A Survey of Context Engineering for Large Language Models
《大规模语言模型的上下文工程综述》

Zero-shot prompting enables task performance without prior examples, relying exclusively on instruction clarity and pre-trained knowledge [1361, 336, 553, 67, 1046]. Few-shot prompting extends this capability by incorporating limited exemplars to guide model responses, demonstrating task execution through strategic example selection [1361, 401, 103, 546, 788, 1371]. In-context learning facilitates adaptation to novel tasks without parameter updates by leveraging demonstration examples within prompts, with performance significantly influenced by example selection and ordering strategies [365, 103, 1287, 1016, 920, 846, 1139, 348, 576].
零样本提示能够在没有先例的情况下完成任务，仅依赖于指令清晰度和预训练知识 [1361, 336, 553, 67, 1046]。少样本提示通过引入有限的示例来引导模型响应，从而通过策略性示例选择展示任务执行能力 [1361, 401, 103, 546, 788, 1371]。上下文学习通过在提示中利用示范示例来实现对新任务的适应，而无需更新参数，其性能显著受到示例选择和排序策略的影响 [365, 103, 1287, 1016, 920, 846, 1139, 348, 576]。

Chain-of-Thought (CoT) prompting decomposes complex problems into intermediate reasoning steps, mirroring human cognition [1138, 401, 336, 939, 603]. Zero-shot CoT uses trigger phrases like “Let’s think step by step,” improving MultiArith accuracy from 17.7% to 78.7% [553, 1099, 472, 662], with Automatic Prompt Engineer refinements yielding additional gains [1215, 526].
链式思维（CoT）提示将复杂问题分解为中间推理步骤，模拟人类认知过程 [1138, 401, 336, 939, 603]。零样本 CoT 使用诸如“让我们一步一步思考”之类的触发短语，将 MultiArith 的准确率从 17.7%提高到 78.7% [553, 1099, 472, 662]，自动提示工程师的进一步优化也带来了额外的提升 [1215, 526]。

Tree-of-Thoughts (ToT) organizes reasoning as hierarchical structures with exploration, lookahead, and backtracking capabilities, increasing Game of 24 success rates from 4% to 74% [1246, 217, 557, 598]. Graph-of-Thoughts (GoT) models reasoning as arbitrary graphs with thoughts as vertices and dependencies as edges, improving quality by 62% and reducing costs by 31% compared to ToT [69, 826, 1366].
思维树（ToT）将推理组织为具有探索、前瞻和回溯能力的层次结构，将 24 点游戏的成功率从 4%提高到 74% [1246, 217, 557, 598]。思维图（GoT）将推理表示为任意图，其中思维是顶点，依赖关系是边，与 ToT 相比，GoT 的质量提高了 62%，成本降低了 31% [69, 826, 1366]。

Cognitive prompting implements structured human-like operations including goal clarification, decomposition, filtering, abstraction, and pattern recognition, enabling systematic multi-step task resolution through deterministic, self-adaptive, and hybrid variants [558, 557, 1205, 1164]. Guilford’s Structure of Intellect model provides psychological foundations for categorizing cognitive operations such as pattern recognition, memory retrieval, and evaluation, enhancing reasoning clarity, coherence, and adaptability [556, 191]. Advanced implementations incorporate cognitive tools as modular reasoning operations, with GPT-4.1 performance on AIME2024 increasing from 26.7% to 43.3% through structured cognitive operation sequences [243, 1030].
认知提示实现了结构化的人类操作，包括目标澄清、分解、过滤、抽象和模式识别，通过确定性、自适应和混合变体，使系统化多步任务解决成为可能[558, 557, 1205, 1164]。吉尔福德的认知能力结构模型为分类认知操作（如模式识别、记忆检索和评估）提供了心理学基础，增强了推理的清晰度、连贯性和适应性[556, 191]。高级实现将认知工具作为模块化推理操作进行整合，GPT-4.1 在 AIME2024 上的表现通过结构化认知操作序列从 26.7%提升到了 43.3%[243, 1030]。

RAG combines parametric knowledge stored in model parameters with non-parametric information retrieved from external sources, enabling access to current, domain-specific knowledge while maintaining parameter efficiency [591, 311, 253]. FlashRAG provides comprehensive evaluation and modular implementation of RAG systems, while frameworks like KRAGEN and ComposeRAG demonstrate advanced retrieval strategies with substantial performance improvements across diverse benchmarks [500, 749, 1159].
RAG 将存储在模型参数中的参数化知识与从外部来源检索的非参数化信息相结合，使模型能够访问当前且领域特定的知识，同时保持参数效率 [591, 311, 253]。FlashRAG 提供了对 RAG 系统的全面评估和模块化实现，而 KRAGEN 和 ComposeRAG 等框架则展示了先进的检索策略，并在多种基准测试中取得了显著的性能提升 [500, 749, 1159]。

Self-RAG introduces adaptive retrieval mechanisms where models dynamically decide when to retrieve information and generate special tokens to control retrieval timing and quality assessment [41]. Advanced implementations include RAPTOR for hierarchical document processing, HippoRAG for memory-inspired retrieval architectures, and Graph-Enhanced RAG systems that leverage structured knowledge representations for improved information access [928, 366, 360].
自适应 RAG 引入了自适应检索机制，其中模型动态决定何时检索信息，并生成特殊标记以控制检索时间与质量评估 [41]。高级实现包括用于分层文档处理的 RAPTOR、受记忆启发的检索架构 HippoRAG，以及利用结构化知识表示以提高信息访问能力的图增强 RAG 系统 [928, 366, 360]。

Knowledge graph integration addresses structured information retrieval through frameworks like KAPING, which retrieves relevant facts based on semantic similarities and prepends them to prompts without requiring model training [48, 673]. KARPA provides training-free knowledge graph adaptation through pre-planning, semantic matching, and relation path reasoning, achieving state-of-the-art performance on knowledge graph question answering tasks [258].
知识图谱集成通过 KAPING 等框架解决结构化信息检索问题，该框架基于语义相似性检索相关事实，并将其添加到提示中，无需模型训练[48, 673]。KARPA 通过预规划、语义匹配和关系路径推理，实现无需训练的知识图谱适应性，并在知识图谱问答任务中达到最先进的性能[258]。

Think-on-Graph enables sequential reasoning over knowledge graphs to locate relevant triples, conducting exploration to retrieve related information from external databases while generating multiple reasoning pathways [1000, 720]. StructGPT implements iterative reading-then-reasoning approaches that construct specialized functions to collect relevant evidence from structured data sources [489].
Think-on-Graph 通过在知识图谱上进行顺序推理来定位相关三元组，并在生成多个推理路径[1000, 720]的同时，进行探索以从外部数据库检索相关的信息。StructGPT 实现了迭代的读取-推理方法，构建专门的功能从结构化数据源中收集相关证据[489]。

Agentic RAG systems treat retrieval as dynamic operations where agents function as intelligent investigators analyzing content and cross-referencing information [648, 162, 965]. These systems incorporate sophisticated planning and reflection mechanisms requiring integration of task decomposition, multi-plan selection, and iterative refinement capabilities [438, 1183].
自主型 RAG 系统将检索视为动态操作，其中代理充当智能调查员，分析内容并进行信息交叉引用 [648, 162, 965]。这些系统整合了复杂的规划和反思机制，需要任务分解、多计划选择和迭代改进能力 [438, 1183]。

Modular RAG architectures enable flexible composition of retrieval components through standardized interfaces and plug-and-play designs. Graph-Enhanced RAG systems leverage structured knowledge representations for improved information access, while Real-time RAG implementations address dynamic information requirements in streaming applications [312, 1391].
模块化 RAG 架构通过标准化接口和即插即用设计，使检索组件的灵活组合成为可能。图增强型 RAG 系统利用结构化的知识表示以提高信息访问能力，而实时 RAG 实现则解决了流式应用中动态信息需求的问题 [312, 1391]。

The assembly function $\mathcal{A}$ encompasses template-based formatting, priority-based selection, and adaptive composition strategies that must adapt to varying task requirements, model capabilities, and resource constraints [702, 1133, 569]. Contemporary orchestration mechanisms manage agent selection, context distribution, and interaction flow control in multi-agent systems, enabling effective cooperation through user input processing, contextual distribution, and optimal agent selection based on capability assessment [894, 53, 171].
组装函数 $\mathcal{A}$ 包含基于模板的格式化、基于优先级的选择以及适应性组合策略，这些策略必须适应不同的任务需求、模型能力和资源限制 [702, 1133, 569]。现代协调机制管理多智能体系统中的代理选择、上下文分布和交互流程控制，通过用户输入处理、上下文分布和基于能力评估的最佳代理选择，实现有效的合作 [894, 53, 171]。

Advanced orchestration frameworks incorporate intent recognition, contextual memory maintenance, and task dispatching components for intelligent coordination across domain-specific agents. The Swarm Agent framework utilizes real-time outputs to direct tool invocations while addressing limitations in static tool registries and bespoke communication frameworks [808, 263, 246].
高级编排框架集成了意图识别、上下文记忆维护和任务调度组件，以实现跨领域代理的智能协调。Swarm Agent 框架利用实时输出来指导工具调用，同时解决了静态工具注册表和定制通信框架的局限性 [808, 263, 246]。

Context assembly must address cross-modal integration challenges, incorporating diverse data types including text, structured knowledge, temporal sequences, and external tool interfaces while maintaining coherent semantic relationships [529, 1221, 496]. Verbalization techniques convert structured data including knowledge graph triples, table rows, and database records into natural language sentences, enabling seamless integration with existing language systems without architectural modifications [12, 782, 1064, 13].
上下文组装必须解决跨模态集成的挑战，整合包括文本、结构化知识、时间序列和外部工具接口在内的多种数据类型，同时保持语义关系的一致性 [529, 1221, 496]。语言化技术将结构化数据（包括知识图谱三元组、表格行和数据库记录）转换为自然语言句子，使这些数据能够无缝集成到现有的语言系统中，而无需进行架构修改 [12, 782, 1064, 13]。

