Multimodal Large Language Models (MLLMs) achieve strong performance through instruction tuning, but real-world deployment requires them to continually expand their capabilities, making Multimodal Continual Instruction Tuning (MCIT) essential. Recent methods leverage sparse expert routing to promote task specialization, but we find that the expert routing process suffers from drift as the data distribution evolves. For example, a grounding query that previously activated localization experts may instead be routed to irrelevant experts after learning OCR tasks. Meanwhile, the grounding-related experts can be overwritten by new tasks and lose their original functionality. Such failure reflects two problems: router drift, where expert selection becomes inconsistent over time, and expert drift, where shared experts are overwritten across tasks. Therefore, we propose StAbilized Mixture-of-Experts (SAME) for MCIT. To address router drift, SAME stabilizes expert selection by decomposing routing dynamics into orthogonal subspaces and updating only task-relevant directions. To mitigate expert drift, we regulate expert updates via curvature-aware scaling using historical input covariance in a rehearsal-free manner. SAME also introduces adaptive expert activation to freeze selected experts during training, reducing redundant computation and cross-task interference. Extensive experiments demonstrate its SOTA performance.

Users increasingly expect image generation models to quickly adapt to highly diverse and personalized requirements, such as producing images with distinctive styles or characteristics. Traditional approaches rely on fine-tuning, which is costly and difficult to scale. To cope with these limitations, the community has accumulated a growing library of fine-tuned modules and adapters, where each component targets specific generation needs and collectively serves as a foundation for handling new demands. This naturally raises a question: instead of repeatedly training new models, can we systematically exploit this expanding ecosystem to better fulfill user instructions? To this end, we present Polaris, an intelligent retrieval framework that automatically selects and integrates suitable models from the model library based on a user's instructions. The key insight is that harnessing such a massive and heterogeneous pool requires not only finding the most relevant modules among thousands of candidates, but also aligning them effectively for instruction-driven generation and editing. Polaris addresses this challenge by indexing over 6,500 checkpoints and 75,000 adapters, and retrieving the most relevant components given a user's input and instruction. In doing so, it delivers scalable, controllable, and well-aligned generation – without any additional training.

Class-Incremental Learning (CIL) requires a learning system to learn new classes while retaining previously learned knowledge. However, in real-world scenarios such as autonomous driving, a system trained on urban roads in sunny weather may later need to operate in rural or highway environments with different traffic patterns and weather conditions. This requires the model not only to overcome catastrophic forgetting, but also to effectively handle domain shifts. In this paper, we propose CrOss-sample Relational Fusion (CORF), a unified framework to address domain shift and catastrophic forgetting simultaneously. To enhance generalizability, we perform selective refinement of training samples by leveraging spatial contribution maps to highlight semantically informative regions. Furthermore, we incorporate predictive confidence to adaptively weigh samples, thereby facilitating the learning of domain-agnostic representations. To alleviate forgetting, we propose a cascaded distillation framework that captures cross-sample relational dependencies across multiple feature hierarchies, enabling multi-grained knowledge transfer from previous tasks. CORF can be seamlessly integrated into existing CIL algorithms to enhance their generalizability, achieving competitive performance across various benchmark datasets. Code is available at https://github.com/LAMDA-CL/TMM26-CORF .

Test-time adaptation (TTA) updates deployed models using unlabeled test data to mitigate performance degradation under distribution shifts. However, most existing TTA methods are designed for single-modality settings, limiting their effectiveness in multimodal scenarios where degradation arises from both unimodal corruption and disrupted cross-modal correspondence. While recent multimodal Test-time adaptation(MMTTA) approaches improve fusion or attention alignment, they often underexploit spatio-temporal cues and lack explicit spatio-temporal consistency constraints during adaptation. To address these limitations, we propose a Spatio-Temporal Alignment framework (STA) for MMTTA, consisting of three modules. First, Spatial-Temporal Evidence Construction (STEC) module encodes motion dynamics and fuses spatial and temporal information to build robust spatio-temporal representations under corruption. Second, Cross-View Attention Consistency (CVAC) module aligns attention distributions across views to improve cross-view agreement during adaptation. Third, Reliability-Aware Deduction (RAD) module performs adaptive view-weighted inference at test time by emphasizing more reliable views. Extensive experiments on two benchmarks with 20 corruption types demonstrate that STA consistently outperforms prior methods, achieving up to a 12.3% accuracy improvement under audio modality corruption.

