Hi insiders — has anyone (publicly, or quietly inside the labs) actually tried the obvious extension of Roger & Greenblatt 2023 and West et al.'s tandem training? Namely: during RL training itself, randomly perturb the CoT mid-rollout — paraphrase sentences, translate chunks, delete "useless-looking" paragraphs, shuffle order, occasionally hand off a few tokens to a weaker model — so the model has to do its reasoning in a form robust to all of this. Eval would presumably use a held-out family of perturbers/paraphrasers not seen during training, to check the model didn't just learn to game the specific training-time ones.

Curious whether this exists as a paper, an internal experiment, a failed experiment, or a known-bad idea I'm missing the obvious flaw of.