Scaffolding Large Language Models (LLMs) into multi-agent systems often improves performance on complex tasks, but the safety impact of such scaffolds has not been as thoroughly explored. In this paper, we introduce AGENTBREEDER a framework for multi-objective evolutionary search over scaffolds. Our REDAGENTBREEDER evolves scaffolds towards jailbreaking the base LLM while achieving high task success, while BLUEAGENTBREEDER instead aims to combine safety with task reward. We evaluate the systems discovered by the different instances of AGENTBREEDER and popular baselines using widely recognized reasoning, mathematics, and safety benchmarks. Our work highlights and mitigates the safety risks due to multi-agent scaffolding.

This paper investigates the utilization of Large Language Models (LLMs) for solving complex linguistic puzzles, a domain requiring advanced reasoning and adept translation capabilities akin to human cognitive processes. We explore specific prompting techniques designed to enhance ability of LLMs to reason and elucidate their decision-making pathways, with a focus on Input-Output Prompting (IO), Chain-of-Thought Prompting (CoT), and Solo Performance Prompting (SPP). Utilizing datasets from the Puzzling Machine Competition and various Linguistics Olympiads, we employ a comprehensive set of metrics to assess the performance of GPT-4 0603, a prominent LLM, across these prompting methods. Our findings illuminate the potential of LLMs in linguistic reasoning and complex translation tasks, highlighting their capabilities and identifying limitations in the context of linguistic puzzles. This research contributes significantly to the broader field of Natural Language Processing (NLP) by providing insights into the optimization of LLM applications for improved reasoning and translation accuracy, thereby enriching the ongoing dialogue in NLP advancements.

This study addresses the growing concern of information asymmetry in consumer contracts, exacerbated by the proliferation of online services with complex Terms of Service that are rarely even read. Even though research on automatic analysis methods is conducted, the problem is aggravated by the general focus on English-language Machine Learning approaches and on major jurisdictions, such as the European Union. We introduce a new methodology and a substantial dataset addressing this gap. We propose a novel annotation scheme with four categories and a total of 20 classes, and apply it on 50 online Terms of Service used in Chile. Our evaluation of transformer-based models highlights how factors like language- and/or domain-specific pre-training, few-shot sample size, and model architecture affect the detection and classification of potentially abusive clauses. Results show a large variability in performance for the different tasks and models, with the highest macro-F1 scores for the detection task ranging from 79% to 89% and micro-F1 scores up to 96%, while macro-F1 scores for the classification task range from 60% to 70% and micro-F1 scores from 64% to 80%. Notably, this is the first Spanish-language multi-label classification dataset for legal clauses, applying Chilean law and offering a comprehensive evaluation of Spanish-language models in the legal domain. Our work lays the ground for future research in method development for rarely considered legal analysis and potentially leads to practical applications to support consumers in Chile and Latin America as a whole.

Reinforcement Learning from Human Feedback (RLHF) has achieved considerable success in aligning large language models (LLMs) by modeling human preferences with a learnable reward model and employing a reinforcement learning algorithm to maximize the reward model's scores. However, these reward models are susceptible to exploitation through various superficial confounding factors, with length bias emerging as a particularly significant concern. Moreover, while the pronounced impact of length bias on preference modeling suggests that LLMs possess an inherent sensitivity to length perception, our preliminary investigations reveal that fine-tuned LLMs consistently struggle to adhere to explicit length instructions. To address these two limitations, we propose a novel framework wherein the reward model explicitly differentiates between human semantic preferences and response length requirements. Specifically, we introduce a Response-conditioned Bradley-Terry (Rc-BT) model that enhances the reward model's capability in length bias mitigating and length instruction following, through training on our augmented dataset. Furthermore, we propose the Rc-DPO algorithm to leverage the Rc-BT model for direct policy optimization (DPO) of LLMs, simultaneously mitigating length bias and promoting adherence to length instructions. Extensive evaluations demonstrate that our approach substantially improves both preference modeling and length instruction compliance, with its effectiveness validated across various foundational models and preference datasets.

Two prominent features of large language models (LLMs) is the presence of large-norm (outlier) features and the tendency for tokens to attend very strongly to a select few tokens. Despite often having no semantic relevance, these select tokens, called attention sinks, along with the large outlier features, have proven important for model performance, compression, and streaming. Consequently, investigating the roles of these phenomena within models and exploring how they might manifest in the model parameters has become an area of active interest. Through an empirical investigation, we demonstrate that attention sinks utilize outlier features to: catch a sequence of tokens, tag the captured tokens by applying a common perturbation, and then release the tokens back into the residual stream, where the tagged tokens are eventually retrieved. We prove that simple tasks, like averaging, necessitate the 'catch, tag, release' mechanism hence explaining why it would arise organically in modern LLMs. Our experiments also show that the creation of attention sinks can be completely captured in the model parameters using low-rank matrices, which has important implications for model compression and substantiates the success of recent approaches that incorporate a low-rank term to offset performance degradation.

The rapid advancement of large language models (LLMs) has led to a surge in both model supply and application demands. To facilitate effective matching between them, reliable, generic and efficient benchmark generators are widely needed. However, human annotators are constrained by inefficiency, and current LLM benchmark generators not only lack generalizability but also struggle with limited reliability, as they lack a comprehensive evaluation framework for validation and optimization. To fill this gap, we first propose an automated and unbiased evaluation framework, structured around four dimensions and ten criteria. Under this framework, we carefully analyze the advantages and weaknesses of directly prompting LLMs as generic benchmark generators. To enhance the reliability, we introduce a series of methods to address the identified weaknesses and integrate them as BenchMaker. Experiments across multiple LLMs and tasks confirm that BenchMaker achieves superior or comparable performance to human-annotated benchmarks on all metrics, highlighting its generalizability and reliability. More importantly, it delivers highly consistent evaluation results across 12 LLMs (0.967 Pearson correlation against MMLU-Pro), while taking only $0.005 and 0.38 minutes per sample.

Scientific data visualization is pivotal for transforming raw data into comprehensible visual representations, enabling pattern recognition, forecasting, and the presentation of data-driven insights. However, novice users often face difficulties due to the complexity of selecting appropriate tools and mastering visualization techniques. Large Language Models (LLMs) have recently demonstrated potential in assisting code generation, though they struggle with accuracy and require iterative debugging. In this paper, we propose PlotGen, a novel multi-agent framework aimed at automating the creation of precise scientific visualizations. PlotGen orchestrates multiple LLM-based agents, including a Query Planning Agent that breaks down complex user requests into executable steps, a Code Generation Agent that converts pseudocode into executable Python code, and three retrieval feedback agents - a Numeric Feedback Agent, a Lexical Feedback Agent, and a Visual Feedback Agent - that leverage multimodal LLMs to iteratively refine the data accuracy, textual labels, and visual correctness of generated plots via self-reflection. Extensive experiments show that PlotGen outperforms strong baselines, achieving a 4-6 percent improvement on the MatPlotBench dataset, leading to enhanced user trust in LLM-generated visualizations and improved novice productivity due to a reduction in debugging time needed for plot errors.

Recently, zeroth-order (ZO) optimization plays an essential role in scenarios where gradient information is inaccessible or unaffordable, such as black-box systems and resource-constrained environments. While existing adaptive methods such as ZO-AdaMM have shown promise, they are fundamentally limited by their underutilization of moment information during optimization, usually resulting in underperforming convergence. To overcome these limitations, this paper introduces Refined Adaptive Zeroth-Order Optimization (R-AdaZO). Specifically, we first show the untapped variance reduction effect of first moment estimate on ZO gradient estimation, which improves the accuracy and stability of ZO updates. We then refine the second moment estimate based on these variance-reduced gradient estimates to better capture the geometry of the optimization landscape, enabling a more effective scaling of ZO updates. We present rigorous theoretical analysis to show (I) the first analysis to the variance reduction of first moment estimate in ZO optimization, (II) the improved second moment estimates with a more accurate approximation of its variance-free ideal, (III) the first variance-aware convergence framework for adaptive ZO methods, which may be of independent interest, and (IV) the faster convergence of R-AdaZO than existing baselines like ZO-AdaMM. Our extensive experiments, including synthetic problems, black-box adversarial attack, and memory-efficient fine-tuning of large language models (LLMs), further verify the superior convergence of R-AdaZO, indicating that R-AdaZO offers an improved solution for real-world ZO optimization challenges.

One-step shortcut diffusion models [Frans, Hafner, Levine and Abbeel, ICLR 2025] have shown potential in vision generation, but their reliance on first-order trajectory supervision is fundamentally limited. The Shortcut model's simplistic velocity-only approach fails to capture intrinsic manifold geometry, leading to erratic trajectories, poor geometric alignment, and instability-especially in high-curvature regions. These shortcomings stem from its inability to model mid-horizon dependencies or complex distributional features, leaving it ill-equipped for robust generative modeling. In this work, we introduce HOMO (High-Order Matching for One-Step Shortcut Diffusion), a game-changing framework that leverages high-order supervision to revolutionize distribution transportation. By incorporating acceleration, jerk, and beyond, HOMO not only fixes the flaws of the Shortcut model but also achieves unprecedented smoothness, stability, and geometric precision. Theoretically, we prove that HOMO's high-order supervision ensures superior approximation accuracy, outperforming first-order methods. Empirically, HOMO dominates in complex settings, particularly in high-curvature regions where the Shortcut model struggles. Our experiments show that HOMO delivers smoother trajectories and better distributional alignment, setting a new standard for one-step generative models.

The advancement of large language models (LLMs) has opened new frontiers in natural language processing, particularly in specialized domains like healthcare. In this paper, we propose the Incremental Curriculum-Based Fine-Tuning (ICFT) framework to enhance the generative capabilities of medical large language models (MLLMs). ICFT combines curriculum-based learning, dual-stage memory coordination, and parameter-efficient fine-tuning to enable a progressive transition from general linguistic knowledge to strong domain-specific expertise. Experimental results across diverse medical NLP tasks, including question answering, preference classification, and response generation, demonstrate that ICFT consistently outperforms state-of-the-art baselines, achieving improvements in both accuracy and efficiency. Further analysis reveals the framework's ability to generalize to unseen data, reduce errors, and deliver diverse, contextually relevant medical responses. These findings establish ICFT as a robust and scalable solution for adapting LLMs to the medical domain, offering practical benefits for real-world healthcare applications.