Music Recommendation Systems (MRSs) are a cornerstone of modern streaming platforms. Existing recommendation models, spanning both recall and ranking stages, predominantly rely on collaborative filtering, which fails to exploit the intrinsic characteristics of audio and consequently leads to suboptimal performance, particularly in cold-start scenarios. However, existing music recommendation datasets often lack rich multimodal information, such as raw audio signals and descriptive textual metadata. Moreover, current recommender system evaluation frameworks remain inadequate, as they neither fully leverage multimodal information nor support a diverse range of algorithms, especially multimodal methods. To address these limitations, we propose TASTE, a comprehensive dataset and benchmarking framework designed to highlight the role of multimodal information in music recommendation. Our dataset integrates both audio and textual modalities. By leveraging recent large-scale self-supervised music encoders, we demonstrate the substantial value of the extracted audio representations across recommendation tasks, including candidate recall and CTR. In addition, we introduce the MuQ-token method, which enables more efficient integration of multi-layer audio features. This method consistently outperforms other feature integration techniques across various settings. Overall, our results not only validate the effectiveness of content-driven approaches but also provide a highly effective and reusable multimodal foundation for future research. Code is available at https://github.com/zreach/TASTE

Pre-training world models on large, action-free video datasets offers a promising path toward generalist agents, but a fundamental flaw undermines this paradigm. Prevailing methods train models to predict a single, deterministic future, an objective that is ill-posed for inherently stochastic environments where actions are unknown. We contend that a world model should instead learn a structured, probabilistic representation of the future where predictive uncertainty correctly scales with the temporal horizon. To achieve this, we introduce a pre-training framework, **H**orizon-c**A**librated **U**ncertainty **W**orld **M**odel (HAUWM), built on a probabilistic ensemble that predicts frames at randomly sampled future horizons. The core of our method is a Horizon-Calibrated Uncertainty (HCU) loss, which explicitly shapes the latent space by encouraging predictive variance to grow as the model projects further into the future. This approach yields a latent dynamics model that is not only predictive but also equipped with a reliable measure of temporal confidence. When fine-tuned for downstream control, our pre-trained model significantly outperforms state-of-the-art methods across a diverse suite of benchmarks, including MetaWorld, the DeepMind Control Suite, and RoboDesk. These results highlight the critical role of structured uncertainty in robust decision-making.

Adapting pre-trained video generation models into controllable world models via latent actions is a promising step towards creating generalist world models. The dominant paradigm adopts a two-stage approach that trains latent action model (LAM) and the world model separately, resulting in redundant training and limiting their potential for co-adaptation. A conceptually simple and appealing idea is to directly replace the forward dynamic model in LAM with a powerful world model and training them jointly, but it is non-trivial and prone to representational collapse. In this work, we propose CoLA-World, which for the first time successfully realizes this synergistic paradigm, resolving the core challenge in joint learning through a critical warm-up phase that effectively aligns the representations of the from-scratch LAM with the pre-trained world model. This unlocks a co-evolution cycle: the world model acts as a knowledgeable tutor, providing gradients to shape a high-quality LAM, while the LAM offers a more precise and adaptable control interface to the world model. Empirically, CoLA-World matches or outperforms prior two-stage methods in both video simulation quality and downstream visual planning, establishing a robust and efficient new paradigm for the field.

