---
format: "aidr-story-markdown/v1"
id: "27ef7f6786c93b5172b849a0c64c4e8eaf11d4bd9d5c66ad93e5aead8c80ce24"
canonical_url: "https://aidr.today/27ef7f67?lang=en"
title: "Unlocking Stable On-Policy RL on TPU with Bitwise-Zero Train–Inference Mismatch"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-16T16:05:39.000Z"
category: "Research"
topics: ["google","llm","rl","tpu","inference"]
source_urls: ["https://developers.googleblog.com/unlocking-stable-on-policy-rl-on-tpu-with-bitwise-zero-traininference-mismatch/"]
summary: "A four-part JAX and Pallas recipe in Tunix was used to eliminate floating-point train–inference logprob drift for on-policy LLM reinforcement learning across 64 Cloud TPU v5p chips."
---

# Unlocking Stable On\-Policy RL on TPU with Bitwise\-Zero Train–Inference Mismatch

> [Open the canonical story](<https://aidr.today/27ef7f67?lang=en>)

**Published:** 2026-11-16T16:05:39.000Z
**Category:** Research
**Topics:** google, llm, rl, tpu, inference

## Summary

A four\-part JAX and Pallas recipe in Tunix was used to eliminate floating\-point train–inference logprob drift for on\-policy LLM reinforcement learning across 64 Cloud TPU v5p chips\.

## Sources

- [Story source](<https://developers.googleblog.com/unlocking-stable-on-policy-rl-on-tpu-with-bitwise-zero-traininference-mismatch/>)

