Challenges and Evaluation Strategies for LLM Agents in 2024

In 2024, agents have made significant progress and become increasingly practical, but they still face several challenges.

Planning ability remains insufficient: current LLMs lack strong complex reasoning; COT/TOT methods do not observe feedback and are only suitable for simple tasks or initialization; ReAct and Reflection observe feedback but lack global thinking and often get stuck in inefficient local oscillations. In practice, a combination of COT planahead + Reflection is widely adopted to balance efficiency and accuracy. Algorithmically, structured thinking memory and OpenAI o1‑like “slow thinking” are needed, while product‑wise, white‑box interaction and domain SOPs are effective supplements.

Implementation and evaluation also present difficulties: a week‑long demo often becomes unusable after half a year, requiring systematic evaluation to guide optimization. Evaluation in the large‑model era is a technical task that must solve dataset generation and metric design. Dataset generation typically needs little or no supervision, leveraging LLMs to produce more and better evaluation data. Metrics must handle the flexibility of LLM answers, using new indicators such as RAGAS instead of strict accuracy.

These points constitute part of the Agent module in Knowledge Map 3.0. Interested readers can book the upcoming release event for a detailed presentation.

2025‑01‑16 19:00 , DataFunTalk will launch the live broadcast of the Data Modeling Knowledge Map release, offering free access to the roadmap; please reserve your spot.