Reactive Optimizers

Reactive Optimizers places an AI agent inside the training loop to rewrite optimizer code as neural networks train.

Reactive Optimizers places an AI agent at the heart of the optimizer to watch training dynamics, react to them, and rewrite the optimizer's source code as training goes. The system reasons on the spot and writes the weight update rule for each parameter group, implementing complex schedules and responding to training accidents. It uses an agent powered by Claude Opus 5.5 and the OPAL package to spend LLM tokens in place of GPU or human hours. The tool handles image generation and language modeling tasks, matching or beating the best off-the-shelf recipes on alternative datasets and new architectures. It operates in a zero-shot manner without retaining data from repeated runs.

Reactive Optimizers at a glance

Best for: Software engineer, Researcher
Runs on: Web
Pricing: Free
Released: October 2026
Category: AI coding tool
Website: xingjianbai.com

Key features

Watches training dynamics
Rewrites optimizer source code on the fly
Writes weight update rules for parameter groups
Implements scheduling and strategies to prevent loss spikes
Extracts training statistics via agent-written probes
Spends LLM tokens to replace manual recipe tuning

Use it for

Training language models on alternative language datasets
Training new deep learning architectures without pre-tuned recipes
Optimizing image generation models like diffusion transformers
Escaping overfitting and suppressing loss spikes

Alternatives to Reactive Optimizers

rLLM — Enables reinforcement learning training for AI agents with minimal code changes.
OpenEvolve — Automates algorithmic discovery and optimization without human guidance.
XTuner — High-efficiency training for 200B+ scale MoE models without expert parallelism.
Tura — Reduces token usage and LLM turns by over 80 percent while improving performance.
Agent Lightning — Train existing artificial intelligence agents with reinforcement learning and zero code changes.
ART — Enhances LLM agent performance and reliability using reinforcement learning techniques.
NeMo Molt — It scales fully-async agentic RL to frontier MoE scale with a small PyTorch stack.
OptiLLM — Increases LLM reasoning accuracy by up to 10x without model training.
Strata — Runs large AI models locally on consumer gaming PCs without server reliance.
AirLLM — Runs massive 70B+ parameter models on small 4GB consumer GPUs.
Optimum — It maximizes efficiency to train and run models on targeted hardware.
Optscale Kompressor — It reduces real provider bills by removing redundant data while preserving the prompt cache.

All Reactive Optimizers alternatives