Programming language representations of structured data, particularly Python implementations for knowledge graphs and SQL for databases, outperform traditional natural language representations in complex reasoning tasks by leveraging inherent structural properties [1166]. Multi-level structurization approaches reorganize input text into layered structures based on linguistic relationships, while structured data representations leverage existing LLMs to extract structured information and represent key elements as graphs, tables, or relational schemas [681, 1125, 1324].
编程语言对结构化数据的表示，特别是用于知识图谱的 Python 实现和用于数据库的 SQL，在利用固有的结构特性进行复杂推理任务时，表现优于传统的自然语言表示[1166]。多级结构化方法根据语言关系将输入文本重新组织成分层结构，而结构化数据表示则利用现有的 LLMs 提取结构化信息，并将关键元素表示为图形、表格或关系模式[681, 1125, 1324]。

Automated prompt engineering addresses manual optimization limitations through systematic prompt generation and refinement algorithms. Automatic Prompt Engineer (APE) employs search algorithms for optimal prompt discovery, while LM-BFF introduces automated pipelines combining prompt-based fine-tuning with dynamic demonstration incorporation, achieving up to 30% absolute improvement across NLP tasks [307, 417, 590]. Promptbreeder implements self-referential evolutionary systems where LLMs improve both task-prompts and mutation-prompts governing these improvements through natural selection analogies [275, 508].
自动化提示工程通过系统化的提示生成和优化算法克服了手动优化的局限性。自动提示工程师（APE）利用搜索算法寻找最优提示，而 LM-BFF 引入了结合基于提示的微调与动态示例集成的自动化管道，实现了 NLP 任务中高达 30%的绝对改进 [307, 417, 590]。Promptbreeder 实现了自参照的进化系统，在这些系统中，LLM 不仅改进任务提示，还通过自然选择类比来控制这些改进的突变提示 [275, 508]。

Self-refine enables iterative output improvement through self-critique and revision across multiple iterations, with GPT-4 achieving approximately 20% absolute performance improvement through this methodology [735, 670]. Multi-agent collaborative frameworks simulate specialized team dynamics with agents assuming distinct roles (analysts, coders, testers), resulting in 29.9-47.1% relative improvement in Pass@1 metrics compared to single-agent approaches [434, 1257].
自我完善通过多次迭代中的自我批判和修订来逐步改进输出，GPT-4 通过这种方法实现了约 20%的绝对性能提升 [735, 670]。多代理协作框架模拟了专业团队动态，其中代理承担不同的角色（分析师、编码员、测试员），与单代理方法相比，在 Pass@1 指标上实现了 29.9%-47.1%的相对改进 [434, 1257]。

Tool integration frameworks combine Chain-of-Thought reasoning with external tool execution, automating intermediate reasoning step generation as executable programs strategically incorporating external data. LangChain provides comprehensive framework support for sequential processing chains, agent development, and web browsing capabilities, while specialized frameworks like Auto-GPT and Microsoft’s AutoGen facilitate complex AI agent development through user-friendly interfaces [963, 1087, 25, 867].
工具集成框架将思维链推理与外部工具执行相结合，自动化生成中间推理步骤，将其作为战略性地结合外部数据的可执行程序。LangChain 提供了一整套框架支持，用于顺序处理链、代理开发和网络浏览能力，而 Auto-GPT 和微软的 AutoGen 等专门框架则通过用户友好的界面促进了复杂 AI 代理的开发 [963, 1087, 25, 867]。

State Space Models (SSMs) maintain linear computational complexity and constant memory requirements through fixed-size hidden states, with models like Mamba offering efficient recurrent computation mechanisms that scale more effectively than traditional transformers [1258, 347, 346]. Dilated attention approaches like LongNet employ exponentially expanding attentive fields as token distance grows, achieving linear computational complexity while maintaining logarithmic dependency between tokens, enabling processing of sequences exceeding one billion tokens [216].
状态空间模型（SSMs）通过固定大小的隐藏状态保持线性计算复杂度和恒定的内存需求，例如 Mamba 模型提供了比传统变压器更有效的递归计算机制，能够更好地扩展[1258, 347, 346]。长网（LongNet）等扩展注意机制通过随着令牌距离增加而指数扩展的注意域，实现了线性计算复杂度，同时保持对数依赖关系，从而能够处理超过十亿个令牌的序列[216]。

Toeplitz Neural Networks (TNNs) model sequences with relative position encoded Toeplitz matrices, reducing space-time complexity to log-linear and enabling extrapolation from 512 training tokens to 14,000 inference tokens [868, 869]. Linear attention mechanisms reduce complexity from O(N²) to O(N) by expressing self-attention as linear dot-products of kernel feature maps, achieving up to 4000× speedup when processing very long sequences [522]. Alternative approaches like non-attention LLMs break quadratic barriers by employing recursive memory transformers and other architectural innovations [547].
Toeplitz 神经网络（TNNs）通过相对位置编码的 Toeplitz 矩阵建模序列，将时空复杂度降低到对数线性，并能够从 512 个训练令牌外推到 14,000 个推理令牌[868, 869]。线性注意机制通过将自注意力表示为核特征图的线性点积，将复杂度从 O(N²)降低到 O(N)，在处理非常长的序列时可实现高达 4000 倍的加速[522]。其他替代方法，如非注意 LLMs，通过使用递归记忆变压器和其他架构创新打破二次复杂度壁垒[547]。

Position interpolation techniques enable models to process sequences beyond original context window limitations by intelligently rescaling position indices rather than extrapolating to unseen positions [150]. Neural Tangent Kernel (NTK) approaches provide mathematically grounded frameworks for context extension, with YaRN combining NTK interpolation with linear interpolation and attention distribution correction [833, 471, 1021].
位置插值技术通过智能重新缩放位置索引而非外推到未见过的位置，使模型能够处理超出原始上下文窗口限制的序列 [150]。神经核方法（Neural Tangent Kernel, NTK）为上下文扩展提供了数学基础框架，YaRN 结合了 NTK 插值和线性插值以及注意力分布校正 [833, 471, 1021]。

LongRoPE achieves 2048K token context windows through two-stage approaches: first fine-tuning models to 256K length, then conducting positional interpolation to reach maximum context length [218]. Position Sequence Tuning (PoSE) demonstrates impressive sequence length extensions up to 128K tokens by combining multiple positional interpolation strategies [1377]. Self-Extend techniques enable LLMs to process long contexts without fine-tuning by employing bi-level attention strategies—grouped attention and neighbor attention—to capture dependencies among distant and adjacent tokens [499].
LongRoPE 通过两阶段方法实现了 2048K 标记的上下文窗口：首先将模型微调到 256K 长度，然后进行位置插值以达到最大上下文长度 [218]。位置序列调优（Position Sequence Tuning, PoSE）通过结合多种位置插值策略实现了高达 128K 标记的序列长度扩展 [1377]。自我扩展技术使 LLMs 能够在不进行微调的情况下处理长上下文，通过使用分层注意力策略——分组注意力和邻近注意力——来捕捉远距离和相邻标记之间的依赖关系 [499]。

Grouped-Query Attention (GQA) partitions query heads into groups that share key and value heads, striking a balance between multi-query attention and multi-head attention while reducing memory requirements during decoding [16, 1341]. FlashAttention exploits asymmetric GPU memory hierarchy to achieve linear memory scaling instead of quadratic requirements, with FlashAttention-2 providing approximately twice the speed through reduced non-matrix multiplication operations and optimized work distribution [196, 195].
Grouped-Query Attention (GQA) 将查询头分为共享键头和值头的组，既平衡了多查询注意力和多头注意力之间的关系，又在解码过程中减少了内存需求 [16, 1341]。FlashAttention 通过利用异构 GPU 内存层次结构，实现了线性内存扩展，而不是二次需求，FlashAttention-2 通过减少非矩阵乘法操作和优化工作分配，大约提高了两倍的速度 [196, 195]。

Ring Attention with Blockwise Transformers enables handling extremely long sequences by distributing computation across multiple devices, leveraging blockwise computation while overlapping communication with attention computation [676]. Sparse attention techniques include Shifted sparse attention (S²-Attn) in LongLoRA and SinkLoRA with SF-Attn, which achieve 92% of full attention perplexity improvement with significant computation savings [1304, 1217].
环形注意力与块状变换器结合使用，通过在多个设备之间分布计算来处理极其长的序列，同时利用块状计算并重叠通信与注意力计算 [676]。稀疏注意力技术包括 LongLoRA 和 SinkLoRA 中的 Shifted 稀疏注意力 (S²-Attn) 以及 SF-Attn，它们在显著节省计算资源的同时，实现了全注意力困惑度改进的 92% [1304, 1217]。

Efficient Selective Attention (ESA) proposes token-level selection of critical information through query and key vector compression into lower-dimensional representations, enabling processing of sequences up to 256K tokens [1084]. BigBird combines local attention with global tokens that attend to entire sequences, plus random connections, enabling efficient processing of sequences up to 8× longer than previously possible [1285].
高效选择性注意力（ESA）通过将查询和键向量压缩到低维表示来在 token 级别选择关键信息，从而能够处理多达 256K 个 token 的序列 [1084]。BigBird 结合了局部注意力与全局 token，后者可以关注整个序列，再加上随机连接，从而能够高效处理比以往长 8 倍的序列 [1285]。

Memory management strategies include Rolling Buffer Cache techniques that maintain fixed attention spans, reducing cache memory usage by approximately 8× on 32K token sequences [1341]. StreamingLLM enables processing infinitely long sequences without fine-tuning by retaining critical “attention sink” tokens together with recent KV cache entries, demonstrating up to 22.2× speedup over sliding window recomputation with sequences up to 4 million tokens [1176].
内存管理策略包括滚动缓冲缓存技术，该技术保持固定的关注跨度，对于 32K 个 token 的序列，可以将缓存内存使用量减少约 8 倍 [1341]。StreamingLLM 通过保留关键的“注意力汇流”token 以及最近的 KV 缓存条目，能够在无需微调的情况下处理无限长的序列，对于多达 400 万个 token 的序列，其速度比滑动窗口重新计算快 22.2 倍 [1176]。

