Recent generations of language models have introduced Large Reasoning Models (LRMs) that generate detailed thinking processes before providing answers. While these models demonstrate improved performance on reasoning benchmarks, their fundamental capabilities, scaling properties, and limitations remain insufficiently understood. Current evaluations primarily focus on established math and coding benchmarks, emphasizing final answer accuracy. However, this evaluation paradigm often suffers from contamination and does not provide insights into the reasoning traces. In this work, we systematically investigate these gaps with the help of controllable puzzle environments that allow precise manipulation of complexity while maintaining consistent logical structures. This setup enables the analysis of not only final answers but also the internal reasoning traces, offering insights into how LRMs think. Through extensive experiments, we show that LRMs face a complete accuracy collapse beyond certain complexities. Moreover, they exhibit a counterintuitive scaling limit: their reasoning effort increases with problem complexity up to a point, then declines despite having remaining token budget. By comparing LRMs with their standard LLM counterparts under same inference compute, we identify three performance regimes: (1) low-complexity tasks where standard models outperform LRMs, (2) medium-complexity tasks where LRMs demonstrates advantage, and (3) high-complexity tasks where both models face complete collapse. We found that LRMs have limitations in exact computation: they fail to use explicit algorithms and reason inconsistently across scales. We also investigate the reasoning traces in more depth, studying the patterns of explored solutions and analyzing the models' computational behavior, shedding light on their strengths, limitations, and raising questions about their reasoning capabilities.

Large Language Models (LLMs) reasoning abilities are increasingly being applied to classical board and card games, but the dominant approach -- involving prompting for direct move generation -- has significant drawbacks. It relies on the model's implicit fragile pattern-matching capabilities, leading to frequent illegal moves and strategically shallow play. Here we introduce an alternative approach: We use the LLM to translate natural language rules and game trajectories into a formal, executable world model represented as Python code. This generated model -- comprising functions for state transition, legal move enumeration, and termination checks -- serves as a verifiable simulation engine for high-performance planning algorithms like Monte Carlo tree search (MCTS). In addition, we prompt the LLM to generate heuristic value functions (to make MCTS more efficient), and inference functions (to estimate hidden states in imperfect information games). Our method offers three distinct advantages compared to directly using the LLM as a policy: (1) Verifiability: The generated CWM serves as a formal specification of the game's rules, allowing planners to algorithmically enumerate valid actions and avoid illegal moves, contingent on the correctness of the synthesized model; (2) Strategic Depth: We combine LLM semantic understanding with the deep search power of classical planners; and (3) Generalization: We direct the LLM to focus on the meta-task of data-to-code translation, enabling it to adapt to new games more easily. We evaluate our agent on 10 different games, of which 4 are novel and created for this paper. 5 of the games are fully observed (perfect information), and 5 are partially observed (imperfect information). We find that our method outperforms or matches Gemini 2.5 Pro in 9 out of the 10 considered games.

We present an automated evaluation system for assessing COBOL-to-Java code translation within IBM's watsonx Code Assistant for Z (WCA4Z). The system addresses key challenges in evaluating LLM-based translators, including model opacity and the complexity of translation quality assessment. Our approach combines analytic checkers with LLM-as-a-judge (LaaJ) techniques to deliver scalable, multi-faceted evaluations. The system supports continuous integration workflows, enables large-scale benchmarking, and reduces reliance on manual review. We describe the system architecture, evaluation strategies, and reporting mechanisms that provide actionable insights for developers and project managers, facilitating the evolution of high-quality, modernized codebases.

Static analysis is a powerful technique for bug detection in critical systems like operating system kernels. However, designing and implementing static analyzers is challenging, time-consuming, and typically limited to predefined bug patterns. While large language models (LLMs) have shown promise for static analysis, directly applying them to scan large codebases remains impractical due to computational constraints and contextual limitations. We present KNighter, the first approach that unlocks practical LLM-based static analysis by automatically synthesizing static analyzers from historical bug patterns. Rather than using LLMs to directly analyze massive codebases, our key insight is leveraging LLMs to generate specialized static analyzers guided by historical patch knowledge. KNighter implements this vision through a multi-stage synthesis pipeline that validates checker correctness against original patches and employs an automated refinement process to iteratively reduce false positives. Our evaluation on the Linux kernel demonstrates that KNighter generates high-precision checkers capable of detecting diverse bug patterns overlooked by existing human-written analyzers. To date, KNighter-synthesized checkers have discovered 70 new bugs/vulnerabilities in the Linux kernel, with 56 confirmed and 41 already fixed. 11 of these findings have been assigned CVE numbers. This work establishes an entirely new paradigm for scalable, reliable, and traceable LLM-based static analysis for real-world systems via checker synthesis.

