I created an LLM post-training method called RPS. Preliminary results show that it improved Qwen3-8b's program synthesis reliability. [R]

2026-05-21 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, quick

Summary

A new LLM post-training method, Regressive Plasticity Schedule (RPS), has been developed, drawing inspiration from neuroscience's two-stage learning model. RPS trains large language models like Qwen3-8b in two distinct stages: an initial phase on "easy" data with a high learning rate, followed by a second phase on "hard" data using 10% of the initial learning rate. This approach combines curriculum learning with learning rate decay. Preliminary results indicate RPS improved Qwen3-8b's program synthesis reliability, achieving 1145/1200 program executions without error, compared to 870/1200 for an Equal Plasticity Schedule (EPS) baseline. On ARC-AGI 1 public eval scores, RPS scored 4%, outperforming EPS's 2.4%. The method aims to enhance model performance by mimicking human neuro-plasticity stages.

Key takeaway

For Machine Learning Engineers optimizing LLM performance on code generation, consider integrating a two-stage training approach like RPS. Your current fine-tuning strategies might benefit from separating "easy" and "hard" data with distinct learning rates, potentially improving program synthesis reliability. Evaluate how you define and split your training data to operationalize "easy" versus "hard" examples, and benchmark against single-stage baselines to quantify the gains.

Key insights

The Regressive Plasticity Schedule (RPS) improves LLM program synthesis by mimicking human learning stages with a two-stage curriculum and learning rate decay.

Principles

• Two-stage training can enhance LLM reliability.
• Curriculum learning benefits from varied learning rates.
• Neuroscience models inspire LLM training schedules.

Method

RPS trains models in two stages: easy data with high learning rate, then hard data with 10% of the initial learning rate, combining curriculum learning and learning rate decay.

In practice

• Apply two-stage training to program synthesis tasks.
• Experiment with learning rate decay schedules.
• Evaluate "easy" vs. "hard" data splits.

Topics

• LLM Training
• Program Synthesis
• Curriculum Learning
• Learning Rate Decay
• Qwen3-8b
• Post-training Methods

Code references

• iamjasonfeng/RPS

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Reformulate LLM Reinforcement Learning for Efficient Training under Black-box Discrepancy — LLM RL training stability improves by explicitly constraining train-inference discrepancy within a tolerance region.
• ❓ Introducing How2Everything—a framework for improving how LLMs generate step-by-step procedures — How2Everything improves LLM procedural generation by +10 points without capability regression.
• A Proactive Reliability Metric for Detecting Failures in Language Model Training — The R-Metric proactively predicts LLM training failures by combining hardware, training dynamics, and model performance signals.
• From Heuristic Selection to Automated Algorithm Design: LLMs Benefit from Strong Priors — Integrating benchmark code examples into prompts significantly enhances LLM-driven algorithm optimization performance and robustness.
• New method could increase LLM training efficiency — TLT doubles LLM training speed by adaptively using idle compute to train a smaller, verifying drafter model.
• Unlocking LLM Code Correction with Iterative Feedback Loops — Iterative feedback loops significantly enhance LLM code correction, especially for reasoning models and specific error types.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.