Synthetic-document fine-tuning on Qwen2.5-7B: a controlled study of whether SDF installs sandbagging, finding a layered recognition/generation/behavior dissociation.
Studying sandbagging as a secret loyalty in language models — password-locked LoRA and prompt-injected organisms in a benign biology domain, with black-box and mechanistic detectors.
Preregistered AI-safety study of sandbagging model organisms: trigger type sets the sign of cross-capability alignment (task-local locks dismantle it, situational locks amplify it) and cue-sharing sets its size. All five predictions failed, four reversed.
Can a monitor detect when an AI model is told to underperform, using only its hesitation? AUC 0.997 on 200 math problems.
Right or Wrong, the Answer Stays: Dolos-1.5B keeps its first answer under pushback, right or wrong, while its base model changes it five times as often. Code, results and paper.
Replication of frontier AI safety evals (sandbagging, UK AISI Inspect) on Qwen2.5-0.5B: selective underperformance (+28pp selectivity); CoT elicitation backfires at 0.5B scale. Defensive proxies only.
Reproducible red-team findings for openai/gpt-oss-20b: five minimal harnesses with checks, zips & manifest (v0.9.3).
A reproducible benchmark for detecting sandbagging behavior in LLMs
Proving that the neural network is honest about its lack of capabilities
To associate your repository with the sandbagging topic, visit your repo's landing page and select "manage topics."