Inference-Time Budget Control for LLM Search Agents

LLM search agents increasingly rely on tools at inference time, but their trajectories are often constrained by hard limits on both tool calls and generated tokens. Under such dual budgets, better answers require not only stronger models, but also explicit control over which search action should receive the next budget unit and when the accumulated evidence is sufficient to commit a final answer. We study this problem in multi-hop question answering (QA) and formulate it as two-stage inference-time budget control. At search time, our controller assigns each feasible action a task-level Value-of-Information (VOI) score, defined as an operational estimate of marginal task value per unit budget under the current search state and remaining dual budget, and uses this score to choose among retrieval, decomposition, and answer commitment. After search, a selective evidence-grounded finalizer compares the trajectory answer with a refined candidate and rewrites only when the residual error appears to be a low-risk answer-form error. Across four multi-hop QA benchmarks, three LLM backbones, and four budget levels, the method yields positive aggregate gains over four audited baselines under the same hard dual-budget protocol. Ablations show that search-time budget control, especially budget-dependent penalty, provides the main performance gain, while answer-time control helps mainly when the retrieval path is already adequate. These results suggest that inference-time budget control for LLM search agents should govern both how budget is spent during search and how the final answer is committed.

Divide and Conquer: Decoupled Representation Alignment for Multimodal World Models

Emerging multi-modal world models attempt to jointly generate videos across diverse modalities (e.g., RGB, depth, and mask), yet they fail to fully exploit the rich priors of existing foundation models. We propose $M^2$-REPA, the first representation alignment method tailored for multi-modal video generation. Our key insight is that foundation models trained on different modality spaces naturally capture distinct domain-specific priors, acting as complementary "experts." Specifically, we first decouple modality-specific features from the diffusion model's intermediate representations, then align each with its corresponding expert foundation model. To this end, we design two synergistic objectives: a multi-modal representation alignment loss that enforces feature-to-expert matching, and a modality-specific decoupling regularization that encourages complementarity across different modalities. This design enables joint optimization, fully exploiting priors from multiple foundation models. Extensive experiments demonstrate that our method significantly outperforms baselines in visual quality and long-term consistency.

AgenticQwen: Training Small Agentic Language Models with Dual Data Flywheels for Industrial-Scale Tool Use

Modern industrial applications increasingly demand language models that act as agents, capable of multi-step reasoning and tool use in real-world settings. These tasks are typically performed under strict cost and latency constraints, making small agentic models highly desirable. In this paper, we introduce the AgenticQwen family of models, trained via multi-round reinforcement learning (RL) on synthetic data and a limited amount of open-source data. Our training framework combines reasoning RL and agentic RL with dual data flywheels that automatically generate increasingly challenging tasks. The reasoning flywheel increases task difficulty by learning from errors, while the agentic flywheel expands linear workflows into multi-branch behavior trees that better reflect the decision complexity of real-world applications. We validate AgenticQwen on public benchmarks and in an industrial agent system. The models achieve strong performance on multiple agentic benchmarks, and in our industrial agent system, close the gap with much larger models on search and data analysis tasks. Model checkpoints and part of the synthetic data: https://huggingface.co/collections/alibaba-pai/agenticqwen. Data synthesis and RL training code: https://github.com/haruhi-sudo/data_synth_and_rl. The data synthesis pipeline is also integrated into EasyDistill: https://github.com/modelscope/easydistill.

Lifecycle-Aware Federated Continual Learning in Mobile Autonomous Systems

Federated continual learning (FCL) allows distributed autonomous fleets to adapt collaboratively to evolving terrain types across extended mission lifecycles. However, current approaches face several key challenges: 1) they use uniform protection strategies that do not account for the varying sensitivities to forgetting on different network layers; 2) they focus primarily on preventing forgetting during training, without addressing the long-term effects of cumulative drift; and 3) they often depend on idealized simulations that fail to capture the real-world heterogeneity present in distributed fleets. In this paper, we propose a lifecycle-aware dual-timescale FCL framework that incorporates training-time (pre-forgetting) prevention and (post-forgetting) recovery. Under this framework, we design a layer-selective rehearsal strategy that mitigates immediate forgetting during local training, and a rapid knowledge recovery strategy that restores degraded models after long-term cumulative drift. We present a theoretical analysis that characterizes heterogeneous forgetting dynamics and establishes the inevitability of long-term degradation. Our experimental results show that this framework achieves up to 8.3\% mIoU improvement over the strongest federated baseline and up to 31.7\% over conventional fine-tuning. We also deploy the FCL framework on a real-world rover testbed to assess system-level robustness under realistic constraints; the testing results further confirm the effectiveness of our FCL design.

