---
title: "Autonomous LLM post-training with Tunix on TPUs"
description: "autofinetune applies autonomous research loops to LLM post-training, covering supervised fine-tuning (SFT) and reinforcement learning with GRPO through Tunix, Gemma, Cloud TPUs, Antigravity CLI, and Gemini Flash 3.7. A human defines the arena in program.md and supplies a self-contained run.py, while an agent edits hyperparameters, launches jobs, evaluates a target metric, keeps winning Git commits, reverts regressions, and logs results in results.tsv. On FunctionGemma 270M with google/mobile-actions, 20 Cloud TPU v5e-1 experiments tested LoRA settings, optimization choices, schedules, batch size, and seeds without changing the dataset, epochs, or architecture. A second study used Gemma 3 1B, GSM8K, and Cloud TPU v6e-1 for 40 GRPO experiments over two to three days, improving the combined numerical-and-format reward by about 10%."
---

# Autonomous LLM post-training with Tunix on TPUs

[Google](https://yomu.fyi/company/google) · Wei Wei · Sep 11, 2026

**Type:** Benchmark

## Summary

autofinetune applies autonomous research loops to LLM post-training, covering supervised fine-tuning (SFT) and reinforcement learning with GRPO through Tunix, Gemma, Cloud TPUs, Antigravity CLI, and Gemini Flash 3.7. A human defines the arena in program.md and supplies a self-contained run.py, while an agent edits hyperparameters, launches jobs, evaluates a target metric, keeps winning Git commits, reverts regressions, and logs results in results.tsv. On FunctionGemma 270M with google/mobile-actions, 20 Cloud TPU v5e-1 experiments tested LoRA settings, optimization choices, schedules, batch size, and seeds without changing the dataset, epochs, or architecture. A second study used Gemma 3 1B, GSM8K, and Cloud TPU v6e-1 for 40 GRPO experiments over two to three days, improving the combined numerical-and-format reward by about 10%.

## Context

Traditional LLM post-training is described as a repetitive manual cycle of forming hypotheses, editing scripts and hyperparameters, launching accelerator jobs, monitoring evaluations, and recording or reverting results. The work applies autonomous research loops to make this process continuous and agent-driven.

## Approach / What changed

The system uses a human-authored program.md to define boundaries, evaluation criteria, and constraints, plus a self-contained run.py for execution. An agent modifies permitted training settings, runs fine-tuning jobs, evaluates the target metric, retains winning Git commits, reverts regressions, and records experiments in results.tsv.

## Takeaways

- The SFT arena allowed changes to LoRA rank and alpha, target projection layers, learning rates, schedules, optimizers, gradient clipping, batch size, and seeds, while disallowing dataset, epoch-count, and model-architecture changes.
- The GRPO study used Post\_RL\_metric, defined as numerical\_accuracy plus format\_accuracy, to simplify the agent’s outer optimization loop.
- In the GRPO trajectory, the agent identified improvements involving LoRA configurations, rollout temperature, KL penalty, and system prompt, raising total reward by approximately 10%.

**Tags:** [Gemma](https://yomu.fyi/topic/gemma), [LoRA](https://yomu.fyi/topic/lora), [Reinforcement Learning](https://yomu.fyi/topic/reinforcement-learning), [TPU](https://yomu.fyi/topic/tpu), [Tunix](https://yomu.fyi/topic/tunix)

- Source: [Google](https://developers.googleblog.com/autonomous-llm-post-training-with-tunix-on-tpus)
- Source URL: https://developers.googleblog.com/autonomous-llm-post-training-with-tunix-on-tpus
- Ingested by Yomu: 2026-09-11T16:01:44.912Z

[Read original post](https://developers.googleblog.com/autonomous-llm-post-training-with-tunix-on-tpus)