We identified three distinct failure modes in modern LLM agents. We show that existing frameworks like Plan-and-Solve, ReAct, and Reflexion each primarily address only one of these. Our proposed Unified Agent Structure is the first to explicitly design components to counteract all three failure modes within a single, coherent architecture. We demonstrate through rigorous ablation studies that removing any one of our core components (Planner, CoT-SC Worker, or Evaluator) leads to a measurable drop in performance, proving that it is the synergistic combination of these elements that drives the superior performance.
We argue that these components must operate within a tightly-coupled, hierarchical control loop to be truly effective. Our contribution is the design of such a architecture, where a Planner provides high-level strategic direction, a Worker ensures reasoning robustness via CoT-SC, and an LLM Evaluator acts as the crucial link between them, monitoring progress and governing a Reflection process that continuously refines the agent's contextual understanding.
I'm not completely sure yet, but I am thinking of benchmarking using HotPotQA, Webshopping, Alfsword, and other benchmarks that research papers used mentioned in the "sources" below.
Role: Defined it as a controller or dispatcher. Its primary job is to check the state of the overall sub-task progress.
Outputs/Decisions:
- [Task Complete]: The sub-task is finished. Stop the current loop and signal success to the main controller.
- [In Progress, No Error]: The task is not finished, but things are proceeding.
- [Stuck / Infinite Loop]: The agent is not making progress and is repeating actions or thoughts.
Action Flow:
- If [Task Complete], the flow stops.
- If [In Progress, No Error], the output is sent to Self-Reflection for summarization (ST Memory update), and the loop continues to the Worker.
- If [Stuck / Infinite Loop], this is an error condition that needs to be handled. You mentioned stopping and moving on, which is one strategy.
The Evaluator isn't judging the correctness of a single action, but rather the momentum of the entire sub-task loop.
Tree of Thoughts: Deliberate Problem Solving with Large Language Models 12/3/2023 · Shunyu Yao, Karthik Narasimhan, Jeffrey Zhao, et al.
Reflexion: Language Agents with Verbal Reinforcement Learning 10/9/2023 · Shunyu Yao, Karthik Narasimhan, Noah Shinn, et al.
Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models 5/26/2023 · Lei Wang, Yunshi Lan, Zhiqiang Hu, et al.
ReAct: Synergizing Reasoning and Acting in Language Models 3/9/2023 · Nan Du, Shunyu Yao, Karthik Narasimhan, et al.
Self-Consistency Improves Chain of Thought Reasoning in Language Models 3/7/2023 · Denny Zhou, Sharan Narang, Jason Wei, et al.
Toolformer: Language Models Can Teach Themselves to Use Tools 2/9/2023 · Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, et al.
FireAct: Toward Language Agent Fine-tuning 10/9/2023 · Shunyu Yao, Karthik Narasimhan, Baian Chen, et al.