Neural scaling laws are observed in a range of domains, to date with no clear understanding of why they occur. Recent theories suggest that loss power laws arise from Zipf's law, a power law observed in domains like natural language. One theory suggests that language scaling laws emerge when Zipf-distributed task quanta are learned in descending order of frequency. In this paper we examine power-law scaling in AlphaZero, a reinforcement learning algorithm, using a theory of language-model scaling. We find that game states in training and inference data scale with Zipf's law, which is known to arise from the tree structure of the environment, and examine the correlation between scaling-law and Zipf's-law exponents. In agreement with quanta scaling theory, we find that agents optimize state loss in descending order of frequency, even though this order scales inversely with modelling complexity. We also find that inverse scaling, the failure of models to improve with size, is correlated with unusual Zipf curves where end-game states are among the most frequent states. We show evidence that larger models shift their focus to these less-important states, sacrificing their understanding of important early-game states.

Due to the remarkable generative potential of diffusion-based models, numerous researches have investigated jailbreak attacks targeting these frameworks. A particularly concerning threat within image models is the generation of Not-Safe-for-Work (NSFW) content. Despite the implementation of security filters, numerous efforts continue to explore ways to circumvent these safeguards. Current attack methodologies primarily encompass adversarial prompt engineering or concept obfuscation, yet they frequently suffer from slow search efficiency, conspicuous attack characteristics and poor alignment with targets. To overcome these challenges, we propose Antelope, a more robust and covert jailbreak attack strategy designed to expose security vulnerabilities inherent in generative models. Specifically, Antelope leverages the confusion of sensitive concepts with similar ones, facilitates searches in the semantically adjacent space of these related concepts and aligns them with the target imagery, thereby generating sensitive images that are consistent with the target and capable of evading detection. Besides, we successfully exploit the transferability of model-based attacks to penetrate online black-box services. Experimental evaluations demonstrate that Antelope outperforms existing baselines across multiple defensive mechanisms, underscoring its efficacy and versatility.

This paper proposes a new method for preventing unsafe or otherwise low quality large language model (LLM) outputs, by leveraging the stochasticity of LLMs. We propose a system whereby LLM checkers vote on the acceptability of a generated output, regenerating it if a threshold of disapproval is reached, until sufficient checkers approve. We further propose estimators for cost and failure rate, and based on those estimators and experimental data tailored to the application, we propose an algorithm that achieves a desired failure rate at the least possible cost. We demonstrate that, under these models, failure rate decreases exponentially as a function of cost when voter count and threshold are chosen according to the algorithm, and that the models reasonably estimate the actual performance of such a system in action, even with limited data.

Large Language Models (LLMs) have shown impressive capabilities but also a concerning tendency to hallucinate. This paper presents RefChecker, a framework that introduces claim-triplets to represent claims in LLM responses, aiming to detect fine-grained hallucinations. In RefChecker, an extractor generates claim-triplets from a response, which are then evaluated by a checker against a reference. We delineate three task settings: Zero, Noisy and Accurate Context, to reflect various real-world use cases. We curated a benchmark spanning various NLP tasks and annotated 11k claim-triplets from 2.1k responses by seven LLMs. RefChecker supports both proprietary and open-source models as the extractor and checker. Experiments demonstrate that claim-triplets enable superior hallucination detection, compared to other granularities such as response, sentence and sub-sentence level claims. RefChecker outperforms prior methods by 6.8 to 26.1 points on our benchmark and the checking results of RefChecker are strongly aligned with human judgments. This work is open sourced at https://github.com/amazon-science/RefChecker

One critical issue for chat systems is to stay consistent about preferences, opinions, beliefs and facts of itself, which has been shown a difficult problem. In this work, we study methods to assess and bolster utterance consistency of chat systems. A dataset is first developed for studying the inconsistencies, where inconsistent dialogue responses, explanations of the inconsistencies, and recovery utterances are authored by annotators. This covers the life span of inconsistencies, namely introduction, understanding, and resolution. Building on this, we introduce a set of tasks centered on dialogue consistency, specifically focused on its detection and resolution. Our experimental findings indicate that our dataset significantly helps the progress in identifying and resolving conversational inconsistencies, and current popular large language models like ChatGPT which are good at resolving inconsistencies however still struggle with detection.

Addressing the gap in understanding visual comprehension in Large Language Models (LLMs), we designed a challenge-response study, subjecting Google Bard and GPT-Vision to 64 visual tasks, spanning categories like "Visual Situational Reasoning" and "Next Scene Prediction." Previous models, such as GPT4, leaned heavily on optical character recognition tools like Tesseract, whereas Bard and GPT-Vision, akin to Google Lens and Visual API, employ deep learning techniques for visual text recognition. However, our findings spotlight both vision-language model's limitations: while proficient in solving visual CAPTCHAs that stump ChatGPT alone, it falters in recreating visual elements like ASCII art or analyzing Tic Tac Toe grids, suggesting an over-reliance on educated visual guesses. The prediction problem based on visual inputs appears particularly challenging with no common-sense guesses for next-scene forecasting based on current "next-token" multimodal models. This study provides experimental insights into the current capacities and areas for improvement in multimodal LLMs.