Tag
#alignment
6 posts tagged alignment.
- bypass
G4-MeroMero-31B: Abliteration Drops Refusal Rate 99% to 15%
A new uncensored fine-tune of Gemma 4 31B achieves a 15/100 refusal rate via Arbitrary-Rank Ablation on attention output projections — KL divergence 0.
- alignment
LLM Alignment Evaluation: Why Benchmarks Don't Predict Safety
Practitioners rely on alignment benchmarks that miss the attack surface that matters: agentic tasks, implicit harm, and low-resource languages.
- deep-dive
KV Cache Compression Is Now an Alignment Problem
A new preprint argues that compressing KV cache during RL rollouts silently biases the policy you ship. For teams treating RLHF as a defensive control
- guardrails
ChatGPT Safety: How OpenAI's Guardrails Work and Fail
ChatGPT safety explained: how RLHF, Rule-Based Rewards, safe-completions, and the Moderation API work, plus the jailbreaks that defeat each layer.
- alignment
LLM Alignment: What It Does, Where It Breaks, How to Deploy
LLM alignment trains models to internalize safety constraints — but every technique has documented bypass paths.
- defense-in-depth
LLM Safety: What It Actually Means and How to Build It
LLM safety spans alignment training, inference-time guardrails, and external filters — each with known failure modes.