- formatting
- images
- links
- math
- code
- blockquotes
- external-services
•
•
•
•
•
•
-
Automating Auditing Games of Sandbagging
Can LLM agents run a full red–blue auditing game for evaluation sandbagging? Agents train sandbagging model organisms (Qwen3-8B) and then audit them with black-box and white-box techniques.
-
Negated Reward Hacking
A research note on negated reward hacking — using negated self-documentation to steer models away from reward-hacking behavior during RL.
-
Selective Inoculation
Investigating how selective inoculation prompts can suppress unwanted behavioral traits in fine-tuned LLMs while preserving intended ones.
-
a post with plotly.js
this is what included plotly.js code could look like
-
a post with image galleries
this is what included image galleries could look like
