Adaptive graph boosts long task success in large language models
Adaptive Consistency Graph for Long-Horizon Agents
Artificial IntelligenceComputation and Language
Summary
Large language models often do well on simple tasks but struggle with long, complicated ones requiring many steps. The authors created a method called the Adaptive Consistency Graph that keeps track of all past decisions and evidence during a task. This graph helps the model focus on the right information when making each choice, improving its success over long sequences. Their approach works by adding context without changing the model's internal decision process, leading to better results on complex tasks.
What this means in practice
- •For software developers: Improve large language model agents in software that require multi-step task completion by managing context more effectively.
- •For automation engineers: Enhance automated agents controlling complex workflows by providing structured, traceable information for long sequence decisions.
Authors
Jiecong Wang, Hao Peng, Zhanyi Wang
Abstract
Large language model agents can often make reasonable local decisions on short tasks, yet their performance degrades when success requires long sequences of dependent actions and tool calls. During execution, task requirements, historical evidence, and the current execution state may gradually become disconnected, so later decisions can drift from the original objective. We study this problem by introducing the Adaptive Consistency Graph (ACG) for long-horizon execution. ACG incrementally organizes execution evidence and its provenance in a persistent graph, then constructs a temporary requirement-centered view for each decision under a bounded context budget. Rather than replacing the base agent's planner or tool executor, ACG provides a structured and traceable context view for each decision. In the matched evaluation, ACG improves GPT-5.6-luna's average success from 44.5\% with ReAct to 50.2\%, with the largest gain on BrowseComp-Plus (73.5\% versus 62.4\%). We further analyze trajectory structure and inference cost to characterize this improvement.