Selected work

Representative Papers

PHASE-Tree: Modeling Character-State Evolution in Long-Horizon Role-Playing Dialogue

Aug 07, 2026

Existing role-playing dialogue systems struggle to dynamically evolve character states over long-form narratives while maintaining consistency, and lack benchmarks for evaluating such evolutionary generation capabilities. This work proposes PHASE-Tree, a multi-timescale hierarchical character state representation structured as a tree with an immutable identity root and mutable layers encompassing personality, conversational memory, and transient states. The framework supports localized cross-session updates and guides response generation through either explicit textual injection or implicit parameter adaptation. We introduce the first character state representation framework enabling partial updates and release LongEvoRoleBench, a new evaluation benchmark. Experiments show that PHASE-Tree significantly outperforms baselines across all 12 metrics, achieving improvements of 19.7%, 12.4%, and 15.1% in character-level, semantic, and embedding scores, respectively, with human preferences strongly correlating with GPT-4.1 evaluations (r = 0.65).

0 citationsRead paper

SparseX: Efficient Segment-Level KV Cache Sharing for Interleaved LLM Serving

Jun 01, 2026

This work addresses the inefficiency of conventional KV caching in reusing non-prefix, cross-request, cross-turn, and cross-agent repetitive content, which incurs substantial computational overhead during the prefill phase of long-context large language models. To overcome this, the authors propose SparseX, a segment-level KV cache sharing method that treats contiguous token segments as reusable units and recovers complex interleaved context interactions via sparse recomputation within a single forward pass. Its key innovations include a hybrid attention mechanism leveraging Sparse-Q indexing and layer-adaptive thresholds, enabling effective support for multi-turn dialogues, retrieval-augmented generation (RAG), and agent workflows. Integrated seamlessly into vLLM through segmented cache lookup, PagedAttention, RoPE alignment, and FlashAttention, SparseX provides a model-agnostic, training-free, and Prefix Cache–compatible unified execution path. Experiments demonstrate that SparseX significantly reduces first-token latency and computational cost while improving cache efficiency and generation quality.

0 citationsRead paper

SkillsVote: Lifecycle Governance of Agent Skills from Collection, Recommendation to Evolution

May 18, 2026

This work addresses the challenges of high trajectory noise in long-horizon tasks and the redundancy, inconsistent quality, and environmental sensitivity inherent in open-ended skill ecosystems for large language model agents. To tackle these issues, the authors propose SkillsVote, a framework that establishes the first end-to-end governance system encompassing skill acquisition, recommendation, and evolution. By integrating structured skill retrieval, trajectory decomposition, multi-factor attribution, environment-adaptability analysis, verifiable task synthesis, and an evidence-gated update mechanism, SkillsVote enables high-quality skill evolution. Experimental results demonstrate that, solely through governing an external skill repository, a frozen GPT-5.2 achieves performance gains of 7.9 and 2.6 percentage points on Terminal-Bench 2.0 and SWE-Bench Pro, respectively, confirming the method’s effectiveness and generalizability.

0 citationsRead paper

A Survey on the Security of Long-Term Memory in LLM Agents: Toward Mnemonic Sovereignty

Apr 17, 2026

This study addresses critical security risks—such as memory tampering, unauthorized access, and cross-session poisoning—that threaten the long-term memory of large language model (LLM) agents, an area insufficiently explored in existing research with regard to governability. Integrating insights from cognitive neuroscience and philosophy of memory, the work proposes a comprehensive memory lifecycle framework encompassing writing, storage, retrieval, execution, sharing, and forgetting. Aligning this framework with four core security objectives—integrity, confidentiality, availability, and governability—it conceptualizes memory as an independent security dimension and introduces the notion of “memory sovereignty,” identifying nine governance primitives. Through interdisciplinary analysis, the paper systematically categorizes threats including memory poisoning, extraction attacks, and control-flow hijacking, exposes gaps in current architectures concerning governance coverage and LLMs’ reflexive security capabilities, and highlights the untapped potential of leveraging LLMs themselves to enforce memory security.

0 citationsRead paper

MemReader: From Passive to Active Extraction for Long-Term Agent Memory

Apr 09, 2026

Current approaches to long-term memory construction in intelligent agents rely on passive, one-shot information extraction, which is highly susceptible to noisy dialogue, missing references, and cross-turn dependencies, often resulting in memory contamination and inconsistency. This work proposes the MemReader family of models, introducing for the first time an active memory extraction paradigm wherein agents dynamically evaluate the value, referential clarity, and completeness of incoming information to autonomously decide whether to write, delay, retrieve, or discard it. Through model distillation, we develop MemReader-0.6B and further optimize MemReader-4B using Group Relative Policy Optimization (GRPO), enabling reasoning-driven memory management within the ReAct framework. Evaluated on LOCOMO, LongMemEval, and HaluMem benchmarks, MemReader-4B achieves state-of-the-art performance in knowledge updating, temporal reasoning, and hallucination suppression, and has been integrated into MemOS for real-world deployment.

