Hangzhan Jin

Hangzhan Jin, Tianwei Ni, Lu Li et al.

Supervised fine-tuning (SFT) improves in-domain performance but can degrade out-of-domain (OOD) generalization. Prior work suggests that this degradation is related to changes in dominant singular subspaces of pretrained weight matrices. However, directly identifying loss-sensitive directions with Hessian or Fisher information is computationally expensive at LLM scale. In this work, we propose preserving projected rotations in pretrained singular subspaces as an efficient proxy for Fisher-sensitive directions, which we call Rotation-Preserving Supervised Fine-Tuning (RPSFT). RPSFT penalizes changes in the projected top-$k$ singular-vector block of each pretrained weight matrix, limiting unnecessary rotation while preserving task adaptation. Across model families and sizes trained on math reasoning data, RPSFT improves the in-domain/OOD trade-off over standard SFT and strong SFT baselines, better preserves pretrained representations, and provides stronger initializations for downstream RL fine-tuning. Code is available at \href{https://github.com/jinhangzhan/RPSFT.git}{https://github.com/jinhangzhan/RPSFT}.

Hangzhan Jin, Sitao Luan, Sicheng Lyu et al.

The two-stage fine-tuning paradigm of Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL) has empirically shown better reasoning performance than one-stage SFT for the post-training of Large Language Models (LLMs). However, the evolution and mechanism behind the synergy of SFT and RL are still under-explored and inconclusive. In our study, we find the well-known claim "SFT memorizes, RL generalizes" is over-simplified, and discover that: (1) OOD performance peaks at the early stage of SFT and then declines (OOD forgetting), the best SFT checkpoint cannot be captured by training/test loss; (2) the subsequent RL stage does not generate fundamentally better OOD capability, instead it plays an \textbf{OOD restoration} role, recovering the lost reasoning ability during SFT; (3) The recovery ability has boundaries, \ie{} \textbf{if SFT trains for too short or too long, RL cannot recover the lost OOD ability;} (4) To uncover the underlying mechanisms behind the forgetting and restoration process, we employ SVD analysis on parameter matrices, manually edit them, and observe their impacts on model performance. Unlike the common belief that the shift of model capacity mainly results from the changes of singular values, we find that they are actually quite stable throughout fine-tuning. Instead, the OOD behavior strongly correlates with the \textbf{rotation of singular vectors}. Our findings re-identify the roles of SFT and RL in the two-stage fine-tuning and discover the rotation of singular vectors as the key mechanism. %reversing the rotations induced by SFT, which shows recovery from forgetting, whereas imposing the SFT parameter directions onto a RL-tuned model results in performance degradation. Code is available at https://github.com/xiaodanguoguo/RL_Heals_SFT

Hangzhan Jin, Sicheng Lv, Sifan Wu et al.

Training large language models (LLMs) from scratch is increasingly impractical, making post-training methods such as supervised fine-tuning (SFT) and reinforcement-learning fine-tuning (RL-FT, e.g., PPO) central to modern practice. Using an out-of-distribution (OOD) variant of the 24-point card game and new spectrum-based diagnostics, we revisit how these two stages reshape model representation and OOD performance. Our key findings are- (1) RL-FT can restore much of the OOD performance loss from SFT (e.g., Llama-11B 8.97% to 15.38%, Qwen-7B 17.09% to 19.66%). But when SFT induces severe overfitting and a clear distribution shift, RL-FT cannot fully recover OOD performance. (2) Direction shifts of singular vectors matter more than singular value magnitudes. These shifts concentrate on directions linked to the largest and smallest singular values, leaving the bulk spectrum intact. (3) Low-rank and shallow recovery is effective: restoring singular vector directions for the top 20% of values or first 25% of layers recovers 70-80% of OOD performance. (4) Stronger SFT checkpoints enable better recovery by RL, while overfitted ones resist restoration. These results reconcile prior reports of RL superior OOD performance: RL primarily counteracts SFT-induced directional drift rather than finding new solutions. Our spectrum-aware analysis highlights inexpensive recovery knobs low-rank UV merging and shallow-layer resets that practitioners can use before costly RL fine-tuning.

Hangzhan Jin, Mohammad Hamdaqa

Unlike code generation, which involves creating code from scratch, code completion focuses on integrating new lines or blocks of code into an existing codebase. This process requires a deep understanding of the surrounding context, such as variable scope, object models, API calls, and database relations, to produce accurate results. These complex contextual dependencies make code completion a particularly challenging problem. Current models and approaches often fail to effectively incorporate such context, leading to inaccurate completions with low acceptance rates (around 30\%). For tasks like data transfer, which rely heavily on specific relationships and data structures, acceptance rates drop even further. This study introduces CCCI, a novel method for generating context-aware code completions specifically designed to address data transfer tasks. By integrating contextual information, such as database table relationships, object models, and library details into Large Language Models (LLMs), CCCI improves the accuracy of code completions. We evaluate CCCI using 289 Java snippets, extracted from over 819 operational scripts in an industrial setting. The results demonstrate that CCCI achieved a 49.1\% Build Pass rate and a 41.0\% CodeBLEU score, comparable to state-of-the-art methods that often struggle with complex task completion.