuring trajectory sampling and batch composition. It introduces (1) Pairwise Trajectory Sampling, which selects high-contrast trajectories with large advantages to improve gradient signal quality, and (2) Advantage-based Trajectory Shuffle, which increases exposure of valuable rollouts through informed batch reshuffling. Experiments across multiple reasoning benchmarks show that our framework consistently outperforms strong RL baselines with minimal overhead. These results highlight the importance of data-centric adaptations for more efficient RL training in MLLM.

We explain how data-driven quantization of a linear unit in a neural network corresponds to solving the closest vector problem for a certain lattice generated by input data. We prove that the GPTQ algorithm is equivalent to Babai's well-known nearest-plane algorithm. We furthermore provide geometric intuition for both algorithms. Lastly, we note the consequences of these results, in particular hinting at the possibility of using lattice basis reduction for improved quantization.

Recent advancements in diffusion models have significantly improved symbolic music generation. However, most approaches rely on transformer-based architectures with self-attention mechanisms, which are constrained by quadratic computational complexity, limiting scalability for long sequences. To address this, we propose Symbolic Music Diffusion with Mamba (SMDIM), a novel diffusion-based architecture integrating Structured State Space Models (SSMs) for efficient global context modeling and the Mamba-FeedForward-Attention Block (MFA) for precise local detail preservation. The MFA Block combines the linear complexity of Mamba layers, the non-linear refinement of FeedForward layers, and the fine-grained precision of self-attention mechanisms, achieving a balance between scalability and musical expressiveness. SMDIM achieves near-linear complexity, making it highly efficient for long-sequence tasks. Evaluated on diverse datasets, including FolkDB, a collection of traditional Chinese folk music that represents an underexplored domain in symbolic music generation, SMDIM outperforms state-of-the-art models in both generation quality and computational efficiency. Beyond symbolic music, SMDIM's architectural design demonstrates adaptability to a broad range of long-sequence generation tasks, offering a scalable and efficient solution for coherent sequence modeling.

To operate effectively in the real world, robots should integrate multimodal reasoning with precise action generation. However, existing vision-language-action (VLA) models often sacrifice one for the other, narrow their abilities to task-specific manipulation data, and suffer catastrophic forgetting of pre-trained vision-language capabilities. To bridge this gap, we introduce InstructVLA, an end-to-end VLA model that preserves the flexible reasoning of large vision-language models (VLMs) while delivering leading manipulation performance with the help of embodied reasoning. InstructVLA introduces a novel training paradigm, Vision-Language-Action Instruction Tuning (VLA-IT), which employs multimodal training with mixture-of-experts adaptation to jointly optimize embodied reasoning and action generation on both standard VLM corpora and a curated 650K-sample VLA-IT dataset. On in-domain SimplerEnv tasks, InstructVLA achieves 33% improvement over SpatialVLA. To evaluate generalization, we introduce SimplerEnv-Instruct, an 80-task benchmark requiring closed-loop control and high-level instruction understanding, where it outperforms a fine-tuned OpenVLA by 96% and an action expert aided by GPT-4o by 29%. Additionally, InstructVLA surpasses baseline VLMs on multimodal tasks and exhibits inference-time scaling by leveraging textual reasoning to boost manipulation performance in both simulated and real-world settings. These results demonstrate InstructVLA's potential for bridging intuitive and steerable human-robot interaction with efficient policy learning.