SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusal (2024 · 87 citations)
SORRY-Bench is a benchmark that measures how reliably aligned LLMs decline unsafe requests. The authors identify three weaknesses in earlier safety evaluations. Their category schemes are coarse and inconsistent, and they concentrate on a few topics such as fraud while barely testing others like self-harm. They…
Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications (2024 · 66 citations)
Aligned LLMs can be jailbroken or lose their safeguards even after harmless fine-tuning. The authors ask whether this fragility can be traced to specific parts of the weights. Using the Llama2-chat family, they attribute behaviour to individual neurons and to low-rank components of weight matrices. They use two…
Scaling Latent Reasoning via Looped Language Models (2025 · 22 citations)
Most large language models learn to reason through explicit chain-of-thought added after pre-training. This lengthens outputs and makes little use of pre-training data. The authors propose Ouro, a family of Looped Language Models (LoopLM) in which a stack of weight-tied layers is applied several times, so hidden…

All works

4 works of theirs in our library, the most cited first; 3 with a page here, the others link their original until we summarise them.

Co-authors here

12 of 21 co-authors with a page, the most papers together first.

Papers of our library cite their works 168 times (of the 149,439 papers whose bibliographies we read); 40 citing papers with a page here →.

Updated 2 November 2026