A new analysis warns that compressing large language models to 4-bit or 8-bit formats for edge deployment can erase safety guardrails, a phenomenon called 'alignment collapse,' with activation safety …
A study by an anonymous researcher, conducted as part of Neel Nanda's MATS 10.0 stream, found that linear 'trust' vectors extracted from the residual streams of Llama-3.2-3B-Instruct and Llama-3.1-8B-…
A new neural network compression method called PIN v2, developed by an unnamed research team, allows model size and speed to be chosen in advance by tying groups of cells to hold identical values, kee…
Redwood Research finds that hint-based chain-of-thought faithfulness evaluations still work on Claude models, contradicting Anthropic system card claims that recent models no longer use hints. The rep…
Ex-OpenAI AI researcher and co-author of the original RLHF paper argues that Chinese labs cannot effectively distill US frontier reasoning models because these models do not return full reasoning trac…
Researchers Chen et al. published a paper on arXiv 2025 introducing Janus-Pro, a unified multimodal model that uses separate encoders for understanding and generation within a single transformer. The …
A BlueDot Technical AI Safety Project researcher reproduced a study from Goodfire demonstrating that language model representations form curved geometric manifolds, not simple linear directions. The w…