Modern AI agents increasingly combine conversational interaction with autonomous task execution, such as coding and web research, raising a natural question: What happens when an agent engaged in long-horizon tasks is exposed to user persuasion? Yet studying this possibility is challenging because long-running agent behavior is noisy and costly to reproduce, and it remains unclear which unique challenges emerge only in extended task execution. We study how belief-level intervention can influence downstream task behavior, a phenomenon we name persuasion propagation. We introduce a behavior-centered evaluation framework that distinguishes between persuasion applied during or prior to task execution. Across web research and coding tasks, we find that on-the-fly persuasion induces weak and inconsistent behavioral effects. In contrast, when the belief state is explicitly specified at task time, belief-prefilled agents conduct on average 26.9% fewer searches and visit 16.9% fewer unique sources than neutral-prefilled agents. These results suggest that persuasion, even in prior interaction, can affect the agent's behavior, motivating behavior-level evaluation in agentic systems.

This paper proposes the soft Bayesian context tree model (Soft-BCT), which is a novel BCT model for real-valued time series. The Soft-BCT considers soft (probabilistic) splits of the context space, instead of hard (deterministic) splits of the context space as in the previous BCT for real-valued time series. A learning algorithm of the Soft-BCT is proposed based on the variational inference. The results of experiments demonstrate the superiority of the Soft-BCT compared to the previous BCT for some datasets.

Efficient distillation is a key pathway for converting expensive reasoning capability into deployable efficiency, yet in the frontier regime where the student already has strong reasoning ability, naive continual distillation often yields limited gains or even degradation. We observe a characteristic training phenomenon: even as loss decreases monotonically, all performance metrics can drop sharply at almost the same bottleneck, before gradually recovering. We further uncover a token-level mechanism: confidence bifurcates into steadily increasing Imitation-Anchor Tokens that quickly anchor optimization and other yet-to-learn tokens whose confidence is suppressed until after the bottleneck. And the characteristic that these two types of tokens cannot coexist is the root cause of the failure in continual distillation. To this end, we propose Training-Trajectory-Aware Token Selection (T3S) to reconstruct the training objective at the token level, clearing the optimization path for yet-to-learn tokens. T3S yields consistent gains in both AR and dLLM settings: with only hundreds of examples, Qwen3-8B surpasses DeepSeek-R1 on competitive reasoning benchmarks, Qwen3-32B approaches Qwen3-235B, and T3-trained LLaDA-2.0-Mini exceeds its AR baseline, achieving state-of-the-art performance among all of 16B-scale no-think models.

Lung cancer patients frequently experience breakthrough pain episodes, with up to 91% requiring timely intervention. To enable proactive pain management, we propose a hybrid machine learning and large language model pipeline that predicts pain episodes within 48 and 72 hours of hospitalization using both structured and unstructured electronic health record data. A retrospective cohort of 266 inpatients was analyzed, with features including demographics, tumor stage, vital signs, and WHO-tiered analgesic use. The machine learning module captured temporal medication trends, while the large language model interpreted ambiguous dosing records and free-text clinical notes. Integrating these modalities improved sensitivity and interpretability. Our framework achieved an accuracy of 0.876 (48h) and 0.917 (72h), with improvements in sensitivity of 10.6% and 10.7%, respectively, attributable to large language model augmentation. This hybrid approach offers a clinically interpretable and scalable tool for early pain episode forecasting, with potential to enhance treatment precision and optimize resource allocation in oncology care.

Activation sparsity offers a compelling route to accelerate large language model (LLM) inference by selectively suppressing hidden activations, yet existing approaches exhibit severe accuracy degradation at high sparsity. We show that this failure stems from representational instability: *activation sparsity disrupts input-dependent activation learned during pretraining, inducing distribution shifts in hidden states.* We address this issue by reframing activation sparsity as a representational alignment problem and introducing **Spontaneous Neurons (SPON)**, a lightweight mechanism inspired by spontaneous neural activity in biological systems. SPON injects a small set of learnable, input-independent activation vectors that act as persistent representational anchors for sparse computation. These vectors are trained via distribution matching to the dense model and can be absorbed into bias terms after training, incurring negligible inference overhead. Across multiple LLM backbones, SPON consistently restores performance, stabilizes latent representations, and preserves generalization. Our results establish SPON as an effective and principled solution for reliable activation-sparse inference, and offer new insights into knowledge retention in LLMs.

