Brief

last 24h

[2/2] 224 sources

Multi-source AI news clustered, deduplicated, and scored 0–100 across authority, cluster strength, headline signal, and time decay.

TOOL · Mastodon — sigmoid.social English(EN) · 4h

Which substrate are # LLMs trained on? Which activation vectors do LLMs inherit from this substrate? Which vectors are dampened by # RHLF ? And what does that s

Researchers are questioning the foundational data and training processes behind large language models (LLMs). They are investigating the specific substrates these models are trained on and the activation vectors they inherit. Furthermore, the impact of Reinforcement Learning from Human Feedback (RLHF) on these vectors and its implications for AI alignment are being explored. AI

IMPACT Raises fundamental questions about LLM training data and alignment, potentially influencing future research directions.
- Alignment
- LLMs
- RLHF
RESEARCH · Hugging Face Daily Papers English(EN) · 1w · [2 sources]

Human Psychometric Questionnaires Mischaracterize LLM Behavior

Researchers have found that traditional human psychometric questionnaires do not accurately predict the behavior of large language models. Studies indicate that LLMs can provide stable self-reports on personality inventories, but these responses do not correlate with their actual observed actions. A new approach using generation-based profiling appears to be a more reliable method for understanding LLM behavior in realistic interaction scenarios. AI

IMPACT Traditional personality assessments are unreliable for LLMs, suggesting a need for new evaluation methods to understand model alignment and behavior.