Using multiple-choice questions (MCQs) has become a standard for assessing LLM capabilities efficiently. A variety of metrics can be employed for this task. However, previous research has not conducted a thorough assessment of them. At the same time, MCQ evaluation suffers from answer fluctuation: models produce different results given slight changes in prompts. We suggest a metric assessment protocol in which evaluation methodologies are analyzed through their connection with fluctuation rates, as well as original performance. Our results show that there is a strong link between existing metrics and the answer changing, even when computed without any additional prompt variants. A novel metric, worst accuracy, demonstrates the highest association on the protocol.

T erahertz (THz) imaging enables non-invasive analysis for applications such as security screening and material classification, but effective image classification remains challenging due to limited annotations, low resolution, and visual ambiguity. W e introduce In-Context Learning (ICL) with Vision-Language Models (VLMs) as a flexible, interpretable alternative that requires no fine-tuning. Using a modality-aligned prompting framework, we adapt two open-weight VLMs to the THz domain and evaluate them under zero-shot and one-shot settings. Our results show that ICL improves classification and interpretability in low-data regimes. This is the first application of ICL-enhanced VLMs to THz imaging, offering a promising direction for resource-constrained scientific domains.

Despite recent progress in large language models (LLMs), evaluation of text generation tasks such as text style transfer (TST) remains a significant challenge. Recent studies (Dementieva et al., 2024; Pauli et al., 2025) revealed a substantial gap between automatic metrics and human judgments. Moreover, most prior work focuses exclusively on English, leaving multilingual TST evaluation largely unexplored. In this paper, we perform the first comprehensive multilingual study on evaluation of text detoxification system across nine languages: English, Spanish, German, Chinese, Arabic, Hindi, Ukrainian, Russian, Amharic. Drawing inspiration from the machine translation, we assess the effectiveness of modern neural-based evaluation models alongside prompting-based LLM-as-a-judge approaches. Our findings provide a practical recipe for designing more reliable multilingual TST evaluation pipeline in the text detoxification case.

Do large language models (LLMs) construct and manipulate internal world models, or do they rely solely on statistical associations represented as output layer token probabilities? We adapt cognitive science methodologies from human mental models research to test LLMs on pulley system problems using TikZ-rendered stimuli. Study 1 examines whether LLMs can estimate mechanical advantage (MA). State-of-the-art models performed marginally but significantly above chance, and their estimates correlated significantly with ground-truth MA. Significant correlations between number of pulleys and model estimates suggest that models employed a pulley counting heuristic, without necessarily simulating pulley systems to derive precise values. Study 2 tested this by probing whether LLMs represent global features crucial to MA estimation. Models evaluated a functionally connected pulley system against a fake system with randomly placed components. Without explicit cues, models identified the functional system as having greater MA with F1=0.8, suggesting LLMs could represent systems well enough to differentiate jumbled from functional systems. Study 3 built on this by asking LLMs to compare functional systems with matched systems which were connected up but which transferred no force to the weight; LLMs identified the functional system with F1=0.46, suggesting random guessing. Insofar as they may generalize, these findings are compatible with the notion that LLMs manipulate internal world models, sufficient to exploit statistical associations between pulley count and MA (Study 1), and to approximately represent system components' spatial relations (Study 2). However, they may lack the facility to reason over nuanced structural connectivity (Study 3). We conclude by advocating the utility of cognitive scientific methods to evaluate the world-modeling capacities of artificial intelligence systems.

This work evaluates the potential of large language models (LLMs) to power digital assistants capable of complex action execution. These assistants rely on pre-trained programming knowledge to execute multi-step goals by composing objects and functions defined in assistant libraries into action execution programs. To achieve this, we develop ASPERA, a framework comprising an assistant library simulation and a human-assisted LLM data generation engine. Our engine allows developers to guide LLM generation of high-quality tasks consisting of complex user queries, simulation state and corresponding validation programs, tackling data availability and evaluation robustness challenges. Alongside the framework we release Asper-Bench, an evaluation dataset of 250 challenging tasks generated using ASPERA, which we use to show that program generation grounded in custom assistant libraries is a significant challenge to LLMs compared to dependency-free code generation.

Phishing emails are a critical component of the cybercrime kill chain due to their wide reach and low cost. Their ever-evolving nature renders traditional rule-based and feature-engineered detectors ineffective in the ongoing arms race between attackers and defenders. The rise of large language models (LLMs) further exacerbates the threat, enabling attackers to craft highly convincing phishing emails at minimal cost. This work demonstrates that LLMs can generate psychologically persuasive phishing emails tailored to victim profiles, successfully bypassing nearly all commercial and academic detectors. To defend against such threats, we propose PiMRef, the first reference-based phishing email detector that leverages knowledge-based invariants. Our core insight is that persuasive phishing emails often contain disprovable identity claims, which contradict real-world facts. PiMRef reframes phishing detection as an identity fact-checking task. Given an email, PiMRef (i) extracts the sender's claimed identity, (ii) verifies the legitimacy of the sender's domain against a predefined knowledge base, and (iii) detects call-to-action prompts that push user engagement. Contradictory claims are flagged as phishing indicators and serve as human-understandable explanations. Compared to existing methods such as D-Fence, HelpHed, and ChatSpamDetector, PiMRef boosts precision by 8.8% with no loss in recall on standard benchmarks like Nazario and PhishPot. In a real-world evaluation of 10,183 emails across five university accounts over three years, PiMRef achieved 92.1% precision, 87.9% recall, and a median runtime of 0.05s, outperforming the state-of-the-art in both effectiveness and efficiency.