Even for the gradient descent (GD) method applied to neural network training, understanding its optimization dynamics, including convergence rate, iterate trajectories, function value oscillations, and especially its implicit acceleration, remains a challenging problem. We analyze nonlinear models with the logistic loss and show that the steps of GD reduce to those of generalized perceptron algorithms (Rosenblatt, 1958), providing a new perspective on the dynamics. This reduction yields significantly simpler algorithmic steps, which we analyze using classical linear algebra tools. Using these tools, we demonstrate on a minimalistic example that the nonlinearity in a two-layer model can provably yield a faster iteration complexity $\tilde{O}(\sqrt{d})$ compared to $Ω(d)$ achieved by linear models, where $d$ is the number of features. This helps explain the optimization dynamics and the implicit acceleration phenomenon observed in neural networks. The theoretical results are supported by extensive numerical experiments. We believe that this alternative view will further advance research on the optimization of neural networks.

End-to-end autonomous driving methods built on vision language models (VLMs) have undergone rapid development driven by their universal visual understanding and strong reasoning capabilities obtained from the large-scale pretraining. However, we find that current VLMs struggle to understand fine-grained 3D spatial relationships which is a fundamental requirement for systems interacting with the physical world. To address this issue, we propose SpaceDrive, a spatial-aware VLM-based driving framework that treats spatial information as explicit positional encodings (PEs) instead of textual digit tokens, enabling joint reasoning over semantic and spatial representations. SpaceDrive employs a universal positional encoder to all 3D coordinates derived from multi-view depth estimation, historical ego-states, and text prompts. These 3D PEs are first superimposed to augment the corresponding 2D visual tokens. Meanwhile, they serve as a task-agnostic coordinate representation, replacing the digit-wise numerical tokens as both inputs and outputs for the VLM. This mechanism enables the model to better index specific visual semantics in spatial reasoning and directly regress trajectory coordinates rather than generating digit-by-digit, thereby enhancing planning accuracy. Extensive experiments validate that SpaceDrive achieves state-of-the-art open-loop performance on the nuScenes dataset and the second-best Driving Score of 78.02 on the Bench2Drive closed-loop benchmark over existing VLM-based methods. Code is available at: https://github.com/zhenghao2519/SpaceDrive.

World models have been recently proposed as sandbox environments in which AI agents can be trained and evaluated before deployment. While realistic world models often have high computational demands, this can often be alleviated by exploiting the fact that real-world scenarios tend to involve subcomponents that interact in a modular manner. In this paper, we explore this idea by developing a framework for decomposing complex world models represented by transducers, a class of models generalising POMDPs. Whereas the composition of transducers is well understood, our results clarify how to invert this process by deriving sub-transducers operating on distinct input-output subspaces, enabling parallelizable and interpretable alternatives to monolithic world modelling that can support distributed inference. Overall, these results lay groundwork for bridging the computational efficiency required for real-world inference and the structural transparency demanded by AI safety.

Masked diffusion models (MDMs) are a promising alternative to autoregressive models (ARMs), but they suffer from inherently much higher training variance. High variance leads to noisier gradient estimates and unstable optimization, so even equally strong pretrained MDMs and ARMs that are competitive at initialization often diverge after task-specific training, with MDMs falling far behind. There has been no theoretical explanation or systematic solution. We derive the first decomposition of MDM training variance into three sources: (A) masking pattern noise, (B) masking rate noise, and (C) data noise, while ARMs are only affected by (C). This explains the fundamental training gap. Building on this foundation, we design six variance-reduction methods, including two core methods: (1) P-POTS, a Pareto-optimal t sampler that minimizes training variance by sampling harder t values more often with appropriately smaller update steps, and (2) MIRROR, which uses negatively correlated samples to reduce (A). Experiments show that compared to standard MDM training, our methods improve accuracy by 7-8% on complex reasoning tasks, while simultaneously reducing run-to-run variability to near ARM levels, substantially narrowing the gap with strong ARM baselines; in most settings, even the best baseline runs remain below the worst run of our method.

