Mingkuan Feng

Jinyang Wu, Mingkuan Feng, Shuai Zhang et al.

In-context learning (ICL) enables large language models (LLMs) to perform downstream tasks through advanced prompting and high-quality demonstrations. However, traditional ICL paradigms encounter significant limitations in complex reasoning tasks, stemming primarily from their dependence on example quality and absence of explicit reasoning guidance. To address these challenges, we introduce HiAR-ICL, a **Hi**gh-level **A**utomated **R**easoning paradigm in **ICL** that shifts focus from specific examples to abstract reasoning patterns, thereby extending the conventional concept of "context" in ICL. Our approach begins by defining five atomic reasoning actions, upon which we employ Monte Carlo Tree Search to systematically construct high-level reasoning patterns. During inference, HiAR-ICL dynamically selects appropriate reasoning patterns based on problem attributes, providing explicit guidance for the model's reasoning process. Experiments demonstrate HiAR-ICL's effectiveness and efficiency: utilizing only 200 prior samples with Qwen2.5-7B-Instruct, our method achieves 80.6% accuracy on MATH and 62.5% on AMC, exceeding GPT-4o's 77.2% and 57.5%. Our approach enhances performance across models of varying sizes while generalizing effectively across domains. Further analysis reveals that HiAR-ICL can also serve as a plug-and-play inference method compatible with post-training techniques like GRPO. Code and data are available at https://github.com/jinyangwu/HiARICL.

Jinyang Wu, Mingkuan Feng, Shuai Zhang et al.

Multimodal large language models excel across diverse domains but struggle with complex visual reasoning tasks. Current approaches aim to incorporate structured thinking via two strategies: explicit search methods and post-training techniques. However, both approaches face significant limitations: Search-based methods suffer from computational inefficiency due to extensive solution space exploration, while post-training methods require substantial data, computational resources, and often encounter training instability. To address these limitations, we propose AStar, an \textbf{A}utomated \textbf{S}tructured \textbf{t}hinking paradigm for multimod\textbf{a}l \textbf{r}easoning. Our method introduces "thought cards", a lightweight library of high-level reasoning patterns abstracted from 500 prior samples using Monte Carlo Tree Search. For each test problem, AStar adaptively retrieves the optimal thought cards and seamlessly integrates these external explicit guidelines with the model's internal implicit reasoning capabilities. Extensive experiments demonstrate AStar's effectiveness and efficiency: using only 500 prior samples and a 7B backbone, our training-free framework achieves 53.9$\%$ accuracy on MathVerse (surpassing GPT-4o's 50.2%) and 32.7% on MathVision (versus GPT-4o's 30.4%). Further analysis reveals that AStar generalizes beyond multimodal reasoning to visual perception and understanding domains, and serves as a plug-and-play test-time inference method compatible with mainstream post-training techniques like GRPO.

Jinyang Wu, Chonghua Liao, Mingkuan Feng et al.

Reinforcement learning (RL) has emerged as an effective paradigm for enhancing model reasoning. However, existing RL methods like GRPO often rely on unstructured self-sampling to fit scalar rewards, often producing inefficient rollouts that fail to capture transferable problem-solving strategies. To address these limitations, we propose **TemplateRL**, a structured template-guided RL framework that augments policy optimization with explicit template guidance. Our approach first constructs a problem-solving template library via MCTS on a small seed set, then seamlessly integrates this high-level structured guidance into RL training. By guiding rollout generation to align with proven template structures, TemplateRL significantly improves high-quality trajectory hit rates while reducing ineffective exploration. This structure-guided design steers the policy toward validated strategic patterns, stabilizing training dynamics, and enhancing RL sampling efficiency. Notably, the explicit template library is interpretable, editable, and supports online updates-enabling continuous updates during both training and inference. Extensive experiments demonstrate that TemplateRL outperforms GRPO by 99% on AIME and 41% on AMC, with superior stability on weak models and remarkable cross-domain generalization, highlighting its potential for broader tasks.

Ruihan Jin, Pengpeng Shao, Zhengqi Wen et al.

The rapid advancements in large language models (LLMs) have led to the emergence of routing techniques, which aim to efficiently select the optimal LLM from diverse candidates to tackle specific tasks, optimizing performance while reducing costs. Current LLM routing methods are limited in effectiveness due to insufficient exploration of the intrinsic connection between user queries and the characteristics of LLMs. To address this issue, in this paper, we present RadialRouter, a novel framework for LLM routing which employs a lightweight Transformer-based backbone with a radial structure named RadialFormer to articulate the query-LLMs relationship. The optimal LLM selection is performed based on the final states of RadialFormer. The pipeline is further refined by an objective function that combines Kullback-Leibler divergence with the query-query contrastive loss to enhance robustness. Experimental results on RouterBench show that RadialRouter significantly outperforms existing routing methods by 9.2\% and 5.8\% in the Balance and Cost First scenarios, respectively. Additionally, its adaptability toward different performance-cost trade-offs and the dynamic LLM pool demonstrates practical application potential.

Mingkuan Feng, Jinyang Wu, Shuai Zhang et al.

Large language models (LLMs) have achieved significant progress across various domains, but their increasing scale results in high computational and memory costs. Recent studies have revealed that LLMs exhibit sparsity, providing the potential to reduce model size through pruning techniques. However, existing pruning methods typically follow a prune-then-finetune paradigm. Since the pruned components still contain valuable information, their direct removal often leads to irreversible performance degradation, imposing a substantial computational burden to recover performance during finetuning. In this paper, we propose a novel paradigm that first applies regularization, then prunes, and finally finetunes. Based on this paradigm, we introduce DReSS, a simple and effective Data-driven Regularized Structured Streamlining method for LLMs. By leveraging a small amount of data to regularize the components to be pruned, DReSS explicitly transfers the important information to the remaining parts of the model in advance. Compared to direct pruning, this can reduce the information loss caused by parameter removal, thereby enhancing its language modeling capabilities. Experimental results demonstrate that DReSS significantly outperforms existing pruning methods even under extreme pruning ratios, significantly reducing latency and increasing throughput.

Mingkuan Feng, Jinyang Wu, Siyuan Liu et al.

The deployment of large language models (LLMs) is largely hindered by their large number of parameters. Structural pruning has emerged as a promising solution. Prior structured pruning methods directly remove unimportant parameters based on certain metrics, which often causes knowledge loss and necessitates extensive retraining. To overcome this, we introduce a novel pruning method TRSP: Two-Stage Regularization-Based Structured Pruning for LLMs. Specifically, we multiply the output of each transformer layer by an initial learnable weight and iteratively learn these weights by adding their $\ell_1$-norm as a regularization term to the loss function, serving as the first-stage regularization. Subsequently, we apply additional regularization to the difference between the output and input of layers with smaller weights, encouraging the shift of knowledge to the preserved layers. This serves as the second-stage regularization. TRSP retains more knowledge and better preserves model performance than direct parameter elimination. Through extensive experimentation we show that TRSP outperforms strong layer-wise structured pruning methods without requiring retraining. As a layer-wise pruning method, it delivers notable end-to-end acceleration, making it a promising solution for efficient LLM deployment.