-д хадгалсан:
| Үндсэн зохиолчид: | , |
|---|---|
| Формат: | Recurso digital |
| Хэл сонгох: | |
| Хэвлэсэн: |
Zenodo
2026
|
| Онлайн хандалт: | https://doi.org/10.5281/zenodo.20123803 |
| Шошгууд: |
Шошго нэмэх
Шошго байхгүй, Энэхүү баримтыг шошголох эхний хүн болох!
|
Агуулга:
- <p>Abstract<br>The evaluation of Tree-of-Thought (ToT) reasoning and self-consistency is important when<br>deploying lightweight language models (LLMs) that can produce capable results while<br>operating on constrained hardware. Self-consistency decoders can enhance the robustness of<br>chain-of-thought style inference for complex reasoning tasks through the use of selfconsistency, by allowing multiple pathways of diverse reasoning to be aggregated through<br>sample-and-vote based methods. By enabling look-ahead, backtracking, and intentional<br>processing of alternate solution paths, the concept of ToT generalizes these ideas to create a<br>search tree that represents all the "thoughts" produced at the intermediate level of processing.<br>This version provides a thorough description of both methods of deriving results from using<br>lightweight language models through self-consistency (SC) and trees of thoughts (ToT)<br>techniques according to an established benchmark suite consisting of three different types of<br>arithmetic problems: word, logic, and combinatorial, and implementing various instances of<br>SC and ToT methodologies (with varying numbers of parameters) on small models that can be<br>checked for performance (accuracy, sample efficiency and latency) against alternative<br>computational environments and the respective experiments were constructed using a<br>combination of algorithmic and statistically-based methods to achieve “faithfulness” to<br>machines' reasoning by employing an agreement between the methods' individual sample<br>trajectories and ground-truth solutions. The result whereby self-tested consistency alone (trial<br>and error) was greater in performance as compared to ToT were due only to circumstances with<br>high “overhead” costs associated to searching tree-based data with additional overheads for<br>ToT data structures permitting “non-trivial” savings achieved through incorrect reasoning, i.e.,<br>by escaping “local consistency” through incorrect use of its rules. Finally, we propose a series<br>of guidelines to assist in the selection and tuning of SC and ToT for use in lightweight LLMs<br>for applications requiring mathematical or logical reasoning, and a framework will be created<br>to serve as an efficiency-based adaptive-toa-of-self-consistency model for future research<br>endeavors.<br>Keywords: Deep Learning, Large Language Model, Tree of Thought Reasoning, Neural<br>Networks, Artificial Intelligence.</p>