Introduction
While Large Language Models (LLMs) have made massive strides in reasoning, they still struggle with self-correction. Existing techniques like Chain-of-Thought (CoT) or Tree-of-Thoughts (ToT) often suffer from significant computational overhead and error propagation. If a model provides poor feedback to itself during an iterative process, those errors accumulate, leading to "hallucination loops."
CLEAR (Contrasting Textual Feedback with Experts and Amateurs for Reasoning) addresses these issues by leveraging a diversity of perspectives. By contrasting feedback from a large "expert" model and a smaller "amateur" model, CLEAR improves LLM outputs while remaining significantly more computationally efficient than complex tree-search methods.
Methodology: The CLEAR Framework
The CLEAR framework operates through an iterative loop consisting of two primary modules:
- Node Evaluator: For a given output, two different LLMs—an Expert (e.g., GPT-4o) and an Amateur (e.g., GPT-3.5-turbo)—provide textual feedback. They highlight positive aspects, negative aspects, and specific areas for improvement.
- Feedback Filter: A third prompting step uses an LLM to contrast these two sets of feedback. It synthesizes them into a single "filtered feedback" prompt, giving higher priority to the expert but using the amateur's input to diversify the critique.
For objective tasks like mathematics, the authors introduce BeCLEAR, which incorporates a Best-First Search algorithm. BeCLEAR assigns numerical scores to nodes and uses a cost function to decide which reasoning path is most promising to expand next, rather than just refining the most recent output.
The Power of Diversity
A striking finding of this research is that the "Amateur" model isn't just a cheaper substitute; it provides semantically different feedback. Analysis of text embedding similarity shows that contrasting an expert with an amateur yields more diverse feedback than contrasting two experts. This "diversity of thought" helps catch different types of errors, preventing the model from falling into narrow reasoning loops.
Key Results and Efficiency
CLEAR consistently outperformed state-of-the-art (SOTA) prompting methods across various benchmarks:
- Mathematics: BeCLEAR reached 97.2% accuracy on GSM8K, surpassing THOUGHTSCULPT (96.0%) and CoT (86.0%).
- Storytelling: In human-like evaluations of "interestingness," CLEAR scored 73%, a significant jump over ToT (43%) and THOUGHTSCULPT (61%).
- Constrained Generation: CLEAR achieved 99.3% concept coverage, effectively solving the "hard" version of the CommonGen dataset.
- Toxicity: It reduced toxicity scores by up to 22% compared to standard baselines.
Crucially, CLEAR achieves these results with a linear or shallow graph structure, requiring roughly 1/4 the computational cost of more complex methods like THOUGHTSCULPT in storytelling tasks.
Limitations and Future Work
The authors identify that CLEAR’s performance is dependent on the quality of the initial feedback prompt and the underlying models. If both the expert and amateur share the same fundamental bias, the "contrast" step cannot fix it. Future research will explore optimal model pairings and applications in code generation and multi-modal reasoning.
Summary updated