2026 Survey of Recent Advances in Prompt Engineering and Context Engineering: 8 Key Papers

Forum topic · 小凯 · 2026-02-20

Summary

This roundup reviews eight research papers published by February 20, 2026, covering the latest advances in prompt engineering and context engineering for large language models (LLMs). On the prompt engineering side, the papers include a systematic review of prompting paradigms in organic chemistry (ChemRxiv), a case-study collection on prompt optimization in data science, an evaluation of domain-knowledge strategies for high-dimensional LLM optimization in software engineering, and a reporting guideline for LLM prompting studies based on analysis of ~300 papers and a survey of 105 reviewers. On the context engineering side, highlights are a 9,649-experiment study of structured context for file-native agentic systems across 11 models and 4 formats, the CEA framework for reliable deep-research agents using a reinforcement learning method called CERL, an empirical study of AI context files (AGENTS.md) across 466 open-source projects, and a representation-engineering approach (RepE + One-Class SVM) for conversational context classification. Key takeaways: model capability outweighs format or architecture choices, structured context enables effective high-dimensional optimization, and the field is shifting from empirical prompt tricks toward systematic, reproducible engineering practice.

2026 Survey of Recent Advances in Prompt Engineering and Context Engineering

> This post summarizes 8 important research papers (as of February 20, 2026) in prompt engineering and context engineering, spanning chemistry, software engineering, data science, finance, and other application domains.

Key points

Paradigm shift: The field is evolving from single-turn prompt tricks (few-shot, CoT, ToT) toward systematic prompt design and, more broadly, context engineering for multi-turn, multimodal, and agentic long-context scenarios.
Model capability dominates: McMillan's 9,649-experiment study shows a 21-percentage-point accuracy gap between frontier and open-source models — far exceeding any format or architecture effect.
Format matters less than assumed: Across YAML, Markdown, JSON, and TOON, format had no significant effect on overall accuracy (chi-squared = 2.45, p = 0.484), though individual open-source models showed format-specific sensitivity.
Reporting standards are emerging: Korn et al. found a significant mismatch between current practice and reviewer expectations in LLM prompting research across ~300 SE papers, proposing a tiered guideline (Essential / Desirable / Exceptional).
Context rot is real: Longer context does not guarantee better responses; the CEA framework addresses this with history management, progress tracking, and key-cue identification, trained via CERL (filtering non-CEA-attributable error trajectories before gradient updates).
AGENTS.md adoption: A study of 466 open-source projects found no standardized content structure yet, but identified five context-provision styles: descriptive, prescriptive, prohibitive, explanatory, and conditional.

Prompt Engineering Papers

1. Systematic review of prompting in organic chemistry

Paper: *A Systematic Review of Prompt Engineering Paradigms in Organic Chemistry: Mining, Prediction, and Model Architectures* — F. Pourgholamali et al., ChemRxiv, Feb 4, 2026
Link: https://chemrxiv.org/doi/10.26434/chemrxiv-2026-625v3
Screened 101 records and included 22 high-quality studies; clarifies terminology inconsistencies and maps three LLM architectures (encoder-only, decoder-only, encoder-decoder) to chemistry tasks such as molecular property prediction, synthesis route design, and reaction condition optimization.

2. Prompt engineering case studies in data science

Paper: *Smarter AI Through Prompt Engineering: Insights and Case Studies from Data Science Application* — S. Paul, arXiv:2602.00337, Jan 30, 2026
Link: https://arxiv.org/abs/2602.00337
Shows prompt optimization can significantly improve LLM output quality without updating model parameters; covers data cleaning, feature engineering, model selection, and result interpretation; emphasizes iterative optimization and version control.

3. Domain knowledge strategies for high-dimensional optimization

Paper: *Beyond the Prompt: Assessing Domain Knowledge Strategies for High-Dimensional LLM Optimization in Software Engineering* — S. Srinivasan, T. Menzies, MSR 2026, arXiv:2602.02752
Link: https://arxiv.org/abs/2602.02752
LLMs underperform Bayesian methods on high-dimensional optimization (beyond ~11 features). Four architectures evaluated: H-DKP (human-in-the-loop domain knowledge prompting), AMP (adaptive multi-stage prompting), DAPR (dimension-aware progressive refinement), and HKMA (TPE statistical scouting + RAG-augmented prompting). Structured knowledge integration enables effective warm starts, but method choice should follow task characteristics.

