Muhammad Talha, Xiaofeng Liu, Yuan Yang, Pang-Ning Tan, A. Pouyan Nejadhashemi
Reliable streamflow prediction across time and in ungauged basins requires representations that capture hydrologically relevant basin characteristics. Existing methods rely on either fixed handcrafted attributes or general-purpose Earth observation embeddings not tailored to basin structure. We introduce \textbf{Basin2Vec}, a self-supervised framework that learns compact basin representations from hydroclimatic and physiographic data without observed streamflow. Basin2Vec organizes representation learning around basin scale and relationships across space and time. Across multiple forecasting architectures, it consistently outperforms predefined attributes and general-purpose geospatial embeddings in temporal generalization and zero-shot spatial transfer. Basin2Vec provides average relative gains of up to 8.3% temporally and achieves 3.5x the median predictive performance of the strongest baselines on unseen basins. These results demonstrate the effectiveness of basin-centric pretraining for transferable hydrologic prediction. Code is available at https://github.com/Muhammad-Talha-MT/basin2vec
Muon has recently shown strong empirical performance in training Transformer-based large language models (LLMs), yet its optimization dynamics remain poorly understood from a theoretical perspective. In this paper, we analyze the training dynamics of Muon for next-token prediction (NTP) in Transformers. The training procedure consists of two phases: the preprocessing phase learns the output-value component, while the main phase learns the key-query component. For the preprocessing phase, we characterize the optimization trajectory exactly and show that the normalized spectral margin is optimal after each update. For the main phase, we establish nonasymptotic convergence rates for the training loss and the normalized margin gap under softmax attention. These results suggest that Muon's update rule can help models efficiently learn which tokens matter for prediction, offering a possible explanation for its strong performance in large language model training. The learned weight structure also yields a generalization guarantee for unseen contexts. Furthermore, we validate our theoretical insights with experiments on next-token prediction tasks.
Yang Wu, Qiang Meng, Zhaojiang Liu, Youquan Liu, Jian Yang, Jin Xie
Current end-to-end autonomous driving models are fundamentally constrained by the behavioral cloning ceiling of imitation learning. While reinforcement learning offers a path to smarter autonomy, it demands two missing pieces of infrastructure: (1) a cognitive foundation that understands traffic semantics and driving intent, and (2) a foresighted physical environment that can anticipate the consequences of candidate actions. To this end, we propose CoPhy, a CognitivePhysical reinforcement learning framework for autonomous driving. To distill to think, we distill VLM knowledge into the BEV encoder and then discard the VLM entirely, retaining cognitive ability at zero inference cost while releasing the cognitive channel as a pluggable interface for optional human language commands. To foresee to act, we build an auto-regressive BEV world model that explicitly predicts future semantic maps conditioned on candidate actions, serving as an interpretable physical sandbox from which safety metrics are directly derived. Built upon this dual infrastructure, we optimize the driving policy via GRPO with a novel dual-reward mechanism: a physical reward derived from BEV rollouts enforces hard safety constraints, while a cognitive reward from a language-aligned scorer ensures intent compliance. Extensive experiments demonstrate that CoPhy not only achieves state-of-the-art results on NAVSIM v1 and v2 benchmarks, but also enables safer driving via cognitively informed scene compliance and flexible intent control through user-defined language instructions. Project page: https://wuyang98.github.io/CoPhy/.
Aggregate scores for continual learning algorithms rest on a distribution over task orders that is rarely stated. We formulate algorithm comparison as an expectation over a declared order distribution, which separates training variation from order variation, and apply it to motion forecasting with four road-layout tasks from the INTERACTION dataset, all 24 task orders, four algorithms, and three training replicates (288 runs). Task order accounts for 65\% to 91\% of the variation in current-task error and backward transfer, and 90\% to 99\% of this order effect comes from the identity of the starting task: the smallest task cannot be learned from a cold start. Beyond the first stage, the order effect is comparable to training variability. In most orders, the leading algorithm in current-task error and backward transfer is within one replicate standard deviation of the runner-up. Backward transfer depends on the order mainly through its reference term, which weights the first task most heavily, so relative improvement and final error identify different leading algorithms for displacement error. A transition-direction descriptor computed from task statistics tracks the position of the smallest task, and weighting orders by it leaves the algorithm ranking unchanged. These results call for benchmarks that declare their order distribution and report order-resolved, starting-task-resolved, warm-start, and final errors alongside aggregate scores. The code and data are available \href{https://anonymous.4open.science/r/task-order-cl-4f7b/}{here}.
