As AI writing tools become widespread, we need to understand how both humans and machines evaluate literary style, a domain where objective standards are elusive and judgments are inherently subjective. We conducted controlled experiments using Raymond Queneau's Exercises in Style (1947) to measure attribution bias across evaluators. Study 1 compared human participants (N=556) and AI models (N=13) evaluating literary passages from Queneau versus GPT-4-generated versions under three conditions: blind, accurately labeled, and counterfactually labeled. Study 2 tested bias generalization across a 14$\times$14 matrix of AI evaluators and creators. Both studies revealed systematic pro-human attribution bias. Humans showed +13.7 percentage point (pp) bias (Cohen's h = 0.28, 95% CI: 0.21-0.34), while AI models showed +34.3 percentage point bias (h = 0.70, 95% CI: 0.65-0.76), a 2.5-fold stronger effect (P$<$0.001). Study 2 confirmed this bias operates across AI architectures (+25.8pp, 95% CI: 24.1-27.6%), demonstrating that AI systems systematically devalue creative content when labeled as "AI-generated" regardless of which AI created it. We also find that attribution labels cause evaluators to invert assessment criteria, with identical features receiving opposing evaluations based solely on perceived authorship. This suggests AI models have absorbed human cultural biases against artificial creativity during training. Our study represents the first controlled comparison of attribution bias between human and artificial evaluators in aesthetic judgment, revealing that AI systems not only replicate but amplify this human tendency.

The increasing adoption of LLM agents with access to numerous tools and sensitive data significantly widens the attack surface for indirect prompt injections. Due to the context-dependent nature of attacks, however, current defenses are often ill-calibrated as they cannot reliably differentiate malicious and benign instructions, leading to high false positive rates that prevent their real-world adoption. To address this, we present a novel approach inspired by the fundamental principle of computer security: data should not contain executable instructions. Instead of sample-level classification, we propose a token-level sanitization process, which surgically removes any instructions directed at AI systems from tool outputs, capturing malicious instructions as a byproduct. In contrast to existing safety classifiers, this approach is non-blocking, does not require calibration, and is agnostic to the context of tool outputs. Further, we can train such token-level predictors with readily available instruction-tuning data only, and don't have to rely on unrealistic prompt injection examples from challenges or of other synthetic origin. In our experiments, we find that this approach generalizes well across a wide range of attacks and benchmarks like AgentDojo, BIPIA, InjecAgent, ASB and SEP, achieving a 7-10x reduction of attack success rate (ASR) (34% to 3% on AgentDojo), without impairing agent utility in both benign and malicious settings.

When learning to code, students often develop misconceptions about various programming language concepts. These can not only lead to bugs or inefficient code, but also slow down the learning of related concepts. In this paper, we introduce McMining, the task of mining programming misconceptions from samples of code from a student. To enable the training and evaluation of McMining systems, we develop an extensible benchmark dataset of misconceptions together with a large set of code samples where these misconceptions are manifested. We then introduce two LLM-based McMiner approaches and through extensive evaluations show that models from the Gemini, Claude, and GPT families are effective at discovering misconceptions in student code.

Large language models (LLMs) have demonstrated impressive reasoning abilities yet remain unreliable on knowledge-intensive, multi-hop questions -- they miss long-tail facts, hallucinate when uncertain, and their internal knowledge lags behind real-world change. Knowledge graphs (KGs) offer a structured source of relational evidence, but existing KGQA methods face fundamental trade-offs: compiling complete SPARQL queries without knowing available relations proves brittle, retrieving large subgraphs introduces noise, and complex agent frameworks with parallel exploration exponentially expand search spaces. To address these limitations, we propose Search-on-Graph (SoG), a simple yet effective framework that enables LLMs to perform iterative informed graph navigation using a single, carefully designed \textsc{Search} function. Rather than pre-planning paths or retrieving large subgraphs, SoG follows an ``observe-then-navigate'' principle: at each step, the LLM examines actual available relations from the current entity before deciding on the next hop. This approach further adapts seamlessly to different KG schemas and handles high-degree nodes through adaptive filtering. Across six KGQA benchmarks spanning Freebase and Wikidata, SoG achieves state-of-the-art performance without fine-tuning. We demonstrate particularly strong gains on Wikidata benchmarks (+16\% improvement over previous best methods) alongside consistent improvements on Freebase benchmarks.

