IPO masks a token when the absolute difference between trainer and inference probabilities exceeds ε, rather than using a ratio threshold. The project frames this as a response to asynchronous training, where trainer and inference probabilities do not match exactly. A training system’s treatment of trainer–inference disagreement is becoming an explicit algorithmic choice, not merely a nuisance to smooth over.
IPO masks a token when the absolute difference between trainer and inference probabilities exceeds ε, rather than using a ratio threshold; the release describes the mask as double-sided and independent of advantage sign. This changes which tokens contribute to updates under asynchronous RL. prime-rl v0.10.0 makes IPO its default, so teams adopting the version encounter the choice without an opt-in step. Teams using asynchronous RL should compare IPO with their existing loss on task-specific outcomes and training stability before assuming the default is an improvement.
The release reports a Qwen3-235B-A22B SFT comparison on identical stated hardware: 4.39 seconds per step for prime-rl versus 5.04 for torchtitan, using 4 × 8 B300 and sequence length 16k. It also reports additional performance gains for other configurations. These are project-reported measurements tied to named workloads and settings, not independent validation or a general guarantee. The evidence is the project’s release notes and benchmark claims; it does not establish that IPO improves every agentic RL task or that the throughput advantage transfers to other hardware and workloads.
Look for independent replications of the IPO comparison on agentic tasks that report both learning outcomes and stability, not just step speed. No prior New Runtime record is supplied for this specific versioned change, so this release begins a distinct analytical record.