Recent progress in LLMs, such as reasoning models, has demonstrated strong abilities to solve complex competitive programming problems, often rivaling top human competitors. However, it remains underexplored whether these abilities generalize to relevant domains that are less seen during training. To address this, we introduce AlgoSimBench, a new benchmark designed to assess LLMs' ability to identify algorithmically similar problems (ASPs)-problems that can be solved using similar algorithmic approaches. AlgoSimBench consists of 1317 problems, annotated with 231 distinct fine-grained algorithm tags, from which we curate 402 multiple-choice questions (MCQs), where each question presents one algorithmically similar problem alongside three textually similar but algorithmically dissimilar distractors. Our evaluation reveals that LLMs struggle to identify ASPs, with the best-performing model (o3-mini) achieving only 65.9% accuracy on the MCQ task. To address this challenge, we propose attempted solution matching (ASM), a novel method for improving problem similarity detection. On our MCQ task, ASM yields an absolute accuracy improvement of 6.7% to 11.7% across different models. We also evaluated code embedding models and retrieval methods on similar problem identification. While the adversarial selection of problems degrades the performance to be less than random, we found that simply summarizing the problem to remove narrative elements eliminates the effect, and combining ASM with a keyword-prioritized method, BM25, can yield up to 52.2% accuracy. Code and data are available at github.com

Spoken Language Models (SLMs) are designed to take speech inputs and produce spoken responses. However, current SLMs lack the ability to perform an internal, unspoken thinking process before responding. In contrast, humans typically engage in complex mental reasoning internally, enabling them to communicate ideas clearly and concisely. Thus, integrating an unspoken thought process into SLMs is highly desirable. While naively generating a complete chain-of-thought (CoT) reasoning before starting to talk can enable thinking for SLMs, this induces additional latency for the speech response, as the CoT reasoning can be arbitrarily long. To solve this issue, we propose Stitch, a novel generation method that alternates between the generation of unspoken reasoning chunks and spoken response chunks. Since the audio duration of a chunk of spoken response is much longer than the time to generate the tokens in a chunk of spoken response, we use the remaining free time to generate the unspoken reasoning tokens. When a chunk of audio is played to the user, the model continues to generate the next unspoken reasoning chunk, achieving simultaneous thinking and talking. Remarkably, Stitch matches the latency of baselines that cannot generate unspoken CoT by design while outperforming those baselines by 15% on math reasoning datasets; Stitch also performs equally well on non-reasoning datasets as those baseline models. Some animations and demonstrations are on the project page: https://d223302.github.io/STITCH.

The performance of Large Language Models (LLMs) is fundamentally determined by the contextual information provided during inference. This survey introduces Context Engineering, a formal discipline that transcends simple prompt design to encompass the systematic optimization of information payloads for LLMs. We present a comprehensive taxonomy decomposing Context Engineering into its foundational components and the sophisticated implementations that integrate them into intelligent systems. We first examine the foundational components: context retrieval and generation, context processing and context management. We then explore how these components are architecturally integrated to create sophisticated system implementations: retrieval-augmented generation (RAG), memory systems and tool-integrated reasoning, and multi-agent systems. Through this systematic analysis of over 1400 research papers, our survey not only establishes a technical roadmap for the field but also reveals a critical research gap: a fundamental asymmetry exists between model capabilities. While current models, augmented by advanced context engineering, demonstrate remarkable proficiency in understanding complex contexts, they exhibit pronounced limitations in generating equally sophisticated, long-form outputs. Addressing this gap is a defining priority for future research. Ultimately, this survey provides a unified framework for both researchers and engineers advancing context-aware AI.

We present Kwaipilot-AutoThink (KAT), an open-source 40B large language model developed to address the overthinking problem in reasoning-intensive tasks, where an automatic thinking training paradigm is proposed to dynamically switch between reasoning and non-reasoning modes based on task complexity. Specifically, first, we construct the dual-regime dataset based on a novel tagging pipeline and a multi-agent synthesis strategy, and then we apply Multi-Token Prediction (MTP)-enhanced knowledge distillation, enabling efficient and fine-grained reasoning transfer with minimal pretraining cost. Besides, we implement a cold-start initialization strategy that introduces mode-selection priors using majority-vote signals and intent-aware prompting. Finally, we propose Step-SRPO, a reinforcement learning algorithm that incorporates intermediate supervision into the GRPO framework, offering structured guidance over both reasoning-mode selection and response accuracy. Extensive experiments across multiple benchmarks demonstrate that KAT consistently matches or even outperforms current state-of-the-art models, including DeepSeek-R1-0528 and Qwen3-235B-A22B, across a wide range of reasoning-intensive tasks while reducing token usage. Notably, KAT outperforms all open-source models and even surpasses o3-mini on the leakage-controlled LiveCodeBench Pro. Beyond academic evaluation, KAT has been successfully deployed in Kwaipilot (i.e., Kuaishou's internal coding assistant), where it improves real-world development workflows with high accuracy, efficiency, and controllable reasoning behaviors. Moreover, we are actively training a 200B Mixture-of-Experts (MoE) model with 40B active parameters, and early results already show significant gains, further demonstrating the scalability of the AutoThink paradigm.