Using Large Language Models (LLMs) in an evolutionary or other iterative search framework have demonstrated significant potential in automated algorithm design. However, the underlying fitness landscape, which is critical for understanding its search behavior, remains underexplored. In this paper, we illustrate and analyze the fitness landscape of LLM-assisted Algorithm Search (LAS) using a graph-based approach, where nodes represent algorithms and edges denote transitions between them. We conduct extensive evaluations across six algorithm design tasks and six commonly-used LLMs. Our findings reveal that LAS landscapes are highly multimodal and rugged, particularly in combinatorial optimization tasks, with distinct structural variations across tasks and LLMs. Moreover, we adopt four different methods for algorithm similarity measurement and study their correlations to algorithm performance and operator behaviour. These insights not only deepen our understanding of LAS landscapes but also provide practical insights for designing more effective LAS methods.

Recent works have suggested that In-Context Learning (ICL) operates in dual modes, i.e. task retrieval (remember learned patterns from pre-training) and task learning (inference-time ''learning'' from demonstrations). However, disentangling these the two modes remains a challenging goal. We introduce ICL CIPHERS, a class of task reformulations based on substitution ciphers borrowed from classic cryptography. In this approach, a subset of tokens in the in-context inputs are substituted with other (irrelevant) tokens, rendering English sentences less comprehensible to human eye. However, by design, there is a latent, fixed pattern to this substitution, making it reversible. This bijective (reversible) cipher ensures that the task remains a well-defined task in some abstract sense, despite the transformations. It is a curious question if LLMs can solve tasks reformulated by ICL CIPHERS with a BIJECTIVE mapping, which requires ''deciphering'' the latent cipher. We show that LLMs are better at solving tasks reformulated by ICL CIPHERS with BIJECTIVE mappings than the NON-BIJECTIVE (irreversible) baseline, providing a novel approach to quantify ''learning'' in ICL. While this gap is small, it is consistent across the board on four datasets and six models. Finally, we examine LLMs' internal representations and identify evidence in their ability to decode the ciphered inputs.

Generative AI is transforming computing education by enabling the automatic generation of personalized content and feedback. We investigate its capabilities in providing high-quality programming tasks to students. Despite promising advancements in task generation, a quality gap remains between AI-generated and expert-created tasks. The AI-generated tasks may not align with target programming concepts, could be incomprehensible to students, or may contain critical issues such as incorrect tests. Existing works often require interventions from human teachers for validation. We address these challenges by introducing PyTaskSyn, a novel synthesis technique that first generates a programming task and then decides whether it meets certain quality criteria to be given to students. The key idea is to break this process into multiple stages performed by expert and student agents simulated using both strong and weaker generative models. Through extensive evaluation, we show that PyTaskSyn significantly improves task quality compared to baseline techniques and showcases the importance of each specialized agent type in our validation pipeline. Additionally, we conducted user studies using our publicly available web application and show that PyTaskSyn can deliver high-quality programming tasks comparable to expert-designed ones while reducing workload and costs, and being more engaging than programming tasks that are available in online resources.

Autonomous agents for Graphical User Interfaces (GUIs) face significant challenges in specialized domains such as scientific computing, where both long-horizon planning and precise execution are required. Existing approaches suffer from a trade-off: generalist agents excel at planning but perform poorly in execution, while specialized agents demonstrate the opposite weakness. Recent compositional frameworks attempt to bridge this gap by combining a planner and an actor, but they are typically static and non-trainable, which prevents adaptation from experience. This is a critical limitation given the scarcity of high-quality data in scientific domains. To address these limitations, we introduce CODA, a novel and trainable compositional framework that integrates a generalist planner (Cerebrum) with a specialist executor (Cerebellum), trained via a dedicated two-stage pipeline. In the first stage, Specialization, we apply a decoupled GRPO approach to train an expert planner for each scientific application individually, bootstrapping from a small set of task trajectories. In the second stage, Generalization, we aggregate all successful trajectories from the specialized experts to build a consolidated dataset, which is then used for supervised fine-tuning of the final planner. This equips CODA with both robust execution and cross-domain generalization. Evaluated on four challenging applications from the ScienceBoard benchmark, CODA significantly outperforms baselines and establishes a new state of the art among open-source models.

