8 papers

Context Engineering

Packing the right tokens into the window: retrieval, compression, and ultra-long context.

Papers
8
Read time
2.6h
Avg length
20m
01
Context Engineering

A Survey of Context Engineering for Large Language Models

This survey argues that what actually controls LLM performance is not the model weights but the *information payload* you feed it at inference time, and it gives that discipline a name, a formal optimization definition, and a taxonomy spanning everything from RAG to memory to multi-agent orchestration — distilled from 1400+ papers.

MEI, YAO, GE, WANG, BI, ET AL. (ICT, CHINESE ACADEMY OF SCIENCES + COLLABORATORS) · 2025
Read
22 min
Source
arXiv:2507.13334
02
Context Engineering

A Survey of Context Engineering for Large Language Models

"Prompt engineering" was only ever the tip of the iceberg — this survey names the whole iceberg "Context Engineering," gives it a formal optimization definition, and maps the entire field (1400+ papers) into a clean taxonomy of components (retrieve, process, manage context) and systems (RAG, memory, tools, multi-agent) you can use as a build checklist.

MEI, YAO, GE, WANG, BI, CAI ET AL. (ICT/CAS) · 2025
Read
22 min
Source
arXiv:2507.13334
03
Context Engineering

Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models

Instead of fine-tuning weights or rewriting one big prompt, ACE grows a structured, ever-expanding "playbook" of bullet-point strategies that an agent edits incrementally from its own execution feedback — beating prompt optimizers and matching a GPT-4.1 production agent while using a smaller open model.

ZHANG, HU, UPASANI ET AL. (STANFORD / SAMBANOVA / UC BERKELEY) · 2025
Read
18 min
Source
arXiv:2510.04618
04
Context Engineering

Context Engineering 2.0: The Context of Context Engineering

Context engineering isn't a 2023 prompt-engineering fad — it's a 30-year discipline of compressing messy human intent into something a machine can act on, and this paper gives you the formal definition, the four-era roadmap, and a concrete design playbook for collecting, managing, and using context in agentic systems.

QISHUO HUA, LYUMANSHAN YE, DAYUAN FU, YANG XIAO, PENGFEI LIU ET AL. (SJTU / SII / GAIR) · OCT 2025
Read
22 min
Source
arXiv:2510.26493
05
Context Engineering

Every Token Counts: Generalizing 16M Ultra-Long Context in Large Language Models

An 8B model trained with only a 32K context window learns a *retrieval-style* attention mechanism (HSA) that lets it pull the right needle out of contexts up to 16 million tokens — 500x longer than it was ever trained on — without falling over.

XIANG HU, ZHANCHAO ZHOU, RUIQI LIANG, ZEHUAN LI, WEI WU, JIANGUO LI (ANT GROUP, WESTLAKE UNIVERSITY) · 2025
Read
18 min
Source
arXiv:2511.23319
06
Context Engineering

Everything is Context: Agentic File System Abstraction for Context Engineering

Treat every piece of an agent's context — memory, tools, history, human notes, even live APIs — as files in a governed virtual file system, so context engineering becomes a versioned, auditable, mountable infrastructure instead of a pile of ad-hoc prompt-stuffing code.

XU, MAO, BAI, LI, ZHU, GU (CSIRO DATA61 · UNSW · UTAS · ARCBLOCK) · DEC 2025
Read
16 min
Source
arXiv:2512.05470
07
Context Engineering

Recursive Language Models

Instead of stuffing a giant prompt into the model's context window, hand the model a REPL where the prompt is just a variable, and let it write code to slice, search, and recursively call *itself* on pieces of that variable — so a model with a 272K window cleanly answers questions over 10M+ tokens at roughly the same cost.

ALEX L. ZHANG, TIM KRASKA, OMAR KHATTAB (MIT CSAIL) · PREPRINT JAN 2026
Read
22 min
Source
arXiv:2512.24601
08
Context Engineering

Solving a Million-Step LLM Task with Zero Errors

Instead of waiting for smarter LLMs, you can chain a *cheap* one across a million dependent steps with literally zero errors by shrinking each agent's job to a single step and voting away mistakes — turning unreliable generation into a reliable computing substrate.

MEYERSON, PAOLO, DAILEY, SHAHRZAD, FRANCON, HAYES, QIU, HODJAT, MIIKKULAINEN (COGNIZANT AI LAB & UT AUSTIN) · NOV 2025
Read
18 min
Source
arXiv:2511.09030