AI Briefing#rlhf
The latest AI and developer news about #rlhf, with the original source and a short summary.
The latest AI and developer news about #rlhf, with the original source and a short summary.
Evaluation of 17 LLMs revealed hallucination rates of 33–60%, with the absence of a verification layer identified as the primary cause.
An Anthropic engineer confirmed that prompt suggestions in Claude Code are designed for user flow, not to collect preference data for reinforcement learning.
Reef Infra has released the OpenClaw-RL recipe and session matching technology that converts user feedback into reinforcement learning signals.
TRL v1.14 introduces Async GRPO functionality that synchronizes LoRA adapters using storage buckets and proxies instead of NCCL.
Paul Christiano joins the OpenAI Foundation Board and the Safety and Security Committee to strengthen AI alignment and governance oversight.
An open-source reproduction project has been released for a watercolor-generating coding model trained with human aesthetic preferences as the reward function, utilizing TRL and OpenEnv.
Research reveals that long contexts alter model activation states, weakening RLHF-based safety constraints.
An analysis of 94 papers on AI value alignment reveals that the majority do not explicitly define 'value,' relying instead on preference and utility maximization.
This introduces the latest research trends and key papers on RLHF and RLAIF presented at NeurIPS 2023.
Introduces the latest research trends in RLHF and RLAIF for improving LLM performance, along with the DPO and Fine-Grained RLHF methodologies.
It introduces the latest research trends for overcoming the limitations of LLMs presented at ACL 2024 and LG AI Research's achievements.
A technology was unveiled that performs RL post-training using 14 Macs distributed across 4 countries worldwide.
ZPPO uses a Replay Buffer to let the model repeatedly learn from difficult questions, improving its ability to solve hard problems without degrading generalization performance.
Meta is forcibly restructuring its engineering organization into a data-labeling-focused ADO organization to secure AI competitiveness, triggering internal conflict.
Cursor unveiled Composer 2.5 and gave all users 10x daily usage.
Anthropic revealed that teaching the 'why' in Claude's alignment training generalizes better.
The author compares SFT, RL, and OPD from a distributional perspective and emphasizes the importance of policy-based sampling.
Hugging Face has released an interactive guide comparing the performance and scalability of various RL frameworks.
P-GRPO reduces GRPO's bias by decoupling individual preferences using per-group reward history.
An analysis of the design patterns and technical differences among 16 open-source libraries aimed at improving the efficiency of Asynchronous RL.
Agentic capabilities were enhanced without degrading Korean language performance by applying mid-training and a parallel RL pipeline.
Marking its 10th anniversary, OpenAI looked back on its past achievements and shared its vision for building superintelligence within the next 10 years.
GSPO reduces GRPO's instability through sequence-level optimization, enabling RL scaling.
Anthropic has raised $580 million in Series B funding to develop safe and reliable AI models.