Separating Stream Stability from Long-Term Recall in Language Models
本文区分了流式语言模型的稳定性与长期记忆问题,通过引入三个时间范围并提出ThreeH评估方法来解决这一区分问题。
本文区分了流式语言模型的稳定性与长期记忆问题,通过引入三个时间范围并提出ThreeH评估方法来解决这一区分问题。
为解决医学转诊中的信息过载和非结构化协作问题,提出多智能体结构化图推理框架MASGR,通过构建临床推理图和引入知识引导仲裁机制提高转诊准确性。
研究了通过反馈的功率受限并行高斯信道传输离散时间LTI向量源状态的问题,提出基于主化方法的线性编码解码器设计。
This work addresses the challenge in zero-shot image captioning where synthetic training data generated by text-to-image models often suffers from fine-grained entity misalignment—such as missing objects or mislocalized attributes—leading to distorted supervision signals. To mitigate this, the authors propose ReCap, a framework that explicitly detects image entities and guides caption rewriting to achieve fine-grained image-text alignment. ReCap further incorporates an adaptive dynamic weighting strategy to downweight unreliable synthetic samples during training. By shifting data refinement from implicit global matching to explicit entity-level realignment, the method introduces a plug-and-play mechanism for fine-grained correction. Experiments demonstrate that ReCap achieves state-of-the-art performance on both in-domain and cross-domain zero-shot image captioning benchmarks, significantly improving caption consistency and supervision fidelity.
This work addresses the challenge that nonlinear operations in large language model (LLM) conversations—such as reprompting, response regeneration, message deletion, and concurrent multi-device interactions—cannot be faithfully captured by traditional linear logs, thereby undermining the reliability of digital forensics and compliance auditing. To resolve this, the paper introduces Verifiable Conversation Transcript (VCT), a novel system that models nonlinear dialogues as serialized state transitions with deletion barriers. VCT ensures integrity and accountable verifiability through a three-layer hash chain structure (spanning QA pairs, sessions, and account-level Merkle roots), joint user–server signatures, and an asynchronous view-fork detection mechanism. Prototype evaluation demonstrates sub-millisecond to low-millisecond cryptographic overhead for core operations and only 0.9% metadata overhead for 21KB transcripts, confirming VCT’s feasibility for high-assurance auditing in production-grade LLM platforms.
本文区分了流式语言模型的稳定性与长期记忆问题,通过引入三个时间范围并提出ThreeH评估方法来解决这一区分问题。
为解决医学转诊中的信息过载和非结构化协作问题,提出多智能体结构化图推理框架MASGR,通过构建临床推理图和引入知识引导仲裁机制提高转诊准确性。
研究了通过反馈的功率受限并行高斯信道传输离散时间LTI向量源状态的问题,提出基于主化方法的线性编码解码器设计。
This work addresses the challenge in zero-shot image captioning where synthetic training data generated by text-to-image models often suffers from fine-grained entity misalignment—such as missing objects or mislocalized attributes—leading to distorted supervision signals. To mitigate this, the authors propose ReCap, a framework that explicitly detects image entities and guides caption rewriting to achieve fine-grained image-text alignment. ReCap further incorporates an adaptive dynamic weighting strategy to downweight unreliable synthetic samples during training. By shifting data refinement from implicit global matching to explicit entity-level realignment, the method introduces a plug-and-play mechanism for fine-grained correction. Experiments demonstrate that ReCap achieves state-of-the-art performance on both in-domain and cross-domain zero-shot image captioning benchmarks, significantly improving caption consistency and supervision fidelity.
This work addresses the challenge that nonlinear operations in large language model (LLM) conversations—such as reprompting, response regeneration, message deletion, and concurrent multi-device interactions—cannot be faithfully captured by traditional linear logs, thereby undermining the reliability of digital forensics and compliance auditing. To resolve this, the paper introduces Verifiable Conversation Transcript (VCT), a novel system that models nonlinear dialogues as serialized state transitions with deletion barriers. VCT ensures integrity and accountable verifiability through a three-layer hash chain structure (spanning QA pairs, sessions, and account-level Merkle roots), joint user–server signatures, and an asynchronous view-fork detection mechanism. Prototype evaluation demonstrates sub-millisecond to low-millisecond cryptographic overhead for core operations and only 0.9% metadata overhead for 21KB transcripts, confirming VCT’s feasibility for high-assurance auditing in production-grade LLM platforms.