Look closely and you'll see that every part of the text is not quite right. A few weeks ago, where I live in Johannesburg, a man ran a stop sign and crashed into my Subaru. At the scene he was frantic, unable to gather his thoughts. Half an hour later, I received a lengthy, perfectly grammatical text from him elegantly explaining how he perceived the crash had happened. For a repair quote, I wrote to a mechanic I know, a man who used to text me in curt phrases riddled with shorthand.

The rapid development of Multimodal Large Reasoning Models (MLRMs) has demonstrated broad application potential, yet their safety and reliability remain critical concerns that require systematic exploration. To address this gap, we conduct a comprehensive and systematic safety evaluation of 11 MLRMs across 5 benchmarks and unveil prevalent safety degradation phenomena in most advanced models. Moreover, our analysis reveals distinct safety patterns across different benchmarks: significant safety degradation is observed across jailbreak robustness benchmarks, whereas safety-awareness benchmarks demonstrate less pronounced degradation. In particular, the long thought process in some scenarios even enhances safety performance. Therefore, it is a potential approach to address safety issues in MLRMs by leveraging the intrinsic reasoning capabilities of the model to detect unsafe intent. To operationalize this insight, we construct a multimodal tuning dataset that incorporates a safety-oriented thought process. Experimental results from fine-tuning existing MLRMs with this dataset effectively enhances the safety on both jailbreak robustness and safety-awareness benchmarks. This study provides a new perspective for developing safe MLRMs. Our dataset is available at https://github.com/xinyuelou/Think-in-Safety.

Models employing long chain-of-thought (CoT) reasoning have shown superior performance on complex reasoning tasks. Yet, this capability introduces a critical and often overlooked inefficiency -- overthinking -- models often engage in unnecessarily extensive reasoning even for simple queries, incurring significant computations without accuracy improvements. While prior work has explored solutions to mitigate overthinking, a fundamental gap remains in our understanding of its underlying causes. Most existing analyses are limited to superficial, profiling-based observations, failing to delve into LLMs' inner workings. This study introduces a systematic, fine-grained analyzer of LLMs' thought process to bridge the gap, TRACE. We first benchmark the overthinking issue, confirming that long-thinking models are five to twenty times slower on simple tasks with no substantial gains. We then use TRACE to first decompose the thought process into minimally complete sub-thoughts. Next, by inferring discourse relationships among sub-thoughts, we construct granular thought progression graphs and subsequently identify common thinking patterns for topically similar queries. Our analysis reveals two major patterns for open-weight thinking models -- Explorer and Late Landing. This finding provides evidence that over-verification and over-exploration are the primary drivers of overthinking in LLMs. Grounded in thought structures, we propose a utility-based definition of overthinking, which moves beyond length-based metrics. This revised definition offers a more insightful understanding of LLMs' thought progression, as well as practical guidelines for principled overthinking management.

Recent advances in self-refinement have demonstrated significant potential for improving the outputs of large language models (LLMs) through iterative refinement. However, most existing self-refinement methods rely on a reactive process with a fixed number of iterations, making it difficult to determine the optimal timing and content of refinement based on the evolving generation context. Inspired by the way humans dynamically refine their thoughts during execution, we propose ProActive Self-Refinement (PASR), a novel method that enables LLMs to refine their outputs during the generation process. Unlike methods that regenerate entire responses, PASR proactively decides whether, when, and how to refine based on the model's internal state and evolving context. We conduct extensive experiments on a diverse set of 10 tasks to evaluate the effectiveness of PASR. Experimental results show that PASR significantly enhances problem-solving performance. In particular, on Qwen3-8B, PASR reduces average token consumption by 41.6% compared to standard generation, while also achieving an 8.2% improvement in accuracy. Our code and baselines used in the paper are available on GitHub.

Large language models (LLMs) are increasingly trusted as automated judges, assisting evaluation and providing reward signals for training other models, particularly in reference-based settings like Reinforcement Learning with Verifiable Rewards (RLVR). However, we uncover a critical vulnerability even in this reference-based paradigm: generative reward models are systematically susceptible to reward hacking. We find that superficial inputs, which we term ''master keys'' such as non-word symbols (e.g., '':'' or ''.'') or generic reasoning openers (e.g., ''Thought process:'' or ''Let's solve this problem step by step.''), can consistently elicit false positive rewards without any substantive reasoning. Our systematic evaluation demonstrates this is a widespread failure affecting a diverse range of models, including leading proprietary systems such as GPT-o1 and Claude-4. These results challenge the assumed robustness of LLM judges and pose a significant threat to their reliability. To address this, we propose a simple yet effective data augmentation strategy using truncated model outputs as adversarial negative examples. The resulting Master Reward Models (Master-RMs) demonstrate state-of-the-art robustness against these ''master key'' attacks while maintaining high performance in standard evaluation settings. We supplement these findings with a comprehensive analysis of the vulnerability across model scales, prompt variations, and common inference-time strategies, offering insights to guide future research on robust LLM evaluation. We release our robust, general-domain reward models and the synthetic training data at https://huggingface.co/sarosavo/Master-RM and https://huggingface.co/datasets/sarosavo/Master-RM.