Infini-attention incorporates compressive memory into vanilla attention, combining masked local attention with long-term linear attention in single Transformer blocks, enabling processing of infinitely long inputs with bounded memory and computation [792]. Heavy Hitter Oracle (H₂O) presents efficient KV cache eviction policies based on observations that small token portions contribute most attention value, improving throughput by up to 29× while reducing latency by up to 1.9× [1333].
Infini-注意力将压缩记忆融入到基础注意力中，在单个 Transformer 块中结合了掩码局部注意力和长期线性注意力，从而能够在有限的内存和计算资源下处理无限长的输入 [792]。Heavy Hitter Oracle（H ₂ O）通过观察到小部分 token 贡献了最多的注意力值，提出了高效的 KV 缓存淘汰策略，吞吐量最多可提高 29 倍，同时延迟最多可降低 1.9 倍 [1333]。

Context compression techniques like QwenLong-CPRS implement dynamic context optimization mechanisms enabling multi-granularity compression guided by natural language instructions [944]. InfLLM stores distant contexts in additional memory units and employs efficient mechanisms to retrieve token-relevant units for attention computation, allowing models pre-trained on sequences of a few thousand tokens to effectively process sequences up to 1,024K tokens [1175].
像 QwenLong-CPRS 这样的上下文压缩技术实现了由自然语言指令引导的动态上下文优化机制，支持多粒度压缩 [944]。InfLLM 在额外的内存单元中存储远程上下文，并采用高效机制检索与 token 相关的单元进行注意力计算，使得预训练于几千个 token 序列的模型能够有效处理多达 1,024K 个 token 的序列 [1175]。

The Self-Refine framework uses the same model as generator, feedback provider, and refiner, demonstrating that identifying and fixing errors is often easier than producing perfect initial solutions [735, 1313, 227]. Reflexion maintains reflective text in episodic memory buffers for future decision-making through linguistic feedback [956], while structured guidance proves essential as simplistic prompting often fails to enable reliable self-correction [672, 587].
Self-Refine 框架使用相同的模型作为生成器、反馈提供者和精炼器，表明识别和修复错误通常比生成完美的初始解决方案更容易 [735, 1313, 227]。Reflexion 通过语言反馈在情景记忆缓冲区中保持反思性文本，以供未来决策使用 [956]，而结构化指导对于确保可靠的自我纠正至关重要，因为简单的提示往往无法实现有效的自我纠正 [672, 587]。

Multi-Aspect Feedback integrates frozen language models and external tools focusing on specific error categories to enable more comprehensive, independent evaluation [799]. The N-CRITICS framework implements ensemble-based evaluation where initial outputs are assessed by both generating LLMs and other models, with compiled feedback guiding refinement until task-specific stopping criteria are fulfilled [789].
多方面反馈将冻结的语言模型和专注于特定错误类别的外部工具结合起来，以实现更全面、独立的评估[799]。N-CRITICS 框架采用基于集成的评估方法，初始输出由生成 LLM 和其他模型共同评估，汇总的反馈指导细化，直到满足特定任务的停止标准[789]。

The A2R framework adopts explicit evaluation across multiple dimensions including correctness and citation quality, formulating natural language feedback for each aspect and iteratively refining outputs [577]. ISR-LLM improves LLM-based planning by translating natural language to formal specifications, creating an initial plan, and then systematically refining it with a validator [1373].
A2R 框架在多个维度上进行显式的评估，包括正确性和引用质量，为每个方面制定自然语言反馈，并迭代细化输出[577]。ISR-LLM 通过将自然语言转换为形式化规范，生成初始计划，然后系统地用验证器进行细化，从而改进基于 LLM 的规划[1373]。

SELF teaches LLMs meta-skills (self-feedback, self-refinement) with limited examples, then has the model continuously self-evolve by generating and filtering its own training data [704]. Self-rewarding mechanisms enable models to improve autonomously through iterative self-judgment, where a single model adopts dual roles as performer and judge, maximizing rewards it assigns itself [1163, 1278].
SELF 通过少量示例教会 LLMs 元技能（自我反馈、自我完善），然后让模型通过生成和过滤自己的训练数据不断自我进化 [704]。自我奖励机制使模型能够通过迭代自我判断来自主改进，其中单个模型同时扮演执行者和评判者的角色，最大化它为自己分配的奖励 [1163, 1278]。

The Creator framework extends this paradigm by enabling LLMs to create and use their own tools through a four-module process encompassing creation, decision-making, execution, and recognition [946, 856]. The Self-Developing framework represents the most autonomous approach, enabling LLMs to discover, implement, and refine their own improvement algorithms through iterative cycles generating algorithmic candidates as executable code [466].
Creator 框架扩展了这一范式，通过一个包含创造、决策、执行和识别四个模块的过程，使 LLMs 能够创建和使用自己的工具 [946, 856]。Self-Developing 框架代表了最自主的方法，使 LLMs 能够通过迭代周期发现、实现和完善自己的改进算法，生成可执行的算法候选代码 [466]。

In-context learning fundamentally represents a form of meta-learning where models learn optimization strategies during pre-training that generalize across diverse tasks, enabling rapid adaptation to novel challenges during inference [179, 1165]. Meta-in-context learning demonstrates that in-context learning abilities can be recursively improved through in-context learning itself, adaptively reshaping model priors over expected tasks and modifying in-context learning strategies [177].
在上下文学习本质上是一种元学习的形式，在预训练过程中，模型学习到的优化策略能够跨多种任务泛化，从而在推理过程中快速适应新的挑战 [179, 1165]。递归的上下文学习表明，可以通过自身进行的上下文学习来递归地提升上下文学习能力，适应性地重塑模型对预期任务的先验，并修改上下文学习策略 [177]。

Memory augmentation represents a powerful approach for implementing meta-learning through frameworks like Memory of Amortized Contexts, which uses feature extraction and memory-augmentation to compress information from new documents into compact modulations stored in memory banks [1011]. Context-aware Meta-learned Loss Scaling addresses outdated knowledge challenges by meta-training small autoregressive models to dynamically reweight language modeling loss for each token during online fine-tuning [430].
记忆增强是一种通过使用如“摊还上下文记忆”框架实现元学习的强大方法，该框架利用特征提取和记忆增强将新文档中的信息压缩成紧凑的模态存储在记忆库中 [10, 11]。上下文感知的元学习损失缩放通过元训练小型自回归模型，在在线微调过程中动态重新加权每个标记的语言建模损失，从而解决过时知识问题 [430]。

Decision-Pretrained Transformers demonstrate how transformers can be trained to perform in-context reinforcement learning, solving previously unseen RL problems by generalizing beyond pretraining distribution [1013, 582]. Context-based meta-reinforcement learning methods enhance performance through direct supervision of context encoders, improving sample efficiency compared to end-to-end training approaches [1072].
决策预训练变换器展示了如何训练变换器执行上下文中的强化学习，通过泛化超出预训练分布来解决以前未见过的 RL 问题 [1013, 582]。基于上下文的元强化学习方法通过直接监督上下文编码器来提升性能，相比端到端训练方法，提高了样本效率 [1072]。

Long Chain-of-Thought has emerged as a significant evolution characterized by substantially longer reasoning traces enabling thorough problem exploration, as implemented in advanced models including OpenAI-o1, DeepSeek-R1, QwQ, and Gemini 2.0 Flash Thinking [147, 718, 1214]. LongCoT effectiveness appears linked to context window capacity, with empirical evidence suggesting larger context windows often lead to stronger reasoning performance [1229].
长链推理(Long Chain-of-Thought)已成为一种重要的演变，其特征是推理痕迹显著加长，从而能够进行彻底的问题探索，这在包括 OpenAI-o1、DeepSeek-R1、QwQ 和 Gemini 2.0 Flash Thinking 在内的先进模型中得到了实现[147, 718, 1214]。长链推理的有效性似乎与上下文窗口容量相关，实验证据表明，更大的上下文窗口往往会导致更强的推理性能[1229]。

Extended reasoning enables self-reflection and error correction mechanisms allowing models to identify and rectify mistakes during problem-solving processes [1334]. The effectiveness of increasing reasoning step length, even without adding new information, considerably enhances reasoning abilities across multiple datasets through test-time scaling [1345].
扩展推理使模型能够进行自我反思和错误修正，使其在解决问题过程中能够识别并纠正错误[1334]。即使不增加新信息，增加推理步骤长度的有效性也通过测试时的扩展显著提升了多种数据集上的推理能力[1345]。

Optimization strategies address computational inefficiencies due to verbose reasoning traces through self-generated shorter reasoning paths via best-of-N sampling, adaptive reasoning modes including Zero-Thinking and Less-Thinking approaches, and explicit compact CoT methods reducing token usage while maintaining reasoning quality [791, 1348, 697]. Auto Long-Short Reasoning enables dynamic adjustment of reasoning path length according to question complexity, helping models decide when longer chains are necessary [715].
优化策略通过自动生成较短的推理路径并利用最佳 N 次采样来解决冗长的推理痕迹带来的计算效率低下问题，采用适应性推理模式，包括零思考和少思考的方法，并通过显式的紧凑型推理方法减少令牌使用量同时保持推理质量 [791, 1348, 697]。自动长短推理能够根据问题的复杂性动态调整推理路径长度，帮助模型决定何时需要更长的推理链 [715]。

Multimodal MLLMs expand upon traditional LLMs by integrating data from diverse modalities like vision, audio, and 3D environments [105, 49, 957]. A primary integration method converts visual inputs into discrete tokens concatenated with text tokens, conditioning the LLM’s generative process on a combined representation [1286]. This is often facilitated by Visual Prompt Generators (VPGs) trained on image-caption pairs to map visual features into the LLM’s embedding space [607]. The dominant architectural paradigm connects specialized, external multimodal encoders—such as CLIP for vision or CLAP for audio—to the LLM backbone via alignment modules like Q-Former or simple MLPs [19, 86, 609, 1130], a modular design that allows for independent encoder updates without retraining the entire model [618].
多模态 LLMs 在传统 LLMs 的基础上，通过整合来自多种模态的数据，如视觉、音频和 3D 环境[105, 49, 957]。主要的整合方法是将视觉输入转换为离散的标记，并将其与文本标记连接起来，使 LLM 的生成过程基于联合表示进行条件化[1286]。这通常通过训练在图像-描述对上进行的视觉提示生成器（VPGs）来实现，以将视觉特征映射到 LLM 的嵌入空间[607]。主流的架构范式是通过对齐模块（如 Q-Former 或简单的 MLP）将专门的外部多模态编码器（如用于视觉的 CLIP 或用于音频的 CLAP）连接到 LLM 的主干上[19, 86, 609, 1130]，这种模块化设计允许独立更新编码器而不重新训练整个模型[618]。