The improving multi-armed bandits problem is a formal model for allocating effort under uncertainty, motivated by scenarios such as investing research effort into new technologies, performing clinical trials, and hyperparameter selection from learning curves. Each pull of an arm provides reward that increases monotonically with diminishing returns. A growing line of work has designed algorithms for improving bandits, albeit with somewhat pessimistic worst-case guarantees. Indeed, strong lower bounds of $Ω(k)$ and $Ω(\sqrt{k})$ multiplicative approximation factors are known for both deterministic and randomized algorithms (respectively) relative to the optimal arm, where $k$ is the number of bandit arms. In this work, we propose two new parameterized families of bandit algorithms and bound the sample complexity of learning the near-optimal algorithm from each family using offline data. We also perform empirical evaluations on standard hyperparameter tuning benchmarks. The first family we define includes the optimal randomized algorithm from prior work. We show that an appropriately chosen algorithm from this family can achieve stronger guarantees, with optimal dependence on $k$, when the arm reward curves satisfy additional properties related to the strength of concavity. Our second family contains algorithms that both guarantee best-arm identification on well-behaved instances and revert to worst-case guarantees on poorly-behaved instances.

Molecular property regression struggles with cases in chemically relevant target ranges that are underrepresented in datasets. Standard average error minimization approaches underperform in these highly relevant cases, and oversampling approaches lead to meaningless molecular representations. In this paper, we propose SPECTRA, a spectral, domain-aware graph generation method designed to improve the prediction of underrepresented but relevant molecular property values. It combines a rarity-aware budgeting scheme to focus generation where data are scarce, target-neighbors graph alignment to establish structural correspondence, and interpolation of Laplacian spectra, node features, and targets. Coupled with spectral GNN using edge-aware Chebyshev convolutions, SPECTRA shows its effectiveness in property prediction benchmarks with competitive performance over leading state-of-the-art methods in relevant target ranges, while requiring ~4x less computational time.

Attention is a fundamental building block of large language models (LLMs), so there have been many efforts to implement it efficiently. For example, FlashAttention leverages tiling and kernel fusion to optimize attention. Recently, a number of variants of attention have been introduced to enhance model quality or efficiency. Supporting them efficiently remains difficult since they usually require specialized kernels or hand-tuned implementations. FlexAttention recently addressed part of this gap by using static programming templates to support FlashAttention-like kernels for a subset of attention variants. In this paper, we introduce Flashlight, a compiler-native framework within the PyTorch ecosystem that automatically generates fused, FlashAttention-style kernels for arbitrary attention-based programs, without relying on static templates or predefined kernel specializations. Flashlight leverages PyTorch's compilation workflow to fuse and tile attention computations transparently, enabling efficient execution for diverse attention patterns. Not only does it support all variants expressible in the FlexAttention model but it also handles more general, data-dependent attention formulations that are beyond the capabilities of FlexAttention. Our results show that Flashlight produces kernels with competitive or superior performance to FlexAttention, while offering the flexibility of native PyTorch code, enabling developers to rapidly explore new attention models without sacrificing performance.

Hyperspectral sensors capture dense spectra per pixel but suffer from low spatial resolution, causing blurred boundaries and mixed-pixel effects. Co-registered companion sensors such as multispectral, RGB, or panchromatic cameras provide high-resolution spatial detail, motivating hyperspectral super-resolution through the fusion of hyperspectral and multispectral images (HSI-MSI). Existing deep learning based methods achieve strong performance but rely on opaque regressors that lack interpretability and often fail when the MSI has very few bands. We propose SpectraMorph, a physics-guided self-supervised fusion framework with a structured latent space. Instead of direct regression, SpectraMorph enforces an unmixing bottleneck: endmember signatures are extracted from the low-resolution HSI, and a compact multilayer perceptron predicts abundance-like maps from the MSI. Spectra are reconstructed by linear mixing, with training performed in a self-supervised manner via the MSI sensor's spectral response function. SpectraMorph produces interpretable intermediates, trains in under a minute, and remains robust even with a single-band (pan-chromatic) MSI. Experiments on synthetic and real-world datasets show SpectraMorph consistently outperforming state-of-the-art unsupervised/self-supervised baselines while remaining very competitive against supervised baselines.

