Releases·Tool Release
AgileRL Arena v1.0: manifest-driven RL training (local or cloud), with LoRA/GRPO LLM finetuning
AgileRL released agilerl-arena v1.0, a PyPI SDK/CLI split from the core AgileRL library that runs RL and LLM finetuning from YAML manifests locally or on a managed cluster. Practitioners can describe a full training run (algorithm, environment, network, mutation, replay buffer) in one validated file and launch it in two lines with LocalTrainer, no account needed.
2 engagement1 sourceFri, Sep 11, 2026, 03:31 PM
AI Overview
- AgileRL Arena v1.0, the SDK/CLI half of AgileRL, ships as a standalone PyPI package (no torch dependency) sharing the agilerl.* namespace
- YAML manifests define algorithm, environment, network, mutation, selection, and replay buffer, and run locally or submit to AgileRL's managed cluster
- Pydantic validation rejects unknown manifest keys and bad configs before training starts; arena manifest validate/schema check configs without submitting
- LLM finetuning accepts any HF model ID with LoRA config (e.g. Qwen/Qwen2.5-0.5B-Instruct, lora_r 16) and supports rollout-env RL, sft, or preference objectives
- Evolutionary HPO trains a population that periodically mutates architecture and hyperparameters, keeping the fittest agents
- v1.0 adds paradigm builders for algorithm construction, strategy-based local training dispatch, and a single-package manifest schema
- PEFT bumped to 0.20, which rejects LoRA on Mamba out_proj/conv1d layers; Arena now excludes them automatically
Related
Releases·Tool Release
Ant Group engineers release AReno, a single-node RL post-training toolkit
Engineers from Ant Group's ASystem Team released AReno, an open-source toolkit for RL post-training of LLMs on a single node. It supports RL, SFT/DPO-style training, serving, and agentic RL without external backends, aiming to make LLM RL accessible to researchers and developers.
Engineers from Ant Group's ASystem Team released AReno, an open-source toolkit for RL post-training of LLMs on a single node. It supports RL, SFT/DPO-style training, serving, and agentic RL without external backends, aiming to make LLM RL accessible to researchers and developers.
1 engagement1 sourceredditTue, Aug 18, 2026, 05:49 AM
Releases·Tool Release
NanoRL: async RL trainer for LLMs in ~1,800 lines, no Ray or TRL
A developer released NanoRL, a minimal async RL training library for LLMs in about 1,800 lines, supporting REINFORCE on CartPole and async GRPO on clusters. It avoids heavy dependencies like Ray, TRL, or DeepSpeed, using stdlib HTTP for worker-trainer communication.
A developer released NanoRL, a minimal async RL training library for LLMs in about 1,800 lines, supporting REINFORCE on CartPole and async GRPO on clusters. It avoids heavy dependencies like Ray, TRL, or DeepSpeed, using stdlib HTTP for worker-trainer communication.
8 engagement1 sourcehackernewsThu, Aug 13, 2026, 02:08 PM
Releases·Tool Release
Developer releases PyTorch-like framework for training model-agnostic harnesses
A developer released 'Harness Training', a PyTorch-like framework for training a frozen harness that can improve any task LLM on any environment. The harness is trained once with a frozen task LLM, then can be swapped out for any model. The project is open-source on GitHub.
A developer released 'Harness Training', a PyTorch-like framework for training a frozen harness that can improve any task LLM on any environment. The harness is trained once with a frozen task LLM, then can be swapped out for any model. The project is open-source on GitHub.
10 engagement1 sourceredditMon, Jul 20, 2026, 04:26 PM
Developments·Paper
Agentic ESOpt: Evolution Strategies for Long-Horizon LLM Agent Fine-Tuning
A new paper proposes Agentic ESOpt, using evolution strategies (ES) to fine-tune long-horizon LLM agents, claiming advantages over RL in scalability and credit assignment. The method aims to reduce GPU requirements for training large models.
A new paper proposes Agentic ESOpt, using evolution strategies (ES) to fine-tune long-horizon LLM agents, claiming advantages over RL in scalability and credit assignment. The method aims to reduce GPU requirements for training large models.
0 engagement1 sourcearxivTue, Aug 18, 2026, 03:03 AM
Products·building
ReasonForge (rl-gym-orchestrator) — reinforcement learning fine-tuning framework for LLMs
An open-source framework that trains language models via reinforcement learning in a gym-style environment, described as a 'cognitive gym' for fine-tuning reasoning. It targets developers and researchers who want to RL fine-tune LLMs with a structured trainer/orchestrator. Distributed as a GitHub repo with a download page, MIT licensed.
An open-source framework that trains language models via reinforcement learning in a gym-style environment, described as a 'cognitive gym' for fine-tuning reasoning. It targets developers and researchers who want to RL fine-tune LLMs with a structured trainer/orchestrator. Distributed as a GitHub repo with a download page, MIT licensed.
55 engagement1 sourcegithubTue, Aug 25, 2026, 09:56 AM