More sophisticated approaches enable deeper modality fusion. Cross-modal attention mechanisms learn fine-grained dependencies between textual and visual tokens directly within the LLM’s embedding space, enhancing semantic understanding for tasks like image editing [564, 901, 102]. To manage lengthy inputs, hierarchical designs process modalities in stages to ensure scalability [155], while the “browse-and-concentrate” paradigm fuses the contexts of multiple images before LLM ingestion to overcome the limitations of isolated processing [1134]. Some research bypasses the adaptation of text-only LLMs, opting for unified training paradigms that jointly pre-train models on multimodal data and text corpora from the start to mitigate alignment challenges [1381, 1224]. Other methods leverage text as a universal semantic space, using LLM in-context learning to improve generalization across diverse modality combinations [1050]. For video, context integration techniques range from prompt tuning to adapter-based methods that transform video content into a sequence for reasoning [1080]. The development of these models is often constrained by the need for vast, high-quality multimodal data and significant computational resources [1295, 609, 211].
更复杂的方案能够实现更深层次的模态融合。跨模态注意力机制直接在 LLM 的嵌入空间中学习文本和视觉标记之间的细微依赖关系，增强语义理解，适用于图像编辑等任务[564, 901, 102]。为了处理长输入，分层设计分阶段处理模态，以确保可扩展性[155]，而“浏览和集中”范式在 LLM 摄入之前将多个图像的上下文融合起来，以克服孤立处理的限制[1134]。一些研究绕过了仅针对文本的 LLM 的适应，选择统一的训练范式，从一开始就联合预训练多模态数据和文本语料库，以减轻对齐挑战[1381, 1224]。其他方法利用文本作为通用的语义空间，使用 LLM 的上下文学习来提高在不同模态组合中的泛化能力[1050]。对于视频，上下文集成技术包括提示调优到基于适配器的方法，后者将视频内容转换为推理所需的序列[1080]。 这些模型的发展往往受限于需要大量高质量的多模态数据和强大的计算资源 [1295, 609, 211]。

A primary obstacle in MLLM development is modality bias, where models favor textual inputs, generating plausible but multimodally ungrounded responses by relying on learned linguistic patterns rather than integrated visual or auditory information [1358, 24, 315, 1325]. This issue is exacerbated by training methodologies; for instance, VPGs trained on simple image-captioning tasks learn to extract only salient features for captions, neglecting other visual details crucial for more complex, instruction-based tasks, which fundamentally limits deep multimodal understanding [607, 504]. Consequently, MLLMs frequently struggle with fine-grained spatial or temporal reasoning, such as precise object localization or understanding detailed event sequences in videos [1031, 957], particularly in complex domains like social media where interpreting the interplay of text and images to understand misinformation or sarcasm is difficult [505]. Effective multimodal reasoning requires not just comprehending each modality but also inferring their combined holistic meaning [385]. Compounding these issues is our limited mechanistic understanding of MLLMs themselves; their internal workings are largely a black box, hindering the development of better architectures [1274].
MLLM 开发中的主要障碍是模态偏见，模型倾向于偏好文本输入，通过依赖于学习到的语言模式而不是整合的视觉或听觉信息，生成合理但跨模态不一致的响应 [1358, 24, 315, 1325]。训练方法加剧了这一问题；例如，用于简单图像-标题任务的 VPGs 仅学会提取对标题重要的特征，而忽略了对更复杂的基于指令的任务至关重要的其他视觉细节，这从根本上限制了深度跨模态理解 [607, 504]。因此，MLLMs 在精细的空间或时间推理方面经常遇到困难，例如精确的对象定位或理解视频中的详细事件序列 [1031, 957]，尤其是在像社交媒体这样的复杂领域中，解读文本和图像之间的相互作用以理解误导信息或讽刺尤为困难 [505]。有效的跨模态推理不仅需要理解每个模态，还需要推断它们的综合整体意义 [385]。 这些问题还受到我们对 MLLM 本身有限的机械理解的影响；它们的内部工作机制大多是一个黑箱，阻碍了更好架构的发展 [1274]。

A key capability of MLLMs is in-context learning, where models adapt to new tasks from multimodal examples in the prompt without weight updates [1397, 1398, 551]. Link-context learning (LCL) enhances this by providing demonstrations with explicit causal links, improving generalization [1012]. However, in-context learning is constrained by fixed context windows, as image tokens consume significant space, limiting many-shot learning [437]. Performance is also sensitive to input order and the relative importance of each modality varies by task [1020, 1197]. Processing long multimodal contexts, crucial for applications like video analysis, remains a major research frontier [1086]. Innovations include adaptive hierarchical token compression for video [1119], variable visual position encoding (V2PE) [1381], specialized modules like ContextQFormer for conversational memory [589], and dynamic, query-aware frame selection for video [581]. MLLMs also show emergent communication efficiency over extended interactions, a phenomenon still under investigation [436].
大语言模型（MLLMs）的一个关键能力是上下文学习，模型可以在不更新权重的情况下，根据提示中的多模态示例适应新任务 [1397, 1398, 551]。链接上下文学习（LCL）在此基础上进一步提供带有明确因果链接的示范，从而提高泛化能力 [1012]。然而，上下文学习受到固定上下文窗口的限制，因为图像标记会占用大量空间，限制了多示例学习 [437]。性能还对输入顺序敏感，每种模态的相对重要性也因任务而异 [1020, 1197]。处理长多模态上下文对于视频分析等应用来说仍然是一个主要的研究前沿 [1086]。创新包括针对视频的自适应分层标记压缩 [1119]、可变视觉位置编码（V2PE）[1381]、专门的模块如用于对话记忆的 ContextQFormer [589]，以及动态、查询感知的帧选择方法 [581]。大语言模型在长时间交互中也表现出新兴的通信效率，这一现象仍在研究之中 [436]。

The ability to process rich multimodal context is unlocking new applications. MLLMs are used for predictive reasoning, such as forecasting human activity from visual scenes [1382], and have demonstrated impressive perception and cognitive capabilities across various multimodal benchmarks [290]. In VQA, context is leveraged for more precise answers, for instance, by prompting the MLLM to generate its own descriptive text context of an image [1346] or by integrating external knowledge via RAG [993, 105]. Other applications include planning digital actions based on sensory inputs [605], enhancing surgical decision support through memory-augmented context comprehension [418], and enabling nuanced video understanding by integrating visual information with speech and audio cues [642, 1193, 7]. Researchers have also extended MLLMs to emerging modalities like tactile information, event data, and graph structures [1358, 1023, 1213]. The growing importance of these real-world use cases has spurred the development of comprehensive evaluation frameworks to assess contextual comprehension [1109]. These advancements enable applications previously impossible with text-only models, such as image captioning and sophisticated multimodal reasoning [1173, 677, 139].
处理丰富多模态上下文的能力正在解锁新的应用。大规模语言模型（MLLM）用于预测推理，例如，从视觉场景预测人类活动[1382]，并在各种多模态基准测试中展示了令人印象深刻的感知和认知能力[290]。在视觉问答（VQA）中，上下文被用于更精确的答案，例如，通过提示 MLLM 生成图像的描述性文本上下文[1346]，或通过 RAG 整合外部知识[993, 105]。其他应用包括基于感官输入规划数字动作[605]，通过记忆增强的上下文理解支持手术决策[418]，以及通过结合视觉信息和语音、音频提示来实现细腻的视频理解[642, 1193, 7]。研究人员还扩展了 MLLM 到新兴模态，如触觉信息、事件数据和图结构[1358, 1023, 1213]。这些实际应用的重要性日益增长，推动了全面评估框架的发展，以评估上下文理解能力[1109]。 这些进步使得以前仅靠文本模型无法实现的应用成为可能，例如图像字幕和复杂的多模态推理 [1173, 677, 139]。

Advanced encoding strategies address structural limitations through knowledge graph embeddings that transform entities and relationships into numerical vectors, enabling efficient processing within language model architectures [12, 1250, 930, 1194]. Graph neural networks capture complex relationships between entities, facilitating multi-hop reasoning across knowledge graph structures through specialized architectures like GraphFormers that nest GNN components alongside transformer blocks [974, 404, 1221, 483].
高级编码策略通过知识图嵌入将实体和关系转换为数值向量，从而克服结构限制，在语言模型架构中实现高效处理 [12, 1250, 930, 1194]。图神经网络捕捉实体之间的复杂关系，通过如 GraphFormers 等特殊架构，将图神经网络组件嵌套在变压器块中，从而在知识图结构上实现多跳推理 [974, 404, 1221, 483]。

GraphToken demonstrates substantial improvements by explicitly representing structural information, achieving up to 73 percentage points enhancement on graph reasoning tasks through parameter-efficient encoding functions [836]. Heterformer and other hybrid GNN-LM architectures perform contextualized text encoding and heterogeneous structure encoding in unified models, addressing the computational challenges of scaling these integrated systems [496, 465, 751].
GraphToken 通过显式表示结构信息，实现了显著改进，在图推理任务中通过参数高效编码函数实现了高达 73 个百分点的提升 [836]。Heterformer 及其他混合 GNN-语言模型架构在统一模型中实现上下文化文本编码和异构结构编码，解决了这些集成系统扩展的计算挑战 [496, 465, 751]。

Verbalization techniques convert structured data including knowledge graph triples, table rows, and database records into natural language sentences, enabling seamless integration with existing language systems without architectural modifications [12, 782, 1064, 13]. Multi-level structurization approaches reorganize input text into layered structures based on linguistic relationships, while structured data representations leverage existing LLMs to extract structured information and represent key elements as graphs, tables, or relational schemas [681, 1125, 1324, 1035, 602].
口头化技术将结构化数据（包括知识图谱三元组、表格行和数据库记录）转换为自然语言句子，从而无需对现有语言系统进行架构修改即可实现无缝集成 [12, 782, 1064, 13]。多层次结构化方法根据语言关系重新组织输入文本为分层结构，而结构化数据表示则利用现有 LLMs 提取结构化信息，并将关键元素表示为图形、表格或关系模式 [681, 1125, 1324, 1035, 602]。

