Improved Supervised Fine-Tuning for Large Language Models to Mitigate Catastrophic Forgetting

Supervised Fine-Tuning (SFT), while enhancing large language models(LLMs)' instruction-following capabilities and domain-specific task adaptability, often diminishes their general capabilities. Moreover, due to the inaccessibility of original pre-training data, catastrophic forgetting tends to be exacerbated when third-party practitioners implement SFT on open-sourced models. To address this challenge, we propose a novel, more cost-effective SFT method which could effectively reduce the risk of catastrophic forgetting without access to original SFT data. Our approach begins by reconstructing the likely SFT instruction distribution of the base model, followed by a multi-model screening process to select optimal data, which is then mixed with new data for SFT. Experimental results demonstrate that our method preserves generalization capabilities in general domains while improving task-specific performance.

Multi-Layer GRPO: Enhancing Reasoning and Self-Correction in Large Language Models

The Group Relative Policy Optimization (GRPO) algorithm has demonstrated considerable success in enhancing the reasoning capabilities of large language models (LLMs), as evidenced by DeepSeek-R1. However, the absence of intermediate supervision in GRPO frequently leads to inefficient exploration dynamics. A single error in a complex reasoning chain can invalidate the entire solution, resulting in abrupt reward vanishing and compromising training stability.To address these challenges, we propose MGRPO (Multi-layer GRPO). MGRPO operates in two layers: the first layer employs standard GRPO to generate an initial response. This response, along with the original query, is then fed into a second-layer GRPO process. This second layer is specifically trained to identify and correct errors in the initial response, effectively creating a self-correction loop. This mechanism provides implicit process-level supervision by rewarding successful error correction, without requiring an explicit, densely-annotated reward model. Experimental results on several mathematical reasoning benchmarks demonstrate that MGRPO significantly outperforms standard GRPO, achieving superior performance by fostering both reasoning and self-correction abilities.