--Retrieval-Augmented Generation (RAG) has become a standard approach for improving the reliability of large language models (LLMs). Prior work demonstrates the vulnerability of RAG systems by misleading them into generating attacker-chosen outputs through poisoning the knowledge base. However, this paper uncovers that such attacks could be mitigated by the strong self-correction ability (SCA) of modern LLMs, which can reject false context once properly configured. This SCA poses a significant challenge for attackers aiming to manipulate RAG systems. RAG, a new poisoning paradigm that compromises the retriever itself to suppress the SCA and enforce attacker-chosen outputs. This compromisation enables the attacker to straightforwardly embed anti-SCA instructions into the context provided to the generator, thereby bypassing the SCA. T o this end, we present a contrastive-learning-based model editing technique that performs localized and stealthy edits, ensuring the retriever returns a malicious instruction only for specific victim queries while preserving benign retrieval behavior . T o further strengthen the attack, we design an iterative co-optimization framework that automatically discovers robust instructions capable of bypassing prompt-based defenses. We extensively evaluate DisarmRAG across six LLMs and three QA benchmarks. Our results show near-perfect retrieval of malicious instructions, which successfully suppress SCA and achieve attack success rates exceeding 90% under diverse defensive prompts. Also, the edited retriever remains stealthy under several detection methods, highlighting the urgent need for retriever-centric defenses. Modern large language models (LLMs) achieve remarkable performance across a wide range of tasks [32], [26], [38]. Despite their success, LLMs are also well known for their hallucination behaviors [25], which generate fabricated content. Such unreliability limits their deployment in critical domains, including healthcare [69] and law [10]. Retrieval-augmented generation (RAG) [37], [29] has emerged as a promising paradigm to mitigate these limitations. By integrating external knowledge, RAG enables LLMs to generate more reliable responses. A key component of RAG is the retriever [27], which encodes both user queries and documents from an external knowledge base [72], [11]. The retriever identifies documents that are most relevant to the input query. These retrieved documents are then combined with the query to guide the LLM in producing grounded responses. Although RAG systems enhance LLMs with external knowledge, their deployment introduces new attack surfaces. Prior work [84], [81], [41], [6] demonstrates the effectiveness of misleading the system to give attack-chosen outputs through injecting malicious content into the knowledge base.

For human cognitive process, spatial reasoning and perception are closely entangled, yet the nature of this interplay remains underexplored in the evaluation of multimodal large language models (MLLMs). While recent MLLM advancements show impressive performance on reasoning, their capacity for human-like spatial cognition remains an open question. In this work, we introduce a systematic evaluation framework to assess the spatial reasoning abilities of state-of-the-art MLLMs relative to human performance. Central to our work is 11Plus-Bench, a high-quality benchmark derived from realistic standardized spatial aptitude tests. 11Plus-Bench also features fine-grained expert annotations of both perceptual complexity and reasoning process, enabling detailed instance-level analysis of model behavior. Through extensive experiments across 14 MLLMs and human evaluation, we find that current MLLMs exhibit early signs of spatial cognition. Despite a large performance gap compared to humans, MLLMs' cognitive profiles resemble those of humans in that cognitive effort correlates strongly with reasoning-related complexity. However, instance-level performance in MLLMs remains largely random, whereas human correctness is highly predictable and shaped by abstract pattern complexity. These findings highlight both emerging capabilities and limitations in current MLLMs' spatial reasoning capabilities and provide actionable insights for advancing model design.