Zhaoyang Wang, Qianhui Wu, Boyu Gou, Demi Wang, Jiacheng Zhang, Peichun Hua, Qize Li, Shi Qiu 等 57 位作者
GUI web agents require realistic and interactive environments that remain stable during both evaluation and training. Live website evaluation suffers from changing content, access restrictions, and limited coverage of authenticated workflows, while existing synthetic environments for training often lack functional depth and visual fidelity. We propose WebHarbor, a suite of 153 executable and dockerized website environments that faithfully reconstruct the visual interfaces, content, and functional workflows of real-world web services under deterministic control. We build WebHarbor through task-driven human-agent collaboration, where environments evolve as new tasks expand their functional scope, coding agents implement the required changes, and human contributors audit visual fidelity, functional correctness, and task validity. The resulting suite comprises 3,518 tasks spanning multi-step and post-authentication workflows, where existing works often fail to cover. Beyond reliable evaluation, WebHarbor provides high-fidelity synthetic web environments for cost-efficient reinforcement learning (RL). We demonstrate that a Qwen 4B-level GUI agent trained with RL in WebHarbor generalizes to unseen tasks and performs competitively with leading open-source GUI web agents across four benchmarks. The code and data are publicly available at https://github.com/aiming-lab/WebHarbor.
Jinyuan Chang, Chenguang Duan, Yuling Jiao, Ruoxuan Li, Yi Xu, Jerry Zhijian Yang
Inference-time alignment for diffusion models aims to adapt a pre-trained reference diffusion model toward a target distribution without retraining the reference score network, thereby preserving the generative capacity of the reference model while enforcing desired properties at inference time. A central mechanism for achieving such alignment is guidance, which modifies the sampling dynamics through an additional drift term. In this work, we introduce variationally stable Doob's matching, a novel framework for provable guidance estimation grounded in Doob's $h$-transform. Our approach formulates guidance as the gradient of the logarithm of an underlying Doob's $h$-function and employs gradient-regularized regression to simultaneously estimate both the $h$-function and its gradient, resulting in a consistent estimator of the guidance. Theoretically, we establish non-asymptotic convergence rates for the estimated guidance. Moreover, we analyze the resulting controllable diffusion processes and prove non-asymptotic convergence guarantees for the generated distributions in the Wasserstein-2 distance. Finally, we show that variationally stable guidance estimators are adaptive to unknown low dimensionality, effectively mitigating the curse of dimensionality under low-dimensional linear subspace assumptions. Numerical experiments demonstrate the effectiveness of our proposed method and show the advantages compared to the baselines.
Félix Chavelli, Arik Ermshaus, Fan Yang, Patrick Schäfer, John Paparrizos, Paul Boniol
Time-series segmentation, studied as either Change Point Detection (CPD) or State Detection (SD), underpins a broad range of monitoring and diagnostic applications. However, the field suffers from a persistent divide between the statistical community (focusing on CPD) and the machine learning and data mining communities (focusing on SD). Consequently, the literature remains fragmented: evaluations typically cover narrow methodological families, datasets are often small or homogeneous, and CPD and SD are rarely evaluated together. We address these gaps with TSB-SEG, a unified benchmark comprising 592 univariate and multivariate time series from 8 heterogeneous datasets and surveying 26 algorithms that span five decades of progress. To ensure a fair comparison, we introduce a two-pool design: a scalable main pool of tractable detectors used for the aggregate ranking, and an extended-scope pool for computationally intensive methods. All methods are unified under tsseg, an open-source library to ensure full reproducibility. Our systematic evaluation yields three insights: (a) no single detector dominates across domains, and per-dataset hyperparameter tuning remains decisive for performance, (b) CPD accuracy is a useful, though imperfect, proxy for SD accuracy, and (c) data characteristics should guide method selection. We provide actionable guidance based on data properties, error tolerances, and tuning budgets, identify impactful hyperparameters, and characterize algorithm robustness to these choices.
Mingchuan Ma, Dongqi Cai, Daqian Yang, Zhenyan Lu, Chen Chen, Teresa Lv, Yan Gao, Shangguang Wang 等 11 位作者
LLM pretraining is increasingly compute-intensive, while substantial inference serving compute remains idle during off-peak periods. However, such serving hardware is difficult to use for conventional first-order training due to its memory, networking and other system requirements. Zeroth-order (ZO) optimization offers a natural interface to this forward-only compute resource, but its gradient estimates break down in pretraining. Unlike fine-tuning, which can confine updates to a small parameter subspace, pretraining must update every parameter, and the alignment of a ZO gradient estimate with the true gradient shrinks as the number of updated parameters grows. We make a different observation: noisy ZO estimates can still provide useful curvature information, even when they are too inaccurate to serve as gradients. Based on this insight, we propose ZO-Curv, which leverages idle forward-only devices to estimate curvature and uses it to adapt and refine the update used during first-order training, thereby accelerating the process. We provide convergence guarantees that characterize the effect of finite-query curvature estimation. Experiments on LLMs from 125M to 1.5B parameters show that ZO-Curv consistently accelerates pretraining. At 770M parameters, ZO-Curv reaches AdamW's 20K-step validation loss in 3.9 hours instead of 7.5 hours, while achieving a lower final loss of 2.70 versus 2.84 under the same step budget. These results show that idle model serving compute can accelerate LLM pretraining when zeroth-order queries are used to estimate curvature rather than gradients directly.
