Thinking Before Thinking: Scaling Agentic Inference Through Meta-Reasoning
Abstract Overview
This paper introduces agentic meta-reasoning, an inference-time control framework that separates task execution from control decisions in long-running language-model agents. A controller maintains a compact state, reads and writes persistent artifact memory, and runs a four-stage cycle—assess, propose, evaluate, and dispatch—to decide what computation to perform next, while workers execute the task-level actions. The authors formalize the resulting run as an artifact graph, which records how later work reuses earlier outputs and supports diagnostic analysis beyond final accuracy. They evaluate the approach on IMO ProofBench-Advanced, ARC-AGI-2, LongCoT-mini, and ProgramBench using Gemini 3.1 Pro, GPT-5.5, and Opus 4.8, comparing it with direct-control and external agent baselines.
Novelty
The distinctive contribution is to make inference control itself an explicit, structured agentic process rather than a single next-action decision over an accumulating history. The framework combines staged meta-control, persistent artifact memory, compact controller state, and artifact-graph diagnostics to scale agent behavior without replaying the full run history at every decision.
Results
At the main evaluated budgets, the Meta-Reasoning Agent achieves the highest point estimate in all 12 matched model-benchmark comparisons against the Direct Control Agent, with average gains of 3.6 to 4.2 points depending on benchmark. On ProgramBench, it reaches 71.5% with GPT-5.5 versus 63.7% for direct control and 58.0% for Codex, and 67.2% with Opus 4.8 versus 65.3% for direct control and 65.5% for Claude Code. The analyses further show that meta-reasoning tends to use larger budgets more effectively, increases reuse of prior work, and improves correct-solution coverage in most settings, though its extra deliberation can hurt at smaller budgets and selection gains are not uniform.
Key Points
- The method separates controller and worker roles, with the controller using assess/propose/evaluate/dispatch stages over compact state and persistent memory.
- Artifact graphs are used both as a runtime representation of dependencies between intermediate outputs and as a diagnostic tool for analyzing budget use, reuse, coverage, and selection.
- Empirically, the approach improves performance across reasoning and coding benchmarks and continues to benefit from larger inference budgets in settings where direct control often plateaus.