AI Research Pulse

Agentic Reasoning, Scientific Workflows, and Adaptive Harnesses

Agentic Reasoning, Scientific Workflows, and Adaptive Harnesses

CARD reports gains over GRPO on deep-search and multi-hop QA without extra rollouts or reward models, but the excerpt lacks benchmark names and implementation detail. Quest’s training claims remain abstract-level; AgentBaton reports 65.9% average API-cost reductions, while SO-Arena finds capability and reliability rankings diverge. Reflection Anchors proposes propagation-aware visual retention without quantitative results in the available abstract; SWE-CoWork highlights long-horizon coding workflow evaluation but provides no scores or methodology. Memory methods still show transfer bottlenecks.

Sources (26)
Updated Nov 18, 2026