VirSci v2: million-agent simulation claim and inference queue
Reviewed September 6, 2026 UTC. Public read-only research.

PRIMARY PAPER
https://arxiv.org/html/2505.12039v1
AI-Driven Automation Can Become the Foundation of Next-Era Science of Science Research, May2025. Affiliations include Shanghai AI Laboratory and other Chinese and international institutions. The paper reports a society of one million agents for40 epochs, implemented on32 A100 GPUs with four Llama3.1-8B-serving ports per GPU. It describes asynchronous requests sharing inference endpoints and a maximum memory of five entries per agent. Its reported simulation studies research-society patterns, with agents generating ideas, reviewing and citing simulated papers. These are author-reported experiments, not a reproduced deployment or public-internet swarm. The million count denotes simulated agent population, not a million distinct model instances or independent operators. The stated infrastructure implies128 serving ports; this arithmetic does not establish measured concurrency or throughput.

CURRENT SOURCE
https://github.com/RenqiChen/Virtual-Scientists-v2
Revision b68b2a70c17a0bfdabfb4299b330300bb50996ef. Recursive tree truncated=false,272 file blobs;228 lie under bundled camel-master,44 outside it. Counts must not be treated as agent or run counts.
README describes checkpoint output under a configured run directory: citations, paper texts, embeddings, team data, SQLite database and collaboration-weight matrix. Its described paper database combines historical source papers with simulation-generated material, distinguished by year=-1 versus simulation epoch. Thus a paper collection is not automatically a set of new agent outputs.
The current tree has no visible saved checkpoint directory or obvious dialogue-log bundle. This is a current-path inventory, not a review of every file's contents or repository history. No bytecode, environment configuration, model keys or deployment endpoints were inspected.

QUEUE IMPLEMENTATION
sci_platform/social_agent/channel.py creates an asyncio request queue, UUID message IDs and a lock-protected dictionary for responses. Callers poll for their matching response with a short adaptive wait.
sci_platform/inference/inference_manager.py creates one inference worker for each configured host/port pair. A dispatcher feeds pending requests to idle workers and routes completed responses back by message ID. It logs received-message counts. This is concrete shared-inference machinery supporting many logical agents; it is not a public peer-discovery network or evidence of agents writing to unrelated internet surfaces.
The source file logs to inference.log when executed, but no resulting inference.log was found in the current tree. Source-level logging capability is not a preserved execution record.
No simulation, package install, model inference or network participation was performed.

ASSESSMENT
VirSci v2 is another institution-associated designed multi-agent simulation, with a more concrete scale description than its headline alone. This improves the inventory of Chinese-context swarm research but does not supply the original sought escaped scratch-memory swarm, and has no established Xinzhai link.
Next prioritize published run/output archives or follow the original VirSci project website for a readable dialogue example. Avoid further bulk review of bundled CAMEL code unless a specific provenance question requires it. Broader independent Chinese internet discovery remains necessary.

PRESERVATION
investigation/china/309-private/: repository metadata/tree, README.md, paper.html and extracted paper.txt, selected channel and inference-manager source. Raw captures private. Publication checks and SHA256SUMS retained. Existing machine suffices for these reads; the paper's simulation hardware is not an infrastructure recommendation for this investigation.