Can LLMs Learn by Teaching for Better Reasoning? A Preliminary Study (2406.14629v3)

Abstract: Teaching to improve student models (e.g., knowledge distillation) is an extensively studied methodology in LLMs. However, for humans, teaching improves not only students but also teachers, by fostering more rigorous and clear reasoning as well as knowledge building. We ask: Can LLMs also learn by teaching (LbT) for better reasoning? If the answer is yes, we can potentially unlock the possibility of continuously advancing the models without solely relying on human-produced data or stronger models. In this paper, we provide a preliminary exploration on this question. We show that LbT ideas can be incorporated into existing LLM training/prompting pipelines and bring improvements. Specifically, we design three methods, each mimicking one of the three levels of LbT: observing students' feedback, learning from the feedback, and learning iteratively, with the goals of improving answer accuracy without training or improving models' inherent capability with fine-tuning. We reveal some findings: (1) Teaching materials that make it easier for students to learn have clearer and more accurate logic when using in-context learning as the student's "learning" method; (2) Weak-to-strong generalization: LbT might help improve strong models by teaching weak models; (3) Diversity in students might help: teaching multiple students could be better than teaching one student or the teacher itself. We hope that our exploration can inspire future research on LbT and more broadly adopting the advanced techniques in education to improve LLMs. The code and website are at https://github.com/imagination-research/lbt and https://sites.google.com/view/LLM-learning-by-teaching.

An Analytical Overview of "Can LLMs Learn by Teaching? A Preliminary Study"

The paper "Can LLMs Learn by Teaching? A Preliminary Study" investigates a potential paradigm shift in training LLMs through a method inspired by human pedagogy: Learning by Teaching (LbT). The authors propose that just as humans can enhance their understanding by teaching others, LLMs might also benefit from teaching, thus facilitating continuous model improvement without relying heavily on external human data or stronger models. This paper explores the feasibility of this concept by developing and implementing strategies that mimic the human LbT process.

Methodologies Explored

The paper introduces three distinct methods corresponding to different levels of the LbT concept:

1. M1 - Observing Students' Feedback: This method seeks to improve answer accuracy without retraining. It employs a "search-based output generation" pipeline, where multiple rationales are generated and scored based on how effectively they can teach student models to solve similar problems through in-context learning.
2. M2 - Learning from Feedback: Here, the focus is on enhancing the inherent capabilities of LLMs through fine-tuning. This method uses a "generating-scoring-finetuning" pipeline where rationales are scored based on their pedagogical impact, rather than merely correct answer alignment, and used to fine-tune the model via Direct Preference Optimization (DPO).
3. M3 - Iteratively Learning from Feedback: This method iterates the student feedback process to improve both the teachers' teaching materials and intrinsic understanding. The improved exemplars arising from diverse student feedback not only enhance student learning outcomes but also refine the teacher model's performance.

Key Results and Observations

The paper reveals several intriguing observations that resonate with human education:

• Weak-to-Strong Generalization: Surprisingly, stronger models still benefit from teaching weaker models, indicating that potential lies in collaboratively leveraging both model types in enhancing capabilities.
• Diversity in Students Enhances Teaching: Teaching multiple diverse students instead of self-teaching appears beneficial. This increase in diversity can cultivate a synergy of knowledge across model types, akin to cross-pollination of ideas among humans.

The experimental results consolidate these insights across various computational benchmarks, including mathematical reasoning and code synthesis tasks. Notably, the LbT paradigm exhibited substantial improvements over traditional self-consistency and self-evaluation strategies, demonstrating higher answer accuracy and improved model capabilities.

Implications and Future Directions

This initial exploration into LbT for LLMs opens avenues for subsequent research, potentially impacting both practical applications and theoretical advancements in AI. The implications are twofold:

• Practical Implication: Employing LbT could reduce dependency on extensive high-quality human annotations by enabling models to enhance their knowledge through internal feedback mechanisms, thereby accelerating the development of superhuman models.
• Theoretical Contribution: This approach may inspire new pedagogic methodologies in AI, fostering a more autonomous and self-evolving machine learning process and diversifying the current reliance on teacher-student model paradigms like knowledge distillation.

Future investigations might probe deeper into optimizing the LbT scoring mechanisms, dynamic student selection, and the automation of exam problem selection to reduce inference costs. Moreover, exploring the alignment of educational techniques in human pedagogy with LLM development, such as collaborative learning environments and diversity-based approach enhancements, paves the way for innovative training pipelines.

In conclusion, the paper suggests promising applications for LbT in advancing LLMs, potentially reshaping how machine learning models learn and evolve, akin to human learning techniques. This preliminary paper acts as a catalyst, inviting further exploration of LbT's full potential in refining LLMs' operational paradigm.