This is a past event
Processing and generating language in ways that align with human expectations—both in communication and reasoning—requires computational representations that approximate how humans understand, abstract, and manipulate knowledge. Generative Large Language Models (LLMs), powered by vast data and extensive parameters, have achieved impressive alignment with human-like outputs. Though not explicitly trained for reasoning, they have become highly effective general-purpose reasoning engines. Yet, their capabilities are limited: they often fail to discern logical inference from content, rely on surface-level patterns instead of structured reasoning, and struggle with genuine abstraction, planning, and multi-step tasks. While scaling model size, training data and reasoning chains are the dominant strategies to improve LLMs reasoning, they have recently shown diminishing returns and escalating computational costs. We propose principled strategies that enhance reasoning capabilities, which are orthogonal to model scaling, through structured, abstraction-driven reasoning with sustainable resource demands.
The discussion will centre on two key questions:
1.Can we improve LLM reasoning by reducing reliance on linguistic patterns and scale? Human reasoning extends beyond language to include visual, spatial, and abstract thought, raising doubts about whether LLMs can fully emulate it.
2.To what extent can we discern genuine reasoning from pattern matching and memorised sequences?
This distinction is often made intuitively, particularly when contrasting human with LLM reasoning, but remains challenging to formalise and measure.
We will explore both conceptual and empirical strategies to enhance inferential capabilities in LLMs, focusing on symbolic and mathematical reasoning tasks across multiple languages, and discuss the implications for real-world applications where robust generalisation is required.
- Speaker
- Leonardo Ranaldi
- Venue
- Meston G05 and Microsoft Teams