Unveiling Over-Memorization in Finetuning LLMs for Reasoning Tasks

The pretrained large language models (LLMs) are finetuned with labeled data for better instruction following ability and alignment with human values. In this paper, we study the learning dynamics of LLM finetuning on reasoning tasks and reveal the uncovered over-memorization phenomenon during a specific stage of LLM finetuning. At this stage, the LLMs have excessively memorized training data and exhibit high test perplexity while maintaining good test accuracy. We investigate the conditions that lead to LLM over-memorization and find that training epochs and large learning rates contribute to this issue. Although models with over-memorization demonstrate comparable test accuracy to normal models, they suffer from reduced robustness, poor out-of-distribution generalization, and decreased generation diversity. Our experiments unveil the over-memorization to be broadly applicable across different tasks, models, and finetuning methods. Our research highlights that overparameterized, extensively finetuned LLMs exhibit unique learning dynamics distinct from traditional machine learning models. Based on our observations of over-memorization, we provide recommendations on checkpoint and learning rate selection during finetuning.

Automatic Robustness Stress Testing of LLMs as Mathematical Problem Solvers

Large language models (LLMs) have achieved distinguished performance on various reasoning-intensive tasks. However, LLMs might still face the challenges of robustness issues and fail unexpectedly in some simple reasoning tasks. Previous works evaluate the LLM robustness with hand-crafted templates or a limited set of perturbation rules, indicating potential data contamination in pre-training or fine-tuning datasets. In this work, inspired by stress testing in software engineering, we propose a novel framework, Automatic Robustness Checker (AR-Checker), to generate mathematical problem variants that maintain the semantic meanings of the original one but might fail the LLMs. The AR-Checker framework generates mathematical problem variants through multi-round parallel streams of LLM-based rewriting and verification. Our framework can generate benchmark variants dynamically for each LLM, thus minimizing the risk of data contamination. Experiments on GSM8K and MATH-500 demonstrate the strong performance of AR-Checker on mathematical tasks. We also evaluate AR-Checker on benchmarks beyond mathematics, including MMLU, MMLU-Pro, and CommonsenseQA, where it also achieves strong performance, further proving the effectiveness of AR-Checker.

Compound-QA: A Benchmark for Evaluating LLMs on Compound Questions

Large language models (LLMs) demonstrate remarkable performance across various tasks, prompting researchers to develop diverse evaluation benchmarks. However, existing benchmarks typically measure the ability of LLMs to respond to individual questions, neglecting the complex interactions in real-world applications. In this paper, we introduce Compound Question Synthesis (CQ-Syn) to create the Compound-QA benchmark, focusing on compound questions with multiple sub-questions. This benchmark is derived from existing QA datasets, annotated with proprietary LLMs and verified by humans for accuracy. It encompasses five categories: Factual-Statement, Cause-and-Effect, Hypothetical-Analysis, Comparison-and-Selection, and Evaluation-and-Suggestion. It evaluates the LLM capability in terms of three dimensions including understanding, reasoning, and knowledge. Our assessment of eight open-source LLMs using Compound-QA reveals distinct patterns in their responses to compound questions, which are significantly poorer than those to non-compound questions. Additionally, we investigate various methods to enhance LLMs performance on compound questions. The results indicate that these approaches significantly improve the models' comprehension and reasoning abilities on compound questions.