AI
DeReAct: Decomposed Reasoning and Acting for Reliable AI Agents
Publication: cs.AI updates on arXiv.org
Author: Ajay Vohra, Tao Chen, Neeti Narayan, Caron Zhang
Source published: 05/10/2026, 04:00:00 UTC
Briefing created: 05/10/2026, 04:09:38 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 05/10/2026, 04:03:24 UTC.
Briefings
DeReAct is a new agent architecture that separates action authorization from completion control by introducing external gating policies. It improves weaker AI models on benchmarks like GAIA and SWE-bench by 6.5 to 7.0 points for Qwen3-Coder-480B and 4.2 to 5.2 points for Claude Sonnet~4.5, while maintaining comparable performance for stronger models like Claude Opus~4.5. The method uses a Critic to validate actions and a Context Manager to reconstruct the environment state, ensuring more evidence-complete trajectories.
Key points
- DeReAct introduces a modular architecture that externalizes two gating policies: a Critic for action validation and a Context Manager for environment state reconstruction.
- DeReAct improves Pass@1 scores by 6.5 to 7.0 points for Qwen3-Coder-480B and 4.2 to 5.2 points for Claude Sonnet~4.5 on GAIA and SWE-bench Verified.
Context and limitations
Performance gains diminish as model capability increases. Trajectory and ablation analyses indicate effectiveness depends on targeted failure prevalence and the sufficiency of the gating policy itself.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 05/10/2026, 04:09:38 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeibatuwbm2533apfqtsst42iaur7dev5rbtnt5pvnjj2ub3wdskwom