Yuling Jiao, Lican Kang, Jerry Zhijian Yang, Jincheng Ying
Diffusion policies represent multimodal action distributions, but a $Q$-weighted update does not specify how to sample from the resulting target distribution. We propose Schrödinger--Föllmer Actor--Critic (SFAC), an offline-to-online reinforcement learning (RL) method for Kullback--Leibler (KL)-regularized policy improvement through conditional diffusion. A minimax Bellman critic estimates the action-value function, which defines an exponentially tilted target policy. A Doob $h$-transform expresses this update as a correction to the reference diffusion drift. The target drift is a weighted conditional mean of Brownian-bridge labels. We fit it by weighted regression using endpoints from the current policy and pointwise action-value evaluations. In the small-update regime, the KL-regularized update follows the natural policy-gradient direction, and the Doob correction represents the same local change in the space of diffusion drifts. Under suitable conditions, we derive finite-sample bounds that separate the errors from critic estimation, weighted neural drift regression, and diffusion discretization through a reverse-KL actor bound. The average-performance bound depends explicitly on the comparison coefficients over the policy iterations. Synthetic experiments assess approximation to prescribed advantage-tilted targets and sensitivity to the diffusion-step count. Across six offline-to-online continuous-control tasks, the $Q$-guided bridge-regression implementation achieves higher final-checkpoint returns than its BC initialization after $500K$ online interactions.
Read, Allocate, Route: Role-Coupled Tokenization for Pixel-Space Generative Transformers
Jiongfan Yang, Weichen Fan, Haiwen Diao, Zhongang Cai, Ziwei Liu
AI 速览：SpecSlot通过将每个空间锚点拆成多个按DCT频带分工的slot，让全局注意力只处理低频slot、高频细节在局部交互，从而在不加长全局序列的前提下提升像素空间生成式Transformer的细节建模能力。
Pixel-space generative Transformers typically use one token per image patch, so capturing finer detail requires more tokens and makes global attention increasingly expensive. We observe that finer detail does not necessarily require more global communication: low-frequency structure benefits from global exchange, while high-frequency detail can be modeled locally. Based on this insight, we introduce SpecSlot, a pixel-space rectified-flow Transformer that decouples output detail from the length of the global attention sequence. SpecSlot expands each spatial anchor into multiple slots, each assigned a disjoint band of DCT coefficients; this separates anchor density from spectral subdivision. Only the lowest-frequency slot from each anchor participates in global attention, while all slots interact locally. This allows the model to increase spectral detail without enlarging the global sequence. A spatial head further adds a local correction to the spectral prediction, while the generative state, target, and loss remain in pixel space. On ImageNet-256, SpecSlot achieves a comparable final FID to a FLOPs-matched JiT control while converging faster. On ImageNet-512, it improves FID by 2.84 over the control with 3.9% fewer FLOPs, and comes within only 0.69 FID of JiT-B/16 at 38.2% lower cost. On text-to-image generation, SpecSlot surpasses JiT-B/8 on DPG-Bench by 1.06 at 40.6% lower forward cost.
Discounted regret measures the performance of an online learner by assigning exponentially decaying weights to past losses, reflecting the greater importance of recent losses in non-stationary online learning. While previous work obtains uniform discounted regret bounds over a continuous range of discount factors using a specialized meta-algorithm, its results are limited to convex functions. In this paper, we demonstrate that strongly adaptive regret bounds can directly imply uniform discounted regret guarantees for convex, strongly convex, and exp-concave loss functions. Our key observation is that discounted regret against a fixed comparator admits an exact decomposition into a convex combination of interval regrets. For convex functions, this reduction yields an $O(\sqrt{\log T/(1-λ)})$ discounted regret bound for all discount factors $λ\in(0,1)$, matching the previous rate over a broader range of discount factors. For strongly convex and exp-concave functions, we provide regret bounds of $O\left(\log T\left(1+\log\frac{1}{1-λ}\right)\right)$.
