Introduction
When a Large Language Model (LLM) provides an incorrect response to a complex reasoning task, a common user reaction is to prompt the model to "try again" or "check your work." However, recent research suggests that this "redundant prompting"—repeatedly asking for a solution without providing specific guidance—can be counterproductive. This paper explores the phenomenon of performance degradation under repetitive feedback and introduces Recursive Chain-of-Feedback (R-CoF), a novel framework designed to help models self-correct by decomposing errors into manageable sub-problems.
The Divergence Trap in Redundant Prompting
The authors identify a significant issue termed the Chain-of-Feedback (CoF) effect. In standard interactions, providing meaningless or repetitive feedback often causes the model's performance to diverge further from the ground truth rather than converge upon it. As the number of trials increases, the model may hallucinate more wildly or eventually "give up," claiming no solution exists.
To quantify this, the researchers developed a penalty function measuring the absolute deviation from correct numerical answers. If a model repeats an old error or fails to provide a solution, the penalty increases exponentially ($10 \times 2^n$). This metric highlights how standard iterative prompting fails to bridge the reliability gap for users who lack the expertise to verify complex reasoning steps themselves.
Methodology: Recursive Chain-of-Feedback (R-CoF)
To address this reliability gap, the authors propose Recursive Chain-of-Feedback (R-CoF). This method moves away from asking the model to re-evaluate the entire problem. Instead, it focuses on isolating and correcting specific failures within the reasoning chain.
The R-CoF process involves four key stages:
- Identify: Pinpoint the exact step where the reasoning logic failed.
- Isolate: Extract that specific incorrect step and transform it into a standalone, simplified sub-problem.
- Recursive Call: Query a separate LLM instance to solve only the isolated sub-problem, removing the distracting context of the original complex prompt.
- Integrate: Reinsert the corrected sub-step into the original reasoning chain to finalize the solution.
This approach effectively bypasses the "ignorance" problem, where models acknowledge a mistake but immediately repeat it because they are overwhelmed by the global context of the problem. By solving the step in a "clean" environment, the model's attention is focused, significantly increasing the likelihood of a correct calculation.
Experimental Results
The framework was evaluated using a subset of 50 difficult problems from the MATH dataset that ChatGPT-3.5 initially failed to solve. The results demonstrate the efficacy of the recursive approach:
- Baseline: 0/50 (0% accuracy) using standard redundant prompting.
- R-CoF (1 Recursive Call): Accuracy jumped to 31/50 (62%).
- R-CoF (2 Recursive Calls): Accuracy reached 37/50 (74%).
These findings suggest that even a single level of recursive decomposition can resolve the majority of errors that previously led to total model failure.
Conclusion and Future Directions
While R-CoF significantly improves reasoning accuracy, it currently relies on manual identification of the incorrect reasoning step. Future work aims to automate this error-detection phase to create a fully autonomous self-correction loop. Additionally, expanding the testing to more advanced models and diverse domains such as code generation will be essential to understanding the broader applicability of recursive feedback mechanisms.
Summary updated