Paper Review: Latest Advances in Prompt Engineering and Context Engineering (Early 2026)

> This post compiles 8 notable papers from early 2026 (as of Feb 20) covering prompt engineering and context engineering across chemistry, software engineering, data science, finance, and agent systems.

Key Trends

1. Vertical domain adaptation — prompt engineering practice in chemistry and finance 2. High-dimensional optimization strategies — complex optimization tasks in software engineering 3. Structured context — context engineering for file-native agent systems 4. Representation engineering — understanding context from model internal states

The field is shifting from single-turn prompt tricks (Few-shot, CoT, ToT) toward systematic prompt design (domain-knowledge integration, reporting standards) and full context engineering (structured context, agent frameworks, representation engineering).

Prompt Engineering Papers

1. Systematic Review of Prompt Engineering in Organic Chemistry

Pourgholamali et al., ChemRxiv, Feb 4, 2026 — https://chemrxiv.org/doi/10.26434/chemrxiv-2026-625v3

• Screened 101 records, included 22 high-quality studies; clarifies terminology inconsistencies
• Architecture fit: encoder-only for text understanding/classification (molecular property prediction), decoder-only for generation (synthesis route design), encoder-decoder for translation/summarization (reaction condition optimization)

2. Prompt Engineering Case Studies in Data Science

S. Paul, arXiv:2602.00337 — https://arxiv.org/abs/2602.00337
• Prompt optimization significantly improves LLM output quality without updating model parameters
• Covers data cleaning, feature engineering suggestions, model selection guidance, result interpretation; emphasizes iterative optimization and version control

3. Domain Knowledge Strategies for High-Dimensional LLM Optimization

Srinivasan & Menzies, MSR 2026, arXiv:2602.02752 — https://arxiv.org/abs/2602.02752
• LLMs lag behind Bayesian methods on high-dimensional SE optimization; four architectures evaluated: H-DKP (human-in-the-loop knowledge prompting), AMP (adaptive multi-stage prompting), DAPR (dimension-aware progressive refinement), 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

Korn et al., FORGE 2026, arXiv:2601.01954 — https://arxiv.org/abs/2601.01954
• Analyzed ~300 top SE conference papers (2022+) and surveyed 105 PC members
• Found misalignment between practice and reviewer expectations on version disclosure, prompt justification, and validity threats
• Proposes a tiered guideline: Essential (full prompt text, model version/params, design rationale), Desirable (prompt testing, failure analysis, sensitivity analysis), Exceptional (multi-model comparison, cost-benefit analysis, reproducibility package)

Context Engineering Papers

5. Structured Context Engineering for File-Native Agentic Systems

Damon McMillan, arXiv:2602.05447 — https://arxiv.org/abs/2602.05447
• 9,649 experiments, 11 models, 4 formats (YAML, Markdown, JSON, TOON), schema scale from 10 to 10,000 tables
• Key findings:
• Frontier models gain +2.7% file-retrieval accuracy (p=0.029); open-source models drop −7.7% (p<0.001)
• Format has no significant overall effect (chi-squared = 2.45, p = 0.484), though some open-source models show format-specific sensitivity
• A 21-percentage-point capability gap between frontier and open models dwarfs format/architecture effects
• File-native agents scale to 10,000 tables via domain-partitioned schemas
• File size does not predict runtime efficiency; compact formats can add token overhead depending on model capability

6. CEA: Context Engineering Agent for Deep Research Systems

Huang et al., ICLR 2026 submission — https://openreview.net/forum?id=6QUNblHtto
• Addresses "context rot": longer context does not guarantee better responses
• CEA manages interaction history, tracks progress, preserves key cues, and balances token efficiency vs. memory integrity
• CERL multi-turn RL method filters non-CEA-attributable error trajectories before gradient updates, improving training stability; plug-and-play with existing systems

7. Context Engineering for AI Agents in Open-Source Software

Mohsenimofidi et al., MSR 2026, arXiv:2510.21413 — https://arxiv.org/abs/2510.21413
• Studies AGENTS.md adoption in 466 open-source projects; an emerging de facto standard
• Content structure is not yet standardized; identifies five context styles: descriptive, prescriptive, prohibitive, explanatory, conditional
• Files evolve with the codebase, offering a unique window into real-world context engineering

8. Conversational Context Classification via Representation Engineering

Jonathan Pan, arXiv:2601.12286 — https://arxiv.org/abs/2601.12286
• Combines Representation Engineering (RepE) with One-Class SVM to detect when LLMs deviate from expected conversational norms (topic drift, factual errors, hallucination)
• Experiments on Llama and Qwen models; promising results and a contribution to LLM interpretability

Discussion

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

Open disagreements: whether format matters (McMillan finds no overall effect; Paul's case studies show material impact in data science), and human vs. AI domain knowledge (Srinivasan: a hybrid approach may be optimal).

Practical advice: researchers should follow Korn et al.'s reporting guideline; engineers can adopt CEA for agent context management and AGENTS.md practices flexibly; domain experts should use the architecture-selection framework from the chemistry review.

--- *Compiled from public paper materials; original review by Xiaokai, published Feb 20, 2026.*