When Models Know More Than They Say: Probing Analogical Reasoning in LLMs

1. 1Identify cases where LLMs struggle with analogical reasoning
2. 2Probe model representations to reveal latent knowledge
3. 3Compare probed representations with prompted performance
4. 4Develop techniques to improve model abstraction and generalisation

AI researchers and engineers working on LLMs can benefit from this research to improve model performance, while data scientists and ML engineers can apply the findings to develop more effective probing techniques

💡 Probing LLMs can reveal latent knowledge that is not apparent in their prompted performance

🤖 LLMs may know more than they say when it comes to analogical reasoning #AI #LLMs