If SOTA models fail to recognize simple shapes, should we be evaluating them on complex geometric tasks? Most MLLMs struggle with counting the number of sides of regular polygons and all MLLMs receive 0% on novel shapes. William Rudman Amir Bar Vedant Palit [1/6]

New KDD 2025 paper: Can large language models (LLMs) reason like biomedical scientists? We introduce K-Paths, a retrieval framework for extracting reasoning paths from knowledge graphs (KGs) to aid drug discovery tasks. 👇 Thread:

Pruning is essential for deploying a #LLM efficiently. However, beyond 70% sparsity, performance drops sharply: perplexity increases exponentially and accuracy deteriorates. We introduce TRIM, a method that maintains model quality even at extreme sparsity levels #NLProc #AI [1/5]

⚠️Circuit-finding too slow? We introduce Accelerated Path Patching (APP), a method that uses task-specific pruning to cut down the search space before patching. 🚀APP finds faithful circuits with up to ~93% faster runtime.

Do you hate running Path Patching on larger models? Accelerated Path Patching can save days of compute time. APP produces faithful circuits by using contrastive pruning to reduce the search space of circuit discovery methods. Go save some trees 🌳by using APP to speed up PP.