DeReAct
A modular AI agent architecture that decouples reasoning from execution. It introduces two independent control components—a Critic and a Context Manager—externalizing action validation and task-completion judgment from the main model to improve agent reliability.
Timeline (last 90 days)
DeReAct is a modular agent architecture whose core idea is to externalize two gating strategies from the main model into independent control components; its name derives from Decomposed ReAct (decoupled reasoning and execution).
DeReAct's Critic component validates an action before it is actually executed, and only actions that pass review are permitted to act on the environment.
DeReAct's Context Manager is responsible for reconstructing the environment-grounded State and certifying task completion.
The research team evaluated DeReAct on two benchmarks: GAIA and SWE-bench Verified.
DeReAct improves the Pass@1 of Qwen3-Coder-480B by 6.5 to 7.0 percentage points.
DeReAct improves the Pass@1 of Claude Sonnet 4.5 by 4.2 to 5.2 percentage points.
For Claude Opus 4.5, DeReAct's Pass@1 is roughly on par with the original ReAct baseline.
External gating yields the most pronounced improvements for weaker main models, with gains gradually diminishing as the main model's capability increases.
For external gating to be effective, two prerequisites must be met: the target type of failure must be sufficiently common in the original trajectories, and the gating strategy itself must be sufficiently strong.
On strong models such as Claude Opus 4.5, DeReAct did not improve Pass@1 but made trajectories more evidence-complete and more compliant with constraints.
2 more timeline events
All Facts (12)
DeReAct is a modular agent architecture whose core idea is to externalize two gating strategies from the main model into independent control components; its name derives from Decomposed ReAct (decoupled reasoning and execution).
50%UnverifiedDeReAct's Critic component validates an action before it is actually executed, and only actions that pass review are permitted to act on the environment.
50%UnverifiedDeReAct's Context Manager is responsible for reconstructing the environment-grounded State and certifying task completion.
50%UnverifiedThe research team evaluated DeReAct on two benchmarks: GAIA and SWE-bench Verified.
50%UnverifiedDeReAct improves the Pass@1 of Qwen3-Coder-480B by 6.5 to 7.0 percentage points.
50%UnverifiedDeReAct improves the Pass@1 of Claude Sonnet 4.5 by 4.2 to 5.2 percentage points.
50%UnverifiedFor Claude Opus 4.5, DeReAct's Pass@1 is roughly on par with the original ReAct baseline.
50%UnverifiedExternal gating yields the most pronounced improvements for weaker main models, with gains gradually diminishing as the main model's capability increases.
50%UnverifiedFor external gating to be effective, two prerequisites must be met: the target type of failure must be sufficiently common in the original trajectories, and the gating strategy itself must be sufficiently strong.
50%UnverifiedOn strong models such as Claude Opus 4.5, DeReAct did not improve Pass@1 but made trajectories more evidence-complete and more compliant with constraints.
50%UnverifiedDeReAct's completion control mechanism involves a trade-off: sacrificing the efficiency of earlier termination in exchange for stronger grounding.
50%UnverifiedDeReAct's modular gating mechanism can be added as a general-purpose reliability enhancement layer on top of existing ReAct agents without retraining the main model.
50%