MLLMs have been widely studied for video question answering recently. However, most existing assessments focus on natural videos, overlooking synthetic videos, such as AI-generated content (AIGC). Meanwhile, some works in video generation rely on MLLMs to evaluate the quality of generated videos, but the capabilities of MLLMs on interpreting AIGC videos remain largely underexplored. To address this, we propose a new benchmark, VF-Eval, which introduces four tasks-coherence validation, error awareness, error type detection, and reasoning evaluation-to comprehensively evaluate the abilities of MLLMs on AIGC videos. We evaluate 13 frontier MLLMs on VF-Eval and find that even the best-performing model, GPT-4.1, struggles to achieve consistently good performance across all tasks. This highlights the challenging nature of our benchmark. Additionally, to investigate the practical applications of VF-Eval in improving video generation, we conduct an experiment, RePrompt, demonstrating that aligning MLLMs more closely with human feedback can benefit video generation.

Transformer-based language models have demonstrated impressive capabilities across a range of complex reasoning tasks. Prior theoretical work exploring the expressive power of transformers has shown that they can efficiently perform multi-step reasoning tasks involving parallelizable computations. However, the learnability of such constructions, particularly the conditions on the data distribution that enable efficient learning via gradient-based optimization, remains an open question. Towards answering this question, in this work we study the learnability of the $k$-fold composition task, which requires computing an interleaved composition of $k$ input permutations and $k$ hidden permutations, and can be expressed by a transformer with $O(\log k)$ layers. On the negative front, we prove a Statistical Query (SQ) lower bound showing that any SQ learner that makes only polynomially-many queries to an SQ oracle for the $k$-fold composition task distribution must have sample size exponential in $k$, thus establishing a statistical-computational gap. On the other hand, we show that this function class can be efficiently learned, with runtime and sample complexity polynomial in $k$, by gradient descent on an $O(\log k)$-depth transformer via two different curriculum learning strategies: one in which data consists of $k'$-fold composition functions with $k' \le k$ presented in increasing difficulty, and another in which all such data is presented simultaneously. Our work sheds light on the necessity and sufficiency of having both easy and hard examples in the data distribution for transformers to learn complex compositional tasks.

Integrating structured information has long improved the quality of abstractive summarization, particularly in retaining salient content. In this work, we focus on a specific form of structure: argument roles, which are crucial for summarizing documents in high-stakes domains such as law. We investigate whether instruction-tuned large language models (LLMs) adequately preserve this information. To this end, we introduce Argument Representation Coverage (ARC), a framework for measuring how well LLM-generated summaries capture salient arguments. Using ARC, we analyze summaries produced by three open-weight LLMs in two domains where argument roles are central: long legal opinions and scientific articles. Our results show that while LLMs cover salient argument roles to some extent, critical information is often omitted in generated summaries, particularly when arguments are sparsely distributed throughout the input. Further, we use ARC to uncover behavioral patterns -- specifically, how the positional bias of LLM context windows and role-specific preferences impact the coverage of key arguments in generated summaries, emphasizing the need for more argument-aware summarization strategies.

Recently evolved large reasoning models (LRMs) show powerful performance in solving complex tasks with long chain-of-thought (CoT) reasoning capability. As these LRMs are mostly developed by post-training on formal reasoning tasks, whether they generalize the reasoning capability to help reduce hallucination in fact-seeking tasks remains unclear and debated. For instance, DeepSeek-R1 reports increased performance on SimpleQA, a fact-seeking benchmark, while OpenAI-o3 observes even severer hallucination. This discrepancy naturally raises the following research question: Are reasoning models more prone to hallucination? This paper addresses the question from three perspectives. (1) We first conduct a holistic evaluation for the hallucination in LRMs. Our analysis reveals that LRMs undergo a full post-training pipeline with cold start supervised fine-tuning (SFT) and verifiable reward RL generally alleviate their hallucination. In contrast, both distillation alone and RL training without cold start fine-tuning introduce more nuanced hallucinations. (2) To explore why different post-training pipelines alters the impact on hallucination in LRMs, we conduct behavior analysis. We characterize two critical cognitive behaviors that directly affect the factuality of a LRM: Flaw Repetition, where the surface-level reasoning attempts repeatedly follow the same underlying flawed logic, and Think-Answer Mismatch, where the final answer fails to faithfully match the previous CoT process. (3) Further, we investigate the mechanism behind the hallucination of LRMs from the perspective of model uncertainty. We find that increased hallucination of LRMs is usually associated with the misalignment between model uncertainty and factual accuracy. Our work provides an initial understanding of the hallucination in LRMs.

The model context protocol (MCP) has been widely adapted as an open standard enabling the seamless integration of generative AI agents. However, recent work has shown the MCP is susceptible to retrieval-based "falsely benign" attacks (FBAs), allowing malicious system access and credential theft, but requiring that users download compromised files directly to their systems. Herein, we show that the threat model of MCP-based attacks is significantly broader than previously thought, i.e., attackers need only post malicious content online to deceive MCP agents into carrying out their attacks on unsuspecting victims' systems. To improve alignment guardrails against such attacks, we introduce a new MCP dataset of FBAs and (truly) benign samples to explore the effectiveness of direct preference optimization (DPO) for the refusal training of large language models (LLMs). While DPO improves model guardrails against such attacks, we show that the efficacy of refusal learning varies drastically depending on the model's original post-training alignment scheme--e.g., GRPO-based LLMs learn to refuse extremely poorly. Thus, to further improve FBA refusals, we introduce Retrieval Augmented Generation for Preference alignment (RAG-Pref), a novel preference alignment strategy based on RAG. We show that RAG-Pref significantly improves the ability of LLMs to refuse FBAs, particularly when combined with DPO alignment, thus drastically improving guardrails against MCP-based attacks.

A significant portion of the textual data used in the field of Natural Language Processing (NLP) exhibits gender biases, particularly due to the use of masculine generics (masculine words that are supposed to refer to mixed groups of men and women), which can perpetuate and amplify stereotypes. Gender rewriting, an NLP task that involves automatically detecting and replacing gendered forms with neutral or opposite forms (e.g., from masculine to feminine), can be employed to mitigate these biases. While such systems have been developed in a number of languages (English, Arabic, Portuguese, German, French), automatic use of gender neutralization techniques (as opposed to inclusive or gender-switching techniques) has only been studied for English. This paper presents GeNRe, the very first French gender-neutral rewriting system using collective nouns, which are gender-fixed in French. We introduce a rule-based system (RBS) tailored for the French language alongside two fine-tuned language models trained on data generated by our RBS. We also explore the use of instruct-based models to enhance the performance of our other systems and find that Claude 3 Opus combined with our dictionary achieves results close to our RBS. Through this contribution, we hope to promote the advancement of gender bias mitigation techniques in NLP for French.

Language models (LMs) perform well on standardized coding benchmarks but struggle with real-world software engineering tasks such as resolving GitHub issues in SWE-Bench, especially when model parameters are less than 100B. While smaller models are preferable in practice due to their lower computational cost, improving their performance remains challenging. Existing approaches primarily rely on supervised fine-tuning (SFT) with high-quality data, which is expensive to curate at scale. An alternative is test-time scaling: generating multiple outputs, scoring them using a verifier, and selecting the best one. Although effective, this strategy often requires excessive sampling and costly scoring, limiting its practical application. We propose Evolutionary Test-Time Scaling (EvoScale), a sample-efficient method that treats generation as an evolutionary process. By iteratively refining outputs via selection and mutation, EvoScale shifts the output distribution toward higher-scoring regions, reducing the number of samples needed to find correct solutions. To reduce the overhead from repeatedly sampling and selection, we train the model to self-evolve using reinforcement learning (RL). Rather than relying on external verifiers at inference time, the model learns to self-improve the scores of its own generations across iterations. Evaluated on SWE-Bench-Verified, EvoScale enables our 32B model, Satori-SWE-32B, to match or exceed the performance of models with over 100B parameters while using a few samples. Code, data, and models will be fully open-sourced.