Video large language models (Vid-LLMs), which excel in diverse video-language tasks, can be effectively constructed by adapting image-pretrained vision-language models (VLMs). However, this adaptation remains challenging, as it requires processing dense and temporally extended visual inputs that exceed the capacity of image-based models. This paper identifies the perception bottleneck and token overload as key challenges in extending image-based VLMs to the video domain. To address these issues, we propose D-CoDe, a training-free adaptation framework that incorporates dynamic compression and question decomposition. Specifically, dynamic compression alleviates the perception bottleneck through adaptive selection of representative frames and content-aware aggregation of spatial tokens, thereby reducing redundancy while preserving informative content. In parallel, question decomposition mitigates token overload by reformulating the original query into sub-questions, guiding the model to focus on distinct aspects of the video and enabling more comprehensive understanding. Experiments demonstrate that D-CoDe effectively improves video understanding across various benchmarks. Furthermore, strong performance on the challenging long-video benchmark highlights the potential of D-CoDe in handling complex video-language tasks. Code is available at https://github.com/hukcc/D-CoDe.

Large Language Models (LLMs) are increasingly deployed across multilingual applications that handle sensitive data, yet their scale and linguistic variability introduce major privacy risks. Mostly evaluated for English, this paper investigates how language structure affects privacy leakage in LLMs trained on English, Spanish, French, and Italian medical corpora. We quantify six linguistic indicators and evaluate three attack vectors: extraction, counterfactual memorization, and membership inference. Results show that privacy vulnerability scales with linguistic redundancy and tokenization granularity: Italian exhibits the strongest leakage, while English shows higher membership separability. In contrast, French and Spanish display greater resilience due to higher morphological complexity. Overall, our findings provide the first quantitative evidence that language matters in privacy leakage, underscoring the need for language-aware privacy-preserving mechanisms in LLM deployments.

Unlike general-purpose coding, scientific workflows require algorithms that are rigorous, interconnected with deep domain knowledge, and incorporate domain-specific reasoning, as well as algorithm iteration without requiring I/O test cases. Many scientific problems also require a sequence of subproblems to be solved, leading to the final desired result. MOSAIC is designed as a training-free framework with specially designed agents to self-reflect, create the rationale, code, and debug within a student-teacher paradigm to address the challenges of scientific code generation. This design facilitates stepwise problem decomposition, targeted error correction, and, when combined with our Consolidated Context Window (CCW), mitigates LLM hallucinations when solving complex scientific tasks involving chained subproblems. We evaluate MOSAIC on scientific coding benchmarks and demonstrate that our specialized agentic framework outperforms existing approaches in terms of accuracy, robustness, and interpretability.

Policy design for various systems controllers has conventionally been a manual process, with domain experts carefully tailoring heuristics for the specific instance in which the policy will be deployed. In this paper, we re-imagine policy design via a novel automated search technique fueled by recent advances in generative models, specifically Large Language Model (LLM)-driven code generation. We outline the design and implementation of PolicySmith, a framework that applies LLMs to synthesize instance-optimal heuristics. We apply PolicySmith to two long-standing systems policies - web caching and congestion control, highlighting the opportunities unraveled by this LLM-driven heuristic search. For caching, PolicySmith discovers heuristics that outperform established baselines on standard open-source traces. For congestion control, we show that PolicySmith can generate safe policies that integrate directly into the Linux kernel.

While Large Language Models (LLMs) have demonstrated advanced reasoning capabilities, their comprehensive evaluation in general Chinese-language contexts remains understudied. To bridge this gap, we propose Chinese Commonsense Multi-hop Reasoning (CCMOR), a novel benchmark designed to evaluate LLMs' ability to integrate Chinese-specific factual knowledge with multi-step logical reasoning. Specifically, we first construct a domain-balanced seed set from existing QA datasets, then develop an LLM-powered pipeline to generate multi-hop questions anchored on factual unit chains. To ensure the quality of resulting dataset, we implement a human-in-the-loop verification system, where domain experts systematically validate and refine the generated questions. Using CCMOR, we evaluate state-of-the-art LLMs, demonstrating persistent limitations in LLMs' ability to process long-tail knowledge and execute knowledge-intensive reasoning. Notably, retrieval-augmented generation substantially mitigates these knowledge gaps, yielding significant performance gains.

Transformer attention scales quadratically with sequence length O(n^2), limiting long-context use. We propose Adaptive Retention, a probabilistic, layer-wise token selection mechanism that learns which representations to keep under a strict global budget M. Retention is modeled with Bernoulli gates trained via a Hard-Concrete/variational relaxation and enforced with a simple top-M rule at inference, making the method differentiable and drop-in for standard encoders. Across classification, extractive QA, and long-document summarization, keeping only 30-50% of tokens preserves >= 95% of full-model performance while cutting peak memory by ~35-45% and improving throughput by up to ~1.8x. This architecture-agnostic approach delivers practical long-context efficiency without modifying base attention or task heads.