TLDR: AutoContext is a new framework that enables Large Language Model (LLM) agents to learn and utilize “instance-level context” – verifiable, reusable facts specific to an environment, like object locations or local rules. This context, often overlooked, is crucial for LLM agents to succeed in complex tasks. AutoContext uses a guided exploration process with a “TODO forest” and a “plan-act-extract” loop to automatically build a high-precision context document. Experiments show significant improvements in success rates and efficiency for various LLM agents across different environments like TextWorld, ALFWorld, and Crafter.
Large Language Model (LLM) agents are becoming increasingly sophisticated, tackling complex tasks in interactive environments. Traditionally, these agents rely on two main types of information: environment-level manuals, which define global rules and interaction interfaces, and task-level guidance, which provides instructions for specific goals. However, new research highlights a critical, often overlooked, third type of context: instance-level context.
The Missing Piece: Instance-Level Context
Instance-level context refers to verifiable and reusable facts that are specific to a particular environment instance. This includes details like the exact locations of objects, specific crafting recipes available, or unique local rules that apply only to that instance. The absence of this precise, persistent information can be a major reason why LLM agents struggle with complex tasks, as success often depends on more than just general rules or task instructions.
Acquiring this instance-specific knowledge is challenging. It requires efficient exploration, careful validation of discovered facts, and formatting them in a way that LLM agents can easily use, all while operating under tight interaction budgets.
Introducing AutoContext: Learning from the Environment
To address this challenge, researchers have formalized the problem as Instance-Level Context Learning (ILCL) and introduced a novel, task-agnostic method called AutoContext. This framework is designed to perform a compact, one-off exploration of an environment instance and distill its findings into a durable, agent-readable document. This document, rich with instance-level facts, can then be reused by various downstream tasks and agents, effectively amortizing the initial exploration cost.
AutoContext operates through a guided exploration process, utilizing a unique ‘TODO forest’ to intelligently prioritize its next actions. It employs a lightweight ‘plan–act–extract’ loop to execute these actions, automatically generating a high-precision context document. This document captures essential information such as object locations, preconditions for actions, and effective interaction rules.
How AutoContext Works
The framework consists of three main components:
-
The Planner: This component identifies gaps in the agent’s knowledge by scanning for ‘Unknown’ markers in the context document and analyzing the TODO forest. It then proposes targeted actions (TODOs) to uncover new information.
-
The Actor: The Actor executes the proposed TODOs. In simpler environments, it performs primitive actions directly. For more complex scenarios, it delegates control to a sub-agent (like a ReAct-style agent) to complete higher-level subtasks, summarizing the outcome into a concise ‘key result’.
-
The Extractor: After actions are executed, the Extractor validates new facts against the agent’s trajectory and a predefined schema. It proposes, checks, and applies edits to update and expand the instance context, ensuring reliability and schema compliance.
A key innovation is the ‘TODO forest’, a data structure that organizes states and subtasks into shallow trees. It records the entire exploration process, including both successful and failed attempts, which serves as a valuable collection of in-context examples for future learning and planning.
Significant Performance Gains
Experiments across diverse environments like TEXTWORLD, ALFWORLD, and CRAFTER demonstrate the consistent effectiveness of AutoContext. For instance, in TEXTWORLD, a ReAct agent’s mean success rate dramatically increased from 37% to 95% when augmented with AutoContext. Similarly, the state-of-the-art IGE method improved from 81% to 95%.
AutoContext not only boosts success rates but also significantly enhances efficiency. It rapidly achieves over 95% coverage of locations and objects in TEXTWORLD within 200 steps, a feat that state-of-the-art baselines struggle to match. Ablation studies further confirm that each component of AutoContext—the TODO forest, Planner, and Extractor—plays a crucial role in its overall performance.
Also Read:
- AutoMaAS: A Self-Evolving Framework for Multi-Agent AI Systems
- Guiding LLM Reasoning: A New Approach to Maintain Focus in Complex Tasks
The Future of LLM Agents
This research highlights that future general-purpose LLM agents may increasingly rely on structured, per-instance context as a foundational component of intelligence. By transforming one-off exploration into persistent, reusable knowledge, AutoContext complements existing context types to enable more reliable and efficient LLM agents. While the current approach has limitations, such as potential issues with extremely large observation spaces and the manual design of schemas, it paves the way for more robust and intelligent AI systems.
You can read the full research paper here: BEYOND MANUALS AND TASKS: INSTANCE-LEVEL CONTEXT LEARNING FOR LLM AGENTS.
