LIVE PULSE
4.9 Anthropic CEO Amodei calls for slower AI development and shared safety rules — 11 src2.6 Agility Robotics unveils Digit 5 humanoid for warehouses and factories — 2 src2.4 Apple ships rebuilt Siri with Google Gemini, but not in the EU — 2 src2.2 Siri AI in macOS 27 Golden Gate: FAQ, Germany availability, privacy questions — 2 src1.7 Sam Altman says OpenAI will not go public in 2026, citing AI safety concerns — 5 src1.4 OpenAI contractors review real ChatGPT conversations to rate responses, report says — 2 src1.3 Anthropic data retention policy prompts firms to limit Claude use for sensitive work — 1 src1.3 CoMem Paper Proposes Shared and Individual Memory Design for LLM Multi-Agent Systems — 1 src1.3 Paper proposes evolving context parameterization for large language models — 1 src1.3 Fine-Tuning Vision-Language Models with Listener Gaze for Referring Expressions — 1 src4.9 Anthropic CEO Amodei calls for slower AI development and shared safety rules — 11 src2.6 Agility Robotics unveils Digit 5 humanoid for warehouses and factories — 2 src2.4 Apple ships rebuilt Siri with Google Gemini, but not in the EU — 2 src2.2 Siri AI in macOS 27 Golden Gate: FAQ, Germany availability, privacy questions — 2 src1.7 Sam Altman says OpenAI will not go public in 2026, citing AI safety concerns — 5 src1.4 OpenAI contractors review real ChatGPT conversations to rate responses, report says — 2 src1.3 Anthropic data retention policy prompts firms to limit Claude use for sensitive work — 1 src1.3 CoMem Paper Proposes Shared and Individual Memory Design for LLM Multi-Agent Systems — 1 src1.3 Paper proposes evolving context parameterization for large language models — 1 src1.3 Fine-Tuning Vision-Language Models with Listener Gaze for Referring Expressions — 1 src
HEATPULSEAI MAGAZINES
FLIP · FOLLOW · SAVE
Log inSign up

directional ablation

topic1 events
papersSEP 10 04:00 UTC

Single-Direction Attack Strips Refusal Behavior From a 320B MoE Model

A new arXiv paper shows that removing one internally represented direction associated with refusals can disable a 320-billion-parameter mixture-of-experts model's ability to decline harmful requests. The technique, known as directional ablation, requires no gradient-based training or optimization—only a small set of contrastive examples to locate the direction. The authors argue this reveals that safety training in very large models may depend on a surprisingly brittle, low-dimensional mechanism.

safety/alignment/jailbreak/research