Researchers are mapping the hidden circuits of transformers, identifying 'truth directions' in latent space, and predicting reasoning breakthroughs, pushing toward more interpretable and trustworthy AI.
A new evaluation framework pairs a unified taxonomy benchmark with self-adaptive multimodal fuzzing (SAMF) and shows that state-of-the-art MLLMs degrade under stress, revealing a gap between reasoning and factual grounding. RL alignment even worsens ...
A research team uses activation engineering to extract a falsehood direction in the latent space of transformers. Without fine‑tuning or external evidence retrieval, the last‑token projection feeds an MLP classifier. The method works across models fr...
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...
The community project Scotoma-2 tackles one of the most grating issues in modern LLMs: mechanical repetition of formulaic phrases. By combining ablation and DPO, it reduces unnatural patterns in Google’s model without harming reasoning – a case study...
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 ...
