Have LLMs Advanced Enough? A Challenging Problem Solving Benchmark For Large Language Models

24 May 2023  ·  Daman Arora, Himanshu Gaurav Singh, Mausam ·

The performance of large language models (LLMs) on existing reasoning benchmarks has significantly improved over the past years. In response, we present JEEBench, a considerably more challenging benchmark dataset for evaluating the problem solving abilities of LLMs. We curate 515 challenging pre-engineering mathematics, physics and chemistry problems from the highly competitive IIT JEE-Advanced exam. Long-horizon reasoning on top of deep in-domain knowledge is essential for solving problems in this benchmark. Our evaluation on various open-source and proprietary models reveals that the highest performance, even after using techniques like self-consistency, self-refinement and chain-of-thought prompting, is less than 40%. The typical failure modes of GPT-4, the best model, are errors in algebraic manipulation, difficulty in grounding abstract concepts into mathematical equations accurately and failure in retrieving relevant domain-specific concepts. We also observe that by mere prompting, GPT-4 is unable to assess risk introduced by negative marking for incorrect answers. For this, we develop a post-hoc confidence-thresholding method over self-consistency, which enables effective response selection. We hope that our challenging benchmark will guide future re-search in problem-solving using LLMs.

PDF Abstract

Datasets


Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Overall - Test JEEBench GPT-4+CoT+Self-Consistency@8 Score 0.389 # 1
Overall - Test JEEBench GPT-4+CoT+Self-Critique Score 0.339 # 3
Overall - Test JEEBench GPT-4+(1-shot CoT) Score 0.292 # 5
Overall - Test JEEBench GPT-4+CoT Score 0.350 # 2
Overall - Test JEEBench GPT-4 Score 0.309 # 4
Overall - Test JEEBench GPT-3.5 Score 0.177 # 7
Overall - Test JEEBench PaLM2 Score 0.153 # 8
Overall - Test JEEBench GPT-3 Score 0.122 # 9
Overall - Test JEEBench Falcon7B-Instruct Score 0.098 # 11
Overall - Test JEEBench Alpaca-LoRA Score 0.089 # 12
Overall - Test JEEBench Random Score 0.105 # 10

Methods