Trap: Mitigating Poisoning-Based Backdoor Attacks by Treating Poison With Poison IEEE Transactions on Dependable and Secure Computing(CCF A) 2026-06-10 #深度学习 #大模型
Hallucination as Exploit: Evidence-Carrying Multimodal Agents (arxiv 2026) 2026-06-01 #深度学习 #多模态 #大模型
Finding the Correct Visual Evidence Without Forgetting: Mitigating Hallucination in LVLMs via Inter-Layer Visual Attention Discrepancy (ICML 2026) 2026-05-29 #深度学习 #大模型
IN-CONTEXT SHARPNESS AS ALERTS: AN INNER REPRESENTATION PERSPECTIVE FOR HALLUCINATION MITIGATION (PMLR 2024) 2026-05-28 #深度学习 #大模型
S0 Tuning: Zero-Overhead Adaptation of Hybrid Recurrent-Attention Models (arxiv 2026) 2026-05-21 #深度学习 #大模型
Where Does Reasoning Break? Step-Level Hallucination Detection via Hidden-State Transport Geometry (arxiv 2026) motivation正确的推理过程会在一个局部连贯的稳定流形(Manifold)中移动 。当模型产生第一个推理错误(幻觉)时,在隐藏状态空间中会表现为一种局部且明显的传输成本偏移(Transport Excursion),即轨迹偏离了正确的流形 。 方法为了捕捉这种几何偏移,作者设计了一个两阶段架构:一个非部署的“教师模型”(用于探索理论上限)和一个可实际部署的“学生 2026-05-20 #深度学习 #大模型
AutoRISE: Agent-Driven Strategy Evolution for Red-Teaming Large Language Models (arxiv 2026) 有点类似openai翁家翌(jiayi weng)提出的Learning Beyond Gradients。 2026-05-11 #深度学习 #大模型
Detecting Contextual Hallucinations in LLMs with Frequency-Aware Attention (ICML 2026) 2026-05-06 #深度学习 #大模型
SwiftSage: A Generative Agent with Fast and Slow Thinking for Complex Interactive Tasks NeurIPS 2023 (spotlight) 2026-04-25 #深度学习 #大模型