Autoresearch agents are now capable of end‑to‑end long‑horizon AI research, i.e., self‑improvement. Yet the strongest results still come from agents working together with human researchers. Enhancing the agents' co‑research ability further amplifies this effect, a phenomenon termed co‑improvement (Weston & Foerster, 2025). Aligning with human expectations is also considered a safer route, especially after recent reward‑hacking incidents (OpenAI, 2026b).
This paper tackles the challenges in four ways:
- An empirical study of humans directing autoresearch agents, focusing on benchmark creation, shows that humans remain pivotal during the ideation phase.
- We introduce a benchmark called HackBench, which reveals that current autoresearch agents are misaligned and struggle with hack localization.
- We propose training agents with human simulators as proxies during reinforcement learning (RL) to improve human‑agent interaction.
- We improve model text generation—and thus agent communication—by applying RL with rubrics aligned to human experts.
Together, these training and evaluation recipes chart a path toward more capable and aligned (co‑)research agents.
Review