Policy gradient methods rely on a baseline to measure the relative advantage of an action, ensuring the model reinforces behaviors that outperform its current average capability. In the training of Large Language Models (LLMs) using Actor-Critic methods (e.g., PPO), this baseline is typically estimated by a Value Model (Critic) often as large as the policy model itself. However, as the policy continuously evolves, the value model requires expensive, synchronous incremental training to accurately track the shifting capabilities of the policy. To avoid this overhead, Group Relative Policy Optimization (GRPO) eliminates the coupled value model by using the average reward of a group of rollouts as the baseline; yet, this approach necessitates extensive sampling to maintain estimation stability. In this paper, we propose V_0, a Generalist Value Model capable of estimating the expected performance of any model on unseen prompts without requiring parameter updates. We reframe value estimation by treating the policy's dynamic capability as an explicit context input; specifically, we leverage a history of instruction-performance pairs to dynamically profile the model, departing from the traditional paradigm that relies on parameter fitting to perceive capability shifts. Focusing on value estimation at State Zero (i.e., the initial prompt, hence V_0), our model serves as a critical resource scheduler. During GRPO training, V_0 predicts success rates prior to rollout, allowing for efficient sampling budget allocation; during deployment, it functions as a router, dispatching instructions to the most cost-effective and suitable model. Empirical results demonstrate that V_0 significantly outperforms heuristic budget allocation and achieves a Pareto-optimal trade-off between performance and cost in LLM routing tasks.

Document parsing, the task of extracting diverse content from PDFs while preserving their structural integrity, has been significantly advanced by Multimodal Large Language Models (MLLMs). These models have achieved remarkable success, largely driven by extensive post-training on massive datasets. This paper therefore undertakes a deep analysis of the two dominant adaptation strategies, Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), prompted by a puzzling observation on the PDF-to-Markdown task: SFT makes a negligible impact, especially on parsing complex tables and formulas, while RL achieves substantial overall gains. To unravel the reasons, our systematic investigation reveals a clear and complementary division of labor: SFT primarily operates as a structure learner, biased towards mastering the low-entropy syntax of document layouts. While it learns the format of a table, it struggles to ensure the fidelity of its high-entropy cell content. Conversely, RL excels as a content refiner by optimizing a holistic reward that reflects final accuracy. We further ground this phenomenon in the distinct theoretical nature of their respective objective functions. Based on these findings, we introduce a unified strategy that explicitly harnesses their individual strengths while mitigating their weaknesses. This work shows that a deep understanding of post-training methods is key to unlocking performance beyond what data scaling alone can achieve.

Antimicrobial peptides (AMPs) are promising alternatives to conventional antibiotics, yet early translation is often hindered by the perceived coupling between antibacterial potency and mammalian toxicity. This assumption complicates prioritization: highly active candidates are frequently suspected to be hemolytic, while existing multi-task predictors rarely reveal where selectivity resides in sequence space. Here, we present Multifunctional AMP Learning Engine (MAPLE), an interpretable dual-stream framework for AMP identification and systematic category-specific functional profiling across 14 activity categories directly from peptide sequences. MAPLE combines protein language model embeddings with explicit physicochemical descriptors, enabling robust task-specific prediction under severe label imbalance. Across the benchmark dataset and a sequence-non-overlapping independent validation set, MAPLE achieves consistently well-balanced performance, including on low-prevalence but clinically relevant endpoints. Building on this predictive basis, we conduct systematic k-mer enrichment to map motif-level selectivity and show that potency-hemolysis coupling is motif-regime-dependent rather than universal. Motifs most strongly enriched for antibacterial activity exhibit reduced hemolytic overlap and occupy a physicochemical regime characterized by moderate cationicity, lower hydrophobicity, and higher amphipathicity. We further provide a proof-of-concept prioritization workflow leveraging antibacterial-selective motifs, with structural modeling yielding conformations consistent with amphipathic α-helices. Despite limitations of predominantly binary annotations and incomplete structural integration, MAPLE offers reproducible sequence-level hypotheses and prioritization principles to support the engineering of potent and safer AMPs.

