AI research and discussion on LessWrong. www.lesswrong.com
Proposes detecting sandbagging on dangerous biosecurity capability evals by comparing accuracy on formal versus casual framings of the same questions; frontier models show over 50% drops under sandbagging prompts.
A capstone reproduction showing that adding weight noise to a prompt-induced sandbagging Qwen2.5-1.5B model can reveal its hidden GSM8k capability, with sandbagging strategies analysed.
Tests whether steering Llama 3.3 70B with SAE features tied to sandbagging or evaluation awareness can recover hidden capabilities; steering recovers up to about 25%, less than a simple prompt suffix.