Jiandong Gao

Yutong Chen, Jiandong Gao, Ji Wu

R1-style Reinforcement Learning (RL) significantly enhances Large Language Models' reasoning capabilities, yet the mechanism behind rule-based RL remains unclear. We found that small-scale SFT has substantial influence on RL but shows poor efficiency. To explain our observations, we propose an analytical framework and compare the efficiency of SFT and RL by measuring \textbf{sample effect}. Our hypothetical analysis shows the potential to improve SFT efficiency. Guided by our analysis, we propose \textbf{Re-distillation}, a technique that aims to boost the effectiveness of small-scale distillation by sampling from the RL-trained policy. Re-distillation shows consistent surprising efficiency on three datasets and both Qwen\&Llama models: Re-distilled models matched RL performance with far fewer samples and less computation. As a result, on K\&K dataset, our re-distilled Qwen-2.5-1.5B model surpasses DeepSeek-V3-0324 with only 1K SFT samples. We demonstrate that re-distillation can be used to efficiently balance multiple goals in RL. Our work explains several interesting phenomena in R1-style RL, shedding light on the mechanisms behind its empirical success. Code is available at: https://github.com/on1262/deep-reasoning.

Yutong Chen, Jiandong Gao, Ji Wu

Training Large Reasoning Model (LRM) is usually unstable and unpredictable, especially on hard problems or weak foundation models. We found that the current post-training scaling strategy can still improve on these cases. We propose CoScale-RL, a novel scaling strategy with better data and computational efficiency. We first scale up solutions to make problems solvable. The core idea is to collect multiple solutions for each problem, rather than simply enlarging the dataset. Then, we scale up rollout computation to stabilize Reinforcement Learning. We further leverage a model merge technique called Re-distillation to sustain or even improve computational efficiency when scaling up. Our method significantly improves data and computational efficiency, with an average 3.76$\times$ accuracy improvement on four benchmarks. CoScale-RL is able to improve an LRM's ability boundary without an extensive SFT dataset. Our method provides a new scaling direction to further improve LRM's reasoning ability.

Mutian Yang, Zisen Zhan, Yutong Chen et al.

Catastrophic forgetting during knowledge injection severely undermines the continual learning capability of large language models (LLMs). Although existing methods attempt to mitigate this issue, they often lack a foundational theoretical explanation. We establish a gradient-based theoretical framework to explain catastrophic forgetting. We first prove that strongly negative gradient similarity is a fundamental cause of forgetting. We then use gradient similarity to identify two types of neurons: conflicting neurons that induce forgetting and account for 50%-75% of neurons, and collaborative neurons that mitigate forgetting and account for 25%-50%. Based on this analysis, we propose a knowledge injection method, Collaborative Neural Learning (CNL). By freezing conflicting neurons and updating only collaborative neurons, CNL theoretically eliminates catastrophic forgetting under an infinitesimal learning rate eta and an exactly known mastered set. Experiments on five LLMs, four datasets, and four optimizers show that CNL achieves zero forgetting in in-set settings and reduces forgetting by 59.1%-81.7% in out-of-set settings.

Mutian Yang, Jiandong Gao, Ji Wu

While large language models (LLMs) leverage both knowledge and reasoning during inference, the capacity to distinguish between them plays a pivotal role in model analysis, interpretability, and development. Inspired by dual-system cognitive theory, we propose a cognition attribution framework to decouple the contribution of knowledge and reasoning. In particular, the cognition of LLMs is decomposed into two distinct yet complementary phases: knowledge retrieval (Phase 1) and reasoning adjustment (Phase 2). To separate these phases, LLMs are prompted to generate answers under two different cognitive modes, fast thinking and slow thinking, respectively. The performance under different cognitive modes is analyzed to quantify the contribution of knowledge and reasoning. This architecture is employed to 15 LLMs across 3 datasets. Results reveal: (1) reasoning adjustment is domain-specific, benefiting reasoning-intensive domains (e.g., mathematics, physics, and chemistry) and potentially imparing knowledge-intensive domains. (2) Parameter scaling improves both knowledge and reasoning, with knowledge improvements being more pronounced. Additionally, parameter scaling make LLMs reasoning significantly more prudent, while moderately more intelligent. (3) Knowledge primarily resides in lower network layers, while reasoning operates in higher layers. Our framework not only helps understand LLMs from a "decoupling" perspective, but also provides new insights into existing research, including scaling laws, hierarchical knowledge editing, and limitations of small-model reasoning.