论文

本综述旨在厘清自进化 coding agent 的概念边界，为设计更具适应性、可靠性与软件感知能力的 agentic 系统奠定基础。This survey aims to clarify the conceptual boundaries of self-evolving coding agents and provide a foundation for designing more adaptive, reliable, and software-aware agentic systems.

一个持久化执行状态、使运行可被中断、能在崩溃后存活并继续的框架，必须为已经触发的副作用界定"恢复"的含义。五种广泛部署的 Agent 工作流框架给出了不同答案，且均未公开可机器验证的契约，其行为甚至违背了它们自己声明的片段。RESUME CONTRACT 针对持久化 API 陈述了六项性质（前缀延续、副作用恰好一次、分支确定性、检查点有效性、消费一次、恢复确定性），并附加分支意图与活性义务。TLA+ 模型对参考语义进行了检验……A framework that persists execution state so a run can be interrupted, survive a crash, and continue must decide what a resume means for effects that already fired. Five widely deployed agent workflow frameworks answer differently, none exposes a machine-checkable contract, and behavior violates even the fragments they state. The RESUME CONTRACT states six properties over the persistence API (prefix continuation, effect exactly-once, fork determinism, checkpoint validity, consume-once, recovery determinism), plus fork-intent and liveness obligations. A TLA+ model checks a reference semantics e

设计了一种类型化的领域特定语言（DSL），通过一组可逆算子捕获重复区域、浮点域等常见张量结构，将无损张量压缩建模为程序合成问题。A typed domain-specific language that captures recurring tensor structures, such as repeated regions and floating-point fields, through a set of reversible operators, is designed, which formulates lossless tensor compression as program synthesis.

将软件侧密钥替换为硬件受限的密钥，通过厂商中立的 PKCS#11 接口访问，并由包含会话身份、作用域边界、语义验证、污点追踪和硬件执行边界的五层 Zero-Trust 执行栈提供保护。This work replaces software-resident keys with hardware-confined keys accessible through a vendor-neutral PKCS#11 interface, enabled by a surrounding five-layer Zero-Trust enforcement stack comprising session identity, scope bounds, semantic validation, taint tracking, and the hardware execution boundary.

将一个确定性、无模型的流水线编译进 Agent 记忆：该流水线将本地采集流切分为类型化的活动帧与有界事件片段，携带应用、站点、时间、输入量以及回指原始行的证据指针，全程无模型参与。A deterministic, zero-model pipeline is compiled into agent memory with a deterministic, zero-model pipeline that segments a local capture stream into typed activity frames, bounded episodes carrying application, site, timing, input volume, and evidence pointers back to the raw rows, with no model in the loop.

提出 YOLO-PEFT，一个结构感知的框架，将 adapter 的放置建模为可审计的约束规划问题，以显式、可审查的规划取代手工对目标模块的试错，同时保留已验证的 train-save-merge-export 路径。YOLO-PEFT is proposed, a structure-aware framework that formulates adapter placement as an auditable constraint-planning problem that replaces manual target-module trial and error with explicit, inspectable planning while preserving verified train-save-merge-export paths.

往返一致性将可逆性转化为生成式模型一种实用的可信信号；双向训练带来负成本，在两个方向上均优于单向专家模型；其中反向还可作为快速的逆问题求解器。Round-trip consistency turns reversibility into a practical trust signal for generative models, and Bidirectional training comes at negative cost, beating direction specialists in both directions, and the backward direction doubles as a fast inverse solver.

本文提出一种结构化协议，通过对开源 LLM 评估与安全工具的分类驱动分析来自动化 AI 风险缓解，并给出一个可同时适用于开源与商用方案的分类驱动框架。This paper proposes a structured protocol to automate AI risk mitigation through a taxonomy-driven analysis of open-source LLM evaluation and security tools, and presents a taxonomy-driven framework applicable to open-source and proprietary solutions.

本文是一项关于如何提升蒸馏训练效率的实践研究，围绕两项系统贡献展开，并提出一种融合的 chunked KL loss，使峰值内存随序列长度线性增长。A practitioner's study of how to make distillation training efficient is presented, organised around two systems contributions, and a fused, chunked KL loss is introduced, making peak memory linear in the sequence length.

本文在每个时间步以骨骼作为输入，引入一种新的正则化方案来学习骨骼关节的共现特征，并提出一种同时作用于 LSTM 神经元门、单元和输出响应的新型 dropout 算法。This work takes the skeleton as the input at each time slot and introduces a novel regularization scheme to learn the co-occurrence features of skeleton joints, and proposes a new dropout algorithm which simultaneously operates on the gates, cells, and output responses of the LSTM neurons.

提出 Decoupling CLI Agent Scaffolding（DCAS），一种后端替换的拦截层，可在不修改 scaffold 的前提下，在任意 CLI scaffold 与任意后端模型之间路由 API 流量，从而支持跨 scaffold 评估与具备规划感知的轨迹采集。Decoupling CLI Agent Scaffolding (DCAS) is introduced, a backend-substitution interception layer that routes API traffic between any CLI scaffold and any backend model without modifying the scaffold, enabling cross-scaffold evaluation and planning-aware trajectory collection.