Programming language representations of structured data, particularly Python implementations for knowledge graphs and SQL for databases, outperform traditional natural language representations in complex reasoning tasks by leveraging inherent structural properties [1166]. Resource-efficient approaches using structured matrix representations offer promising directions for reducing parameter counts while maintaining performance on structured data tasks [343].
结构化数据的编程语言表示，特别是用于知识图谱的 Python 实现和用于数据库的 SQL，在利用固有的结构特性进行复杂推理任务时表现优于传统的自然语言表示 [1166]。利用结构化矩阵表示的资源高效方法为在保持结构化数据任务性能的同时减少参数数量提供了有前景的方向 [343]。

The integration of knowledge graphs with language models follows distinct paradigms characterized by different implementation strategies and performance trade-offs [817, 1140]. Pre-training integration methods like K-BERT inject knowledge graph triples during training to internalize factual knowledge, while inference-time approaches enable real-time knowledge access without requiring complete model retraining [690, 1237, 712].
将知识图谱与语言模型集成遵循不同的范式，这些范式具有不同的实现策略和性能权衡[817, 1140]。预训练集成方法如 K-BERT 在训练过程中注入知识图谱三元组以内化事实知识，而推理时的方法则允许实时访问知识，无需重新训练整个模型[690, 1237, 712]。

KG-enhanced LLMs incorporate structured knowledge to improve factual grounding through retrieval-based augmentation methods like KAPING, which retrieves relevant facts based on semantic similarities and prepends them to prompts without requiring model training [48, 673, 591]. More sophisticated implementations embed KG-derived representations directly into model latent spaces through adapter modules and cross-attention mechanisms, with Text2Graph mappers providing linking between input text and KG embedding spaces [132, 1066, 428].
KG 增强的 LLMs 通过检索增强方法（如 KAPING）整合结构化知识，以提高事实基础，这些方法基于语义相似性检索相关事实，并将它们添加到提示中，而无需进行模型训练[48, 673, 591]。更复杂的实现方式直接将 KG 衍生的表示嵌入到模型的潜在空间中，通过适配器模块和跨注意力机制，文本到图的映射器提供输入文本与 KG 嵌入空间之间的链接[132, 1066, 428]。

Synergized approaches create unified systems where both technologies play equally important roles, addressing fundamental limitations through bidirectional reasoning driven by data and knowledge [817, 853, 1111]. GreaseLM facilitates deep interaction across all model layers, allowing language context representations to be grounded by structured world knowledge while linguistic nuances inform graph representations [1321]. QA-GNN implements bidirectional attention mechanisms connecting question-answering contexts and knowledge graphs through joint graph formation and mutual representation updates via graph-based message passing [1250, 974].
协同方法创建了统一系统，在这些系统中，两种技术都扮演着同等重要的角色，通过数据和知识驱动的双向推理来解决根本性限制 [817, 853, 1111]。GreaseLM 促进了模型各层的深度交互，使语言上下文表示能够基于结构化世界知识，而语言学细微差别则指导图表示 [1321]。QA-GNN 通过联合图构建和基于图的消息传递进行的相互表示更新，实现了问题回答上下文与知识图谱之间的双向注意机制 [1250, 974]。

Structured data integration significantly enhances LLM capabilities across multiple dimensions, with knowledge graphs providing structured information that reduces hallucinations by grounding responses in verifiable facts and improving factual accuracy through clearly defined information sources [1002, 1342, 200, 565]. Knowledge graphs enhance reasoning capabilities by providing structured entity relationships that enable complex multi-hop reasoning and logical inferences, with their rich repository of hierarchical knowledge significantly improving precision and reliability of inferences [1166, 208, 1018].
结构化数据集成在多个维度上显著增强了 LLM 的能力，知识图谱提供了结构化的信息，通过将响应扎根于可验证的事实，减少了幻觉现象，并通过明确的信息来源提高了事实准确性 [1002, 1342, 200, 565]。知识图谱通过提供结构化的实体关系，增强了推理能力，使复杂的多跳推理和逻辑推断成为可能，其丰富的层次知识库显著提高了推断的精确性和可靠性 [1166, 208, 1018]。

Real-world applications demonstrate substantial improvements across specialized domains. Healthcare systems combine structured medical knowledge with contextual understanding through Retrieval-Augmented Generation frameworks to improve disease progression modeling and clinical decision-making [842, 583]. Scientific research platforms organize findings into structured knowledge supporting hypothesis generation and research gap identification, while business analytics systems balance rule-based precision with AI pattern recognition for more actionable insights [1326, 1062].
现实世界的应用展示了在专业领域中的显著改进。医疗系统通过检索增强生成框架，结合结构化的医学知识和上下文理解，以提高疾病进展建模和临床决策的效果[842, 583]。科研平台将研究成果组织成结构化的知识，支持假设生成和研究空白识别，而商业分析系统则在基于规则的精确性和 AI 模式识别之间取得平衡，以获得更具行动指导意义的洞察[1326, 1062]。

Question answering systems benefit from natural language interfaces over structured data sources, with integration creating more robust systems capable of handling multimodal queries and providing personalized responses that overcome static knowledge base limitations [1317, 1116, 914, 1206]. Research demonstrates that structured knowledge representations can improve summarization performance by 40% and 14% across public datasets compared to unstructured memory approaches, with Chain-of-Key strategies providing additional performance gains through dynamic structured memory updates [459].
问答系统从自然语言接口而非结构化数据源中受益，集成可以创建更 robust 的系统，能够处理多模态查询并提供个性化响应，从而克服静态知识库的局限性 [13, 17, 116, 914, 1206]。研究表明，结构化知识表示在公共数据集上相比非结构化记忆方法可提高总结性能 40%和 14%，而通过动态结构化记忆更新的 Chain-of-Key 策略还能提供额外的性能提升 [459]。

LLM memory systems can be organized into multiple classification frameworks. The primary temporal classification divides memory into three categories: sensory memory (input prompts), short-term memory (immediate context processing), and long-term memory (external databases or dedicated structures) [935]. From a persistence perspective, short-term memory includes key-value caches and hidden states existing only within single sessions, while long-term memory encompasses text-based storage and knowledge embedded in model parameters, persisting across multiple interaction cycles [935, 818].
LLM 的内存系统可以组织成多种分类框架。主要的时间分类将记忆分为三类：感觉记忆（输入提示）、短期记忆（即时上下文处理）和长期记忆（外部数据库或专用结构）[935]。从持久性角度来看，短期记忆包括仅存在于单个会话中的键值缓存和隐藏状态，而长期记忆则包括基于文本的存储和嵌入在模型参数中的知识，这些知识可以在多个交互周期中持续存在[935, 818]。

Implementation-based classifications identify parametric memory (knowledge encoded in model weights), ephemeral activation memory (context-limited runtime states), and plaintext memory accessed through Retrieval-Augmented Generation methods [637]. Current implementations lack sophisticated lifecycle management and multi-modal integration, limiting long-term knowledge evolution. Feed-forward network layers serve as key-value tables storing memory, functioning as “inner lexicon” for word retrieval and creating mechanisms analogous to human associative memory [518, 325, 326, 764, 464]. These classification schemes reflect attempts to develop LLM memory architectures paralleling human cognitive systems [1167].
基于实现的分类识别参数化记忆（知识编码在模型权重中）、瞬态激活记忆（受上下文限制的运行时状态）以及通过检索增强生成方法访问的明文记忆 [637]。当前的实现缺乏复杂的生命周期管理和多模态集成，限制了长期知识的进化。前馈网络层充当存储记忆的键值表，作为“内部词典”用于单词检索，并创建类似于人类联想记忆的机制 [518, 325, 326, 764, 464]。这些分类方案反映了试图开发与人类认知系统相平行的 LLM 记忆架构的努力 [1167]。

Short-term memory in LLMs operates through the context window, serving as working memory maintaining immediate access to previously processed tokens [1282]. This functionality is implemented through key-value caches storing token representations but disappearing when sessions terminate [891]. Architectural variations demonstrate significant differences: transformer-based models implement working memory systems flexibly retrieving individual token representations across arbitrary delays, while LSTM architectures maintain coarser, rapidly-decaying semantic representations weighted toward earliest items [40].
LLMs 的短期记忆通过上下文窗口运作，作为工作记忆保持对已处理令牌的即时访问[1282]。这种功能是通过键值缓存存储令牌表示实现的，但在会话结束时消失[891]。架构变体显示出显著差异：基于变压器的模型灵活地在任意延迟下检索单个令牌表示，而 LSTM 架构则保持较粗略、快速衰减的语义表示，侧重于最早项目[40]。

Modern LLM short-term memory frequently manifests as in-context learning, reflecting models’ ability to acquire and process information temporarily within context windows [1180, 103]. This enables few-shot learning and task adaptation without parameter updates. Research identifies three primary memory configurations: full memory (utilizing entire context history), limited memory (using context subsets), and memory-less operation (without historical context) [1044]. Despite advances expanding context windows to millions of tokens, LLMs struggle with effective reasoning over extended contexts, particularly when relevant information appears in middle positions [891, 685].
现代 LLM 短期记忆经常表现为上下文学习，反映了模型在上下文窗口内临时获取和处理信息的能力 [1180, 103]。这使得模型能够在不更新参数的情况下实现少样本学习和任务适应。研究识别了三种主要的记忆配置：全记忆（利用整个上下文历史）、有限记忆（使用上下文子集）和无记忆操作（没有历史上下文） [1044]。尽管通过扩展上下文窗口到数百万个标记取得了进展，但 LLM 在处理长上下文的有效推理方面仍然存在困难，尤其是在相关信息出现在中间位置时 [891, 685]。

LLMs face significant challenges maintaining long-term memory due to context window limitations and catastrophic forgetting [114]. External memory-based methods address these limitations by utilizing physical storage to cache historical information, allowing relevant history retrieval without maintaining all information within constrained context windows [682, 1362]. These approaches contrast with internal memory-based methods focusing on reducing self-attention computational costs to expand sequence length [682, 287].
LLMs 因为上下文窗口限制和灾难性遗忘，面临着长期记忆维护的重大挑战 [114]。基于外部存储的方法通过利用物理存储来缓存历史信息，从而在不保持所有信息在受限上下文窗口内的前提下检索相关历史信息 [682, 1362]。这些方法与专注于通过减少自注意力计算成本来扩展序列长度的基于内部存储的方法形成对比 [682, 287]。

