context management

Current AI coding agents often produce code requiring extensive debugging due to insufficient contextual understanding, thereby diminishing development efficiency. This work proposes a three-phase preparation methodology inspired by the culinary concept of “mise en place”—termed MEP—comprising context anchoring, collaborative specification formulation, and task decomposition, which enhances agent coding performance through structured contextualization. The study introduces “contextual fluency” as a novel developer competency, integrating backward design principles and theories of tacit knowledge externalization. It employs structured documentation, human–AI collaborative dialogues, and dependency-aware task logging to operationalize this approach. In a hackathon setting, just two hours of preparatory work enabled multiple AI agents to concurrently and effectively construct a complete educational platform, substantially reducing the overall development cycle.

This study addresses the frequent inefficiencies in human-AI collaboration caused by incomplete contextual information, which often leads to excessive iteration and suboptimal output quality. To mitigate this, the authors propose a structured context construction framework that integrates a five-role context package—comprising authority, exemplars, constraints, evaluation criteria, and metadata—within a four-stage workflow encompassing review, design, construction, and audit. Notably, this work pioneers the incorporation of information theory and reliability engineering principles into context quality assessment, yielding a reusable and auditable collaboration framework. Empirical results from 200 interaction trials demonstrate that the approach reduces the average number of iterations from 3.8 to 2.0, increases first-pass success rates from 32% to 55%, and achieves a final task success rate of 91.5%.

This work addresses inaccurate planning in tool invocation caused by the absence of explicit dependency modeling. We propose a dual-graph collaborative framework integrating a tool knowledge graph with a domain-specific document knowledge graph. Methodologically: (1) we automatically construct a tool dependency graph from tool schemas; (2) we extract procedural knowledge from SOP documents to build a domain workflow graph; and (3) we introduce a novel deep sparse ensemble strategy to align and fuse the two graphs at both structural and semantic levels. Our key contribution lies in explicitly modeling implicit tool dependencies and organically coupling tool capabilities with domain workflow logic. Experiments demonstrate that our framework significantly improves planning rationality, tool invocation accuracy, and contextual consistency in example artifact generation, outperforming existing tool-augmented reasoning methods across multiple benchmark tasks.

This work proposes Meta Context Engineering (MCE), a novel framework that overcomes the structural biases and limited optimization space inherent in existing context engineering approaches, which rely on manually designed, fixed pipelines. MCE introduces a bilevel agent co-evolution paradigm that dynamically optimizes both the structure and content of prompts through agent crossover, skill history tracing, execution feedback, and programmable context representations. By moving beyond static heuristics, the method significantly enhances context adaptability, transferability, and training efficiency. Empirical evaluations across five tasks demonstrate consistent improvements, yielding relative performance gains ranging from 5.6% to 53.8% (average 16.9%) over state-of-the-art agent-based context engineering methods.

Knowledge engineering (KE) faces significant challenges in constructing large-scale, dynamic, multilingual, and multimodal knowledge graphs (KGs). Method: This study employs a hackathon-style mixed-methods approach—including interviews, ethnographic observation, artifact analysis, and empirical LLM experiments—to investigate how large language models (LLMs) can serve as effective collaborative assistants for knowledge engineers. Contribution/Results: We identify prompt engineering as a critical yet underappreciated core competency in KE practice. We introduce “KG Cards,” the first responsible AI framework specifically designed for KG construction, addressing ethical implementation gaps. Empirical results demonstrate that LLMs substantially improve KG construction efficiency; however, new bottlenecks emerge in trustworthiness assessment, cross-lingual alignment, and accountability governance. Collectively, this work provides empirically grounded guidelines and methodological foundations for human-AI collaboration in knowledge engineering.

This study addresses the challenge of capturing and sharing users’ cognitive traces in complex knowledge work, which often leaves AI systems without a meaningful understanding of authentic task context. The authors propose a novel paradigm centered on in-situ, user-generated snippet memoing—fragmentary notes that serve as primary sources of contextual information for AI. For the first time, this approach treats such cognitive traces as first-class inputs and grants users direct editorial control over their contextual representations. Through a probe system comprising interaction mechanisms and a context-organizing interface, users can instantly record, inspect, and restructure their thought processes, thereby guiding AI to produce responses more aligned with their intent. Experimental results demonstrate that this method significantly enhances users’ task awareness and ability to structure thinking, with 78.1% of participants preferring AI responses grounded in snippet-based context, leading to greater perceived agency and ownership in collaborative creation.

This work addresses the challenge that large language models struggle to effectively extract and apply novel knowledge from complex task contexts during in-context learning. To overcome this limitation, the authors propose a high-fidelity chain-of-thought synthesis mechanism that enhances the model’s understanding and utilization of task-specific context through the generation of high-quality synthetic reasoning paths. The proposed approach substantially improves model performance on context-dependent tasks, achieving a significant increase in average solve rate—from 17.2% to a markedly higher level—on the CL-Bench benchmark. This advancement effectively narrows the performance gap between state-of-the-art models on such challenging tasks.

This study addresses the challenge of designing effective inference-time harnesses to improve the long-term execution success of large language model agents in complex tasks. Recognizing that excessive decomposition or guidance can degrade performance, the work formalizes the harness design as a trajectory alignment problem and decouples it into two mechanisms: task decomposition and guided execution. It systematically investigates the impact of workflow granularity, retry budgets, and action reweighting. The authors propose a “partial harness” strategy—specifying only the initial steps—which effectively steers execution while avoiding failure modes such as over-decomposition, over-pruning, and hallucinated actions. Empirical validation in both synthetic environments and real-world terminal agent tasks demonstrates that this approach significantly enhances task completion rates, outperforming fully structured workflows.

This work addresses the limitations of current large language models in tool invocation, which rely heavily on in-context documentation and examples, leading to high inference overhead and susceptibility to hallucination, while conventional fine-tuning struggles to effectively internalize specific tool knowledge. To overcome these challenges, the authors propose ParaTool, a novel framework that shifts tool representation from the context into the model’s parameter space. Through a three-stage process—parameterized pretraining, gated network-driven soft tool selection, and parameterized joint fine-tuning—ParaTool enables dynamic, lightweight tool calling without dependence on contextual cues. Experiments demonstrate that this approach significantly outperforms strong in-context learning baselines on the Stable ToolBench and BFCL benchmarks, achieving higher tool invocation accuracy while reducing computational complexity.

This work addresses the challenge of formalizing human tacit knowledge—such as implicit assumptions, contextual constraints, and experiential judgments—in procedural tasks into machine-reasonable representations. To this end, the authors propose a novel neuro-symbolic framework that integrates logic-augmented generation with active inference to automatically extract tacit knowledge from instructional videos depicting assembly and maintenance procedures, and to construct an ontology-aligned knowledge graph. By synergistically combining neural perception with symbolic reasoning, the approach significantly enhances the semantic completeness and reusability of the resulting knowledge graph. Empirical validation in manufacturing and maintenance scenarios demonstrates its effectiveness, thereby advancing the application of neuro-symbolic systems in industrial knowledge engineering.