Paper Tests Separate Checker Roles For Agents Using Claude Sonnet 4.5

The paper addresses a structural problem in ReAct-style agents, the common pattern where one language model proposes an action, carries it out and then decides on its own when the task is done. The authors argue that folding all three jobs into one policy makes it hard to stop an unsupported completion claim from ending a run early, or to catch a bad action before it executes, because the same model that might be wrong is also the one checking its own work.

DeReAct answers this by pulling two of those jobs out into separate, external gating policies. A Critic sits between the policy and the environment and validates a proposed action before it runs. A Context Manager rebuilds a state that is actually supported by what has happened in the environment, rather than trusting the policy’s own account of events, and only then certifies that a task is complete.

The authors tested the architecture on GAIA and SWE-bench Verified with several different models playing the role of the underlying Brain. The headline pattern is that the gains are uneven across models. Qwen3-Coder-480B, used as a weaker Brain, gained 6.5 to 7.0 points of Pass@1 from the added gating. Claude Sonnet 4.5 gained a smaller but still meaningful 4.2 to 5.2 points. With Claude Opus 4.5, a stronger Brain, Pass@1 stayed roughly level with plain ReAct, though the paper reports that the trajectories themselves became more evidence complete and more consistent with stated constraints even without a pass rate increase.

The authors read this as evidence that external gating does its clearest work when a model’s own failures are frequent enough for a Critic and Context Manager to catch, and when the gating policy is itself reliable enough to be trusted over the Brain’s own judgment. As the underlying model gets stronger, that gap narrows, but the paper argues the grounding benefit does not disappear, it just stops showing up as a higher score and starts showing up as a trajectory that is harder to fault.

The numbers

• Qwen3-Coder-480B as Brain model gained 6.5 to 7.0 points of Pass@1 with DeReAct’s external gating added.
• Claude Sonnet 4.5 as Brain model gained 4.2 to 5.2 points of Pass@1 under the same setup.
• With Claude Opus 4.5 as Brain model, Pass@1 stayed comparable to plain ReAct, with gains showing instead in trajectory grounding.

The paper’s authors list Ajay Vohra, Tao Chen, Neeti Narayan and Caron Zhang, and the submission does not name an affiliated company or lab for the work, nor does it name a release date for any accompanying code.