Reproduce “ALERT: Adapting Language Models to Reasoning Tasks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ping Yu, Tianlu Wang, Olga Golovneva, Badr AlKhamissi, Siddharth Verma, Zhijing Jin, Gargi Ghosh, Mona Diab and 1 more
Researchers at Meta AI introduce ALERT, a benchmark and analytical framework for evaluating fine-grained reasoning skills in large language models. Their analysis reveals that finetuning with Chain-of-Thought rationales can impart specific reasoning abilities like textual entailment and abductive reasoning, boosting performance by up to 15.22% ROUGE-L on a 13B model, though it simultaneously reduces the model's robustness to varying prompt templates.
Key insights
- Finetuning with Chain-of-Thought rationales can impart specific reasoning skills, such as textual entailment, abductive reasoning, and analogical reasoning, which are harder to acquire during pretraining.
- Performance gains from finetuning do not strongly correlate with data memorization, suggesting genuine skill acquisition rather than rote learning.
- Finetuning, especially without explicit rationales, can reduce a model's robustness to variations in prompt templates, leading to overfitting to specific input formats.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.