Large Reasoning Models (LRMs) significantly improve the reasoning ability of Large Language Models (LLMs) by learning to reason, exhibiting promising performance in complex task-solving. However, their deliberative reasoning process leads to inefficiencies in token usage, memory consumption, and inference time. Thus, this survey provides a review of efficient inference methods designed specifically for LRMs, focusing on mitigating token inefficiency while preserving the reasoning quality. First, we introduce a taxonomy to group the recent methods into two main categories: (a) explicit compact Chain-of-Thought (CoT), which reduces tokens while keeping the explicit reasoning structure, and (b) implicit latent CoT, which encodes reasoning steps within hidden representations instead of explicit tokens. Meanwhile, we discuss their strengths and weaknesses. Then, we conduct empirical analyses on existing methods from performance and efficiency aspects. Besides, we present open challenges in this field, including human-centric controllable reasoning, trade-off between interpretability and efficiency of reasoning, ensuring safety of efficient reasoning, and broader applications of efficient reasoning. In addition, we highlight key insights for enhancing LRMs' inference efficiency via techniques such as model merging, new architectures, and agent routers. We hope this work serves as a valuable guide, helping researchers overcome challenges in this vibrant field\footnote{https://github.com/yueliu1999/Awesome-Efficient-Inference-for-LRMs}.

Scaling the effective context length is essential for advancing large language models (LLMs) toward artificial general intelligence (AGI). However, the quadratic increase in computational complexity inherent in traditional attention mechanisms presents a prohibitive overhead. Existing approaches either impose strongly biased structures, such as sink or window attention which are task-specific, or radically modify the attention mechanism into linear approximations, whose performance in complex reasoning tasks remains inadequately explored. In this work, we propose a solution that adheres to the ``less structure'' principle, allowing the model to determine where to attend autonomously, rather than introducing predefined biases. We introduce Mixture of Block Attention (MoBA), an innovative approach that applies the principles of Mixture of Experts (MoE) to the attention mechanism. This novel architecture demonstrates superior performance on long-context tasks while offering a key advantage: the ability to seamlessly transition between full and sparse attention, enhancing efficiency without the risk of compromising performance. MoBA has already been deployed to support Kimi's long-context requests and demonstrates significant advancements in efficient attention computation for LLMs. Our code is available at https://github.com/MoonshotAI/MoBA.

Language model pretraining with next token prediction has proved effective for scaling compute but is limited to the amount of available training data. Scaling reinforcement learning (RL) unlocks a new axis for the continued improvement of artificial intelligence, with the promise that large language models (LLMs) can scale their training data by learning to explore with rewards. However, prior published work has not produced competitive results. In light of this, we report on the training practice of Kimi k1.5, our latest multi-modal LLM trained with RL, including its RL training techniques, multi-modal data recipes, and infrastructure optimization. Long context scaling and improved policy optimization methods are key ingredients of our approach, which establishes a simplistic, effective RL framework without relying on more complex techniques such as Monte Carlo tree search, value functions, and process reward models. Notably, our system achieves state-of-the-art reasoning performance across multiple benchmarks and modalities -- e.g., 77.5 on AIME, 96.2 on MATH 500, 94-th percentile on Codeforces, 74.9 on MathVista -- matching OpenAI's o1. Moreover, we present effective long2short methods that use long-CoT techniques to improve short-CoT models, yielding state-of-the-art short-CoT reasoning results -- e.g., 60.8 on AIME, 94.6 on MATH500, 47.3 on LiveCodeBench -- outperforming existing short-CoT models such as GPT-4o and Claude Sonnet 3.5 by a large margin (up to +550%).

The swift advancement in the scales and capabilities of Large Language Models (LLMs) positions them as promising tools for a variety of downstream tasks. In addition to the pursuit of better performance and the avoidance of violent feedback on a certain prompt, to ensure the responsibility of the LLM, much attention is drawn to the robustness of LLMs. However, existing evaluation methods mostly rely on traditional question answering datasets with predefined supervised labels, which do not align with the superior generation capabilities of contemporary LLMs. To address this issue, we propose a novel rational evaluation approach that leverages pre-trained reward models as diagnostic tools to evaluate the longer conversation generated from more challenging open questions by LLMs, which we refer to as the Reward Model for Reasonable Robustness Evaluation (TREvaL). Longer conversations manifest the comprehensive grasp of language models in terms of their proficiency in understanding questions, a capability not entirely encompassed by individual words or letters, which may exhibit oversimplification and inherent biases. Our extensive empirical experiments demonstrate that TREvaL provides an innovative method for evaluating the robustness of an LLM. Furthermore, our results demonstrate that LLMs frequently exhibit vulnerability to word-level perturbations that are commonplace in daily language usage. Notably, we are surprised to discover that robustness tends to decrease as fine-tuning (SFT and RLHF) is conducted. The code of TREval is available in https://github.com/Harry-mic/TREvaL.

In this work, we explore a Multilingual Information Retrieval (MLIR) task, where the collection includes documents in multiple languages. We demonstrate that applying state-of-the-art approaches developed for cross-lingual information retrieval to MLIR tasks leads to sub-optimal performance. This is due to the heterogeneous and imbalanced nature of multilingual collections -- some languages are better represented in the collection and some benefit from large-scale training data. To address this issue, we present KD-SPD, a novel soft prompt decoding approach for MLIR that implicitly "translates" the representation of documents in different languages into the same embedding space. To address the challenges of data scarcity and imbalance, we introduce a knowledge distillation strategy. The teacher model is trained on rich English retrieval data, and by leveraging bi-text data, our distillation framework transfers its retrieval knowledge to the multilingual document encoder. Therefore, our approach does not require any multilingual retrieval training data. Extensive experiments on three MLIR datasets with a total of 15 languages demonstrate that KD-SPD significantly outperforms competitive baselines in all cases. We conduct extensive analyses to show that our method has less language bias and better zero-shot transfer ability towards new languages.

Benefiting from transformer-based pre-trained language models, neural ranking models have made significant progress. More recently, the advent of multilingual pre-trained language models provides great support for designing neural cross-lingual retrieval models. However, due to unbalanced pre-training data in different languages, multilingual language models have already shown a performance gap between high and low-resource languages in many downstream tasks. And cross-lingual retrieval models built on such pre-trained models can inherit language bias, leading to suboptimal result for low-resource languages. Moreover, unlike the English-to-English retrieval task, where large-scale training collections for document ranking such as MS MARCO are available, the lack of cross-lingual retrieval data for low-resource language makes it more challenging for training cross-lingual retrieval models. In this work, we propose OPTICAL: Optimal Transport distillation for low-resource Cross-lingual information retrieval. To transfer a model from high to low resource languages, OPTICAL forms the cross-lingual token alignment task as an optimal transport problem to learn from a well-trained monolingual retrieval model. By separating the cross-lingual knowledge from knowledge of query document matching, OPTICAL only needs bitext data for distillation training, which is more feasible for low-resource languages. Experimental results show that, with minimal training data, OPTICAL significantly outperforms strong baselines on low-resource languages, including neural machine translation.