Hao Liu

The paper theorizes that aligned LLMs remain jailbreakable due to 'Refusal-Escape Directions' (RED), which are continuous perturbation paths that shift model behavior from refusal to answering, and shows this vulnerability is linked to specific operator-level sources within the model architecture.

The paper proposes Safety Context Injection (SCI), an inference-time framework that prepends a structured external risk report to protect Large Reasoning Models (LRMs) against sophisticated jailbreaks, significantly reducing attack success rates.

The paper introduces a new benchmark and decomposition method, Sufficiency-Tightness Decomposition, demonstrating that current coding agents struggle to accurately infer least-privilege authorization, and that this decomposition significantly improves both security and task success.

The paper introduces POLARIS, a novel framework that systematically generates comprehensive and verifiable safety tests for LLMs by formalizing natural language policies into First-Order Logic and exploring the resulting Semantic Policy Graph.

VFEAgent is a novel multi-agent framework that automates the entire Finite Element Analysis (FEA) workflow, achieving high success rates in generating complete and physically valid simulations directly from multimodal inputs.

The paper introduces HRBench, a unified and comprehensive evaluation framework for systematically benchmarking and comparing various thinking-mode switching strategies in hybrid-reasoning LLMs.

The paper introduces OmniMatBench, a comprehensive, human-calibrated multimodal reasoning benchmark covering 19 materials science subfields, revealing that current multimodal language models (MLLMs) have significant gaps in complex materials-science reasoning.

SkillsInjector proposes a two-stage adaptive method to dynamically optimize skill selection, quantity, and presentation for LLM agents, significantly improving task performance over static injection methods.

The paper introduces AnyMo, a unified multimodal framework that enables high-quality, scalable conditional human motion generation by leveraging a massive, cross-modal dataset and a masked modeling transformer.

The paper proposes a novel zeroth-order optimization framework to enhance the robustness of LLM safety alignment, showing that few refinement steps can significantly improve safety while maintaining utility.

TRACE proposes a novel method to mitigate catastrophic forgetting in continual LLM fine-tuning by identifying and isolating a small, task-specific subset of essential parameters for each task.

The paper introduces Dr. DocBench, a difficulty-aware, comprehensive benchmark designed to rigorously test expert-level and challenging document parsing capabilities for VLMs, demonstrating that current state-of-the-art models fail on complex, domain-specific structures.

The paper introduces Iteris, an agentic research system, demonstrating its capability to generate numerical evidence, constructions, and proof drafts for open problems in computational mathematics, requiring human expert validation.

The paper proposes using a mask-conditioned latent diffusion model to generate synthetic, labeled TEM images for data augmentation, achieving small but measurable performance improvements in defect detection and classification.

MOSS-Audio is a unified audio-language model designed for comprehensive understanding of speech, environmental sounds, and music, achieving strong performance across various audio-grounded tasks.

The paper introduces TRON, an online, rule-verifiable environment substrate that generates an unbounded stream of fresh, controllable visual reasoning training instances, significantly improving RL performance on external multimodal benchmarks.

The paper introduces TVIR, a new benchmark and multi-agent framework for deep research, to evaluate and improve the generation of factually reliable, text-visual interleaved reports.

The paper proposes Skill-RM, a unified framework that treats reward modeling as an agentic task to consistently integrate diverse evaluation criteria, achieving superior performance over traditional methods.

The paper proposes OneReason, a framework that enhances the reasoning capability of generative recommendation models by focusing on improving item perception and structuring user behavior into coherent latent interests.

This paper proposes a training-free framework called ReasonAlloc to mitigate inference bottlenecks in large language models by recasting decoding-time key-value compression as a hierarchical budget allocation problem.