AI paper roundup for October 5–11 spans agent harnesses, verification and memory

In its October 5–11 roundup, dair_ai describes Harness Learning, which trains a model to revise an agent’s harness without changing the solver model’s weights; CorpusMap, which links documents through recurring entities; and VeriHarness, which checks disagreements and claims shared across agent runs. Other picks cover updating models alongside their harnesses, post-training trade-offs, memory and simulated users.