Weakly-Supervised Referring Video Object Segmentation through Text Supervision

Referring video object segmentation (RVOS) aims to segment the target instance in a video, referred by a text expression. Conventional approaches are mostly supervised learning, requiring expensive pixel-level mask annotations. To tackle it, weakly-supervised RVOS has recently been proposed to replace mask annotations with bounding boxes or points, which are however still costly and labor-intensive. In this paper, we design a novel weakly-supervised RVOS method, namely WSRVOS, to train the model with only text expressions. Given an input video and the referring expression, we first design a contrastive referring expression augmentation scheme that leverages the captioning capabilities of a multimodal large language model to generate both positive and negative expressions. We extract visual and linguistic features from the input video and generated expressions, then perform bi-directional vision-language feature selection and interaction to enable fine-grained multimodal alignment. Next, we propose an instance-aware expression classification scheme to optimize the model in distinguishing positive from negative expressions. Also, we introduce a positive-prediction fusion strategy to generate high-quality pseudo-masks, which serve as additional supervision to the model. Last, we design a temporal segment ranking constraint such that the overlaps between mask predictions of temporally neighboring frames are required to conform to specific orders. Extensive experiments on four publicly available RVOS datasets, including A2D Sentences, J-HMDB Sentences, Ref-YouTube-VOS, and Ref-DAVIS17, demonstrate the superiority of our method. Code is available at https://github.com/viscom-tongji/WSRVOS.

Bridging What the Model Thinks and How It Speaks: Self-Aware Speech Language Models for Expressive Speech Generation

Speech Language Models (SLMs) exhibit strong semantic understanding, yet their generated speech often sounds flat and fails to convey expressive intent, undermining user engagement. We term this mismatch the semantic understanding-acoustic realization gap. We attribute this gap to two key deficiencies: (1) intent transmission failure, where SLMs fail to provide the stable utterance-level intent needed for expressive delivery; and (2) realization-unaware training, where no feedback signal verifies whether acoustic outputs faithfully reflect intended expression. To address these issues, we propose SA-SLM (Self-Aware Speech Language Model), built on the principle that the model should be aware of what it thinks during generation and how it speaks during training. SA-SLM addresses this gap through two core contributions: (1) Intent-Aware Bridging, which uses a Variational Information Bottleneck (VIB) objective to translate the model's internal semantics into temporally smooth expressive intent, making speech generation aware of what the model intends to express; and (2) Realization-Aware Alignment, which repurposes the model as its own critic to verify and align acoustic realization with intended expressive intent via rubric-based feedback. Trained on only 800 hours of expressive speech data, our 3B parameter SA-SLM surpasses all open-source baselines and comes within 0.08 points of GPT-4o-Audio in overall expressiveness on the EchoMind benchmark.

BCMDA: Bidirectional Correlation Maps Domain Adaptation for Mixed Domain Semi-Supervised Medical Image Segmentation

In mixed domain semi-supervised medical image segmentation (MiDSS), achieving superior performance under domain shift and limited annotations is challenging. This scenario presents two primary issues: (1) distributional differences between labeled and unlabeled data hinder effective knowledge transfer, and (2) inefficient learning from unlabeled data causes severe confirmation bias. In this paper, we propose the bidirectional correlation maps domain adaptation (BCMDA) framework to overcome these issues. On the one hand, we employ knowledge transfer via virtual domain bridging (KTVDB) to facilitate cross-domain learning. First, to construct a distribution-aligned virtual domain, we leverage bidirectional correlation maps between labeled and unlabeled data to synthesize both labeled and unlabeled images, which are then mixed with the original images to generate virtual images using two strategies, a fixed ratio and a progressive dynamic MixUp. Next, dual bidirectional CutMix is used to enable initial knowledge transfer within the fixed virtual domain and gradual knowledge transfer from the dynamically transitioning labeled domain to the real unlabeled domains. On the other hand, to alleviate confirmation bias, we adopt prototypical alignment and pseudo label correction (PAPLC), which utilizes learnable prototype cosine similarity classifiers for bidirectional prototype alignment between the virtual and real domains, yielding smoother and more compact feature representations. Finally, we use prototypical pseudo label correction to generate more reliable pseudo labels. Empirical evaluations on three public multi-domain datasets demonstrate the superiority of our method, particularly showing excellent performance even with very limited labeled samples. Code available at https://github.com/pascalcpp/BCMDA.

