Studies reveal LLMs' poor self‑assessment of difficulty, predictable chain‑of‑thought patterns, and novel metrics like the Ignition Index, while automated circuit annotation reduces manual labor.
A new framework uses mean-field approximation to statistically describe chain-of-thought reasoning. 'Clue' tokens are identified via surprisal, and the emergent regularities are reproducible and modelable with a differential equation. Implications fo...
A research team has developed a scalar metric that captures the moment a language model shifts from gradual processing to a switch-like understanding. The concrete finding: 9.6 times more selectivity for genuine linguistic structure over spurious pat...
A new benchmark tests LLMs' ability to infer search spaces and algorithms from natural-language descriptions of black-box optimization tasks. Current models can select algorithms based on the evaluation budget but struggle with search space design wh...
Researchers have developed a pipeline that uses an LLM to automatically group features and neurons into supernodes, producing attribution graph annotations as interpretable as those made by humans. Tested on a two-hop reasoning task, it recovers the ...
A study shows even the most advanced language models struggle to estimate item difficulty in educational tests. GPT-4.1 is outperformed by a traditional encoder, while GPT-5.4 tends to see everything as easy. A wake-up call for anyone using LLMs to a...