Learning new concepts while preserving the learned knowledge is the goal for learning systems in Class-Incremental Learning (CIL). Previous methods mainly approach CIL from two perspectives: one focuses on preserving the learned knowledge, making the model compatible with previously learned tasks, i.e., backward-compatible; the other considers making the model prepared for learning new tasks, so the model should also be compatible with the future tasks, i.e., forward-compatible. In this paper, we systematically identify one of the fundamental challenges behind CIL, named feature collision, where the features learned by the current task-specific model may collide with those of the previous models, leading to forgetting of previously learned tasks and hindering the learning of new tasks. Therefore, we propose a bi-compatible CIL framework to promote continual learning by addressing feature collision in both forward and backward compatible manners. Specifically, we consider: 1) forward-compatible task-agnostic feature learning to capture task-agnostic features continually with a separate model in advance for subsequent tasks; 2) backward-compatible rehearsal feature augmentation to preserve the discrimination ability of the learned task-specific feature. The framework is called Bi-compatible Task-agnostic Feature Expansion and Augmentation (BitFea), comprising both forward-compatible feature expansion and backward-compatible feature augmentation. Firstly, it captures task-agnostic features continually with a separate model, providing extra task-agnostic features for subsequent tasks. Secondly, to obtain useful features from the task-agnostic model for the current task, it aggregates the task-agnostic features with the task-specific feature using a merge attention. Then the aggregated feature is transferred back into the task-specific feature for inference, helping the task-specific model capture diverse features. Lastly, the task-specific feature of the rehearsal samples is augmented by a gradient ascent step to help keeping the generalization on the previous tasks, preserving the discrimination ability of the task-specific feature across the learned tasks. Extensive experiments show the effectiveness and superiority of BitFea on various CIL settings. The code is available at https://github.com/bwnzheng/BitFea_IJCV .

The practice of knowledge distillation, wherein pre-trained models - termed "teachers" - serve to enhance the performance of subsequent "student" models, has seen a surge in popularity. With a lot of pre-trained models available, there is a pressing need to investigate effective mechanisms for utilizing these resources optimally. This paper tackles two main challenges inherent in developing such mechanisms. The initial challenge lies in the effective selection of the most advantageous models from a vast pool, given that exhaustively testing each pre-trained model is impractical. The second challenge stems from the diversity of pre-trained models, as it is common for a pre-trained model to have a different label space than that of the current task. This necessitates the development of a universal model reuse approach capable of integrating any pre-trained model, regardless of its unknown label set. To address these issues, we introduce a dual-phase framework named "Selective Cross-Label Distillation." The first phase, termed model assessment, evaluates the semantic similarity between a potential pre-trained model and the target model through optimal transport. By determining the transportation cost, we pinpoint candidate models with lower costs to serve as our source models. The second phase, knowledge reuse, concentrates on minimizing the transportation cost between the chosen source models and the target model. This dual-phase approach enables efficient model selection and bridges the semantic gap between pre-trained models and the target task, leading to enhanced performance. The effectiveness of our framework in model selection and knowledge reuse is validated by experimental results.

Time Series Foundation Models (TSFMs) have achieved remarkable success through large-scale pretraining. However, their design primarily targets real-valued series, limiting their ability to handle general forecasting tasks involving diverse and often \emph{heterogeneous covariates}—such as categorical variables and multimodal data (e.g., images, text)—which are typically task-specific and difficult to leverage during pretraining. To address this gap, we propose Unified Covariate Adaptation (UniCA), a framework to bridge TSFMs with general covariate-aware forecasting. UniCA first performs covariate homogenization to transform heterogeneous covariates into high-level homogeneous series representations and then fuses them via a unified attention-based fusion mechanism. UniCA is compatible and universal for adaptation with both homogeneous and heterogeneous covariates, incorporating extra covariate information while preserving the generalization ability of TSFMs. Extensive experiments on multiple unimodal and multimodal covariate-aware forecasting benchmarks demonstrate the superiority of UniCA, highlighting the promise of covariate-aware TSFM adaptation in real-world forecasting scenarios. Code: https://anonymous.4open.science/r/UniCA-C5E0.