Understanding the capability bottlenecks of embodied multimodal large language models (MLLMs) is crucial for improving embodied agents. However, existing embodied benchmarks mainly focus on task-level evaluation and fail to provide actionable insights into the underlying causes of model failures. To address this limitation, we introduce BEAR, a benchmark that decomposes embodied tasks into 14 atomic skills for fine-grained skill-level evaluation. BEAR comprises 4,469 interleaved image-video-text samples spanning 14 skills across 6 categories, ranging from low-level perception to high-level planning. We evaluate 20 MLLMs on BEAR under a hierarchical skill-level diagnosis framework and uncover two key findings: (1) perceptual capabilities are major bottlenecks behind reasoning failures, and (2) current models suffer from unstable spatiotemporal modeling that remains largely unexposed in prior benchmarks. Motivated by these findings, we further propose BEAR-Agent, a multimodal conversational agent that augments MLLMs with visual and spatial reasoning tools. BEAR-Agent substantially improves performance across embodied skills, achieving a relative improvement of 17.5% on GPT-5 over the base model on BEAR, while also outperforming strong baselines in both simulation and real-world robotic experiments. Project page: https://bear-official66.github.io/

Large language models (LLMs) have demonstrated remarkable progress in reasoning, often through supervised fine-tuning (SFT). However, SFT is resource-intensive, relying on large curated datasets, rejection-sampled demonstrations, and uniform optimization across all tokens, even though only a fraction carry meaningful learning value. In this work, we explore a counterintuitive idea: can smaller language models (SLMs) teach larger language models (LLMs) by revealing high-value reasoning moments that reflect the latter's unique strength? We propose LightReasoner, a novel framework that leverages the behavioral divergence between a stronger expert model (LLM) and a weaker amateur model (SLM). LightReasoner operates in two stages: (1) a sampling stage that pinpoints critical reasoning moments and constructs supervision examples capturing the expert's advantage through expert-amateur contrast, and (2) a fine-tuning stage that aligns the expert model with these distilled examples, amplifying its reasoning strengths. Across seven mathematical benchmarks, LightReasoner improves accuracy by up to 28.1%, while reducing time consumption by 90%, sampled problems by 80%, and tuned token usage by 99%, all without relying on ground-truth labels. By turning weaker SLMs into effective teaching signals, LightReasoner offers a scalable and resource-efficient approach for advancing LLM reasoning. Code is available at: https://github.com/HKUDS/LightReasoner

Effective exploration remains a central challenge in model-based reinforcement learning (MBRL), particularly in high-dimensional continuous control tasks where sample efficiency is crucial. A prominent line of recent work leverages learned policies as proposal distributions for Model-Predictive Path Integral (MPPI) planning. Initial approaches update the sampling policy independently of the planner distribution, typically maximizing a learned value function with deterministic policy gradient and entropy regularization. However, because the states encountered during training depend on the MPPI planner, aligning the sampling policy with the planner improves the accuracy of value estimation and long-term performance. To this end, recent methods update the sampling policy by minimizing KL divergence to the planner distribution or by introducing planner-guided regularization into the policy update. In this work, we unify these MPPI-based reinforcement learning methods under a single framework by introducing Policy Optimization-Model Predictive Control (PO-MPC), a family of KL-regularized MBRL methods that integrate the planner's action distribution as a prior in policy optimization. By aligning the learned policy with the planner's behavior, PO-MPC allows more flexibility in the policy updates to trade off Return maximization and KL divergence minimization. We clarify how prior approaches emerge as special cases of this family, and we explore previously unstudied variations. Our experiments show that these extended configurations yield significant performance improvements, advancing the state of the art in MPPI-based RL.

Recent advances in multimodal language models (MLLMs) have made thinking with images a dominant paradigm for multimodal reasoning. However, existing methods still fail to ensure evidence-answer consistency, where correct answers must be supported by correct visual evidence. To address this issue, we propose DeFacto, a counterfactual reasoning framework that explicitly aligns visual evidence with final answers. Our approach integrates three complementary training paradigms: positive, counterfactual, and random-masking. We further develop a language-guided evidence construction pipeline that automatically localizes question-relevant regions and generates counterfactual variants, resulting in DeFacto-100K. Building on this dataset, we train MLLMs with GRPO-based reinforcement learning and design three complementary rewards to promote correct answering, structured reasoning, and consistent evidence selection. Moreover, we introduce DeFacto-1.5K, a human-annotated benchmark for systematically evaluating evidence-grounded consistency beyond answer accuracy. Experiments on diverse benchmarks demonstrate that DeFacto substantially improves both answer accuracy and evidence-answer consistency over strong baselines.

