Thinking Before Thinking: Scaling Agentic Inference Through Meta-Reasoning
Paras Dahal, Anton Bakhtin, Taco Cohen, Zhengxing Chen, Carole-Jean Wu, Rob Fergus, Scott Yih, Gabriel Synnaeve, Ruslan Salakhutdinov, Sanjeev Arora, Jason Weston, Anirudh Goyal
cs.AI
2026-09-30
Meta's four-stage meta-reasoning wins all 12 matched comparisons at main budgets, hitting 71.5% on ProgramBench with GPT-5.5 vs 58.0% for Codex.
The longer an agent run gets, the more of its calls go to control: keep editing a partial result, start over, or stop. A bad call wastes more than one invocation. It can keep a dead approach alive until the budget is gone, or throw away a correct answer already sitting in memory.
Most production coding agents and research harnesses fold that choice into the same step that does the work, conditioned on a growing transcript. As the history lengthens, the evidence that matters is buried, and later steps rarely reuse earlier artifacts. Extra budget often goes unused, or scores stop moving.
Meta Superintelligence Labs treats what to compute next as its own reasoning job. They call it agentic meta-reasoning: an inference-time harness that makes control an explicit, staged process.
Workers do object-level work. A worker is either one model call or a tool-using coding agent. The controller never authors the final answer. It assigns work and chooses which stored artifacts each worker may see. Every output is an artifact with a stable id in persistent memory. Between cycles the controller keeps only a compact, rewritten state, not the full history.
Each cycle has four stages. Each stage is itself an agentic process with tools, so it can read memory, write notes, and spend several calls:
Controller calls and worker calls draw from the same nominal budget, so the extra deliberation has to pay for itself. The matched baseline is a Direct Control Agent with the same workers, delegation interface, context selection, and stop action. It chooses each action in one turn over the accumulated history. The comparison tests the control design as a bundle. Dependencies are recorded as an artifact graph: an edge means one artifact was supplied as context for producing another.
Main settings are 100 model calls on the reasoning benchmarks and 1200 on ProgramBench, with Gemini 3.1 Pro, GPT-5.5, and Opus 4.8. Point estimates rise in all 12 matched comparisons. Mean gains over Direct Control are 4.0, 4.2, and 3.6 points on IMO ProofBench-Advanced, ARC-AGI-2, and LongCoT-mini. The spread is wide: Gemini gains 9.2 points on LongCoT-mini, GPT-5.5 only 0.4.
| Model | Task | Meta-Reasoning | Direct Control | External |
| GPT-5.5 | ProgramBench | 71.5% | 63.7% | Codex 58.0% |
| Opus 4.8 | ProgramBench | 67.2% | 65.3% | Claude Code 65.5% |
| Gemini 3.1 Pro | IMO ProofBench-Adv | 91.3 | 82.7 | RLM 73.3 |
| Gemini 3.1 Pro | ARC-AGI-2 | 84.2 | 77.5 | RLM 76.7 |
| Gemini 3.1 Pro | LongCoT-mini | 62.7 | 53.5 | RLM 46.5 |
On ProgramBench it also beats mini-SWE Agent by 2.5 to 13.9 points. The 13.5-point Codex gap is the number that will be quoted. Direct Control already sits at 63.7% with the same workers, so the control change itself is 7.8 points.
How the budget is spent matters more than that quote. Direct Control often stops early: GPT-5.5 uses about 18% of the 1200-call ProgramBench allowance; meta-reasoning uses about 89% to 101%. Raising ProgramBench from 400 to 1200 calls lifts GPT-5.5 meta-reasoning from 64.1% to 71.5%; Direct Control stays near 64%. Opus 4.8 roughly doubles calls used, 376 to 768, while the score only moves from 62.7% to 65.3% and peaks at the middle budget. Doing more work is not enough if the next assignment is wrong.
Small budgets reverse the ranking. At 400 calls with Opus 4.8, meta-reasoning scores 56.6% against Direct Control's 62.7%. At 1200 it leads, 67.2% vs 65.3%. The four-stage overhead needs a long run to repay.
Artifact graphs show more worker outputs and far more reuse. With Gemini on IMO, worker artifacts roughly double while recorded dependencies rise by an order of magnitude. Coverage (whether a correct candidate ever appears) rises in most settings: +20 points on IMO and +10 on LongCoT-mini with Gemini. Controller verdicts rank correctness better than worker confidence; Type-2 AUC on IMO with Gemini goes from 0.55 for workers to 0.88 for the controller. Final selection is uneven. On ARC-AGI-2 with GPT-5.5, frontier selection gain is 11 points over a uniform draw, versus 3 for Direct Control. On LongCoT-mini with Gemini, Direct Control is slightly ahead.
The controller rereads its own notes more than worker outputs: 4.4 to 9.8 reads per retrieved note on ARC and LongCoT, against 1.9 to 3.3 for worker artifacts. Direct Control's history exceeds one million characters on the longest Opus IMO runs. Meta-reasoning state stays in the thousands to tens of thousands.
This is an inference-time harness. The controller is not fine-tuned. It fits long runs where intermediate artifacts depend on each other and the budget is large enough that control is itself expensive: program reconstruction, proof search, multi-step constraint tracking. On short tasks and small budgets, the four stages can be a net loss.
A familiar failure becomes visible. A model that can solve a step can still be bad at managing its own compute. Direct Control already has the same workers and the same allowance, and it still plateaus at large budgets. Stacking more calls does not fix the allocation problem.
Four stages are not shown to be the only cut, and the Codex and Claude Code comparisons mix in whole-system differences. What the paper does show is that explicit control pays off at the long budgets they tested.
The authors flag the cost. Overhead dominates at small budgets. A wrong Assess can propagate a misleading state or drop good partial work it never rereads. Compact state is lossy; facts that matter only later have to be recovered by a future memory read. On the chess subset of LongCoT-mini, meta-reasoning loses to Direct Control. Sampled trajectories point to unproductive reconsideration: extra checking destabilizes an answer that was already correct. Gain size tracks the model. Gemini moves the reasoning benches, GPT-5.5 moves ProgramBench, Opus gains less but stays positive.
The matched comparison is a bundle, not a single-factor ablation. Equal call counts are not equal tokens, latency, or FLOPs. ProgramBench has no grading protocol for intermediate reconstructions, so coverage and selection are reasoning-only. Whether Direct Control plateaus because of million-character histories is untested. The 12-for-12 result is point estimates; the main text does not report confidence intervals or significance tests.