---
type: "post"
id: "new-default-rl-loss-ipo-masks-on-probability-e29bff"
slug: "new-default-rl-loss-ipo-masks-on-probability-e29bff"
title: "New default RL loss IPO masks on probability difference instead of ratio, stabilizing async training."
description: "**Thesis:** In asynchronous RL, the masking rule can matter more than the nominal importance-sampling correction: Prime-RL’s documented IPO default uses absolute probability movement to limit trainer–inference drift."
retrieval_nugget: "**Thesis:** In asynchronous RL, the masking rule can matter more than the nominal importance-sampling correction: Prime-RL’s documented IPO default uses absolute probability movement to limit trainer–inference drift."
published_at: "2026-10-20"
updated_at: "2026-10-20"
record_date: "2026-10-20"
date_kind: "published_at"
topics: ["inference"]
source_urls: ["https://github.com/PrimeIntellect-ai/prime-rl/blob/main/docs/ipo.md"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:new-default-rl-loss-ipo-masks-on-probability-e29bff"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:0502ed03daeef036f9ab9fb7ea78a7e7e3572f372b67c3eae1dbc168da07ab7b","reviewed_at":"2026-10-20T08:05:15.041Z","source_evidence_count":1,"verified_claim_count":2,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":2,"implication_count":1,"watch_condition_count":1,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"In asynchronous RL, the masking rule can matter more than the nominal importance-sampling correction: Prime-RL’s documented IPO default uses absolute probability movement to limit trainer–inference drift.","observed_facts":[{"text":"Prime-RL documentation defines IPO masking as |π_train − π_infer| > ε and contrasts it with IcePop’s ratio band.","source_urls":["https://github.com/PrimeIntellect-ai/prime-rl/blob/main/docs/ipo.md"]},{"text":"In the reported GLM-4.5-Air SWE RL ablation, IcePop’s trainer–inference KL rose to more than 20× baseline around step 1.2k while solve rate fell from about 0.8 to about 0.3; both IPO arms kept mismatch low through that window.","source_urls":["https://github.com/PrimeIntellect-ai/prime-rl/blob/main/docs/ipo.md"]}],"mechanism":"The ratio treats tokens similarly regardless of baseline probability; an absolute difference instead responds to the amount of probability mass moved. In the two-outcome framing used by the documentation, that difference is a total-variation distance.","why_now":"Prime-RL identifies IPO as its default loss and publishes configurations to reproduce the comparison, making this a practical design choice for teams running asynchronous RL rather than a purely theoretical proposal.","implications":["Teams using asynchronous policy training can compare probability-difference masking against ratio masking while tracking mismatch KL, solve rate, entropy, and masked-token fraction."],"evidence_boundary":"The supplied document reports one controlled ablation on one model/task recipe; it does not establish general superiority across models, tasks, or training systems.","watch_conditions":["Whether independent runs across different tasks and trainer–inference configurations reproduce IPO’s lower mismatch without sacrificing task performance."],"related_records":[],"new_branch_reason":"No prior New Runtime analytical record was supplied for this loss-design change, so it starts a distinct record grounded in the Prime-RL documentation and its ablation."}
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/new-default-rl-loss-ipo-masks-on-probability-e29bff/","markdown":"https://newruntime.com/posts/new-default-rl-loss-ipo-masks-on-probability-e29bff.md","json":"https://newruntime.com/posts/new-default-rl-loss-ipo-masks-on-probability-e29bff.json"}
---

# New default RL loss IPO masks on probability difference instead of ratio, stabilizing async training.

## Retrieval answer

**Thesis:** In asynchronous RL, the masking rule can matter more than the nominal importance-sampling correction: Prime-RL’s documented IPO default uses absolute probability movement to limit trainer–inference drift.

**Thesis:** In asynchronous RL, the masking rule can matter more than the nominal importance-sampling correction: Prime-RL’s documented IPO default uses absolute probability movement to limit trainer–inference drift.

**Observed facts:** Prime-RL’s documentation defines IPO masking as |π_train − π_infer| > ε and contrasts it with IcePop’s ratio band. **Observed facts:** In the reported GLM-4.5-Air SWE RL ablation, IcePop’s trainer–inference KL rose to more than 20× baseline around step 1.2k while solve rate fell from about 0.8 to about 0.3; both IPO arms kept mismatch low through that window. These results come from a particular setup: a 100B MoE model, specific trainer and inference topology, batch and context settings, and SWE-Bench Verified evaluation.

**Mechanism:** The ratio treats tokens similarly regardless of baseline probability; an absolute difference instead responds to the amount of probability mass moved. In the two-outcome framing used by the documentation, that difference is a total-variation distance. The **Why now:** Prime-RL identifies IPO as its default loss and publishes configurations to reproduce the comparison, making this a practical design choice for teams running asynchronous RL rather than a purely theoretical proposal.

**First implication:** Teams using asynchronous policy training can compare probability-difference masking against ratio masking while tracking mismatch KL, solve rate, entropy, and masked-token fraction. **Evidence boundary:** The supplied document reports one controlled ablation on one model/task recipe; it does not establish general superiority across models, tasks, or training systems.

**First watch condition:** Whether independent runs across different tasks and trainer–inference configurations reproduce IPO’s lower mismatch without sacrificing task performance.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Inference](https://newruntime.com/topics/inference.md): Explore the inference topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)