Long-term memory implementations categorize into knowledge-organization methods (structuring memory into interconnected semantic networks), retrieval mechanism-oriented approaches (integrating semantic retrieval with forgetting curve mechanisms), and architecture-driven methods (implementing hierarchical structures with explicit read-write operations) [515, 1362, 444]. Memory storage representations can be further divided into token-level memory (information stored as structured text for direct retrieval) and latent-space memory (utilizing high-dimensional vectors for abstract and compact information representation) [1216, 1124]. Advanced approaches incorporate psychological principles, with MemoryBank implementing Ebbinghaus Forgetting Curve theory for selective memory preservation based on temporal factors [1362], emotion-aware frameworks employing Mood-Dependent Memory theory [444], and memorization mechanisms balancing performance advantages with privacy concerns through extraction vulnerability analysis [1041, 122, 123].
长期记忆实现方法可以分为知识组织方法（将记忆组织成相互连接的语义网络）、检索机制导向的方法（将语义检索与遗忘曲线机制相结合）以及架构驱动的方法（采用分层结构并明确进行读写操作）[515, 1362, 444]。记忆存储表示可以进一步分为标记级记忆（以结构化文本形式存储信息，以便直接检索）和潜在空间记忆（利用高维向量进行抽象和紧凑的信息表示）[1216, 1124]。高级方法结合了心理学原理，MemoryBank 通过实施艾宾浩斯遗忘曲线理论，根据时间因素选择性地保存记忆[1362]，情绪感知框架则采用情绪依赖记忆理论[444]，而记忆机制则通过提取漏洞分析平衡性能优势与隐私问题[1041, 122, 123]。

LLMs exhibit characteristic memory access patterns with notable similarities to human cognitive processes, demonstrating clear primacy and recency effects when recalling information lists [477]. Memory retrieval operates through sequential access (retrieving content in consecutive order) and random access (accessing information from arbitrary points without processing preceding content) [1387]. Memory persistence studies employ recognition experiments, recall experiments, and retention experiments to quantify information accessibility duration and retrieval conditions [810], with cognitive psychology concepts like semantic and episodic memory integration improving LLM information synthesis capabilities [240].
LLMs 表现出与人类认知过程显著相似的记忆访问模式，在回忆信息列表时表现出明显的首因效应和近因效应[477]。记忆检索通过顺序访问（按连续顺序获取内容）和随机访问（从任意点访问信息而不处理先前内容）进行[1387]。记忆持久性研究通过识别实验、回忆实验和保持实验来量化信息可访问时间和检索条件[810]，而认知心理学概念如语义记忆和情景记忆的整合则提高了 LLM 的信息合成能力[240]。

Memory organization encompasses diverse structural approaches including textual-form storage (complete and recent agent-environment interactions, retrieved historical interactions, external knowledge), knowledge representation structures (chunks, knowledge triples, atomic facts, summaries, mixed approaches), hierarchical systems with library-enhanced reasoning components, and functional patterns organized by tasks, temporal relevance, or semantic relationships [1329, 1290, 1027]. Core memory operations include encoding (transforming textual information into latent space embeddings), retrieval (accessing relevant information based on semantic relevance, importance, and recency), reflection (extracting higher-level insights), summarization (condensing texts while highlighting critical points), utilization (integrating memory components for unified outputs), forgetting (selective information discarding), truncation (formatting within token limitations), and judgment (assessing information importance for storage prioritization) [1331]. These structures offer varying trade-offs between comprehensiveness, retrieval efficiency, and computational requirements.
内存组织涵盖了多种结构方法，包括文本形式存储（完整的和最近的代理-环境交互，检索的历史交互，外部知识）、知识表示结构（块、知识三元组、原子事实、摘要、混合方法）、带有图书馆增强推理组件的分层系统，以及按任务、时间相关性或语义关系组织的功能模式 [1329, 1290, 1027]。核心内存操作包括编码（将文本信息转换为潜在空间嵌入）、检索（基于语义相关性、重要性和最近性访问相关信息）、反思（提取高层次见解）、总结（压缩文本并突出关键点）、利用（将内存组件整合为统一输出）、遗忘（选择性地丢弃信息）、截断（在标记限制内格式化）以及判断（评估信息的重要性以确定存储优先级） [1331]。这些结构在全面性、检索效率和计算需求之间提供了不同的权衡。

Contemporary LLM agents employ memory systems analogous to computer memory hierarchies, with short-term memory functioning as primary storage for contextual understanding within context windows, while long-term memory serves as persistent storage for extended information retention [770]. From object-oriented perspectives, AI systems generate personal memories related to individual users and system memories containing intermediate task results [1167]. Structured frameworks like MemOS classify memory into Parametric Memory (knowledge encoded in model weights), Activation Memory, and Plaintext Memory, with parametric memory representing long-term knowledge embedded within feedforward and attention layers enabling zero-shot generation [637].
当代 LLM 代理使用类似于计算机内存层次结构的内存系统，短期记忆作为上下文窗口内语境理解的主要存储，而长期记忆则作为扩展信息保留的持久存储 [770]。从面向对象的角度来看，AI 系统生成与个别用户相关的个人记忆，以及包含中间任务结果的系统记忆 [1167]。结构化的框架如 MemOS 将内存分类为参数化内存（模型权重中编码的知识）、激活内存和明文内存，参数化内存代表嵌入在前馈和注意力层中的长期知识，使零样本生成成为可能 [637]。

Memory integration frameworks have evolved to address LLM limitations through sophisticated architectures. The Self-Controlled Memory (SCM) framework enhances long-term memory through LLM-based agent backbones, memory streams, and memory controllers managing updates and utilization [649]. The REMEMBERER framework equips LLMs with experience memory exploiting past episodes across task goals, enabling success/failure learning without parameter fine-tuning through verbal reinforcement and self-reflective feedback mechanisms [1299]. Advanced systems like MemLLM implement structured read-write memory modules addressing challenges in memorizing rare events, updating information, and preventing hallucinations [779]. Autonomous agents leveraging LLMs rely on four essential components—perception, memory, planning, and action—working together to enable environmental perception, interaction recall, and real-time planning and execution [614, 38].
记忆整合框架已经通过复杂的架构来解决 LLM 的局限性。Self-Controlled Memory（SCM）框架通过基于 LLM 的代理后端、记忆流和管理更新与利用的记忆控制器来增强长期记忆[649]。REMEMBERER 框架为 LLM 配备了经验记忆，利用跨任务目标的过往经历，通过口头强化和自我反思反馈机制实现成功/失败学习，而无需参数微调[1299]。先进的系统如 MemLLM 实现了结构化的读写记忆模块，以应对记忆稀有事件、更新信息和防止幻觉的挑战[779]。利用 LLM 的自主代理依赖于四个关键组件——感知、记忆、规划和行动，这些组件共同工作以实现环境感知、交互回忆和实时规划与执行[614, 38]。

Memory-enhanced LLM agents have demonstrated transformative impact across diverse application domains. In conversational AI, memory systems enable more natural, human-like interactions by recalling past experiences and user preferences to deliver personalized, context-aware responses. Commercial implementations include Charlie Mnemonic (combining Long-Term, Short-Term, and episodic memory using GPT-4), Google Gemini (leveraging long-term memory for personalized experiences across Google’s ecosystem), and ChatGPT Memory (remembering conversations across sessions) [578]. User simulation applications employ LLM-powered conversational agents mimicking human behavior for cost-effective dialogue system evaluation, adapting flexibly across open-domain dialogues, task-oriented interactions, and conversational recommendation [204], with systems like Memory Sandbox enabling user control over conversational memories through data object manipulation [455].
增强记忆的 LLM 代理已在多种应用领域展示了变革性的影响。在对话 AI 中，记忆系统通过回忆过去的经历和用户偏好，提供更加自然、人性化的交互，从而实现个性化、上下文感知的响应。商业实现包括 Charlie Mnemonic（结合长期记忆、短期记忆和情景记忆，使用 GPT-4），Google Gemini（利用长期记忆为 Google 生态系统中的个性化体验提供支持），以及 ChatGPT Memory（在会话之间记住对话内容）[578]。用户模拟应用使用基于 LLM 的对话代理，模仿人类行为以经济高效地评估对话系统，灵活适应开放领域对话、任务导向交互和对话推荐[204]，如 Memory Sandbox 系统允许用户通过数据对象操作控制对话记忆[455]。

Task-oriented agents utilize memory to perform complex autonomous operations with minimal human intervention, employing LLMs as controllers extended through multimodal perception, tool utilization, and external memory [1160]. Applications span recommendation systems (RecMind providing personalized recommendations through planning and external knowledge, InteRecAgent employing LLMs with recommender models as tools), autonomous driving (DiLu instilling human-like knowledge through reasoning, reflection, and memory), scientific research (ChemCrow automating chemical synthesis design and execution), and social simulation (generative agents exhibiting believable behavior through memory storage and synthesis) [1019, 647, 92, 825]. Proactive conversational agents address challenges in strategic dialogue scenarios requiring goal-oriented conversation steering through prompt-based policy planning methods and AI feedback generation based on dialogue history [204, 203].
面向任务的代理利用记忆来执行复杂的自主操作，同时减少人类干预，通过多模态感知、工具利用和外部记忆将 LLMs 作为控制器扩展[1160]。应用范围包括推荐系统（RecMind 通过规划和外部知识提供个性化推荐，InteRecAgent 利用 LLMs 和推荐模型作为工具）、自动驾驶（DiLu 通过推理、反思和记忆灌输人类般知识）、科学研究（ChemCrow 自动化化学合成设计和执行）以及社会模拟（通过记忆存储和合成表现出可信行为的生成代理）[1019, 647, 92, 825]。主动对话代理通过基于提示的策略规划方法和基于对话历史的 AI 反馈生成解决战略对话场景中的挑战，实现目标导向的对话引导[204, 203]。