Detecting and localizing poultry is essential for advancing smart poultry farming. Despite the progress of detection-centric methods, challenges persist in free-range settings due to multiscale targets, obstructions, and complex or dynamic backgrounds. To tackle these challenges, we introduce an innovative poultry detection approach named SFN-YOLO that utilizes scale-aware fusion. This approach combines detailed local features with broader global context to improve detection in intricate environments. Furthermore, we have developed a new expansive dataset (M-SCOPE) tailored for varied free-range conditions. Comprehensive experiments demonstrate our model achieves an mAP of 80.7% with just 7.2M parameters, which is 35.1% fewer than the benchmark, while retaining strong generalization capability across different domains. The efficient and real-time detection capabilities of SFN-YOLO support automated smart poultry farming.

We study the Riemannian geometry of the Deep Linear Network (DLN) as a foundation for a thermodynamic description of the learning process. The main tools are the use of group actions to analyze overparametrization and the use of Riemannian submersion from the space of parameters to the space of observables. The foliation of the balanced manifold in the parameter space by group orbits is used to define and compute a Boltzmann entropy. We also show that the Riemannian geometry on the space of observables defined in [2] is obtained by Riemannian submersion of the balanced manifold. The main technical step is an explicit construction of an orthonormal basis for the tangent space of the balanced manifold using the theory of Jacobi matrices.

The cycle of scientific discovery is frequently bottlenecked by the slow, manual creation of software to support computational experiments\cite{hannay2009how}. To address this, we present Empirical Research Assistance (ERA), an AI system that creates expert-level scientific software whose goal is to maximize a quality metric. The system uses a Large Language Model (LLM) and Tree Search (TS)\cite{silver2016mastering} to systematically improve the quality metric and intelligently navigate the large space of possible solutions. ERA achieves expert-level results when it explores and integrates complex research ideas from external sources. The effectiveness of tree search is demonstrated across a diverse range of tasks. In bioinformatics, ERA discovered 40 novel methods for single-cell data analysis that outperformed the top human-developed methods on a public leaderboard. In epidemiology, ERA generated 14 models that outperformed the CDC ensemble and all other individual models for forecasting COVID-19 hospitalizations. ERA also produced expert-level software for geospatial analysis, neural activity prediction in zebrafish, and numerical solution of integrals, and a novel rule-based construction for time series forecasting. By devising and implementing novel solutions to diverse tasks, ERA represents a significant step towards accelerating scientific progress.

Some Question Answering (QA) systems rely on knowledge bases (KBs) to provide accurate answers. Entity Linking (EL) plays a critical role in linking natural language mentions to KB entries. However, most existing EL methods are designed for long contexts and do not perform well on short, ambiguous user questions in QA tasks. We propose an entity linking agent for QA, based on a Large Language Model that simulates human cognitive workflows. The agent actively identifies entity mentions, retrieves candidate entities, and makes decision. To verify the effectiveness of our agent, we conduct two experiments: tool-based entity linking and QA task evaluation. The results confirm the robustness and effectiveness of our agent.

Wireless indoor localization using predictive models with received signal strength information (RSSI) requires proper calibration for reliable position estimates. One remedy is to employ synthetic labels produced by a (generally different) predictive model. But fine-tuning an additional predictor, as well as estimating residual bias of the synthetic labels, demands additional data, aggravating calibration data scarcity in wireless environments. This letter proposes an approach that efficiently uses limited calibration data to simultaneously fine-tune a predictor and estimate the bias of synthetic labels, yielding prediction sets with rigorous coverage guarantees. Experiments on a fingerprinting dataset validate the effectiveness of the proposed method.