DSCSNet: A Dynamic Sparse Compression Sensing Network for Closely-Spaced Infrared Small Target Unmixing

Due to the limitations of optical lens focal length and detector resolution, distant clustered infrared small targets often appear as mixed spots. The Close Small Object Unmixing (CSOU) task aims to recover the number, sub-pixel positions, and radiant intensities of individual targets from these spots, which is a highly ill-posed inverse problem. Existing methods struggle to balance the rigorous sparsity guarantees of model-driven approaches and the dynamic scene adaptability of data-driven methods. To address this dilemma, this paper proposes a Dynamic Sparse Compressed Sensing Network (DSCSNet), a deep-unfolded network that couples the Alternating Direction Method of Multipliers (ADMM) with learnable parameters. Specifically, we embed a strict $\ell_1$-norm sparsity constraint into the auxiliary variable update step of ADMM to replace the traditional $\ell_2$-norm smoothness-promoting terms, which effectively preserves the discrete energy peaks of small targets. We also integrate a self-attention-based dynamic thresholding mechanism into the reconstruction stage, which adaptively adjusts the sparsification intensity using the sparsity-enhanced information from the iterative process. These modules are jointly optimized end-to-end across the three iterative steps of ADMM. Retaining the physical logic of compressed sensing, DSCSNet achieves robust sparsity induction and scene adaptability, thus enhancing the unmixing accuracy and generalization in complex infrared scenarios. Extensive experiments on the synthetic infrared dataset CSIST-100K demonstrate that DSCSNet outperforms state-of-the-art methods in key metrics such as CSO-mAP and sub-pixel localization error.

SGMatch: Semantic-Guided Non-Rigid Shape Matching with Flow Regularization

Establishing accurate point-to-point correspondences between non-rigid 3D shapes remains a critical challenge, particularly under non-isometric deformations and topological noise. Existing functional map pipelines suffer from ambiguities that geometric descriptors alone cannot resolve, and spatial inconsistencies inherent in the projection of truncated spectral bases to dense pointwise correspondences. In this paper, we introduce SGMatch, a learning-based framework for semantic-guided non-rigid shape matching. Specifically, we design a Semantic-Guided Local Cross-Attention module that integrates semantic features from vision foundation models into geometric descriptors while preserving local structural continuity. Furthermore, we introduce a regularization objective based on conditional flow matching, which supervises a time-varying velocity field to encourage spatial smoothness of the recovered correspondences. Experimental results on multiple benchmarks demonstrate that SGMatch achieves competitive performance across near-isometric settings and consistent improvements under non-isometric deformations and topological noise.

Mock Worlds, Real Skills: Building Small Agentic Language Models with Synthetic Tasks, Simulated Environments, and Rubric-Based Rewards

Small LLMs often struggle to match the agentic capabilities of large, costly models. While reinforcement learning can help, progress has been limited by two structural bottlenecks: existing open-source agentic training data are narrow in task variety and easily solved; real-world APIs lack diversity and are unstable for large-scale reinforcement learning rollout processes. We address these challenges with SYNTHAGENT, a framework that jointly synthesizes diverse tool-use training data and simulates complete environments. Specifically, a strong teacher model creates novel tasks and tool ecosystems, then rewrites them into intentionally underspecified instructions. This compels agents to actively query users for missing details. When handling synthetic tasks, an LLM-based user simulator provides user-private information, while a mock tool system delivers stable tool responses. For rewards, task-level rubrics are constructed based on required subgoals, user-agent interactions, and forbidden behaviors. Across 14 challenging datasets in math, search, and tool use, models trained on our synthetic data achieve substantial gains, with small models outperforming larger baselines.