面向小规模语言模型智能体的鲁棒强化学习

摘要

使用强化学习对参数量在70-500M范围内的小型语言模型（SLM）进行对齐通常被认为是不稳定的，但其潜在失效机制尚未得到系统性研究。在最新前沿研究中，采用近端策略优化（PPO）方法对十五种（模型，语料库）配置进行了训练。实验基于TinyStories、CNN/DailyMail和Wikitext-103语料库，涵盖了Pythia-70M/160M/410M及SmolLM2-135M/360M等模型。研究识别出小型语言模型中三种可复现的失效模式：标准PEFT/TRL流程中的静默LoRA参数冻结、使用bf16浮点格式时重要性比率数值溢出、以及奖励模型误差导致的灾难性策略崩溃。针对这些问题，分别采用了合并重初始化适配器技术、PPO更新时使用float32精度、以及包含奖励白化、重要性比率守护和权重回滚的三层级安全机制。本文提出容量余量假设，认为SLM规模的PPO性能同时依赖于流畅的有监督模型（困惑度<20）和具备区分性的奖励信号，而非模型参数量。所提系统在所有实验中均能稳定收敛，并在具备流畅先验和有效奖励信号的配置中较SFT基线提升了偏好胜率。此外，该系统以显著更少的训练数据超越了指令微调基线。所有检查点、偏好数据集及训练脚本均已公开发布。

English

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 (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^§.