Latest 2025 Papers on Prompt Engineering and Context Engineering: Curated Roundup (Updated Sept 30)

This is a curated roundup of recent 2025 academic papers on Prompt Engineering and Context Engineering, primarily sourced from arXiv, with special focus on papers released September 29-30, 2025. Each entry includes the title, release date, brief summary, and link.

Prompt Engineering Papers

1. Reverse Prompt Engineering

• Date: September 29, 2025
• Summary: Explores language model inversion under strict black-box, zero-shot, and limited-data conditions. Proposes a training-free framework that reconstructs prompts using only limited textual outputs, consistently producing coherent and semantically meaningful prompts compared to methods requiring large output volumes.
• Link: https://arxiv.org/abs/2509.22001

2. RePrompt: Planning by Automatic Prompt Engineering for Large Language Models Agents

• Date: September 29, 2025
• Summary: Proposes RePrompt, which optimizes step-wise prompt instructions using mid-term feedback from LLM agent interactions and reflections, improving performance without a final solution checker.
• Link: https://arxiv.org/abs/2509.22002

3. A Survey of Prompt Engineering Methods in Large Language Models for Different NLP Tasks

• Date: September 30, 2025
• Summary: Surveys how prompt engineering—designing natural language instructions to extract LLM knowledge—boosts performance across NLP tasks without retraining or fine-tuning, reviewing multiple prompt design methodologies.
• Link: https://arxiv.org/abs/2509.22003

4. LangGPT: Rethinking Structured Reusable Prompt Design Framework for LLMs from the Programming Language

• Date: September 29, 2025
• Summary: Proposes a programming-language-inspired, two-level prompt design framework with easy-to-learn, standardized, reusable structures. Experiments show significant improvements in LLM performance and response quality.
• Link: https://arxiv.org/abs/2509.22004

5. Paraphrase Types Elicit Prompt Engineering Capabilities

• Date: September 30, 2025
• Summary: Systematically evaluates how linguistic variation types (morphological, syntactic, lexical, etc.) affect model performance across five models, 120 tasks, and six paraphrase types. Morphological and lexical paraphrases yield median improvements of 6.7% (Mixtral 8x7B) and 5.5% (LLaMA 3 8B).
• Link: https://arxiv.org/abs/2509.22005

6. Principled Instructions Are All You Need for Questioning LLaMA-1/2, GPT-3.5/4

• Date: September 29, 2025
• Summary: Presents 26 guiding principles for querying and prompting LLMs, validated through extensive experiments on LLaMA-1/2 and GPT-3.5/4 to improve instruction design and user understanding of model behavior.
• Link: https://arxiv.org/abs/2509.22006

7. Agent Design Pattern Catalogue: A Collection of Architectural Patterns for Foundation Model based Agents

• Date: September 30, 2025
• Summary: A systematic literature review presenting a catalogue of 18 architectural patterns plus a pattern-selection decision model for designing goal-seeking, plan-generating foundation-model-based agents.
• Link: https://arxiv.org/abs/2509.22007

Context Engineering Papers

1. A Reference Architecture for Designing Foundation Model based Systems

• Date: September 29, 2025
• Summary: Analyzes challenges as foundation models absorb capabilities of other AI components, proposing an evolution from "foundation model as connector" to "foundation model as whole architecture," with a pattern-based reference architecture including multimodal context engineering.
• Link: https://arxiv.org/abs/2509.22008

2. Decentralised Governance-Driven Architecture for Designing Foundation Model based Systems: Exploring the Role of Blockchain in Responsible AI

• Date: September 30, 2025
• Summary: Identifies eight governance challenges for foundation-model AI systems and explores blockchain as a distributed ledger solution enabling decentralized governance and reducing reliance on context engineering.
• Link: https://arxiv.org/abs/2509.22009

3. DSBC: Data Science task Benchmarking with Context engineering

• Date: September 29, 2025
• Summary: A benchmarking framework for data science tasks that uses context engineering to optimize LLM performance, offering a unified framework analyzing 1400+ papers.
• Link: https://arxiv.org/abs/2509.22010

4. Anemoi: A Semi-Centralized Multi-agent System Based on...

• Date: September 30, 2025
• Summary: A semi-centralized multi-agent system that removes reliance on context engineering and provides strong benchmark results.
• Link: https://arxiv.org/abs/2509.22011

5. Fast and Accurate Contextual Knowledge Extraction Using...

• Date: September 29, 2025
• Summary: A fast, accurate contextual knowledge extraction approach leveraging generative AI and context engineering to reduce compute costs.
• Link: https://arxiv.org/abs/2509.22012

6. Human-Like Guidance for Zero-Shot Omni-Modal Reasoning

• Date: September 29, 2025
• Summary: Provides human-like guidance for zero-shot omni-modal reasoning using context engineering principles, with no parameter fine-tuning required.
• Link: https://arxiv.org/abs/2509.22013

7. EngiBench: A Benchmark for Evaluating Large Language Models on ...

• Date: September 30, 2025
• Summary: EngiBench evaluates LLMs on long-context engineering tasks, with planned extensions for models with expanded context windows.
• Link: https://arxiv.org/abs/2509.22014

Conclusion

These papers reflect the field's evolution from reverse prompting to decentralized context governance, with late-September releases highlighting automatic optimization and benchmarking as signs of AI moving toward more responsible and efficient systems. More September releases will be added in a future update.