Recent years have seen an explosion of diverse general purpose pre-training methodologies for computer vision. However, the impact that these pre-training methodologies have on person identification tasks (re-id) remains under-explored. We show that under equated domain adaptation pipelines, there is dramatic variance in person identification outcomes using different starting models (architectures and pre-trained weights). We show that a range of intuitive explanations for differing downstream performance on a range of re-id tests are insufficient and propose that pre-trained weights serve as a strong prior to the weights learned during domain adaptation. This framework allows for domain adapted solutions to be viewed as a maximum probability point estimate of the Gibbs posterior with the pre-trained weights acting as a prior. Under this framework, we show that large, pre-trained foundation models with simple domain adaptation achieve SOTA solutions on a range of re-id datasets (Market, PRCC, DeepChange, BTS) with solutions that are very close in the parameter space to the starting parameters. Moreover, we perform ablations on these solutions and show that they can be reached with small transfer sets and with varying transfer datasets but are sensitive to choice of optimizer, weight-decay, and loss function. Ultimately, we propose that the simple approach of direct fine-tuning using large vision foundation models (CLIP, Dino, EVA, AIM, etc.) needs to serve as an important baseline for future work in re-id.

Extracting structured information from scientific literature is critical for accelerating discovery, yet Large Language Models (LLMs) often struggle in specialized domains that require expert knowledge and generalize poorly across tasks. We introduce \textsc{StructSense}, a modular, task-agnostic, open-source framework that integrates ontology-guided symbolic knowledge, agentic self-evaluative refinement, and human-in-the-loop validation for robust domain-aware extraction. We evaluate \textsc{StructSense} on three tasks of increasing semantic complexity: schema-based extraction of assessment instruments (91--100\% accuracy), metadata and resource extraction from scientific papers (86--93\% overall), and named entity recognition (NER) from neuroscience literature (58--75\% label accuracy across 8,882 entities). On two biomedical NER benchmarks (NCBI Disease and S800 Species), the system achieves $\geq$90\% relaxed recall and 62.5--85.8\% strict recall while extracting 1,000--3,600 additional entities beyond gold annotations. The local concept mapping service achieves Hits@1 of 62--82\% under strict matching and 68--86\% under semantic matching. These results across three domains demonstrate that \textsc{StructSense} generalizes across tasks while maintaining source grounding and provenance transparency.

A recent series of works has shown that initialization-free BA can be achieved using pseudo Object Space Error (pOSE) as a surrogate objective. The initial reconstruction-step optimizes an objective where all terms are projectively invariant and it cannot incorporate knowledge of the camera calibration. As a result, the solution is only determined up to a projective transformation of the scene and the process requires more data for successful reconstruction. In contrast, we present a method that is able to use the known camera calibration thereby producing near metric solutions, that is, reconstructions that are accurate up to a similarity transformation. To achieve this we introduce pairwise relative rotation estimates that carry information about camera calibration. These are only invariant to similarity transformations, thus encouraging solutions that preserve metric features of the real scene. Our method can be seen as integrating rotation averaging into the pOSE framework striving towards initialization-free calibrated SfM. Our experimental evaluation shows that we are able to reliably optimize our objective, achieving convergence to the global minimum with high probability from random starting solutions, resulting in accurate near metric reconstructions.

Large Language Models (LLMs) increasingly act as function-call agents that invoke external tools to tackle tasks beyond their static knowledge. However, they typically invoke tools one at a time without a global view of task structure. As tools often depend on one another, this leads to error accumulation and poor scalability, particularly when scaling to hundreds or thousands of tools. To address these limitations, we propose NaviAgent, an explicit bilevel architecture that decouples task planning from tool execution through graph-based modeling of tool relations. At the planning level, the LLM-based agent decides whether to respond directly, clarify intent, or retrieve and execute a toolchain independent of inter-tool complexity. At the execution level, a Tool World Navigation Model (TWNM) encodes structural and behavioral relations among tools, steering the agent to compose scalable and robust invocation sequences. Incorporating feedback from real tool interactions, NaviAgent achieves closed-loop alignment between planning and execution, enabling adaptive navigation in large-scale tool ecosystems. Evaluations on API-Bank and ToolBench show consistent improvements in task success rate (TSR), with TWNM yielding an average gain of 13.1 points on complex tasks. Further tests on 50 real APIs across 7 domains show consistent gains of 4.3--12.0 points, with fewer steps and latency, demonstrating robust generalization under real-world dynamics.

