Task-specific LoRA adapters offer a modular way to specialize large language and vision-language models. However, existing adapter composition methods are mostly static and cannot adapt to individual test inputs. To address these issues, we propose LiST, a label-free test-time LoRA fusion framework that converts an existing LoRA bank into a target-conditioned local simplex and searches sample-specific fusion weights at inference time. LiST builds joint task representations from LoRA parameter anchors and prompt-level behavior vectors, retrieves neighboring adapters as a local search space, and performs branch-preserving fusion without updating the backbone or adapters. Candidate weights are selected by a prompt-level energy with prior, geometric, and stochastic-consistency constraints, and are deployed only when they pass a safe acceptance rule. Otherwise, LiST falls back to a target-conditioned prior. Experiments on multimodal and language benchmarks show that LiST outperforms static LoRA merging and conventional test-time adaptation baselines, while preserving task-specific adapter utility and improving robustness on unseen tasks.

Large language model (LLM) agents are increasingly organized into multi-agent systems for complex reasoning and decision-making tasks. However, scaling such systems can substantially increase inference cost, especially when multiple agents repeatedly process overlapping or role-irrelevant context. Many existing multi-agent frameworks focus on orchestration and role specialization, but still rely on full-context sharing or static context routing, leading to redundant token consumption, increased runtime, and inefficient memory utilization. We present Role-Conditioned Context Routing (RCR), an execution-time layer that allocates role-relevant shared-memory items to agents under token budgets. Given a structured shared memory and a set of role-specialized agents, RCR selectively routes task-relevant memory items to each agent according to its role, optional task-stage metadata, and token budget. RCR integrates role-conditioned relevance scoring, token-budgeted memory selection, and iterative memory updates to reduce redundant context exposure during multi-agent execution. Experiments on multi-hop QA, embodied reasoning (ALFWorld), and web-based decision-making (WebShop) show that RCR reduces token usage by 35% on average and runtime by 38% on average relative to Full-Context, while maintaining or improving task performance.

Text-to-motion (T2M) generation has achieved impressive realism but poses significant safety risks by faithfully executing unsafe textual prompts. Existing motion unlearning methods face dual challenges: brittle keyword filtering and discrete codebook manipulation, which may degrade benign generation quality and introduce jerky transitions. To address these challenges, we propose SafeMo, a linguistically grounded framework that leverages Large Language Models (LLMs) to align motion generation with safety constraints. First, we introduce SafeMoEngine, an LLM-agent pipeline that autonomously classifies unsafe intents and performs semantic rewriting to construct SafeMoVAE-29K, the first safety-aligned text-to-motion dataset. Second, we propose Minimal Motion Unlearning (MMU), a continuous-space unlearning strategy that projects unsafe semantic concepts into a task vector for precise negation, avoiding the quantization artifacts of discrete methods. Experiments demonstrate SafeMo’s effective unlearning performance by more strongly suppressing unsafe text–motion retrieval than the prior SOTA motion unlearning baseline LCR, reducing forget-set R@1 by 35.3% and achieving 2.61× normalized retrieval trade-off score (NRTS), while showing better or competitive benign utility on safe prompts, with similar trends on Motion-X. Code and checkpoints are publicly available at https://github.com/AIGeeksGroup/SafeMo. Project page: https://aigeeksgroup.github.io/SafeMo.

LLM agents are moving from single-prompt use to long task streams in which reusable memory becomes a core capability for terminal, software-engineering, and web tasks. Such memory is useful only when stored experience remains reliable across hundreds of interactions, but two failure modes break that assumption in practice. The first is unreliable admission: failed trajectories, accidental successes, and misleading observations enter memory because they appear relevant, then mislead later decisions. The second is memory drift: long-running banks accumulate duplicate, stale, and conflicting records that retrieval alone cannot repair. MemGuard’s key distinction is to treat verifier output not as a one-shot filter, but as persistent lifecycle metadata. It converts multi-criteria score-token verification into reward, confidence, label, and uncertainty descriptors that are attached to every candidate before activation and reused during retrieval, conflict resolution, summarization, and archival. We evaluate MemGuard on Terminal-Bench 2.0, SWE-Bench Verified, WebArena, and Mind2Web across four backbones, comparing against four memory baselines plus a verifier-only control under matched runtime budgets. Averaged over five seeds, MemGuard achieves the best success metric and lowest average steps in all 16 backbone–benchmark settings, improving over ReasoningBank, the strongest prior baseline among the memory methods we evaluate, with a largest gain of 7.9 success-rate points on WebArena, 5.6 step-success-rate points on Mind2Web, and 2.4–3.5 points on terminal and software-engineering benchmarks. Code is available at https://github.com/whyyyyy123/MemGuard.