This paper investigates the ability of large language models (LLMs) to recognise and solve tasks which have been obfuscated beyond recognition. Focusing on competitive programming and benchmark tasks (LeetCode and MATH), we compare performance across multiple models and obfuscation methods, such as noise and redaction. We demonstrate that all evaluated LLMs can solve tasks obfuscated to a level where the text would be unintelligible to human readers, and does not contain key pieces of instruction or context. We introduce the concept of eager pattern matching to describe this behaviour, which is not observed in tasks published after the models' knowledge cutoff date, indicating strong memorisation or overfitting to training data, rather than legitimate reasoning about the presented problem. We report empirical evidence of distinct performance decay patterns between contaminated and unseen datasets. We discuss the implications for benchmarking and evaluations of model behaviour, arguing for caution when designing experiments using standard datasets. We also propose measuring the decay of performance under obfuscation as a possible strategy for detecting dataset contamination and highlighting potential safety risks and interpretability issues for automated software systems.

A restaurant dinner or a hotel stay may lead to memorable experiences when guests encounter unexpected aspects that also match their interests. For example, an origami-making station in the waiting area of a restaurant may be both surprising and enjoyable for a customer who is passionate about paper crafts. Similarly, an exhibit of 18th century harpsichords would be atypical for a hotel lobby and likely pique the interest of a guest who has a passion for Baroque music. Motivated by this insight, in this paper we introduce the new task of engineering serendipity through recommendations of items with atypical aspects. We describe an LLM-based system pipeline that extracts atypical aspects from item reviews, then estimates and aggregates their user-specific utility in a measure of serendipity potential that is used to rerank a list of items recommended to the user. To facilitate system development and evaluation, we introduce a dataset of Yelp reviews that are manually annotated with atypical aspects and a dataset of artificially generated user profiles, together with crowdsourced annotations of user-aspect utility values. Furthermore, we introduce a custom procedure for dynamic selection of in-context learning examples, which is shown to improve LLM-based judgments of atypicality and utility. Experimental evaluations show that serendipity-based rankings generated by the system are highly correlated with ground truth rankings for which serendipity scores are computed from manual annotations of atypical aspects and their user-dependent utility. Overall, we hope that the new recommendation task and the associated system presented in this paper catalyze further research into recommendation approaches that go beyond accuracy in their pursuit of enhanced user satisfaction. The datasets and the code are made publicly available at https://github.com/ramituncc49er/ATARS .

Recent advancements in large language model (LLM) agents have significantly accelerated scientific discovery automation, yet concurrently raised critical ethical and safety concerns. To systematically address these challenges, we introduce \textbf{SafeScientist}, an innovative AI scientist framework explicitly designed to enhance safety and ethical responsibility in AI-driven scientific exploration. SafeScientist proactively refuses ethically inappropriate or high-risk tasks and rigorously emphasizes safety throughout the research process. To achieve comprehensive safety oversight, we integrate multiple defensive mechanisms, including prompt monitoring, agent-collaboration monitoring, tool-use monitoring, and an ethical reviewer component. Complementing SafeScientist, we propose \textbf{SciSafetyBench}, a novel benchmark specifically designed to evaluate AI safety in scientific contexts, comprising 240 high-risk scientific tasks across 6 domains, alongside 30 specially designed scientific tools and 120 tool-related risk tasks. Extensive experiments demonstrate that SafeScientist significantly improves safety performance by 35\% compared to traditional AI scientist frameworks, without compromising scientific output quality. Additionally, we rigorously validate the robustness of our safety pipeline against diverse adversarial attack methods, further confirming the effectiveness of our integrated approach. The code and data will be available at https://github.com/ulab-uiuc/SafeScientist. \textcolor{red}{Warning: this paper contains example data that may be offensive or harmful.}