Research Papers Worth Reading

ARC-AGI-2 Reasoning Race. Gemini 3.1 Pro hit 77.1% and Claude Opus 4.6 hit 68.8% — both more than doubling their predecessors' scores in a single generation. The entire field moved ~10 points over the prior two years; each model jumped 30-40 points in one release. The top three coding models (Opus 4.6 80.8%, Codex 80.0%, Gemini 3.1 Pro 80.6%) are in a statistical tie on SWE-bench. Source: Vellum AI

Arcee Trinity Large: Open 400B Sparse MoE. Largest open-weight US model. 400B total, 13B active per inference. MMLU 87.2 (vs Llama 4 Maverick 85.5). Trained in 33 days for ~$20M. Apache 2.0. 2-3x faster inference due to extreme sparsity. Source: Arcee AI Blog | arXiv

PCAS: Policy Compiler for Agent Security. Deterministic enforcement via Datalog-derived policy language. Improved compliance from 48% to 93% with zero violations. Fills the "enforcement gap" that threat models identified but didn't solve. Source: arXiv

SpargeAttention2: 95% Sparsity, 16.2x Speedup. Hybrid masking for robust sparse attention. Applicable for video generation and long-context inference. Source: arXiv

CUWM: Computer-Using World Model. Microsoft's desktop world model predicts UI state before acting. Test-time action search for GUI agents. Source: arXiv

VESPO: Stable Off-Policy LLM Training. Top paper on HuggingFace (152 upvotes). Stable at 64x staleness. Applicable for RLHF/RLAIF. Code on GitHub. Source: arXiv

SAGE: Self-Aware Guided Efficient Reasoning. Reasoning models know when to stop thinking, but sampling obscures it. Practical technique for reducing inference costs. Source: arXiv