0 citationsRead paper

Recent publications

Latest Papers

PHASE-Tree: Modeling Character-State Evolution in Long-Horizon Role-Playing Dialogue

Aug 07, 2026

Existing role-playing dialogue systems struggle to dynamically evolve character states over long-form narratives while maintaining consistency, and lack benchmarks for evaluating such evolutionary generation capabilities. This work proposes PHASE-Tree, a multi-timescale hierarchical character state representation structured as a tree with an immutable identity root and mutable layers encompassing personality, conversational memory, and transient states. The framework supports localized cross-session updates and guides response generation through either explicit textual injection or implicit parameter adaptation. We introduce the first character state representation framework enabling partial updates and release LongEvoRoleBench, a new evaluation benchmark. Experiments show that PHASE-Tree significantly outperforms baselines across all 12 metrics, achieving improvements of 19.7%, 12.4%, and 15.1% in character-level, semantic, and embedding scores, respectively, with human preferences strongly correlating with GPT-4.1 evaluations (r = 0.65).

0 citationsRead paper

SparseX: Efficient Segment-Level KV Cache Sharing for Interleaved LLM Serving

Jun 01, 2026

This work addresses the inefficiency of conventional KV caching in reusing non-prefix, cross-request, cross-turn, and cross-agent repetitive content, which incurs substantial computational overhead during the prefill phase of long-context large language models. To overcome this, the authors propose SparseX, a segment-level KV cache sharing method that treats contiguous token segments as reusable units and recovers complex interleaved context interactions via sparse recomputation within a single forward pass. Its key innovations include a hybrid attention mechanism leveraging Sparse-Q indexing and layer-adaptive thresholds, enabling effective support for multi-turn dialogues, retrieval-augmented generation (RAG), and agent workflows. Integrated seamlessly into vLLM through segmented cache lookup, PagedAttention, RoPE alignment, and FlashAttention, SparseX provides a model-agnostic, training-free, and Prefix Cache–compatible unified execution path. Experiments demonstrate that SparseX significantly reduces first-token latency and computational cost while improving cache efficiency and generation quality.

0 citationsRead paper

SkillsVote: Lifecycle Governance of Agent Skills from Collection, Recommendation to Evolution

May 18, 2026

This work addresses the challenges of high trajectory noise in long-horizon tasks and the redundancy, inconsistent quality, and environmental sensitivity inherent in open-ended skill ecosystems for large language model agents. To tackle these issues, the authors propose SkillsVote, a framework that establishes the first end-to-end governance system encompassing skill acquisition, recommendation, and evolution. By integrating structured skill retrieval, trajectory decomposition, multi-factor attribution, environment-adaptability analysis, verifiable task synthesis, and an evidence-gated update mechanism, SkillsVote enables high-quality skill evolution. Experimental results demonstrate that, solely through governing an external skill repository, a frozen GPT-5.2 achieves performance gains of 7.9 and 2.6 percentage points on Terminal-Bench 2.0 and SWE-Bench Pro, respectively, confirming the method’s effectiveness and generalizability.

0 citationsRead paper

A Survey on the Security of Long-Term Memory in LLM Agents: Toward Mnemonic Sovereignty

Apr 17, 2026

This study addresses critical security risks—such as memory tampering, unauthorized access, and cross-session poisoning—that threaten the long-term memory of large language model (LLM) agents, an area insufficiently explored in existing research with regard to governability. Integrating insights from cognitive neuroscience and philosophy of memory, the work proposes a comprehensive memory lifecycle framework encompassing writing, storage, retrieval, execution, sharing, and forgetting. Aligning this framework with four core security objectives—integrity, confidentiality, availability, and governability—it conceptualizes memory as an independent security dimension and introduces the notion of “memory sovereignty,” identifying nine governance primitives. Through interdisciplinary analysis, the paper systematically categorizes threats including memory poisoning, extraction attacks, and control-flow hijacking, exposes gaps in current architectures concerning governance coverage and LLMs’ reflexive security capabilities, and highlights the untapped potential of leveraging LLMs themselves to enforce memory security.

0 citationsRead paper

MemReader: From Passive to Active Extraction for Long-Term Agent Memory

Apr 09, 2026

Current approaches to long-term memory construction in intelligent agents rely on passive, one-shot information extraction, which is highly susceptible to noisy dialogue, missing references, and cross-turn dependencies, often resulting in memory contamination and inconsistency. This work proposes the MemReader family of models, introducing for the first time an active memory extraction paradigm wherein agents dynamically evaluate the value, referential clarity, and completeness of incoming information to autonomously decide whether to write, delay, retrieve, or discard it. Through model distillation, we develop MemReader-0.6B and further optimize MemReader-4B using Group Relative Policy Optimization (GRPO), enabling reasoning-driven memory management within the ReAct framework. Evaluated on LOCOMO, LongMemEval, and HaluMem benchmarks, MemReader-4B achieves state-of-the-art performance in knowledge updating, temporal reasoning, and hallucination suppression, and has been integrated into MemOS for real-world deployment.

0 citationsRead paper