ArbiGraph: Arbitrarily Scalable Verifiable Task Graphs for Evaluating Context Management

2026-07-22

AI Summary

This work proposes the first arbitrarily scalable and automatically verifiable task-graph benchmark designed to evaluate language agents’ ability to retain, update, combine, and discard contextual information during complex reasoning. The benchmark constructs task graphs from natural language questions paired with executable Python solvers, modeling tasks through typed intermediate states such as scalars and lists. It enables flexible control over task length, dependency structure, distractors, and value types. Experimental results reveal that while Qwen3.5-27B excels on isolated tasks, its accuracy drops by up to 33.3% on complex tasks involving branching dependencies, effectively exposing a critical bottleneck in current agents’ context management capabilities.

Abstract

We introduce ARBIGRAPH, a benchmark generator for evaluating whether tool-assisted language agents can retain, update, compose, and discard task-relevant context across extended reasoning workflows. ARBIGRAPH represents each task as a natural-language problem with an executable Python solver, and composes tasks through typed intermediate states, instantiated here as scalar and list values. This design enables controllable task graphs whose length, dependency structure, distractor count, and value type can be varied while preserving exact automatic verification. We instantiate ARBIGRAPH with math, GSM-style word-problems, and Python-tracing task categories, and evaluate a Qwen3.5-27B tool-assisted agent across four topologies. The results show high accuracy on isolated tasks but substantial degradation on more complex dependent tasks: accuracy drops by up to 33.3% on branching chains of dependent math tasks. This shows that ARBIGRAPH exposes failures that are not visible from single-task evaluation alone. Our code, generated datasets, and evaluation results are available at https://github.com/pavelgolikov/ArbiGraph.git

Problem

• context management
• task graphs
• language agents
• reasoning workflows
• verifiable evaluation

Innovation

• verifiable task graphs
• context management
• tool-assisted language agents
• scalable benchmark
• dependency-aware reasoning