RP-OPSD targets 'reasoning pivots' across 17 languages

A new arXiv preprint takes a small but interesting angle on a problem that keeps showing up in multilingual LLM work: getting a model that reasons well in English to actually reason well in the other languages it can speak.

The method, RP-OPSD, sits inside the broader family of on-policy self-distillation techniques, where a model teaches itself from its own rollouts. The twist is where the authors focus the distillation signal. Rather than treating every token equally, they try to isolate what they call reasoning pivots, the decisions in a chain of thought that advance or redirect the reasoning process. The pivots are inferred by comparing what a teacher produces with and without an English reference solution alongside the target-language prompt, then concentrating the privileged distillation signal on the tokens where those two views disagree the most.

Reported results cover mathematical reasoning benchmarks across 17 languages and multiple difficulty levels, and the authors claim the approach outperforms strong multilingual reasoning baselines and OPSD variants. The framing is that target-language reasoning is not one uniform stream of tokens, it is a mix of surface text and a smaller number of high-leverage decisions, and only the latter really need the extra supervision.

The honest caveat is that this is an unreviewed preprint, and the abstract does not name which specific languages, which base models, or how large the margin over baselines actually is. Those are the details that decide whether the technique holds up outside a math benchmark and whether it generalizes to code, dialogue, or agentic reasoning in the same languages.

For teams shipping reasoning assistants beyond English, the direction is what matters more than the specific numbers here. If focusing supervision on the tokens that actually change the trajectory of a reasoning trace is what unlocks cross-lingual transfer, that is a cheaper knob to turn than gathering more pretraining data in every target language.