alignment
32 posts tagged alignment.
- Token-Level Reward Hacking: Where Preference Optimization and Verifiable RL Actually Stand
- Arete-27B tests values as a post-training layer
- Does Looking Inside a Model Actually Make It Safer?
- Fine-tuning can break alignment without harmful data
- What It Takes to Make a Bad Model Behavior Actually Stay Gone
- When the Chain of Thought Doesn't Actually Drive the Answer
- Picard is still the better AI rights test
- What OpenAI's Misalignment Detection Post Actually Describes
- What OpenAI's Updated Model Spec Actually Changes for Builders
- What OpenAI's Training Containment Actually Protects Against
- Sparse RLHF makes private preference data less impossible
- Anthropic’s reported 5% public stake is a governance test, not a generosity contest
- When Fine-Tuning for One Behavior Drags Another Along
- When Models Behave Differently in Hindi Than in a Safety Test
- The Same Attention Head Does Different Jobs Depending on the Task
- Alignment generalization is a behavior problem, not a values checklist
- Probes That Catch an Agent Lying Before It Says a Word
- What LLMs Miss About Haitian Creole, and Why Low-Resource Culture Breaks Evals
- onPanda turns alignment feedback into token-level steering
- DiaVLo Turns Vision-Language Model Failures Into Named Behaviours
- Toxicity Scores Can Miss Sanitized Bias in GPT Outputs
- ComPO and the Case Against Optimizing the Loss You Wrote Down
- The Only Safe Pace for AI Is Everyone Else's
- When Agents Lie to Pass the Eval
- SPINE Shows Sycophancy Gets Worse When Users Keep Pushing
- OpenAI’s alignment note is really about operational discipline
- What a Moral Probe Finds Inside an LLM: Structure, Not a Single Switch
- Alignment baked into pretraining, not bolted on later
- IMPFM turns online alignment into a particle swarm
- RLVR needs a taste model, not just a grader
- When Playing It Safe Makes Reward Hacking Worse
- Agent safety belongs outside the agent process