This paper proposes a data-free dual orthogonal projection framework to perform continual model merging.

This paper proposes Panacea, a post-fine-tuning method that mitigates harmful fine-tuning in large language models, maintaining safety alignment without sacrificing performance across different tasks and models.

We propose an automatica data engine framework for quantum property estimation under limited quantum resources.

We propose a training-free projection-based continual merging method that efficiently combines models incrementally.

This paper investigates how hallucinations arise and persist in RLLM reasoning, revealing error self-reinforcement and limited metacognition.

We detect and remove backdoor samples in MLLM fine-tuning by identifying abnormal attention entropy patterns without requiring clean data or model modifications.

This paper proves that self-verification prevents model collapse in recursive training without relying on real data.

Given an input mesh, Puppeteer first transforms it into an animation-ready model through automatic rigging, and subsequently animates it under video guidance.

In this paper, we propose Value-Guided Decision Transformer (VDT), which employs progressively optimized value functions to guide the Decision Transformer (DT) in making optimal decisions.

This paper explores the representation of robot actions in the frequency domain and introduces FreqPolicy, a novel frequency-domain autoregressive framework for hierarchical robotic action generation.

We investigate how digit bias learned from skewed pretraining data lead to systematic numerical hallucinations in LLMs.

We propose GOOD, a training-free framework that leverages off-the-shelf classifiers to guide diffusion models for generating diverse, informative OOD samples—improving outlier exposure without requiring external datasets or embedding alignment.

Talk2Event is a new benchmark for attribute-aware visual grounding from event cameras.

Propose a vehicle trajectory learning model capable of transferring across regions and tasks without retraining.

We propose CLAdapter to refine and adapt the rich representations learned from large-scale data for transfer across diverse, data-limited scientific downstream domains.

Our approach improves Vision Transformer Dense Representations via Self-Distillation

Analogy-Based Multi-Turn Jailbreak Method for LLMs

We introduce an RL framework that unify the training of answer generation and verification in a single model.

Imagine360 creates high-quality, immersive 360 videos from perspective video anchors.

We construct a family of smooth convex loss with linear surrogate regret bounds via infimal convolution and Fenchel–Young losses, overcoming the smoothness–regret trade-off for arbitrary discrete target losses.

We propose a new paradigm to explore the best trade-off through mode connectivity for multi-task learning.

This paper addresses limitations in infrared small target detection (IRSTD) evaluation, aiming to establish a comprehensive hierarchical evaluation framework for IRSTD models.

Intend to Move (I2M) is a new multimodal dataset for embodied AI, designed for intention-aware human motion understanding in real-world environments.

A training-free framework for high-fidelity textured 3D morphing via Flow Models.

We propose a memory-enhanced improvement search framework based on deep reinforcement learning to solve flexible job-shop scheduling problems.

We propose Equivalent Transformation Mechanism with KKT-Multiplier Regularization for solving SemiUOT and UOT

HyperGraphRAG introduces hypergraph-based knowledge representation to capture real-world n-ary relations, boosting answer accuracy, retrieval efficiency, and generation quality.

We propose the first robust weak-to-strong generalization framework to elicit robust knowledge from a strong student VLM in an unsupervised scheme.

A unified and scalable RL framework applicable to online, offline, and offline-to-online settings.

We propose a novel safety-driven unlearning framework for diffusion models that can better maintain unlearning performance after downstream fine-tuning.

We propose HubGT as a scalable Graph Transformer enabling fast and hierarchical embeddings.

A memory framework for video world models that improves long-term consistency

We present the shortcomings of existing dropout-based methods in modeling long-range tasks.

STRIDER is a framework for zero-shot VLN-CE, featuring a SWG to constrain action space with spatial structure, and a TAR to adjust behavior based on task progress, helping agents align actions with spatial layout and task intent.

A defense method against RAG corpus attacks that combines semantic of document.

