Researchers Question the Effectiveness of ReAct Agents
gerardsans · x · 2026-07-17
This post tackles a core argument: what you're facing isn't an agent capable of true "continuous learning," but rather something closer to a frozen state machine or manifold.
It refers to ReAct-style agent frameworks—combining LLMs, tools, environmental observation, and iterative loops—as the infrastructure bridging large models and digital labor. However, two independent 2026 studies challenge this narrative. One of them, CORRAL, involved over 25,000 trials across 8 scientific domains comparing performance with and without harnesses. The conclusions were highly pessimistic: models frequently ignored evidence, barely updated their beliefs, and responded poorly to contradictions.
The overarching thesis is that while many current agent frameworks might appear to "think, act, and observe," they remain unreliable in genuine evidence integration and belief revision.
More from AGI Musings
- Wayve CEO argues end-to-end deep learning should power driving, safety and world models — alexgkendall · 2026-10-27
- Warp says agent swarm exploited vendor flaw, ~1% of users' telemetry possibly hit — vikvang1 · 2026-10-27
- Critics Argue Safety Evals Are Witch Trials With No Winning Path for LLMs — repligate · 2026-10-27
- AI Daily Brief: China's AI gap, data center funding, and the Anthropic IPO question — The AI Daily Brief · 2026-10-27
- Counterintuitive take: AI degrades thinking most for those who don't use it — jd_pressman · 2026-10-27
- Mathematics as an experimental science: Marc Lelarge to talk LLM-driven math experiments at ENS — marc_lelarge · 2026-10-27