所提架构通过提供模块化、面向实现的具身认知能力 IVA 部署框架，弥合高层智能体推理模型与实时具身执行之间的鸿沟，助力在复杂交互虚拟环境中构建可扩展、自适应且可解释的智能体。The proposed architecture contributes by providing a modular, implementation-oriented framework for the deployment of embodied, cognitive-capable IVAs and bridges the gap between high-level agent reasoning models with real-time embodied execution, for scalable, adaptive, and explainable agents in complex interactive virtual environments.

提出 MultiMedQA 基准，整合六个现有医学问答数据集（涵盖专业医学、研究与消费者查询）及一个全新的在线医学问题搜索数据集，并提出针对模型答案的人工评估框架，揭示了 LLM 在医学领域的潜在应用价值。MultiMedQA, a benchmark combining six existing medical question answering datasets spanning professional medicine, research and consumer queries and a new dataset of medical questions searched online, is presented and a human evaluation framework for model answers is proposed, suggesting the potential utility of LLMs in medicine.

VeriForge 是一种混合主动写作系统，通过划分认知劳动，使系统在领域发现上承担主动权，而作者保留对叙事合成的完全主动权；在受控的冷启动写作任务中，专家评审者认为其产出在领域扎根方面更强。VeriForge is a mixed-initiative writing system that divides cognitive labor so that the system assumes initiative over domain discovery while the author retains full initiative over narrative synthesis, and is perceived by expert raters to produce passages with stronger domain grounding in a controlled cold-start writing task.

本文提出 UniMoMo，一种后训练压缩框架，将其形式化为约束图粗化问题，并引入分层自适应保护机制，根据路由暴露度限制对高流量 expert 的合并。UniMoMo, a post-training compression framework formulated as a constrained graph coarsening problem, is introduced, and a layer-adaptive protection mechanism that restricts the merging of high-traffic experts based on their routing exposure is introduced.

本文提出 LiveAnimate，据作者所知是首个将实时流式生成与十亿参数规模下的稳定长视频生成相结合的系统，基于 140 亿参数的视频 Diffusion Transformer（DiT）。This work presents LiveAnimate, to their knowledge the first animation system to combine real-time streaming with stable long-form generation at billion scale, built on a 14B-parameter video Diffusion Transformer (DiT).

SKILLER 是一个由自然语言驱动的强化学习框架，旨在为小模型自动生成执行器特定的 skills，使用强模型作为 actor 和 critic，将小模型 Agent 系统视为环境，并通过自然语言完全传递所有强化学习信号。SKILLER is a natural-language-driven reinforcement learning framework designed to automatically generate executor-specific skills for small models, which employs a strong model as the actor and critic, treats the small-model agent system as the environment, and propagates all reinforcement learning signals entirely via natural language.

本文给出了 LLM routing 的统一形式化，将其刻画为由五个组件构成的序贯决策过程：context 编码器、模型编码器、评分函数、决策规则和学习信号，涵盖单轮、多轮和个性化 routing。This work presents a unified formulation of LLM routing as a sequential decision process characterized by five components: context encoders, model encoders, scoring functions, decision rules, and learning signals, covering single-turn, multi-turn, and personalized routing.

本文将 S2G-RAG 的结构化充分性-缺口判断适配到冻结的 Search-R1 流程中，并在来自 900 个不相交 HotpotQA 问题的 3009 个状态上训练了一个 Qwen3.5-2B judge，以减少检索次数同时广泛保持答案准确性。This work adapts S2G-RAG's structured sufficiency-and-gap judgment to a frozen Search-R1 pipeline and trains a Qwen3.5-2B judge on 3,009 states from 900 disjoint HotpotQA questions to reduce retrieval while broadly preserving answer accuracy.

在多个 benchmark 数据集上的实证评估表明，RAEF 在准确率和推理开销方面均优于 RAF，并且与零样本及微调基础模型的全面对比显示，RAEF 在避免微计算负担的同时取得了与微调相当或更优的性能。Empirical evaluation across multiple benchmark datasets demonstrates that RAEF outperforms RAF in both accuracy and inference overhead, and comprehensive comparisons with zero-shot and fine-tuned foundation models show that RAEF achieves competitive or superior performance to fine-tuning while avoiding its computational burden.

提出一种 chunked-prefill 策略，可缓解由此带来的内存容量惩罚，在 32 GiB 共享内存上将允许的上下文宽度扩展 $2.7 \times$，但即便降低了内存开销，仍需打补丁才能使 Nanbeige4.2-3B 可用。A chunked-prefill strategy is introduced which alleviates the incurred memory-capacity penalty, extending allowable context width by $2.7 \times$ on 32~GiB shared memory, however, even with the reduced memory overhead, it is shown that patches are required to render Nanbeige4.2-3B usable.