In this paper, we investigate decentralized online convex optimization with compressed communication, where $n$ learners over a network jointly minimize a sequence of global loss functions based only on their local losses and compressed messages from their neighbors. The best known regret bounds for this problem are $O(ω^{-1/2}nρ^{-1}\sqrt{T\ln n})$ and $O(ω^{-1}nρ^{-2}\ln n\ln T)$ for convex and strongly convex loss functions, respectively, where $ρ\in(0,1)$ is the spectral gap of the communication matrix and $ω\in(0,1]$ is the compression quality factor. Compared with the corresponding lower bounds of $Ω(ω^{-1/2}nρ^{-1/4}\sqrt{T})$ and $Ω(ω^{-1}nρ^{-1/2}\ln T)$, these upper bounds exhibit polynomial gaps in their dependence on $ρ^{-1}$. To close these gaps, we first develop Ascent-gossip, which combines an online compression strategy with a Chebyshev recursion and attains a faster consensus rate than existing compressed gossip methods. Furthermore, by adopting Ascent-gossip, we propose an online algorithm that achieves optimal regret bounds of $\widetilde O(ω^{-1/2}nρ^{-1/4}\sqrt{T})$ and $\widetilde O(ω^{-1}nρ^{-1/2}\ln T)$ for convex and strongly convex loss functions, respectively. Up to logarithmic factors in $n$ and $ρ^{-1}$, these results match the lower bounds in terms of $T$, $n$, $ρ$, and $ω$.
Offline reinforcement learning seeks to learn decision policies from previously collected data, but deployment environments may differ from the domains in which the data were collected, and logged variables may omit confounders that affected both historical actions and outcomes. We study offline policy learning under the joint challenges of domain discrepancy and unmeasured confounding, where multiple heterogeneous source datasets provide complementary information about a target domain. We develop a causally adjusted transportability framework combining transportability theory with instrumental variables, negative controls, and proxies to identify target-domain causal quantities. For confounded contextual bandits, identification reduces to systems of bridge-function integral equations, enabling minimax estimation from finite samples. Confidence sets for the identified quantities are then used for pessimistic policy optimization, yielding CAT-CCB. We extend the framework to episodic confounded Markov decision processes, where complementary sources identify target reward and transition mechanisms under structural and completeness conditions. Under Gaussian dynamics and linear function approximation, this yields CAT-CMDP. We establish high-probability finite-sample suboptimality guarantees that quantify the effects of approximation, estimation, source-to-target shift, offline coverage, and sample size. For bounded finite-dimensional linear classes, the contextual-bandit guarantee recovers the standard $\widetilde{\mathcal O}(N^{-1/2})$ rate. Numerical experiments support the effectiveness of the framework for learning target-domain policies from heterogeneous confounded data. Together, the results provide a principled approach when no single source dataset suffices for causal identification and policy optimization.
Sihan Yang, Weichen Bai, Ruichuan An, Yiming Cheng, Zhiyuan Feng, Tianyi Jiang, Renye Yan, Ning Ding 等 12 位作者
As executable environments for language agents scale, generating and verifying their rollouts becomes increasingly expensive. Under finite rollout budgets, this raises an important question: can data management at the environment level make agentic reinforcement learning (RL) more efficient by organizing, evaluating, and scheduling task instances together with the environments that determine their tools, states, transitions, and verification? To address this question, we introduce EnvSched (Environment-aware Scheduling), an online scheduler that learns from rollout history. EnvSched assigns scheduling value through two complementary modeling choices: exploitation prioritizes tasks likely to provide informative reward differences, while coverage spreads each batch across underobserved environments and interaction patterns. EnvSched organizes environment--task pairs as nodes in a dual-relation graph, with within-environment edges capturing task semantics and shared tool, state, and verifier structure, and cross-environment edges capturing abstract execution and reward topology. By propagating historical rollout statistics along both edge types, it predicts within-task reward variance. These predictions guide a scheduler that allocates each batch to tasks with high predicted reward variance and underobserved environments and tasks. Environment quotas and a similarity penalty further limit concentration and redundancy, yielding informative and diverse batches without additional pilot rollouts. In RL on Qwen3-8B, EnvSched achieves an average score of 48.81 across BFCL-v3, tau-Bench, and ACEBench-Agent after 40 training steps, exceeding the highest score attained by random task selection after 120 steps. It also improves downstream performance across model sizes at the same training budget.