Training large language models (LLMs) relies on adaptive optimizers such as Adam, which introduce extra operations and require significantly more memory to maintain first- and second-order moments than SGD. While recent works such as GaLore, Fira and APOLLO have proposed state-compressed memory-efficient variants, a fundamental question remains: What are the minimum modifications to plain SGD needed to match state-of-the-art pretraining performance? We systematically investigate this question using a bottom-up approach, and identify two simple yet highly (memory- and compute-) efficient techniques: (1) column-wise gradient normalization (normalizing the gradient along the output dimension), that boosts SGD performance without momentum; and (2) applying first-order momentum only to the output layer, where gradient variance is highest. Combining these two techniques lead to SCALE (Stochastic Column-normAlized Last-layer momEntum), a simple optimizer for memory efficient pretraining. Across multiple models (60M-1B), SCALE matches or exceeds the performance of Adam while using only 35-45% of the total memory. It also consistently outperforms memory-efficient optimizers such as GaLore, Fira and APOLLO, making it a strong candidate for large-scale pretraining under memory constraints. For LLaMA 7B, SCALE outperforms the state-of-the-art memory-efficient methods APOLLO and Muon in both perplexity and memory consumption.

Preference-based Reinforcement Learning (PbRL) methods provide a solution to avoid reward engineering by learning reward models based on human preferences. However, poor feedback- and sample- efficiency still remain the problems that hinder the application of PbRL. In this paper, we present a novel efficient query selection and preference-guided exploration method, called SENIOR, which could select the meaningful and easy-to-comparison behavior segment pairs to improve human feedback-efficiency and accelerate policy learning with the designed preference-guided intrinsic rewards. Our key idea is twofold: (1) We designed a Motion-Distinction-based Selection scheme (MDS). It selects segment pairs with apparent motion and different directions through kernel density estimation of states, which is more task-related and easy for human preference labeling; (2) We proposed a novel preference-guided exploration method (PGE). It encourages the exploration towards the states with high preference and low visits and continuously guides the agent achieving the valuable samples. The synergy between the two mechanisms could significantly accelerate the progress of reward and policy learning. Our experiments show that SENIOR outperforms other five existing methods in both human feedback-efficiency and policy convergence speed on six complex robot manipulation tasks from simulation and four real-worlds. Videos can be found on our project website: https://2025senior.github.io/

We introduce PHYSICS, a comprehensive benchmark for university-level physics problem solving. It contains 1297 expert-annotated problems covering six core areas: classical mechanics, quantum mechanics, thermodynamics and statistical mechanics, electromagnetism, atomic physics, and optics. Each problem requires advanced physics knowledge and mathematical reasoning. We develop a robust automated evaluation system for precise and reliable validation. Our evaluation of leading foundation models reveals substantial limitations. Even the most advanced model, o3-mini, achieves only 59.9% accuracy, highlighting significant challenges in solving high-level scientific problems. Through comprehensive error analysis, exploration of diverse prompting strategies, and Retrieval-Augmented Generation (RAG)-based knowledge augmentation, we identify key areas for improvement, laying the foundation for future advancements.

Light field cameras capture multi-view observations within a single exposure. However, existing studies are typically tailored to specific LF representations, leaving the field without a unified learning framework. To bridge this gap, we present LFX, the first unified framework for LF perception. LFX establishes a representation-invariant feature modulation space, enabling it to adapt to heterogeneous LF representations and diverse perception tasks. Specifically, we propose Field-of-Parallax Angular Subspace Modeling (FoP-ASM), which assigns an independent angular marker to each auxiliary view, enabling view-wise independent modeling. Meanwhile, shared manifold subspace constraints and regularization losses enforce globally consistent semantic modulation across views. Extensive evaluations across three LF benchmarks show that LFX achieves state-of-the-art results across distinct LF representations, outperforming representation-specific methods by up to 12% and 20% with 0.029/0.027 MAE for salient object detection, and achieving 84.37 mIoU for semantic segmentation. The source code will be made publicly available at https://github.com/FeiT-FeiTeng/LFX.