Long-horizon multimodal agents depend on external memory. However, similarity-based retrieval often surfaces stale, low-credibility, or conflicting items, which can trigger overconfident errors. We propose Multimodal Memory Agent (MMA), which assigns each retrieved memory item a dynamic reliability score by combining source credibility, temporal decay, and conflict-aware network consensus, and uses this signal to reweight evidence and abstain when support is insufficient. We also introduce MMA-Bench, a programmatically generated benchmark for belief dynamics with controlled speaker reliability and structured text–vision contradictions. Using this framework, we uncover the "Visual Placebo Effect," revealing how RAG-based agents inherit latent visual biases from foundation models. On FEVER, MMA matches baseline accuracy while reducing accuracy standard deviation by 35.2% and improving selective prediction utility; on LoCoMo, a safety-oriented configuration improves actionable accuracy and reduces wrong answers; on MMA-Bench, MMA reaches 41.18% Type-B accuracy in Vision mode, while the baseline collapses to 0.0% under the same protocol. Code is available at https://github.com/AIGeeksGroup/MMA.

Recent diffusion and autoregressive video generators have substantially improved short-form synthesis, yet composing multiple independently generated clips into a coherent audiovisual story remains challenging. Character identities may drift across scenes, narrative events can become inconsistent, and dialogue or sound may not align with the evolving visual context. We present **DramaAgent**, a hierarchical, agentic, and model-agnostic framework for multi-scene long-horizon audiovisual story generation. Rather than modifying the underlying video backbone, DramaAgent introduces a stateful control layer that combines hierarchical story planning, persistent character conditioning, scene-wise generation, and reflection-guided targeted repair. The framework maintains reusable story and character states across scenes, diagnoses failures along identity, semantic, temporal, and audio-visual dimensions, and maps each diagnosed failure to a corresponding localized repair operation. Experiments on 50 narrative scripts across five video-generation backbones show consistent improvements in semantic alignment, character consistency, narrative fidelity, and scene-level audio-visual compatibility. Controlled same-backbone comparisons further show that these gains cannot be explained solely by stronger generators or additional sampling. Our results suggest that persistent state, failure diagnosis, and targeted repair provide a practical control paradigm for coherent multi-scene audiovisual generation.

Large language models (LLMs) have recently shown promise as planners across web, spatial, and embodied domains. However, existing agents often handle information seeking, spatial grounding, and action planning as separate capabilities, limiting their ability to solve long-horizon compositional tasks. To address these challenges, we present **WebWorld**, a web-simulated world model for long-horizon reasoning over web evidence, scene observations, and action plans. WebWorld maintains structured intermediate states across heterogeneous modalities and verifies them through schema-guided hierarchical planning. The framework introduces a Mixture-of-Transformers architecture that coordinates website, scene, and motion experts via multi-modal self-attention. To support training and evaluation, we develop WebWorld-CoT-55K, a large-scale dataset that integrates heterogeneous modalities into schema-guided reasoning traces, and **WebWorldBench**, a curated benchmark for evaluating long-horizon compositional tasks. Training combines cold-start supervised fine-tuning with offline reinforcement learning based on structural, semantic, and planning rewards; inference uses test-time scaling to aggregate multiple reasoning trajectories. Experiments on WebWorldBench show that WebWorld improves the overall score by up to 8.4% over strong LLM-agent baselines, with gains in outdoor reasoning, geolocation, and travel planning, while still leaving substantial room for progress toward human-level performance.

Large Vision Language Models (LVLMs) achieve great performance on visual-language reasoning tasks, however, the black-box nature of LVLMs hinders in-depth research on the reasoning mechanism. As all images need to be converted into image tokens to fit the input format of large language models (LLMs) along with natural language prompts, sequential visual representation is essential to the performance of LVLMs, and the information flow analysis approach can be an effective tool for determining interactions between these representations. In this paper, we propose integrating attention analysis with LLaVA-CAM, concretely, attention scores highlight relevant regions during forward propagation, while LLaVA-CAM captures gradient changes through backward propagation, revealing key image features. By exploring the information flow from the perspective of visual representation contribution, we observe that it tends to converge in shallow layers but diversify in deeper layers. To validate our analysis, we conduct comprehensive experiments with truncation strategies across various LVLMs for visual question answering and image captioning tasks, and experimental results not only verify our hypothesis but also reveal a consistent pattern of information flow convergence in the corresponding layers, and the information flow cliff layer will be different due to different contexts.