Personalized assistant applications leverage memory to maintain coherent long-term relationships with users, with memory components serving as structured repositories storing contextually relevant information including user preferences and historical interactions [438]. Domain-specific implementations include healthcare assistants employing memory coordination for medical interactions [1316, 1307], recommendation agents leveraging external knowledge bases [1316, 1293], educational agents providing context-aware support through memory-enabled progress tracking [647], and specialized frameworks like MARK enhancing personalized AI assistants through user preference memory [299].
个性化助理应用利用记忆来维持与用户的长期关系，记忆组件作为结构化的存储库，存储与用户偏好和历史交互相关的信息 [438]。特定领域的实现包括医疗助手利用记忆协调进行医疗交互 [1316, 1307]、推荐代理利用外部知识库 [1316, 1293]、教育代理通过记忆启用的进度跟踪提供上下文感知支持 [647]，以及像 MARK 这样的专门框架通过用户偏好记忆增强个性化 AI 助理 [299]。

Memory technology evolution addresses fundamental context window limitations through RAG, which combines parametric and non-parametric memory for language generation using pre-trained seq2seq models and dense vector indices [1209, 591]. This approach enables access to information beyond parameter storage without requiring retraining, significantly extending knowledge capabilities. Advanced memory mechanisms including vector databases and retrieval-augmented generation enable vast information storage with quick relevant data access, incorporating short-term contextual memory and long-term external storage [38, 367, 1184, 507].
内存技术演进通过 RAG（检索增强生成）解决了基本的上下文窗口限制，RAG 结合了参数化和非参数化内存，使用预训练的 seq2seq 模型和密集向量索引[1209, 591]。这种方法能够在不重新训练的情况下访问超出参数存储的信息，显著扩展了知识能力。先进的内存机制包括向量数据库和检索增强生成，能够存储大量信息并快速访问相关数据，结合了短期上下文记忆和长期外部存储[38, 367, 1184, 507]。

Non-parametric approaches maintain frozen LLM parameters while leveraging external resources like RAG to enrich task contexts [934]. Systems like Reflexion implement verbal reinforcement through self-reflective feedback in episodic memory buffers, while REMEMBERER incorporates persistent experience memory enabling learning from past successes and failures. Advanced architectures like MemoryBank enable memory retrieval, continuous evolution through updates, and personality adaptation by integrating previous interaction information [1202, 1362].
非参数化方法保持 LLM 参数不变，同时利用外部资源如 RAG 来丰富任务上下文[934]。例如，Reflexion 系统通过自省反馈在情景记忆缓冲区中实现口头强化，而 REMEMBERER 系统则整合持久的经验记忆，使系统能够从过去的成功和失败中学习。先进的架构如 MemoryBank 能够实现内存检索、通过更新持续进化以及通过整合先前交互信息进行个性适应[1202, 1362]。

Specialized memory architectures address particular agent requirements through sophisticated organization and retrieval mechanisms. While early systems required predefined storage structures and retrieval timing, newer systems like Mem0 incorporate graph databases following RAG principles for more effective memory organization and relevance-based retrieval [1202]. Commercial and open-source implementations including OpenAI ChatGPT Memory, Apple Personal Context, mem0, and MemoryScope demonstrate widespread adoption of memory systems for enhanced personalization capabilities [1167]. Tool-augmentation paradigms validate effectiveness in complex task decomposition while leveraging world interaction tools, with memory-enhanced agents becoming central to modern AI systems performing complex tasks through natural language integration of planning, tool use, memory, and multi-step reasoning [247, 356, 1091, 34].
专门化的内存架构通过复杂的组织和检索机制来满足特定的代理需求。早期的系统需要预定义的存储结构和检索时间，而像 Mem0 这样的新系统则采用了遵循 RAG 原则的图数据库，以更有效地组织记忆并基于相关性进行检索[1202]。商业和开源实现，包括 OpenAI ChatGPT Memory、Apple Personal Context、mem0 和 MemoryScope，展示了记忆系统在增强个性化能力方面的广泛应用[1167]。工具增强范式在复杂任务分解中验证了其有效性，同时利用世界交互工具，记忆增强的代理已成为现代 AI 系统的核心，这些系统通过自然语言整合规划、工具使用、记忆和多步推理来执行复杂任务[247, 356, 1091, 34]。

Contemporary memory evaluation employs specialized metrics extending beyond traditional NLP performance indicators to capture nuanced memory functionality aspects [1330]. Effectiveness metrics focus on factual information storage and utilization through accuracy measures (correctness of responses based on historical messages) and recall@5 indicators (percentage of relevant messages retrieved within top-5 results). Efficiency metrics examine temporal aspects through response time (duration for information retrieval and utilization) and adaptation time (period required for new information storage) [1330].
当代的记忆评估采用了超越传统 NLP 性能指标的专业化指标，以捕捉记忆功能的细微方面 [1330]。有效性指标侧重于事实信息的存储和利用，通过准确度指标（基于历史消息的响应正确性）和召回@5 指标（在前 5 个结果中检索到的相关消息的百分比）来衡量。效率指标则通过响应时间（信息检索和利用所需的时间）和适应时间（新信息存储所需的时间段）来考察时间方面的因素 [1330]。

Extensive benchmarks such as LongMemEval assess five fundamental long-term memory capabilities: information extraction, temporal reasoning, multi-session reasoning, knowledge updates, and abstention through 500 carefully selected questions, demonstrating 30% accuracy degradation in commercial assistants throughout prolonged interactions, while automated memory evaluation frameworks facilitate thorough assessment extending beyond passkey search methodologies [1171]. Dedicated frameworks target episodic memory via benchmarks assessing temporally-situated experiences, with research demonstrating that cutting-edge models including GPT-4, Claude variants, and Llama 3.1 encounter difficulties with episodic memory challenges involving interconnected events or intricate spatio-temporal associations even in comparatively brief contexts [457]. Contemporary LLM benchmarks predominantly concentrate on assessing models’ retention of factual information and semantic relationships while substantially overlooking episodic memory assessment—the capacity to contextualize memories with temporal and spatial occurrence details [841].
广泛的基准测试，如 LongMemEval 评估了五种基本的长期记忆能力：信息提取、时间推理、多会话推理、知识更新和弃权，通过 500 个精心挑选的问题，展示了在长时间交互中商业助手的准确率下降 30%。自动化记忆评估框架则促进了超越密码搜索方法的全面评估[1171]。专门的框架通过评估时间定位体验的基准测试来针对情景记忆，研究表明，包括 GPT-4、Claude 变体和 Llama 3.1 在内的最新模型，在涉及相互关联事件或复杂时空关联的情景记忆挑战中，即使在相对简短的上下文中也遇到困难[457]。当前的 LLM 基准测试主要集中在评估模型对事实信息和语义关系的保留能力，而严重忽视了情景记忆的评估——即能够用时间与空间发生的细节来情境化记忆的能力[841]。

Task-specific evaluations encompass long-context passage retrieval (locating specific paragraphs within extended contexts), long-context summarization (developing comprehensive understanding for concise summaries), NarrativeQA (answering questions based on lengthy narratives), and specialized benchmarks like MADail-Bench evaluating both passive and proactive memory recall in conversational contexts with novel dimensions including memory injection, emotional support proficiency, and intimacy assessment [1329, 1380, 550, 386]. Additional task-specific frameworks include QMSum for meeting summarization, QuALITY for reading comprehension, DialSim for dialogue-based QA requiring spatiotemporal memory, and MEMENTO for personalized embodied agent evaluation using two-stage processes to assess memory utilization in physical environment tasks [1380, 566].
任务特定的评估包括长上下文段落检索（在扩展的上下文中定位特定段落）、长上下文总结（发展全面理解以生成简洁摘要）、NarrativeQA（基于长篇叙述回答问题）以及像 MADail-Bench 这样的专门基准，评估在对话上下文中被动和主动记忆召回，包括记忆注入、情感支持能力和亲密性评估等新维度[1329, 1380, 550, 386]。其他任务特定框架包括 QMSum 用于会议总结、QuALITY 用于阅读理解、DialSim 用于基于对话的问答需要空间时间记忆，以及 MEMENTO 用于个性化实体代理评估，采用两阶段过程评估物理环境任务中的记忆利用情况[1380, 566]。

Memory evaluation faces substantial challenges limiting effective assessment of capabilities. Fundamental limitations include absence of consistent, rigorous methodologies for assessing memory performance, particularly regarding generalization beyond training data [284]. The lack of standardized benchmarks specifically designed for long-term memory evaluation represents another significant obstacle, with existing frameworks often failing to capture the full spectrum of memory capabilities needed for human-like intelligence [1071].
记忆评估面临着重大挑战，限制了对能力的有效评估。基本的限制包括缺乏一致且严谨的评估方法，特别是在评估记忆性能以及其在训练数据之外的泛化能力方面 [284]。缺乏专门设计用于长期记忆评估的标准基准测试是另一个重大障碍，现有的框架往往无法捕捉到实现人类智能所需的所有记忆能力的完整范围 [1071]。

Architectural constraints significantly complicate evaluation efforts, as most contemporary LLM-based agents operate in fundamentally stateless manners, treating interactions independently without truly accumulating knowledge incrementally over time [1355, 1354], despite advances in working memory through attentional tagging mechanisms enabling flexible memory representation control [864]. This limitation prevents genuine lifelong learning assessment—a cornerstone of human-level intelligence involving continuous knowledge acquisition, retention, and reuse across diverse contexts and extended time horizons.
架构限制显著复杂化了评估工作，因为大多数基于 LLM 的代理在本质上是以无状态的方式运行的，它们将交互视为独立事件，而没有真正随着时间的推移逐步积累知识 [1355, 1354]，尽管通过注意力标记机制在工作记忆方面取得了进展，使得灵活的记忆表示控制成为可能 [864]。这一限制阻碍了真正的终身学习评估——这是人类水平智能的核心，涉及在多种不同情境和长时间跨度内持续的知识获取、保持和重用。

Methodological issues arise when isolating memory-specific performance from other intelligence aspects, challenging determination of whether failures stem from inadequate memory mechanisms or reasoning limitations [284]. Dynamic memory usage in real-world applications poses evaluation challenges, as controlled laboratory tests inadequately capture memory system performance in complex scenarios where information relevance changes unpredictably [1071].
在隔离与记忆相关的性能与其他智能方面时，方法论问题会浮现，这使得确定失败是源于不足的记忆机制还是推理限制变得困难 [284]。在实际应用中，动态内存使用带来的评估挑战在于，受控的实验室测试无法充分捕捉到在信息相关性不可预测变化的复杂场景中内存系统的表现 [1071]。

