Thinking Before Thinking: Scaling Agentic Inference Through Meta-Reasoning

简体中文

L1 core-mechanism diagnostic. Implements budget feasibility, Evaluate/Dispatch selection, and separation of persistent artifacts from compact state. No frontier workers run; ProgramBench scores are not local.

Paper information

Field	Details
Paper	arXiv 2609.38147
Organization	First-author affiliation: Meta Superintelligence Labs
First public release	2026-09-29 (arXiv v1)
Authors' code	No public author repository was found as of 2026-10-01.
Adapter	meta-reasoning
Local reproduction code	src/auto_research/agent_research/latest_20261001.py

Original paper

Motivation and method

Workers handle object-level tasks, while a controller splits control into Assess, Propose, Evaluate, and Dispatch. The controller carries compact state, reuses work via persistent artifact memory, and decides whether to continue, branch, or stop within a shared call budget.

flowchart LR
  I[Public input and current state] --> M[Meta-reasoning mechanism]
  M --> A[Auditable intermediates]
  A --> O[Output or state update]

Key figure from the paper

Figure 1: Core architecture, modules, and their connections. From the original paper; copyright remains with the authors.

Core equations

\(s_t=Assess(x,s_{t-1},\Delta M_t;M_t)\), \(\tilde a_t=Evaluate(x,s_t,b_t,\mathcal A_t;M_t)\), then \(a_t=Dispatch(\cdot)\). The local implementation ranks feasible actions by expected value per call cost and keeps complete artifacts in persistent memory.

Paper results

On ProgramBench, GPT-5.5 scored 71.5% versus Codex at 58.0%, while Opus 4.8 scored 67.2% versus Claude Code at 65.5%. Other long-horizon tasks gained 3.6–4.2 points on average over direct control.

Local reproduction

Local comparison: The baseline disables the mechanism or uses the default state, and the treatment enables the core operator. This checks invariants and state transitions only; cross-model relative changes are not applicable.

See the three-seed diagnostic. diagnostic_only=true indicates runnable transitions, gradients, or scheduling invariants only, not eligibility for capability ranking.

Reproduction boundaries

Budget feasibility, Evaluate/Dispatch selection, and separation of persistent artifacts from compact state are implemented. No frontier workers are called; ProgramBench results are not local.