Top postsTop post
TL;DR Recent work from Goodfire & UK AISI – Verbalized Eval Awareness Inflates Measured Safety – shows that newer open-weight models verbalize evaluation-awareness (VEA) more often, and that this inflates measured safety. Between OLMo-3-32B-Think and OLMo-3.1-32B-Think – identical base, SFT, DPO, and RL data, differing only in an additional ~3...
Scaling laws for neural language models show that cross-entropy loss follows predictable power-law trends with dataset size, model size, and training compute. Since lower cross-entropy corresponds to better next-token prediction, and prediction is closely linked to compression, we should expect improvements in language-model loss to translate into better lossless compression....
The Platonic Representation Hypothesis (PRH) proposes that models trained with different objectives and across different modalities can converge toward a shared statistical representation of reality. However, the original experiments focused primarily on image models trained on the same underlying data distribution (ImageNet), leaving open the question of whether this convergence...
TL;DR: I investigate whether transformer reasoning hidden behind meaningless dot chains, inspired by “Let’s Think Dot by Dot”, can be decoded, using 3SUM as a testbed. Background Prior work such as “Let’s Think Dot by Dot” shows that models can sometimes preserve reasoning performance even when normal Chain-of-Thought tokens are...
Consider an adversarial training scheme for solving goal misgeneralization, similar in spirit to Redwood Research's work on “Adversarial Training for High-Stakes Reliability”. Suppose we have trained a model to perform some task. We would like to obtain worst-case guarantees that the model will not pursue a misgeneralized objective in adversarial...
Imagine that you are a trained mathematician and you have been assigned the job of testing an arbitrarily intelligent chatbot for its intelligence. You being knowledgeable about a fair amount of computer-science theory won’t test it with the likes of Turing-test or similar, since such a bot might not have...