Memory optimization encompasses diverse techniques enhancing utilization while minimizing computational overhead and maximizing efficiency. Biologically-inspired forgetting mechanisms provide effective optimization approaches, with frameworks like MemoryBank implementing Ebbinghaus forgetting curves to selectively preserve and discard information based on temporal factors and significance [1362]. Reflection-based optimization through systems like Reflexion enables performance assessment through integrated evaluation and self-reflection, creating dual feedback systems refining memory and behavior through continuous learning [300].
内存优化涵盖了多种技术，这些技术在提高利用率的同时，尽量减少计算开销并最大化效率。受生物学启发的遗忘机制提供了有效的优化方法，例如 MemoryBank 框架通过实现艾宾浩斯遗忘曲线来根据时间因素和重要性选择性地保存和丢弃信息 [1362]。基于反思的优化通过如 Reflexion 这样的系统，能够通过集成评估和自我反思进行性能评估，创建双重反馈系统，通过持续学习不断优化记忆和行为 [300]。

Hierarchical memory structures optimize information organization through multi-level formats enabling efficient retrieval, demonstrated by Experience-based Hierarchical Control frameworks with rapid memory access modules [862], memory consolidation processes through bidirectional fast-slow variable interactions [63], and Adaptive Cross-Attention Networks dynamically ranking memories based on query relevance [406].
分层内存结构通过多级格式优化信息组织，实现高效检索，例如经验驱动的分层控制框架，具有快速记忆访问模块 [862]；通过双向快慢变量交互进行记忆巩固过程 [63]；以及自适应交叉注意力网络，根据查询相关性动态排名记忆 [406]。

Future research directions encompass hybrid memory frameworks combining parametric precision with non-parametric efficiency [934], automated feedback mechanisms for scalable response evaluation [885], multi-agent memory systems enabling collaborative learning through shared external memories [302], enhanced metadata learning with knowledge graph integration [888, 382], domain-specific memory architectures for specialized applications [501], cognitive-inspired optimization incorporating memory consolidation during inactive periods [752], and parameter-efficient memory updates through techniques like Low-Rank Adaptation for efficient knowledge integration [424, 252]. These developments promise advancing memory-enhanced LLM agents toward sophisticated, human-like cognitive capabilities while addressing computational and architectural limitations, with applications extending to long-term robotic planning, real-world decision-making systems, and collaborative AI assistants through streaming learning scenarios and continuous feedback integration [1150, 1336, 1269].
未来的研究方向包括结合参数精度与非参数效率的混合记忆框架[934]、可扩展响应评估的自动化反馈机制[885]、多智能体记忆系统，通过共享外部记忆实现协作学习[302]、与知识图谱集成的元数据学习增强[888, 382]、针对特定应用的认知专用记忆架构[501]、受认知启发的优化，在非活跃期整合记忆巩固[752]、以及通过低秩适应等技术实现参数高效的记忆更新，以高效整合知识[424, 252]。这些发展有望推动增强记忆的 LLM 代理向高级、类人的认知能力迈进，同时解决计算和架构限制，应用范围扩展到长期机器人规划、现实世界决策系统以及通过流式学习场景和持续反馈整合的协作 AI 助手[1150, 1336, 1269]。

Training methodologies evolved from basic prompt-based approaches to sophisticated multi-task learning frameworks, with fine-tuning on specialized datasets through systems like ToolLLM and Granite-20B-FunctionCalling, beginning with synthetic single-tool data followed by human annotations [388, 5, 353, 771, 1226].
训练方法从基本的提示基方法发展到复杂的多任务学习框架，通过 ToolLLM 和 Granite-20B-FunctionCalling 等系统在专门的数据集上进行微调，最初使用合成的单工具数据，随后进行人工标注[388, 5, 353, 771, 1226]。

Data generation strategies include Weaver’s GPT-4-based environment synthesis, APIGen’s hierarchical verification pipelines (format checking, function execution, semantic verification), generating 60,000+ high-quality entries across thousands of APIs [1104, 1177, 1259, 1156, 65, 1393, 743].
数据生成策略包括 Weaver 基于 GPT-4 的环境合成、APGen 的分层验证管道（格式检查、函数执行、语义验证），生成了超过 60,000 个高质量的 API 条目 [1104, 1177, 1259, 1156, 65, 1393, 743]。

Tool selection enhancement involves irrelevance-aware data augmentation, with Hammer’s function masking techniques, oracle tool mixing for increased difficulty, tool intent detection synthesis for over-triggering mitigation, emphasizing high-quality data through stringent filtering and format verification [664, 10, 353, 467, 1291, 214].
工具选择增强涉及无相关性数据增强，使用 Hammer 的功能掩码技术，通过 Oracle 工具混合增加难度，通过工具意图检测合成减轻过度触发，强调高质量数据通过严格的过滤和格式验证 [664, 10, 353, 467, 1291, 214]。

Self-improvement paradigms reduce external supervision dependence through JOSH algorithm’s sparse reward simulation environments and TTPA’s token-level optimization with error-oriented scoring, demonstrating improvements while preserving general capabilities [573, 440, 362, 1262].
自我改进范式通过 JOSH 算法的稀疏奖励模拟环境和 TTPA 的基于错误的 token 级优化减少对外部监督的依赖，同时展示改进并保留通用能力 [573, 440, 362, 1262]。

Sophisticated benchmarks include API-Bank (73 APIs, 314 dialogues), StableToolBench (API instability solutions), NesTools (nested tool evaluation), ToolHop (995 queries, 3,912 tools), addressing single-tool to multi-hop scenarios [615, 359, 373, 1255, 821, 987, 1248, 979].
复杂的基准测试包括 API-Bank（73 个 API，314 个对话），StableToolBench（API 不稳定性解决方案），NesTools（嵌套工具评估），ToolHop（995 个查询，3,912 个工具），解决单工具到多跳场景 [615, 359, 373, 1255, 821, 987, 1248, 979]。

Single-tool frameworks established foundational principles of tool-integrated reasoning through specialized implementations targeting specific computational domains. Program-Aided Language Models (PAL) pioneered problem decomposition strategies by generating executable code while delegating mathematical computations to Python interpreters [305]. ToolFormer demonstrated that language models could learn external API usage with minimal demonstrations, incorporating calculators, search engines, and diverse tools to enhance computational capabilities [931]. ToRA advanced mathematical reasoning by integrating natural language processing with computational libraries and symbolic solvers, while ReTool applied reinforcement learning to optimize code interpreter usage, demonstrating improvements in self-correction patterns [341, 1311, 965]. Self-Edit utilizes execution results of generated code to improve code quality for competitive programming tasks, employing a fault-aware code editor to correct errors based on test case results [1309].
单工具框架通过专门实现，为工具集成推理奠定了基础原则，针对特定的计算领域。程序辅助语言模型（PAL）通过生成可执行代码并委托数学计算给 Python 解释器，率先提出了问题分解策略[305]。ToolFormer 展示了语言模型在最少示范下学习外部 API 使用的能力，将计算器、搜索引擎和各种工具整合进来，以增强计算能力[931]。ToRA 通过将自然语言处理与计算库和符号求解器集成，推进了数学推理，而 ReTool 则利用强化学习优化代码解释器的使用，展示了自我纠正模式的改进[341, 1311, 965]。Self-Edit 利用生成代码的执行结果来提高代码质量，适用于编程竞赛任务，采用一个基于测试案例结果的故障感知代码编辑器来纠正错误[1309]。

Multi-tool coordination systems address the complexity of orchestrating heterogeneous tools within integrated reasoning architectures. ReAct pioneered the interleaving of reasoning traces with task-specific actions, enabling models to think and act complementarily where reasoning supports plan tracking while actions interface with external information sources [1245]. Chameleon introduced plug-and-play compositional reasoning by synthesizing programs combining vision models, search engines, and Python functions with an LLM-based planner core [709]. AutoTools established automated frameworks transforming raw tool documentation into executable functions, reducing manual engineering requirements in tool integration [419, 952]. Chain-of-Agents (CoA) trains models to decode reasoning chains with abstract placeholders, subsequently calling domain-specific tools to fill knowledge gaps [594, 1327].
多工具协调系统解决了在集成推理架构中协调异构工具的复杂性。ReAct 首次提出了将推理痕迹与任务特定操作交错的方法，使模型在推理支持计划跟踪的同时，能够与外部信息源进行互补操作 [1245]。Chameleon 通过结合视觉模型、搜索引擎和基于 LLM 的规划核心的 Python 函数，引入了插拔式组合推理 [709]。AutoTools 建立了自动化框架，将原始工具文档转换为可执行函数，减少了工具集成中的手动工程需求 [419, 952]。Chain-of-Agents（CoA）训练模型解码抽象占位符的推理链，随后调用特定领域的工具来填补知识空白 [594, 1327]。

Agent-based frameworks represent the most sophisticated evolution of TIR systems, moving beyond static prompting approaches to create autonomous and adaptive AI systems. Unlike conventional tool-use that follows rigid patterns, agent models learn to couple Chain-of-Thought (CoT) and Chain-of-Action (CoA) patterns into their core behavior, resulting in stronger logical coherence and natural transitions between reasoning and action [1328]. These systems build upon foundational agent architectures including reactive systems that map perceptions directly to actions, deliberative systems implementing Belief-Desire-Intention (BDI) models, and hybrid architectures combining multiple subsystems in hierarchical structures [728].
基于代理的框架代表了 TIR 系统最复杂的进化，超越了静态提示方法，创建了自主且适应性强的 AI 系统。与遵循固定模式的传统工具使用不同，代理模型学会将思维链（CoT）和行动链（CoA）模式整合到其核心行为中，从而增强了逻辑连贯性并在推理与行动之间实现了更自然的过渡[1328]。这些系统建立在基础代理架构之上，包括直接将感知映射到行动的反应性系统、实现信念-欲望-意图（BDI）模型的审思性系统，以及将多个子系统以分层结构结合的混合架构[728]。