Skip to main content
AI News List

List of AI News about GRPO

Time Details
2026-09-17
00:21 UTC
RULER Reinvents RL rewards with GRPO

According to @_avichawla, RULER in OpenPipe ART uses LLM reviews to rank trajectories and convert comparisons into GRPO rewards for training agents.

Source
2026-07-31
00:43 UTC
LLM training stages Explained: 4-Step Guide

According to @_avichawla, LLMs evolve via pre-training, instruction, preference, and reasoning fine-tuning, each adding distinct behavior and accuracy.

Source
2026-06-06
02:44 UTC
GRPO Training Boosts RULER Results

According to @_avichawla, GRPO with RULER rankings in OpenPipe ART streamlines LLM fine-tuning and replaces brittle reward functions for RAG and support.

Source
2026-05-21
00:38 UTC
RULER Reinvents RL rewards with natural language

According to @_avichawla, RULER lets LLMs score trajectories from plain English criteria, easing brittle reward design for agents, as reported on X.

Source
2026-04-20
16:35 UTC
OpenMythos Breakthrough: Looped Transformer MoE Rebuild of Claude Mythos Shows 2.67x Faster Validation Steps

According to Kye Gomez (@KyeGomezB), OpenMythos is an open-source, first-principles reconstruction of Claude Mythos that implements a looped transformer with Mixture-of-Experts routing to enable iterative depth via weight sharing and conditional expert activation, targeting improved efficiency and multi-step reasoning (as reported on X/Twitter). According to Kye Gomez, a community training run indicated OpenMythos achieved its best validation in 2.67× fewer steps than nanoGPT, suggesting faster convergence in early experiments (as reported on X/Twitter). According to Kye Gomez, the team is pretraining 3B and exploring 5B parameter models on the FineWeb-Edu dataset on Hugging Face, followed by GRPO and high-quality RL fine-tuning, with all artifacts to be open-sourced and training scripts available on GitHub (as reported on X/Twitter). According to Kye Gomez, this is an early-stage research effort and a theoretical hypothesis of how Claude Mythos may function, inviting community contributions to evaluate looped transformer models and MoE routing impacts on reasoning (as reported on X/Twitter).

Source