Hanqing Wang, Zhaotian Cai, Ruichuan An, Bohan Zeng, Tao Zhang, Yilin Chen, Xu Jiang, Kaiquan Chen 等 15 位作者
Vision--Language--Action (VLA) models generate robot actions from visual observations and language instructions, but their understanding of physical dynamics remains limited. Video world models can capture scene dynamics, yet generating future videos introduces high latency into the control loop. We introduce PDRVLA, whose lightweight Physical Dynamics Module extracts physical tokens from vision--language features to condition action generation without online video generation. The module learns through direct prediction of future RGB, depth, and optical-flow features, with auxiliary supervision through a frozen pretrained video world model. Training relies entirely on publicly available data, expanded through human-to-robot synthesis, subtask annotation, and embodied chain-of-thought (ECoT) annotation. Across four simulation benchmarks, PDRVLA achieves success rates of 99.3% on LIBERO, 89.9% on LIBERO-Plus, 93.6% on RoboTwin~2.0, and 81.5% on SimplerEnv. Ablations show improved future prediction with world-model supervision and evaluate how subtask conditioning, future object-pose supervision, and ECoT training affect action success.
Scientific reasoning in specialized engineering fields depends on knowledge scattered across the literature, often proprietary and absent from frontier LLM training data. In such domains, frontier models tend to give generic answers. Knowledge graphs (KGs) can ground LLM reasoning. Because a graph is too large to provide in full, typical LLM+KG systems preselect a portion, deciding relevance before reasoning begins. Here we train a compact model, Prospector, to reason over a domain knowledge graph by traversing it step by step, seeking evidence in light of what it has already found. Prospector calls a small set of tools one at a time to progressively build its context, and answers with a chain of evidence that can be checked edge by edge against what its tools retrieved from the graph. On a knowledge graph built from the semiconductor-manufacturing literature, Prospector approaches the strongest off-the-shelf models given the same graph and tools, and it far exceeds off-the-shelf models answering closed-book or from a pre-selected portion of the graph. Although trained only on the semiconductor graph, Prospector transfers to unseen knowledge graphs from three other distinct fields: bio-inspired materials, per- and polyfluoroalkyl substances (PFAS) and materials properties. Served on a single GPU, Prospector answers faster and at a fraction of the cost of a hosted service, while keeping proprietary data on the owner's hardware. We find that the policy's latent state is indicative of its next move on both seen and unseen graphs, and modulating the latent state can steer the policy from one behavior mode to another. These results suggest a practical LLM + KG paradigm for scientific and engineering reasoning over private data, in which a compact model trained on a graph from published literature transfers its traversal to unseen graphs that stay with their owners.
Many everyday tasks require controlling contact force as well as motion. On a humanoid, this force must often be coordinated across both hands and the whole body, and cannot be specified by kinematic trajectories alone. Therefore, we present FLORE, a framework for learning force-reactive humanoid loco-manipulation. We collect robot-free demonstrations using two force-sensing UMI-style handheld grippers for hand motion and contact wrenches, and wearable trackers for body motion. A force-reactive diffusion policy takes vision, proprioception, and measured wrenches as input, then predicts actions and per-hand force commands. A hybrid compliance whole-body controller with force tracking executes these commands while maintaining balance: an integral of the force error measured by real force/torque (F/T) sensors lets each hand track force accurately along the commanded direction and remain compliant along the others. In simulation, the controller tracks per-hand force commands of varied magnitude and direction, with little loss in whole-body motion accuracy. Across three contact-rich tasks on the robot, the force-reactive policy succeeds in 80% of rollouts, compared with 26.7% for a position-only policy trained on the same demonstrations and executed by the same controller. For more details, please visit our website: https://louis-leee.github.io/FLoRe/.
Xuyuan Xiong, Siyu Chen, Zihan Shao, Yue Su, Weiyang Jin, Ping Luo, Zhuoran Yang, Xihui Liu
Looped transformers reuse a shared block of layers to increase computational depth without increasing the number of trainable parameters. They are trained with backpropagation through time (BPTT). To save training compute, BPTT is often truncated to the last few loops, but truncation yields biased gradients that may weaken the benefit of looping. In this work, we study the value of truncation in looped transformer training in a systematic way: we jointly vary model size, training tokens, and forward and backward loop counts across nearly 600 training runs and fit a unified scaling law that captures the effects of looping and truncation through effective parameters. Despite the training-compute savings of truncation, our fitted law favors full BPTT when forward and backward loop counts are jointly optimized. In matched-compute runs, full BPTT through fewer loops consistently outperforms truncated settings of the same compute. The law further shows that at fixed model size, the compute-optimal loop count and training-token budget both grow with compute. When model size is also free, the law identifies two fully trained loops as the best looped setting. Experimentally, this setting achieves lower validation loss than the compute-optimal non-looped model with about 27% fewer non-embedding parameters and only 8% more training compute.