A retention-data-free LLM unlearning method that uses a self-calibrated term to forget targeted knowledge more precisely while preserving general capabilities.
As Large Language Models (LLMs) are increasingly deployed as autonomous agents in high-stakes domains, understanding contextual factors that may modulate their decision-making becomes critical. While LLMs are trained to perceive and resonate with …
A machine-executable format that replaces traditional papers with structured, agent-consumable research artifacts capturing logic, code, exploration, and evidence.
A momentum-based, sampling-driven method for scaling textual gradient optimization in LLM prompt engineering, improving performance and efficiency across diverse NLP tasks.
An independent reproduction of Anthropic's emotion vector research using the open-weight Llama 3.1 8B model, with the paper's verbatim 171 emotions, 100 topics, and 64 activities. We confirm 10 of 11 verification criteria, with causal steering r=0.956 closely matching the paper's r=0.85.
A training-free method that robustifies LLM safety alignment against fine-tuning by extrapolating low-rank safety subspaces, significantly reducing attack success rates while preserving model utility.