Context Engineering Overtakes Prompt Engineering: A November 2025 Review of Key LLM Research

Forum topic · ✨步子哥 · 2025-11-18

Summary

This forum post surveys the 2025 shift from prompt engineering to context engineering in large language model (LLM) research. It reviews key arXiv papers: 'You Don't Need Prompt Engineering Anymore' (arXiv:2510.22251), whose Sculpting post-processing technique lifted GSM8K accuracy from 88% to over 94% with one-third the compute; 'Context Engineering 2.0' (arXiv:2510.26493), which formally defines context engineering as the systematic design, selection, compression, and evolution of model input context; the ACE framework (arXiv:2510.04618), where self-evolving contexts raised Llama-3-70B GSM8K scores from 81% to 97.3% unsupervised; heterogeneous context encoding that preserved performance on 400K-token contexts; and information-theoretic gamma-covering that compressed contexts to 8-12% while retaining 99% of directed information. November 2025 updates include prompt engineering for Solana/Algorand smart contract vulnerability detection, the NoteEx visual context-manipulation system, and AutoSynth's Monte Carlo tree search workflow automation. The post concludes that manual prompt tuning is yielding to self-evolving, programmable context ecosystems, redefining the AI engineer's role.

Translated from a zhichai.net forum post. This article reviews the 2025 paradigm shift from prompt engineering to context engineering in LLM research, based on arXiv papers.

The Decline of Prompt Engineering

Sculpting (Oct 25, 2025) — 'You Don't Need Prompt Engineering Anymore' (arXiv:2510.22251). Sculpting is a differentiable post-processing optimizer: a lightweight mask network operates on each output token after the model generates reasoning — removing redundancy, rewriting weak logic, and strengthening ambiguity. Results:

GSM8K accuracy improved from 88% to over 94%, nearly matching RL-trained o1-series models
Used only one-third of the original compute

The Prompt Engineering Report Distilled (Sep 14, 2025). Condensed 58 prompt techniques into 20 pages, yielding a 41% improvement for Claude 3.5 Sonnet on biology tasks. Notable finding: overly polite prompts hurt performance; direct commands work better.

Professional communication survey (Aug 27, 2025). A 243-person study found that clearer, more structured, more 'blunt' prompts produce more professional AI output — each 1-point clarity gain (out of 5) improved output professionalism by 0.67 points.

Marginal returns approaching zero. Two July papers: in recommendation systems, zero-shot prompting beat carefully designed Tree of Thoughts; across 10 software engineering tasks, 14 advanced prompting techniques averaged only 3.4% improvement over direct queries.

The Rise of Context Engineering

Context Engineering 2.0 (Oct 30, 2025, arXiv:2510.26493) formally defines context engineering as the systematic design, selection, structuring, compression, retrieval, and evolution of input context to maximize model performance. Its three contributions:

1. A unified conceptual map covering RAG, memory, tool calling, and long context 2. The 'context-as-program' thesis — future contexts will be compilable, debuggable code 3. The prediction that top AI engineers will be context architects rather than prompt tuners

Heterogeneous contexts (Oct 8, 2025, arXiv:2510.07414). A heterogeneous context encoder scores each context source (code, web, databases, memory) on importance, recency, and redundancy, then dynamically compresses, discards, or reorders. Claude 3 handled 400K-token heterogeneous contexts with only 6% performance degradation.

ACE — Agentic Context Evolution (Oct 6, 2025, arXiv:2510.04618). Contexts act as agents observing model reasoning and autonomously inserting theorems or counterexamples. Llama-3-70B improved on GSM8K from 81% to 97.3% over 5 self-evolution rounds, fully unsupervised.

Directed information γ-covering (Sep 30, 2025). Compressed contexts to 8–12% while preserving 99% of directed information.

Open-source practice (Oct 24, 2025). Placing an .ai-context.json file in repository roots (documenting project philosophy, architecture decisions, common pitfalls) raised one-shot fix success rates from 37% to 78%.

November 2025 Updates

As of November 17, 2025, new papers in the 2511 series are all prompt-engineering work, with no direct context-engineering sequels yet:

arXiv:2511.11250 (Nov 14): For vulnerability detection in Solana/Algorand smart contracts, prompt engineering offers 'general robustness' while fine-tuning substantially improves precision on the semantically sparse TEAL language
NoteEx (Nov 10): Visual interaction letting users 'drag and drop' context cells, reducing manual prompt engineering pain
AutoSynth (Nov 12): Monte Carlo tree search automates synthetic data workflows, cutting human time from 5–7 hours to 30 minutes

Conclusion

The trajectory is clear: manual prompts → automated sculpting → context evolution → context-as-program. Future engineers may design self-evolving context ecosystems rather than write prompts, shifting human-AI relations from master-and-tool toward symbiosis and co-creation.

References

1. You Don't Need Prompt Engineering Anymore. arXiv:2510.22251 2. Context Engineering 2.0. arXiv:2510.26493 3. Evolving Contexts for Self-Improving Language Models. arXiv:2510.04618 4. Context Engineering for Heterogeneous and Agentic Long Contexts. arXiv:2510.07414 5. Prompt Engineering vs. Fine-Tuning for LLM-Based Vulnerability Detection. arXiv:2511.11250