AI Model2026
Compresses the model by pruning half its experts and quantizing the retained weights, targeting code, agentic workflows, long context, and vision. It retains 94.8%–98.7% of the base model’s scores on two coding benchmarks, at the cost of broader capabilities.
AI Model2026
A 65B dense language model trained on 12T text and code tokens for reasoning and agentic coding, with context support up to 1M tokens. Its Terminal-Bench 2.1 result stands out, though benchmark scores depend on evaluation setup.
Packages repaired code, environment setup, experiment launchers, and reproduction guidance to help coding agents reproduce ML experiments. Across 36 papers, these refined artifacts raised mean success from 76.4% to 92.6% while reducing downstream API spending.
AI Agent Papers2026
Haibo Jin, Peng Kuang +12University of Illinois Urbana-Champaign, Georgia Institute of Technology +1
Equips reusable resources with resident LLMs so agents can apply them in context. HEELO composes the needed units into task-specific workflows, checks results against environmental evidence, and revises only the work implicated by failures.
AI Dataset2026
Collects multi-step tool-use sessions from Codex Desktop, including 437 trajectories and 314 subagent delegation examples. The work is concentrated on Windows-based C#/C++ development and video-codec reverse engineering.