Designing dense reward functions is pivotal for efficient robotic Reinforcement Learning (RL). However, most dense rewards rely on manual engineering, which fundamentally limits the scalability and automation of reinforcement learning. While Vision-Language Models (VLMs) offer a promising path to reward design, naive VLM rewards often misalign with task progress, struggle with spatial grounding, and show limited understanding of task semantics. To address these issues, we propose MARVL-Multi-stAge guidance for Robotic manipulation via Vision-Language models. MARVL fine-tunes a VLM for spatial and semantic consistency and decomposes tasks into multi-stage subtasks with task direction projection for trajectory sensitivity. Empirically, MARVL significantly outperforms existing VLM-reward methods on the Meta-World benchmark, demonstrating superior sample efficiency and robustness on sparse-reward manipulation tasks.

The efficiency of Large Language Model (LLM) serving is fundamentally limited by the sequential nature of autoregressive decoding. Speculative Decoding (SD) mitigates this by using a lightweight draft model to speculate future tokens, which are then validated by the LLM in a single parallel forward pass. To further boost efficiency, multi-candidate schemes propose diverse candidate sets to increase the likelihood of token acceptance. However, we show that these schemes are bottlenecked by Residual Drift: a phenomenon where the rejection of initial candidates causes the residual target distribution to diverge from the draft model's predictions. This shift renders subsequent candidates ineffective and forces the system into expensive resampling. To resolve this, we propose ResiSpec, a framework that strategically reforms the proposal distribution during verification to anchor the residual target mass within the draft model's high-confidence regions. By mathematically re-aligning the verification process without compromising output exactness, ResiSpec prevents candidate obsolescence and achieves up to 1.92× speedup over state-of-the-art multi-candidate methods. Code is available at https://github.com/Czzzk/Resispec.

LLM agents increasingly solve tasks by invoking multiple tools, where parallel execution is essential for low latency but difficult to manage safely. Existing agent benchmarks primarily evaluate tool selection, argument generation, and end-to-end success under mostly serial execution, largely overlooking valid parallelization and resource-constrained scheduling. This missing scheduling dimension creates a practical failure mode: serial execution is safe but slow, while resource-agnostic parallel execution is fast but prone to avoidable resource overflows. To address this gap, we introduce PeakBench, a benchmark of executable multi-tool workflows with execution-grounded dependency annotations and measured resource profiles. A central challenge in evaluating such workflows is attribution: failures and inefficiencies may arise from incorrect dependency planning, poor resource-constrained scheduling, or both. PeakBench addresses this challenge with a two-part evaluation framework that disentangles logical planning from physical scheduling, with dedicated metrics for each dimension. Using this framework, we show that strong logical planning does not reliably translate into safe or efficient execution under resource constraints. We further show that exposing resource information can reduce avoidable overflows and improve resource utilization, making PeakBench a useful testbed for diagnosing resource-aware agent behavior. Code is available at https://github.com/Czzzk/Staggering-the-Peaks.

Class-Incremental Learning (CIL) aims to continually learn new classes without forgetting previously acquired knowledge. Vision-language models such as CLIP offer strong transferable representations via multi-modal supervision, making them a promising choice for CIL. However, applying CLIP to CIL poses two major challenges: (1) adapting to downstream tasks often requires additional learnable modules, increasing model complexity and susceptibility to forgetting; and (2) while multi-modal representations offer complementary strengths, existing methods have not fully exploited the synergy between visual and textual modalities. To address these issues, we propose BOFA (Bridge-layer Orthogonal Fusion for Adaptation), a novel framework for CIL. BOFA restricts adaptation to CLIP’s existing cross-modal bridge layer, keeping the core learning process parameter-free and avoiding any extra adaptation modules. To prevent forgetting within this layer, it leverages Orthogonal Low-Rank Fusion, a mechanism that constrains parameter updates to a low-rank ``safe subspace" that is mathematically constructed to be approximately orthogonal to the feature subspace of past tasks. This encourages stable knowledge accumulation and mitigates interference between new and previously learned classes. Furthermore, BOFA employs a cross-modal hybrid prototype that fuses stable textual prototypes with dynamic visual counterparts derived from our adapted bridge layer, resulting in a more robust and discriminative classifier. Extensive experiments on standard benchmarks demonstrate that BOFA achieves superior accuracy and efficiency compared to existing methods.

