Skip to content
RSI Closed-Loop Pre-Training Reaches Frontier Scale, Surpasses Qwen3-30B-A3B

RSI Closed-Loop Pre-Training Reaches Frontier Scale, Surpasses Qwen3-30B-A3B

The article details the first public case of applying recursive self-improvement (RSI) closed loops to full frontier-scale pre-training, producing the 801-30B-A3B MoE model trained on 11.8T tokens, which outperforms Qwen3-30B-A3B and Nemotron-3.5-Lightning-30B-A3B through three types of feedback loops and agent-generated training data.

PaperAgent
PaperAgent
PaperAgent

RSI Moves from Post-Training to Frontier Pre-Training

SII & Xinxu AI claim the first publicly detailed application of RSI-style closed loops to a complete frontier-scale pre-training project. The resulting model, 801-30B-A3B (30B total parameters, 3B active per token, Mixture-of-Experts), was trained from scratch on 11.8T tokens in three months and beats Qwen3-30B-A3B and Nemotron-3.5-Lightning-30B-A3B .

RSI closed-loop pre-training overview
RSI closed-loop pre-training overview

Unlike post-training RSI — where a ready-made base is iterated in hours for a verifiable score — frontier pre-training has two distinct constraints:

A full run costs hundreds of thousands of GPU hours; there is usually no second chance.

The goal is to leave potential for post-training, not to maximize an immediate leaderboard score.

Figure 1:RSI 从后训练走向预训练
Figure 1:RSI 从后训练走向预训练
Table 1:后训练与前沿预训练的 RSI 差异
Table 1:后训练与前沿预训练的 RSI 差异

The continuously evolving object is not the model weights themselves, but the production system composed of agents, researchers, data pipelines, and training infrastructure .

Core Solution: Three Categories, Six Closed Loops

Figure 2:三类六条 RSI 闭环
Figure 2:三类六条 RSI 闭环

The framework defines three categories of closed loops, each containing two instances (six total), all sharing a common context infrastructure that preserves failure reasons so the next iteration does not repeat mistakes — recursion happens in organizational memory .

1. Debugging Loops

Infrastructure debugging and data debugging trace anomalies to components that can be re-run at low cost. Agents perform large-scale search over 130+ infrastructure code versions, 106 hyperparameter fields, and 1,053 data subsets ; humans retain triage priority and fix approval.

2. Potential Loops

Tiny-SFT, learning-capacity probes, and multi-protocol evaluation judge the base model's post-training headroom, then adjust the training recipe accordingly.

3. Experience-to-Metric Loops

Classifier development, site discovery, and crawler coverage — traditionally experience-dependent — are turned into quantifiable, iterative processes.

Table 3:五项跨栈诊断与修复
Table 3:五项跨栈诊断与修复

Cannot re-run training? Re-run diagnostics first. Infrastructure loops start from internal signals (loss, gradients, expert load); data loops reverse-trace credential leaks, template jumps, low-diversity repeats. In about one month the team completed five cross-stack diagnostics covering load balancing, optimizer, data quality, weight decay, and data loading. On the same 1T-token budget, the 22-benchmark average rose from 47.78 → 51.96 → 56.83 , though data and recipe changed simultaneously so no single fix can be isolated.

Figure 10:基座停滞与 SFT 后持续提升并存
Figure 10:基座停滞与 SFT 后持续提升并存

Pre-Training Data: Agent-Automated Production

A second self-improvement path: AI not only consumes training data but also manufactures it.

Figure 14:数据分类器的迭代开发闭环
Figure 14:数据分类器的迭代开发闭环

The team converted 260k raw PRs into 220k executable tasks , generated ~2M trajectories across nine agent frameworks, verified and kept 1.1M trajectories (>100B tokens, avg ~91K tokens each) . Combined with other sources, agent-produced data reached 134B tokens and entered long-context pre-training.

Key to scale: generate executable, verifiable trajectories — environment, instructions, and tests must align; prevent answer downloading, test tampering, or reference-patch reading.

论文 Figure 18:智能体数据生产漏斗
论文 Figure 18:智能体数据生产漏斗

Strong Results, But Not Unattended Evolution

Base-model evaluation: 13-task average 73.13 , second only to the larger Qwen3.5-35B-A3B, and above Qwen3-30B-A3B's 66.16. After identical light SFT, general-benchmark average still ranks second.

Table 7:基座模型评测结果
Table 7:基座模型评测结果

Agent tasks: four-benchmark average 58.43 (highest in the paper's comparison), KernelBench 77.60 .

Table 9:智能体基准评测结果
Table 9:智能体基准评测结果

Caveats: 801 received only 6.33B tokens of light SFT, no RL ; competitors are fully post-trained models — not an apples-to-apples comparison. Results come from a single run; Tiny-SFT lacks multi-seed replication; small gaps between adjacent checkpoints should be treated cautiously.

RSI has entered full frontier pre-training for the first time via engineered, organizational small loops — not because AI can already self-upgrade without humans. The model is a by-product; the methodology is the main contribution. Shared context preserves experience, small loops repair links, verifiable feedback drives the next update — this may be the real starting point of RSI on the LLM production line.

No Second Run: Does RSI Stop at Frontier Pre-training?
arXiv:2611.10719v1
Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

Mixture of ExpertsBenchmarkingRSIagent-generated dataclosed-loop pre-trainingfrontier pre-training
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.