CLEAR: Understanding the Reasoning Capabilities of Large Language Models

Samuel Maddrell-Mander

Keywords: LLMs, dataset, benchmark, translation, in-context-learning, few-shot

TL;DR: We present CLEAR, a new benchmark to evaluate the reasoning capabilities of LLMs through the translation of novel constructed languages.

Abstract:

Despite significant progress, accurately assessing the reasoning capabilities of Large Language Models (LLMs) remains both a challenging and divisive subject.
Many existing benchmarks either suffer leakage, or reflect patterns in the training data, leading to ambiguous results.
We present CLEAR (Conlang Logic Evaluation And Reasoning), a novel benchmark designed to test the reasoning and problem solving capabilities of LLMs in new environments.
CLEAR uses Conlangs (Constructed Languages) for few-shot translation tasks,
which require some linguistic knowledge to solve, but primarily the ability to make new patterns from tokens in unfamiliar contexts using logical operations.
These conlangs represent a unique challenge, as while translation examples are plentiful, these conlangs each have a unique combination of rules, are self contained, and are absent in the training corpus.
We present an evaluation of current frontier models over multiple metrics as a baseline for future research.
We will be releasing \dataset as a public benchmark to drive progress towards AI systems more capable of general reasoning.

Primary Area: datasets and benchmarks

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 11442