Xiang Li

This paper introduces the concept of 'Sleeper Attack,' demonstrating that adversarial content can persist across multiple interactions with an LLM agent, posing a more subtle and difficult-to-detect safety threat than single-interaction attacks.

The paper proposes Reasoning-Conditioned Direct Preference Optimization (RC-DPO) to effectively mitigate hallucinations in multimodal large reasoning models by explicitly conditioning the preference optimization on the Chain-of-Thought (CoT) process.

The paper introduces Crafter, a multi-agent harness that significantly improves the generation of editable, publication-quality scientific figures from diverse inputs, addressing the limitations of existing single-purpose systems.

The paper introduces Agora, a domain-aware multi-agent framework that successfully detects deep, previously unknown logic bugs in complex consensus protocols, outperforming existing LLM-based analysis methods.

ESPO is a novel reinforcement learning algorithm that detects trajectory failure in large language models and terminates rollouts early, significantly improving performance on mathematical reasoning benchmarks while reducing computational cost.

This paper introduces the concept of Budget-Aware Agents (BAGEN), showing that current LLM agents often fail to manage resources proactively, and proposes that incorporating early stop and interval estimation significantly improves efficiency.

This paper introduces LVCG, a novel self-supervised framework that learns unified, view-invariant latent representations of cardiac electrical activity directly in the physically grounded Vectorcardiogram (VCG) space, improving generalization over traditional ECG-space methods.

The paper introduces MASA, a model-aware skill alignment framework that adaptively rewrites general and task-specific skills for LLM agents, achieving superior performance across diverse backbones and environments.

The paper introduces Diversity-inducing Initialization (DivIn), a novel method that improves image diversity by re-weighting the initial noise selection based on the guidance potential, thereby mitigating mode collapse.

The eMoT framework enhances multi-step reasoning in LLMs by treating reasoning as an evolving memory, stabilizing performance through symbolic computation and structured refinement.

The paper introduces CASTER, a new human-centric task for evaluating User-Generated Content (UGC) resonance, and proposes MEDEA, an architecture that uses a Social Chain-of-Thought mechanism to simulate community reactions for quality assessment.

This study successfully demonstrates that federated learning can achieve prediction accuracy comparable to centralized modeling for multi-center sepsis prediction while fundamentally preserving patient data privacy.

This paper proposes a four-layer technical architecture for large model inference optimization, including Multi-model Fusion, Model Optimization, Compute-Model Fusion, and Compute-Network-Model Fusion.

This paper proposes a two-stage algorithm, AC-IHT, for high-dimensional regression with contamination, achieving near-optimal estimation and strong oracle property.

KernelFlume is a decode-centric architecture that disaggregates the stable projection/FFN path from core-attention computation to improve efficiency and reduce cost in serving long-context demand.

This paper introduces Chronos, a physics-informed framework for non-Markovian long-horizon manipulation, which elevates observation history to the latent state of the policy dynamics and achieves higher success rates and fewer parameters than Markovian VLA baselines in both simulated and real-world experiments.

This paper proposes Trellis, a data foundation that treats experience graphs from long-horizon agentic tasks as first-class, governed, queryable database state.

This paper proposes CRISP, a model-agnostic framework for source-only medical image segmentation under distribution shift, which uses rank stability of positive regions to derive robust spatial priors.

The paper proposes SALMONN-2, an ALLM built on a unified SSL encoder, and presents a multi-layer feature fusion adapter to better exploit hierarchical SSL encoder representations. It also explores multimodal in-context learning in ALLMs and shows that a general-purpose SSL encoder achieves comparable performance to specialized audio encoders.

This paper proposes RL-MACRO, a cybernetic closed-loop intelligence framework for autonomous robotic craniotomy, which includes a CNN-LSTM observer for temperature reconstruction, an offline Implicit Q-Learning policy, and a novel dual-head Actor for coordinating cutting parameters.