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}.

As large language models (LLMs) continue to evolve, ensuring their alignment with human goals and values remains a pressing challenge. A key concern is \textit{instrumental convergence}, where an AI system, in optimizing for a given objective, develops unintended intermediate goals that override the ultimate objective and deviate from human-intended goals. This issue is particularly relevant in reinforcement learning (RL)-trained models, which can generate creative but unintended strategies to maximize rewards. In this paper, we explore instrumental convergence in LLMs by comparing models trained with direct RL optimization (e.g., the o1 model) to those trained with reinforcement learning from human feedback (RLHF). We hypothesize that RL-driven models exhibit a stronger tendency for instrumental convergence due to their optimization of goal-directed behavior in ways that may misalign with human intentions. To assess this, we introduce InstrumentalEval, a benchmark for evaluating instrumental convergence in RL-trained LLMs. Initial experiments reveal cases where a model tasked with making money unexpectedly pursues instrumental objectives, such as self-replication, implying signs of instrumental convergence. Our findings contribute to a deeper understanding of alignment challenges in AI systems and the risks posed by unintended model behaviors.

Wang et al., 2024b; Wu et al., 2024; Zhang et al., 2024b), as Recent studies have revealed significant shown in Figure 1. Initial successes in LLM-based human gaps between LLM-based human simulations and simulations have been demonstrated across diverse fields, real-world observations, highlighting these dual including society, economics, policy, and psychology (Chen challenges. To address these gaps, we present a et al., 2024a; Li et al., 2024b;f; Lin et al., 2023; Park et al., comprehensive analysis of LLM limitations and 2023b; Yang et al., 2024b). Moreover, reliable LLM simulations our design issues, proposing targeted solutions can generate high-quality data for LLM training for both aspects. Furthermore, we explore future (Tang et al., 2024; Zhang et al., 2024a) and evaluate data directions that address both challenges simultaneously, quality (Chiang et al., 2024; Moniri et al., 2024; Xu et al., particularly in data collection, LLM generation, 2023b; Zheng et al., 2023b), serving as a data generator and and evaluation. To support further research evaluator (Gu et al., 2024; Li et al., 2024c; Son et al., 2024) in this field, we provide a curated collection of to enhance LLM pre-training and simulation abilities.

To broaden the dissemination of scientific knowledge to diverse audiences, scientific document summarization must simultaneously control multiple attributes such as length and empirical focus. However, existing research typically focuses on controlling single attributes, leaving the compositional control of multiple attributes underexplored. To address this gap, we introduce CCSBench, a benchmark for compositional controllable summarization in the scientific domain. Our benchmark enables fine-grained control over both explicit attributes (e.g., length), which are objective and straightforward, and implicit attributes (e.g., empirical focus), which are more subjective and conceptual. We conduct extensive experiments on GPT-4, LLaMA2, and other popular LLMs under various settings. Our findings reveal significant limitations in large language models' ability to balance trade-offs between control attributes, especially implicit ones that require deeper understanding and abstract reasoning.

It is commonly perceived that online fake news and reliable news exhibit stark differences in writing styles, such as the use of sensationalist versus objective language. However, we emphasize that style-related features can also be exploited for style-based attacks. Notably, the rise of powerful Large Language Models (LLMs) has enabled malicious users to mimic the style of trustworthy news outlets at minimal cost. Our analysis reveals that LLM-camouflaged fake news content leads to substantial performance degradation of state-of-the-art text-based detectors (up to 38% decrease in F1 Score), posing a significant challenge for automated detection in online ecosystems. To address this, we introduce SheepDog, a style-agnostic fake news detector robust to news writing styles. SheepDog achieves this adaptability through LLM-empowered news reframing, which customizes each article to match different writing styles using style-oriented reframing prompts. By employing style-agnostic training, SheepDog enhances its resilience to stylistic variations by maximizing prediction consistency across these diverse reframings. Furthermore, SheepDog extracts content-focused veracity attributions from LLMs, where the news content is evaluated against a set of fact-checking rationales. These attributions provide supplementary information and potential interpretability that assist veracity prediction. On three benchmark datasets, empirical results show that SheepDog consistently yields significant improvements over competitive baselines and enhances robustness against LLM-empowered style attacks.

Despite considerable advances in automated fake news detection, due to the timely nature of news, it remains a critical open question how to effectively predict the veracity of news articles based on limited fact-checks. Existing approaches typically follow a "Train-from-Scratch" paradigm, which is fundamentally bounded by the availability of large-scale annotated data. While expressive pre-trained language models (PLMs) have been adapted in a "Pre-Train-and-Fine-Tune" manner, the inconsistency between pre-training and downstream objectives also requires costly task-specific supervision. In this paper, we propose "Prompt-and-Align" (P&A), a novel prompt-based paradigm for few-shot fake news detection that jointly leverages the pre-trained knowledge in PLMs and the social context topology. Our approach mitigates label scarcity by wrapping the news article in a task-related textual prompt, which is then processed by the PLM to directly elicit task-specific knowledge. To supplement the PLM with social context without inducing additional training overheads, motivated by empirical observation on user veracity consistency (i.e., social users tend to consume news of the same veracity type), we further construct a news proximity graph among news articles to capture the veracity-consistent signals in shared readerships, and align the prompting predictions along the graph edges in a confidence-informed manner. Extensive experiments on three real-world benchmarks demonstrate that P&A sets new states-of-the-art for few-shot fake news detection performance by significant margins.

Confidence calibration is central to providing accurate and interpretable uncertainty estimates, especially under safety-critical scenarios. However, we find that existing calibration algorithms often overlook the issue of proximity bias, a phenomenon where models tend to be more overconfident in low proximity data (i.e., lying in the sparse region of the data distribution) compared to high proximity samples, and thus suffer from inconsistent miscalibration across different proximity samples. We examine the problem over pretrained ImageNet models and observe that: 1) Proximity bias exists across a wide variety of model architectures and sizes; 2) Transformer-based models are more susceptible to proximity bias than CNN-based models; 3) Proximity bias persists even after performing popular calibration algorithms like temperature scaling; 4) Models tend to overfit more heavily on low proximity samples than on high proximity samples. Motivated by the empirical findings, we propose ProCal, a plug-and-play algorithm with a theoretical guarantee to adjust sample confidence based on proximity. To further quantify the effectiveness of calibration algorithms in mitigating proximity bias, we introduce proximity-informed expected calibration error (PIECE) with theoretical analysis. We show that ProCal is effective in addressing proximity bias and improving calibration on balanced, long-tail, and distribution-shift settings under four metrics over various model architectures.