Beyond Examples: High-level Automated Reasoning Paradigm in In-Context Learning via MCTS
Jinyang23
Abstract
In-context Learning (ICL) enables large language models (LLMs) to tackle downstream tasks through sophisticated prompting and high-quality demonstrations. However, this traditional ICL paradigm shows limitations when facing complex mathematical reasoning tasks, primarily due to its heavy dependence on example quality and the necessity for human intervention in challenging scenarios. To address these limitations, this paper presents HiAR-ICL, a High-level Automated Reasoning paradigm in ICL that shifts focus from specific examples to abstract thinking patterns, extending the conventional concept of context in ICL. HiAR-ICL introduces five atomic reasoning actions as fundamental components for constructing chain-structured patterns. Using Monte Carlo Tree Search, we explore reasoning paths and construct thought cards to guide subsequent inference. We then develop a cognitive complexity framework that dynamically matches problems with appropriate thought cards. Experimental results demonstrate HiAR-ICL's effectiveness, achieving state-of-the-art accuracy (79.6%) on the MATH benchmark with Qwen2.5-7B-Instruct, surpassing GPT-4o (76.6%) and Claude 3.5 (71.1%).
Community
🚀 We are pleased to share our latest research paper, "Beyond Examples: High-level Automated Reasoning Paradigm in In-Context Learning via MCTS". This work introduces HiAR-ICL, a novel paradigm to enhance the complex reasoning capabilities of large language models.
🌟 Unlike traditional in-context learning, HiAR-ICL shifts the focus from example-based analogical learning to abstract thinking patterns. It employs Monte Carlo Tree Search to explore reasoning paths and creates "thought cards" to guide inferences. By dynamically matching test problems with appropriate thought cards through a proposed cognitive complexity framework, HiAR-ICL achieves state-of-the-art accuracy of 79.6% with 7B model on the challenging MATH benchmark, surpassing both GPT-4o and Claude 3.5.
📑 Paper: https://arxiv.org/pdf/2411.18478
🌐 Project Page: https://jinyangwu.github.io/hiar-icl/
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
- Teaching-Inspired Integrated Prompting Framework: A Novel Approach for Enhancing Reasoning in Large Language Models (2024)
- Reversal of Thought: Enhancing Large Language Models with Preference-Guided Reverse Reasoning Warm-up (2024)
- Markov Chain of Thought for Efficient Mathematical Reasoning (2024)
- AtomThink: A Slow Thinking Framework for Multimodal Mathematical Reasoning (2024)
- Reasoning Paths Optimization: Learning to Reason and Explore From Diverse Paths (2024)
- Improving LLM Reasoning through Scaling Inference Computation with Collaborative Verification (2024)
- Auto-Evolve: Enhancing Large Language Model's Performance via Self-Reasoning Framework (2024)
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment:
@librarian-bot
recommend
I think there is a typo in equation (2), the Q value should not be divided by N(s) (you apply this when the numerator is the sum of rewards, e.g. the number of wins from a given node, but not if you have a value function).
Thanks for your attention and pointing out this typo. In fact, in our specific implementation, the estimated reward value for each node is dynamically updated through back-propagation, which essentially involves dividing the number of wins from a given node by N(s). We'll clarify it in the next version for improved presentation.
Impressive work! Just curious and for the record, any plan to opensource by any chance ? Thanks !
Thank you for your interest in our work! We do plan to fully open source this project. Currently, we are working on supplementing and refining the experimental details. Once that is complete, we will organize the code and release it publicly.
Wow. That will be really really helpful for the community! Since I saw in the comments above that you guys are preparing the next arxiv version, if you don't mind, can I suggest to add a few concrete examples for the thought cards in the appendix ? If I did not misinterepret your paper, the boost of performance lies on the smart design of using 'thought cards' to elicit/instruct the LLMs to think in a specific pattern, I am really curious about how the prompts are designed and presented. I believe that will also serve to clarify your papers a lot to some readers : )
Impressive experimental result! I still have two questions after reading this paper.
- How can I determine which atomic action a reasoning step in the model's output belongs to while doing MCTS?
- And how to prompt the model to reason guided by action-chain from selected thought card?
Insightful questions!
- For the first question, we randomly select from the entire action space. This ensures, theoretically, a deep search space, which is beneficial for constructing more comprehensive thought cards.
- For the second, we provide a description and functionality for each action, allowing the model to perform reasoning based on this information.
I remember that the MCTS realized in rStar has the restrict like One trajectory has only one action2(CoT) and action4(Re-answer subquestion)... SO you don't have restricts like above when exploring the action space?
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper