LLMs Struggle with Paraphrasing: A Deep Dive into Autoformalization

Large Language Models (LLMs) are the new giants in AI, turning heads with their ability to handle autoformalization tasks. Yet, their performance isn't without flaws. Recent investigations showcase that these models struggle when faced with paraphrased natural language inputs. The question is: can we truly rely on them for consistent formal proofs?

The Issue with Paraphrasing

LLMs have seen significant use in areas like text-to-SQL, but this new study reveals that even minor changes in paraphrasing can lead to notable performance variability. This isn't just a technical hiccup, it's a fundamental issue. If LLMs are to be trusted in critical applications, they must handle paraphrased inputs with semantic fidelity.

The research evaluates LLMs' ability to generate formal proofs from paraphrased NL statements. The study used formal benchmarks like MiniF2F and the Lean 4 version of ProofNet, testing two modern LLMs. It found a clear sensitivity to paraphrasing, with performance swinging significantly based on how a statement was worded.

Why It Matters

This isn't just academic theory. In real-world applications, LLMs must interpret a wide variety of inputs. Their current sensitivity to paraphrasing suggests they're not yet ready for prime time in scenarios demanding high reliability.

The paper's key contribution: a demonstration of how subtle changes in language can disrupt a model's output. That's a big deal. If minor tweaks to input can lead to major shifts in output, the potential for error in applications where precision is important can't be ignored.

Looking Forward

So, where do we go from here? It's imperative to refine these models to ensure more consistent outputs across varied language inputs. The ablation study reveals the potential pathways for improvement, but it's clear there's much work to be done.

It's worth asking: should we be placing such heavy reliance on models that falter with basic language variation? Until these issues are addressed, skepticism around LLMs' reliability in critical autoformalization tasks may well be warranted.

, while LLMs have shown impressive capability, their struggles with paraphrasing highlight a critical area for improvement. The research provides a clear call to action: more strong models are needed to truly capitalize on the promise of AI in formalization tasks.