Recent advances in large language models (LLMs) have yielded impressive gains on mathematical reasoning benchmarks via supervised fine-tuning (SFT). However, the brittleness of these models under input perturbations has cast doubt on whether such improvements reflect genuine reasoning abilities or merely superficial alignment with expected output formats. We investigate the mechanisms behind SFT improvements in small-scale LLMs, addressing four key questions: (1) Are performance gains primarily due to format alignment rather than reasoning? (2) Can high-quality supervision encourage genuine reasoning? (3) Does scaling data shift learning from format alignment to deeper reasoning? (4) Are format alignment gains consistent across model sizes and architectures? Through controlled experiments, we find that most performance improvements arise from format alignment rather than genuine reasoning enhancement. Moreover, SFT’s effectiveness is strongly influenced by the alignment between the base model’s inductive biases and the teacher model’s output distribution, rather than the teacher’s raw strength. Finally, scaling up training data offers diminishing returns and does not fundamentally alter the model’s reasoning behavior. These findings suggest that current SFT practices may overestimate the reasoning abilities of LLMs and underscore the need for more rigorous evaluation methods.

Headline generation, a key task in abstractive summarization, strives to condense a full-length article into a succinct, single line of text. Notably, while contemporary encoder-decoder models excel based on the ROUGE metric, they often falter when it comes to the precise generation of numerals in headlines. We identify the lack of datasets providing fine-grained annotations for accurate numeral generation as a major roadblock. To address this, we introduce a new dataset, the NumHG, and provide over 27,000 annotated numeral-rich news articles for detailed investigation. Further, we evaluate five well-performing models from previous headline-generation tasks using human evaluation in terms of numerical accuracy, reasonableness, and readability. Our study reveals a need for improvement in numerical accuracy, demonstrating the potential of the NumHG dataset to drive progress in number-focused headline generation and stimulate further discussions in numeral-focused text generation.

Numbers are frequently utilized in both our daily narratives and professional documents, such as clinical notes, scientific papers, financial documents, and legal court orders. The ability to understand and generate numbers is thus one of the essential aspects of evaluating large language models. In this vein, we propose a collection of datasets in SemEval-2024 Task 7 - NumEval. This collection encompasses several tasks focused on numeral-aware instances, including number prediction, natural language inference, question answering, reading comprehension, reasoning, and headline generation. This paper offers an overview of the dataset and presents the results of all subtasks in NumEval. Additionally, we contribute by summarizing participants’ methods and conducting an error analysis. To the best of our knowledge, NumEval represents one of the early tasks that perform peer evaluation in SemEval’s history. We will further share observations from this aspect and provide suggestions for future SemEval tasks.