Grammatical error correction is a significant task in NLP. Traditional methods based on encoder-decoder models have achieved certain success, but the application of LLMs in this field is still underexplored. Current research predominantly relies on supervised fine-tuning to train LLMs to directly generate the corrected sentence, which limits the model's powerful reasoning ability. To address this limitation, we propose a novel framework based on Rule-Based RL. Through experiments on the Chinese datasets, our Rule-Based RL framework achieves \textbf{state-of-the-art }performance, with a notable increase in \textbf{recall}. This result clearly highlights the advantages of using RL to steer LLMs, offering a more controllable and reliable paradigm for future development in GEC.

Large language models have advanced rapidly, but no single model excels in every area -- each has its strengths and weaknesses. Instead of relying on one model alone, we take inspiration from gossip protocols in distributed systems, where information is exchanged with peers until they all come to an agreement. In this setup, models exchange answers and gradually work toward a shared solution. Each LLM acts as a node in a peer-to-peer network, sharing responses and thought processes to reach a collective decision. Our results show that this "gossip-based consensus" leads to robust, resilient, and accurate multi-agent AI reasoning. It helps overcome the weaknesses of individual models and brings out their collective strengths. This approach is similar to how humans build consensus, making AI seem more collaborative and trustworthy instead of just a black-box program.

This paper elucidates that current state-of-the-art Large Language Models (LLMs) are fundamentally incapable of making decisions or developing "thoughts" within the feature space due to their architectural constraints. We establish a definition of "thought" that encompasses traditional understandings of that term and adapt it for application to LLMs. We demonstrate that the architectural design and language modeling training methodology of contemporary LLMs inherently preclude them from engaging in genuine thought processes. Our primary focus is on this theoretical realization rather than practical insights derived from experimental data. Finally, we propose solutions to enable thought processes within the feature space and discuss the broader implications of these architectural modifications.

Recent advances in LLM-based role-playing language agents (RPLAs) have attracted broad attention in various applications. While chain-of-thought reasoning has shown importance in many tasks for LLMs, the internal thinking processes of RPLAs remain unexplored. Understanding characters' inner thoughts is crucial for developing advanced RPLAs. In this paper, we introduce ROLETHINK, a novel benchmark constructed from literature for evaluating character thought generation. We propose the task of inner thought reasoning, which includes two sets: the gold set that compares generated thoughts with original character monologues, and the silver set that uses expert synthesized character analyses as references. To address this challenge, we propose MIRROR, a chain-of-thought approach that generates character thoughts by retrieving memories, predicting character reactions, and synthesizing motivations. Through extensive experiments, we demonstrate the importance of inner thought reasoning for RPLAs, and MIRROR consistently outperforms existing methods. Resources are available at https://github.com/airaer1998/RPA_Thought.

Longer thought, better performance: large language models with deep reasoning capabilities, particularly o1-like models, have demonstrated remarkable performance by generating extensive thought processes during inference. This trade-off reveals a potential vulnerability: adversaries could compromise model performance by forcing immediate responses without thought processes. To this end, in this paper, we introduce a novel attack scenario targeting the long thought processes of o1-like models and propose BoT (Break CoT), which can selectively break intrinsic reasoning mechanisms through backdoor attacks. BoT constructs poisoned datasets with designed triggers and injects backdoor by either supervised fine-tuning or direct preference optimization. When triggered, the model directly generates answers without thought processes, while maintaining normal reasoning capabilities for clean inputs. Extensive experiments on open-source o1-like models, including recent DeepSeek-R1, demonstrate that BoT nearly achieves high attack success rates while maintaining clean accuracy, highlighting the critical safety risk in current models. Furthermore, the relationship between task difficulty and helpfulness reveals a potential application for good, enabling users to customize model behavior based on task complexity. Code is available at \href{https://github.com/zihao-ai/BoT}{https://github.com/zihao-ai/BoT}.