---
title: Compounding Inference
type: vocabulary
url: "https://www.envisioning.com/vocab/compounding-inference"
summary: A paradigm where intermediate agent work persists in a shared, branchable memory so later computation can inspect, extend, or challenge it instead of discarding it after each attempt.
year: 2026
generality: 0.45
---

# Compounding Inference

A paradigm where intermediate agent work persists in a shared, branchable memory so later computation can inspect, extend, or challenge it instead of discarding it after each attempt.
Compounding inference is a paradigm for organizing inference-time computation so that intermediate work, partial solutions, experimental findings, and even failed approaches persist and can be inspected, extended, combined, or challenged by later computation, rather than being discarded once a single trajectory ends. Vedant Shah, Ankur Samanta, and twelve coauthors, primarily at Meta Superintelligence Labs with collaborators from Mila, Université de Montréal, Columbia, Carnegie Mellon, and Princeton, named the paradigm and introduced its first implementation in the October 2026 paper "GitSwarm: Decentralized Compounding Inference" (arXiv:2610.04862). The paper identifies a gap in most inference-time scaling methods: computation is organized around individual trajectories or candidate solutions that are scored and discarded independently, with no mechanism for one attempt's partial progress to inform the next.

The paper's implementation, GitSwarm, is an asynchronous multi-agent system in which homogeneous workers independently decide how to advance a long-horizon task while collaborating through a shared, branchable Git repository that acts as structured persistent memory. Workers explore, experiment, verify, refine, and synthesize each other's prior work. Each contribution is published as an atomic commit, and workers declare explicit semantic dependencies beyond the direct Git parent, so a worker can extend one branch while drawing on discoveries made on others. A decentralized readout process replaces a single final judge: workers nominate candidates, and the one with the most nominations is selected.

On IMOProofBench-Advanced, GitSwarm solved all 30 problems in a single run using GPT-5.5-class workers. On ProgramBench it reached a 79.4 percent mean score, against 65.1 percent for the strongest reported single-agent baseline under a comparable compute budget. On three neural architecture research tasks (Residual Matrix Transformer, Looped Transformer, and NanoChat), swarms of workers improved on the starting architectures through sustained, GPU-backed experimentation.

The authors also measured accumulation directly. On ProgramBench, 94.7 percent of contributions were later built upon, and the selected solution's ancestry covered 82 to 93 percent of the full contribution graph on average across budgets.

---
Source: Envisioning — Technology Research Institute (https://www.envisioning.com/vocab/compounding-inference)