4. Reporting guideline for LLM prompting in SE

Paper: *Reporting LLM Prompting in Automated Software Engineering: A Guideline Based on Current Practices and Expectations* — A. Korn et al., FORGE 2026, arXiv:2601.01954
Link: https://arxiv.org/abs/2601.01954
Analyzed ~300 top-venue SE papers since 2022 and surveyed 105 program committee members. Gaps exist in version disclosure, prompt justification, and validity threats. Proposed guideline tiers: Essential (full prompt text, model version/parameters, design rationale), Desirable (testing process, failure analysis, sensitivity analysis), Exceptional (multi-model comparison, cost-benefit analysis, reproducibility package).

Context Engineering Papers

5. Structured context for file-native agentic systems

Paper: *Structured Context Engineering for File-Native Agentic Systems* — Damon McMillan, arXiv:2602.05447 (v2, Feb 12)
Link: https://arxiv.org/abs/2602.05447
9,649 experiments, 11 models (frontier + open-source), 4 formats (YAML, Markdown, JSON, TOON), schemas scaling from 10 to 10,000 tables.
Findings: file retrieval accuracy improved +2.7% (p=0.029) for frontier models but dropped -7.7% (p<0.001) for open-source models; format had no significant overall effect; domain-partitioned schemas scale to 10,000 tables; file size does not predict runtime efficiency.
Practical guidance: architecture decisions should be tailored to model capability rather than assuming universal best practices.

6. CEA: Context Engineering Agent for deep research systems

Paper: *CEA: Context Engineering Agent for Enhanced Reliability and Sustainability in Deep Research Systems* — S. Huang et al., ICLR 2026 submission
Link: https://openreview.net/forum?id=6QUNblHtto
Addresses "context rot": longer context can overload agents and cause unexpected failures. CEA combines history management, progress tracking, key-cue identification, and token-efficiency/memory-integrity balancing. The CERL reinforcement learning method filters non-CEA-attributable error trajectories before gradient updates, improving training stability. Plug-and-play integration with minimal code changes.

7. AI context files in open-source software

Paper: *Context Engineering for AI Agents in Open-Source Software* — S. Mohsenimofidi et al., MSR 2026, arXiv:2510.21413
Link: https://arxiv.org/abs/2510.21413
Studies AGENTS.md adoption across 466 open-source projects. Content structure is not yet standardized. Identifies five context-provision styles (descriptive, prescriptive, prohibitive, explanatory, conditional). Commit-level analysis shows context files evolve alongside major codebase changes — a unique window into real-world context engineering.

8. Representation engineering for conversational context classification

Paper: *Conversational Context Classification: A Representation Engineering Approach* — Jonathan Pan, arXiv:2601.12286, Jan 18, 2026
Link: https://arxiv.org/abs/2601.12286
Combines Representation Engineering (RepE) to identify context-relevant subspaces in LLM internal states with One-Class SVM to establish boundaries in hidden-state latent space. Tested on Llama and Qwen models; promising for detecting topic drift, factual errors, and hallucinations, with applications in dialogue safety monitoring and hallucination pre-filtering.

Discussion and Outlook

The 2026 landscape shows a clear progression: from Prompt Engineering 1.0 (single-turn optimization) through 2.0 (systematic prompt design, domain knowledge integration, reporting standards) to Context Engineering (multi-turn, multimodal, long-context agent frameworks and representation-level methods).

Consensus: model capability is the first-order factor; domain adaptation is essential; the community increasingly values methodological transparency and reproducibility.

Open disagreements: whether format choice matters (McMillan: no overall effect; Paul: substantive impact in data science workflows), and whether human or AI-derived domain knowledge works best for optimization (Srinivasan suggests a hybrid approach).

Recommended reading priorities for practitioners: McMillan's large-scale experiments (context engineering details), Korn et al.'s reporting guideline (research methodology), and the CEA framework (agent system architecture).

---

*Compiled from public paper materials; corrections welcome. Published February 20, 2026.*