Recently, Meta's artificial intelligence research team collaborated with the University of Edinburgh to develop a new technology that can predict the correctness of reasoning in large language models (LLMs) and fix errors when they are detected. This method, called Circuit-Based Reasoning Verification (CRV), aims to deeply observe the internal "reasoning circuits" of LLMs, detecting signs of computational errors when the model solves problems.
Studies show that CRV can detect reasoning errors in LLMs with high accuracy by building and observing the computational graph of internal activations. This breakthrough means researchers can use deep internal information to intervene in the model's erroneous reasoning in a targeted manner.
Chain-of-thought reasoning (CoT), as a method to improve the performance of LLMs on complex tasks, has been widely applied, but its reliability remains an issue. Existing verification methods are mainly divided into two categories: "black-box" methods verify by analyzing the final generated tokens or confidence scores; "gray-box" methods attempt to observe the internal state of the model, but these methods cannot explain the root cause of computational failures.
CRV adopts a "white-box" verification approach, assuming that the model uses specific subgraphs of neurons when performing tasks. By replacing standard dense layers with trained "decoders," researchers make the target LLM interpretable, allowing observation of its internal workings. Then, CRV builds an "attribution graph" mapping the causal flow of information between different parts of the model and extracts "structural fingerprints" describing the characteristics of the graph. Finally, a "diagnostic classifier" is trained to predict whether the reasoning steps are correct.
The research team conducted experiments on the Llama3.1 model, and the results showed that CRV outperformed other verification methods on various datasets and metrics, demonstrating its strong effectiveness. Additionally, the study found that error signatures in different reasoning tasks are domain-specific, meaning that different types of reasoning rely on different internal circuits.
Most importantly, CRV is not just a correlation analysis; it provides a transparent computational view, enabling failed predictions to be traced back to specific components. Researchers can thus suppress error features in real-time to correct the model's reasoning path.
Key Points:
🌟 CRV technology can effectively predict and repair reasoning errors in LLMs, improving the reliability of AI.
🧠 The study adopted a "white-box" verification approach, revealing the internal reasoning circuits of LLMs.
🔧 The successful application of CRV lays the foundation for the development of future AI model debugging tools.