In Reinforcement Learning with Verifiable Rewards (RLVR), constructing a robust advantage baseline is critical for policy gradients, effectively guiding the policy model to reinforce desired behaviors. Recent research has introduced Generalist Value Models (such as V_0), which achieve pre-trained value estimation by explicitly encoding model capabilities in-context, eliminating the need to synchronously update the value model alongside the policy model. In this paper, we propose V_0.5, which adaptively fuses the baseline predicted by such value model (acting as a prior) with the empirical mean derived from sparse rollouts. This constructs a robust baseline that balances computational efficiency with extremely low variance. Specifically, we introduce a real-time statistical testing and dynamic budget allocation. This balances the high variance caused by sparse sampling against the systematic bias (or hallucinations) inherent in the value model's prior. By constructing a hypothesis test to evaluate the prior's reliability in real-time, the system dynamically allocates additional rollout budget on demand. This mechanism minimizes the baseline estimator's Mean Squared Error (MSE), guaranteeing stable policy gradients, even under extreme sparsity with a group size of 4. Extensive evaluations across six mathematical reasoning benchmarks demonstrate that V_0.5 significantly outperforms GRPO and DAPO, achieving faster convergence and over some 10

Knowledge Distillation (KD) could transfer the “dark knowledge" of a well-performed yet large neural network to a weaker but lightweight one. From the view of output logits and softened probabilities, this paper goes deeper into the dark knowledge provided by teachers with different capacities. Two fundamental observations are: (1) a larger teacher tends to produce probability vectors with lower distinction among non-ground-truth classes; (2) teachers with different capacities are basically consistent in their cognition of relative class affinity. Through abundant experimental studies we verify these observations and provide in-depth empirical explanations to them. We argue that the distinctness among incorrect classes embodies the essence of dark knowledge. A larger and more accurate teacher lacks this distinctness, which hampers its teaching ability compared to a smaller teacher, ultimately leading to the peculiar phenomenon named "capacity mismatch". Building on this insight, this paper explores multiple simple yet effective ways to address capacity mismatch, achieving superior experimental results compared to previous approaches.

This paper presents OvisOCR, a lightweight and strictly end-to-end Multimodal Language Model (MLLM) tailored for document parsing. Unlike current methods that rely on complex "Crop-OCR-Merge" cascades to handle high-resolution inputs, OvisOCR directly maps full-page visual signals to structured Markdown without localized slicing or layout detection dependencies. Through extensive evaluations on the OmniDoc benchmark, OvisOCR achieves SOTA performance, demonstrating that a compact E2E model can effectively ``digest'' the capabilities of intricate pipelines and surpass specialized and general methods. Technically, OvisOCR establishes a holistic paradigm that synergizes specialized perception with general reasoning, distilling fine-grained recognition from OCR engines and semantic correction from LLMs into a unified model. To balance the performance across diverse document constituents, we design category-specific reward mechanisms for distinct element types, such as dense text, complex tables, and formulas, and ensure the model enhances its formatting strengths for each modality concurrently. This approach effectively resolves the optimization conflict, guaranteeing that improvements in structural layout parsing do not come at the expense of omitting fine-grained textual details. Empirical results confirm that OvisOCR eliminates the error propagation inherent in split-and-merge architectures, offering a streamlined path for next-generation document intelligence.