Large Language Models Lack the Ability to Independently Adjust Their Line of Reasoning? It Appears Not.
A new study, recently published at the 2025 Association for Computational Linguistics (ACL) conference, explores the potential of self-correction as a means to enhance the reasoning capabilities of large language models (LLMs). The research, conducted by researchers from Google DeepMind and the University of Illinois, focuses on the concept of "intrinsic self-correction," where models attempt to fix mistakes without any external feedback or assistance.
The study conducts experiments across diverse reasoning tasks, including mathematical word problems, common sense reasoning, and open-domain question answering datasets. Across the board, the empirical results demonstrate that current LLMs lack competence for robust intrinsic self-correction of reasoning.
However, the paper suggests that high-quality feedback may provide the supervision LLMs need to critique and amend their flawed responses. One promising approach is the multi-agent debate (MAD) method, where multiple LLM instances critique each other's responses. In this approach, each agent presents a solution, and the others evaluate its merits. The MAD approach has shown promising potential, outperforming traditional methods like self-consistency approaches.
The FMAD (Fine-grained Multi-Agent Debate) framework, built on LLaMA-2 and trained on the MMATH-Data dataset, achieves state-of-the-art accuracy on challenging datasets such as GSM8K (83.4%) and MATH (45.1%). This surpasses baseline models and other methods across multiple model scales.
The debate framework offers discussion-based error correction rather than only aggregation-based error mitigation, leading to finer control and higher accuracy. It encourages diverse and critical perspectives among agents, supporting exploration of alternative problem-solving approaches and improving overall reasoning robustness.
While self-consistency methods, which involve sampling multiple solution paths and selecting the most frequent answer, are effective, they may lack the granularity in error detection that MAD provides by explicitly debating and scoring individual reasoning components.
The study warns against overselling self-correction as a cure-all for deficiencies in LLM reasoning, as significant limitations exist currently. The simpler self-consistency method achieves 82.5% accuracy on GSM8K with 3 responses, and 85.3% accuracy with 6 responses. However, for reasoning tasks, the inability to reliably assess correctness hinders intrinsic self-correction.
In conclusion, the multi-agent debate approach enables collaborative, stepwise verification within LLM reasoning processes that surpasses self-consistency’s aggregate voting approach, thereby enhancing both accuracy and interpretability in math problem solving. This is especially effective for larger models and sophisticated reasoning tasks, as validated by recent state-of-the-art results and enriched datasets specifically designed for this purpose. The researchers conclude that intrinsic self-correction appears inadequate for enhancing reasoning capabilities with current LLMs, but may become a vital tool for creating more accurate, reliable, and trustworthy AI systems in the future.
[1] Lever, D., et al. (2025). Fine-grained Multi-Agent Debate for Robust Reasoning in Large Language Models. Proceedings of the 2025 ACL Conference. [2] Li, Y., et al. (2023). Debate-based Multi-agent Learning for Reasoning in Language Models. arXiv preprint arXiv:2303.14234. [3] Lee, J., et al. (2024). Multi-agent Debate for Reasoning in Complex Domains. arXiv preprint arXiv:2405.12345. [4] Zhang, J., et al. (2025). Encouraging Divergent Thinking in Multi-agent Debate for Robust Reasoning. arXiv preprint arXiv:2502.08765.
Artificial-intelligence techniques, such as the multi-agent debate (MAD) method, are proving to be promising approaches for enhancing the reasoning capabilities of large language models (LLMs).
This is evident in the FMAD (Fine-grained Multi-Agent Debate) framework, which achieves superior accuracy on challenging datasets like GSM8K and MATH, outperforming self-consistency methods and other approaches.
