Towards Robust Reinforcement Learning for Small-Scale Language Model Agents

2026-07-27

AI Summary

This work addresses the instability commonly observed in aligning small-scale language models (70M–500M parameters) via PPO-based reinforcement learning, which manifests as LoRA freezing, importance ratio overflow, and policy collapse. The study systematically identifies and reproduces three distinct failure modes for the first time, and proposes a stable training framework integrating adapter reinitialization, float32-precision updates, and a three-tier safety mechanism comprising reward whitening, ratio clipping, and weight rollback. It further introduces the “capability margin” hypothesis, arguing that a fluent policy prior combined with discriminative reward signals is more critical than model size. The proposed method consistently achieves stable convergence across all experimental settings, significantly outperforming supervised fine-tuning and instruction-tuning baselines when effective reward signals are available, while requiring substantially less data.

Abstract

The alignment of Small Language Models (SLMs) in the 70--500M parameter range using reinforcement learning is often considered unstable, though the underlying failure mechanisms have not been systematically investigated. In the State-of-the-Art (SOTA) research, fifteen (model, corpus) configurations were trained using Proximal Policy Optimization (PPO). The experiments included Pythia-70M, 160M, 410M and SmolLM2-135M, 360M on the TinyStories, CNN/DailyMail, and Wikitext-103 corpora. Three reproducible failure modes were identified in small-scale language models: silent LoRA parameter freezing in standard PEFT/TRL pipelines, numerical overflow in importance ratios when using bfloat16, and catastrophic policy collapse due to reward-model error. These issues were addressed using a merge-and-reinitialize adapter technique, float32 precision during PPO updates, and a three-layer safety mechanism comprising reward whitening, importance-ratio guarding, and weight rollback. In this paper, a capacity-headroom hypothesis is proposed, which states that PPO performance at the SLM scale depends on both a fluent supervised model ($\text{PPL}<20$) and a discriminative reward signal, rather than on the number of model parameters. The proposed system converged stably in all experiments and improved preference win rate over the SFT baseline in configurations with a fluent prior and an informative reward signal. Furthermore, it outperformed instruction-tuned baselines while requiring significantly less training data. All checkpoints, preference datasets, and training scripts are publicly released$^§$.

Problem

• Small Language Models
• Reinforcement Learning
• Alignment
• Failure Modes
• Stability

Innovation

• small language models
• reinforcement learning alignment
• failure mode analysis
• adapter reinitialization
• capacity-headroom hypothesis