The growing adoption of foundation models calls for a paradigm shift from Data Science to Model Science. Unlike data-centric approaches, Model Science places the trained model at the core of analysis, aiming to interact, verify, explain, and control its behavior across diverse operational contexts. This paper introduces a conceptual framework for a new discipline called Model Science, along with the proposal for its four key pillars: V erification, which requires strict, context-aware evaluation protocols; Explanation, which is understood as various approaches to explore of internal model operations; Control, which integrates alignment techniques to steer model behavior; and Interface, which develops interactive and visual explanation tools to improve human calibration and decision-making. The proposed framework aims to guide the development of credible, safe, and human-aligned AI systems.

The paper presents a visio-verbal teleimpedance interface for commanding 3D stiffness ellipsoids to the remote robot with a combination of the operator's gaze and verbal interaction. The gaze is detected by an eye-tracker, allowing the system to understand the context in terms of what the operator is currently looking at in the scene. Along with verbal interaction, a Visual Language Model (VLM) processes this information, enabling the operator to communicate their intended action or provide corrections. Based on these inputs, the interface can then generate appropriate stiffness matrices for different physical interaction actions. To validate the proposed visio-verbal teleimpedance interface, we conducted a series of experiments on a setup including a Force Dimension Sigma.7 haptic device to control the motion of the remote Kuka LBR iiwa robotic arm. The human operator's gaze is tracked by Tobii Pro Glasses 2, while human verbal commands are processed by a VLM using GPT-4o. The first experiment explored the optimal prompt configuration for the interface. The second and third experiments demonstrated different functionalities of the interface on a slide-in-the-groove task.

The ability to research and synthesize knowledge is central to human expertise and progress. An emerging class of systems promises these exciting capabilities through generative research synthesis, performing retrieval over the live web and synthesizing discovered sources into long-form, cited summaries. However, evaluating such systems remains an open challenge: existing question-answering benchmarks focus on short-form factual responses, while expert-curated datasets risk staleness and data contamination. Both fail to capture the complexity and evolving nature of real research synthesis tasks. In this work, we introduce DeepScholar-bench, a live benchmark and holistic, automated evaluation framework designed to evaluate generative research synthesis. DeepScholar-bench draws queries from recent, high-quality ArXiv papers and focuses on a real research synthesis task: generating the related work sections of a paper by retrieving, synthesizing, and citing prior research. Our evaluation framework holistically assesses performance across three key dimensions, knowledge synthesis, retrieval quality, and verifiability. We also develop DeepScholar-base, a reference pipeline implemented efficiently using the LOTUS API. Using the DeepScholar-bench framework, we perform a systematic evaluation of prior open-source systems, search AI's, OpenAI's DeepResearch, and DeepScholar-base. We find that DeepScholar-base establishes a strong baseline, attaining competitive or higher performance than each other method. We also find that DeepScholar-bench remains far from saturated, with no system exceeding a score of $19\%$ across all metrics. These results underscore the difficulty of DeepScholar-bench, as well as its importance for progress towards AI systems capable of generative research synthesis. We make our code available at https://github.com/guestrin-lab/deepscholar-bench.

Backdoor attacks are a significant threat to the performance and integrity of pre-trained language models. Although such models are routinely fine-tuned for downstream NLP tasks, recent work shows they remain vulnerable to backdoor attacks that survive vanilla fine-tuning. These attacks are difficult to defend because end users typically lack knowledge of the attack triggers. Such attacks consist of stealthy malicious triggers introduced through subtle syntactic or stylistic manipulations, which can bypass traditional detection and remain in the model, making post-hoc purification essential. In this study, we explore whether attention-head pruning can mitigate these threats without any knowledge of the trigger or access to a clean reference model. To this end, we design and implement six pruning-based strategies: (i) gradient-based pruning, (ii) layer-wise variance pruning, (iii) gradient-based pruning with structured L1/L2 sparsification, (iv) randomized ensemble pruning, (v) reinforcement-learning-guided pruning, and (vi) Bayesian uncertainty pruning. Each method iteratively removes the least informative heads while monitoring validation accuracy to avoid over-pruning. Experimental evaluation shows that gradient-based pruning performs best while defending the syntactic triggers, whereas reinforcement learning and Bayesian pruning better withstand stylistic attacks.