Vibeleaderboard
← All Intel
Intel / article

DeReAct: Decomposed Reasoning and Acting for Reliable AI Agents

Source
arxiv.org
Author
Ajay Vohra, Tao Chen, Neeti Narayan, Caron Zhang
Date

Context

ReAct-style agents let one language model propose each action, run it against the environment and decide when the task is finished. The DeReAct paper, submitted on 1 October 2026, argues that this coupling lets errors spread and lets unsupported claims of success end a run. It adds two separate checkers: a Critic that validates each proposed action before it executes, and a Manager that rebuilds a record of the environment's state from evidence and certifies that the task is complete.

On GAIA (general assistant tasks) and (real software bugs), the authors report first-attempt success (Pass@1) gains of 6.5 to 7.0 points for Qwen3-Coder-480B from Qwen and 4.2 to 5.2 points for Claude Sonnet 4.5 from Anthropic. With Claude Opus 4.5, Pass@1 stayed comparable to plain ReAct, though the abstract says its trajectories were more evidence-complete. The supported reading is narrow: external gating helped weaker models most. Intel's earlier coverage of ECLoop describes a different system that blocks coding- actions until evidence conditions are met.

Key takeaways · AI-distilled
  • DeReAct splits the usual single ReAct policy into gated roles: a Critic validates each proposed action before it runs, and a Context Manager rebuilds an environment-supported state and certifies when the task is actually complete.
  • The motivation: when one LLM proposes actions and also decides it is done, unsupported completion claims can end a run early and errors propagate, because authorization and completion cannot be enforced independently.
  • On GAIA and SWE-bench Verified, gains were largest for weaker base models, about 6.5 to 7.0 Pass@1 points for Qwen3-Coder-480B and 4.2 to 5.2 for Claude Sonnet 4.5, and shrank as the base model got stronger.
  • With Claude Opus 4.5, Pass@1 stayed roughly level with ReAct, but the authors report more evidence-complete, constraint-satisfying trajectories, trading earlier termination for stronger .
Terms in this piece · Glossary
  • context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
  • SWE-bench — The standard benchmark for AI coding agents: real GitHub issues from real repositories, scored by whether the agent's patch passes the project's own tests.
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
  • grounding — Tying a model's answers to checkable sources — retrieved documents, live data, tool results — instead of letting it answer from memory alone.
Recommended reads
Comments

Checking sign-in…

Loading comments…