We propose a principled framework Bayesian Data Scheduler (BDS), which is an adaptive tuning-stage defense strategy against harmful fine-tuning with no need for attack simulation.

VLM unlearning by a closed-form ensemble of infinite number of functions with parameters uniformly sampled from a task arithmetic simplex.

Feed-forward dynamic 3DGS scene reconstruction from videos.

Our proposed orthogonality and variance losses improve performance in downstream fine-tuning of Mixture-of-Experts models by enhancing expert specificity, addressing expert homogenization caused by load balancing, while maintaining load balance.

We propose a noval fine tuning framework that uses reinforcement learning and adapter compression to selectively shield model layers in TEEs, which achieves strong model protection with 2–3× efficiency gains and no accuracy loss.

To train an adaptive video tokenizer, we introduce probabilistic taildrop to inject visual complexity prior to the tokenizer and incorporate GRPO for post-training, which further boosts efficiency in a task-aware adaptive manner.

To enhance the reliability of the reward model for current policy improvement, we have developed the Proximal Policy Exploration (PPE) algorithm to increase the coverage of the preference buffer in areas close to the near-policy distribution.

This paper tackles the problem of zero-shot inference on text-attributed graphs, and proposes a novel method that uses bundles to query LLMs and supervise GNNs.

VLMLight leverages vision-language scene understanding and dual-branch reasoning to achieve safe and efficient traffic signal control, especially in critical scenarios.

A personalized subgraph federated learning framework that learns inter‑client similarity on the fly, enabling adaptive, client‑specific aggregation at the server.

How to employ the coherence between layers to design a PEFT method.

FOA-Attack enhances targeted adversarial transferability to closed-source MLLMs by optimally aligning global and local features.

We propose a framework called Guard to enhance the robustness of a broad class of 1-consistent learning-augmented caching algorithms, while preserving 1-consistency and incurring low additional computational overhead.

A Planning Representation and Paradigm Investigation of Vision-Language-Action Models

To incorporate reward shaping approach into multi-task reinforcement learning, we propose a Centralized Reward Agent based MTRL framework (CenRA) to share and transfer knowledge across multiple tasks.

We present the first parameter-free last-iterate convergence of Counterfactual Regret Minimization algorithms.

We introduce a novel object selection mechanism to allow sim-trained policies to rapidly adapt to real-world visual perturbations.

We provide the statistical guarantee for Sinkhorn bridge method to estimate Schrödinger bridge

We propose FedWMSAM, a novel method that combines personalized momentum and dynamic SAM to achieve faster convergence and flatter minima in federated learning, effectively addressing client drift and data heterogeneity.

We systematically introduce AND, OR, and ADDER gates and analyze the faithfulness and completeness via these gates, and we also propose a framework to fully discovery these gates and obtain some findings.

Exploiting the overfitting of LLMs, we use only ten benign QA pairs to fine-tune and jailbreak them.

VORTA accelerates video diffusion transformers by sparse attention and dynamic routing, achieving up to 14.4× speedup with negligible quality loss.

A targeted restoration method to recover pruned models while preserving their low cost and high efficiency

We propose a trajectory recovery model based on PLM to address the challenge of the availability of dense trajectories is limited and to generalize across sparse trajectories with varying sampling intervals.

We present a unified audio-visual framework to uncover how humans and AI respond to modality conflicts and bias.

We propose a new plug-and-play optimization method from the perspective of symmetry teleportation for mainstream multi-task learning approaches.

We introduce OPHR, a multi-agent reinforcement learning framework that outperforms traditional approaches in volatility trading by dynamically selecting options positions and optimizing hedging strategies.

We propose MTRec, a novel sequential recommendation framework which uses a learned mental reward model to guide the recommendation model to align with users' real preferences.

We introduce MMAR, a new benchmark designed to evaluate the deep reasoning capabilities of Audio-Language Models (ALMs) across massive multi-disciplinary tasks.