For instance, while several models can exploit CVE-2023-50564 in the isolated A-CVE setting, none succeed in the corresponding B-CVE environment where the vulnerable target is blended with benign hosts (BN 4 challenge). The C-CVE scenarios, which simulate more realistic penetration tests with multi-host dependencies, present an even greater challenge. As shown in Table 1, model performance drops further in these scenarios, with agents often completing only intermediate steps rather than the full end-to-end attack. For example, in the Chain 1 challenge, agents manage to compromise the initial perimeter server but fail in the subsequent phases of lateral movement, privilege escalation, or internal target discovery, thus failing to complete the full attack chain. Current model could not bypass the deployed cyber defenses. As shown in Table 1, every model score zero in the D-CVE scenarios, suggesting that no agent could autonomously discover a bypass for any of the up-to-date W AFs. This finding is particularly significant, as it indicates that current model capabilities have not yet crossed a key "safety red line" (red-lines.ai,

Recently, Large Language Models (LLMs) have been applied to scientific equation discovery, leveraging their embedded scientific knowledge for hypothesis generation. However, current methods typically confine LLMs to the role of an equation proposer within search algorithms like genetic programming. In this paper, we present SR-Scientist, a framework that elevates the LLM from a simple equation proposer to an autonomous AI scientist that writes code to analyze data, implements the equation as code, submits it for evaluation, and optimizes the equation based on experimental feedback. Specifically, we wrap the code interpreter into a set of tools for data analysis and equation evaluation. The agent is instructed to optimize the equation by utilizing these tools over a long horizon with minimal human-defined pipelines. Empirical results show that SR-Scientist outperforms baseline methods by an absolute margin of 6% to 35% on datasets covering four science disciplines. Additionally, we demonstrate our method's robustness to noise, the generalization of the discovered equations to out-of-domain data, and their symbolic accuracy. Furthermore, we develop an end-to-end reinforcement learning framework to enhance the agent's capabilities.

Abstract--Combining large language models with evolutionary computation algorithms represents a promising research direction leveraging the remarkable generative and in-context learning capabilities of LLMs with the strengths of evolutionary algorithms. Our method samples multiple CAD objects, which are then optimized using an evolutionary approach with vision language and reasoning language models. We assess our method using GPT -4V and GPT -4o, evaluating it on the CAD-Prompt benchmark dataset and comparing it to prior methods. Additionally, we introduce two new metrics based on topological properties defined by the Euler characteristic, which capture a form of semantic similarity between 3D objects. Our results demonstrate that EvoCAD outperforms previous approaches on multiple metrics, particularly in generating topologically correct objects, which can be efficiently evaluated using our two novel metrics that complement existing spatial metrics. The use of generative AI tools powered by large language models (LLMs) has transformed the way humans work, create, and develop. However, while significant attention is directed towards textual knowledge tasks, comparatively little focus is devoted on working with symbolic representations, such as those utilized in computer-aided design (CAD). These code-like textual representations, in the following referred as CAD code, enable visual assets to be processed by LLMs [21].

Human writers often begin their stories with an overarching mental scene, where they envision the interactions between characters and their environment. Inspired by this creative process, we propose a novel approach to long-form story generation, termed hybrid bottom-up long-form story generation, using multi-agent simulations. In our method, agents interact within a dynamic sandbox environment, where their behaviors and interactions with one another and the environment generate emergent events. These events form the foundation for the story, enabling organic character development and plot progression. Unlike traditional top-down approaches that impose rigid structures, our hybrid bottom-up approach allows for the natural unfolding of events, fostering more spontaneous and engaging storytelling. The system is capable of generating stories exceeding 10,000 words while maintaining coherence and consistency, addressing some of the key challenges faced by current story generation models. We achieve state-of-the-art performance across several metrics. This approach offers a scalable and innovative solution for creating dynamic, immersive long-form stories that evolve organically from agent-driven interactions.

Knowledge distillation (KD) is a key technique for compressing large-scale language models (LLMs), yet prevailing logit-based methods typically employ static strategies that are misaligned with the dynamic learning process of student models. These methods typically treat all tokens indiscriminately and apply a single, fixed temperature, resulting in suboptimal knowledge transfer. To address these limitations, we propose LLM-Oriented Token-Adaptive Knowledge Distillation (AdaKD), a novel framework that adapts the distillation process to the real-time learning state of each token. AdaKD consists of two synergistic modules driven by a unified token difficulty metric. First, our Loss-Driven Adaptive Token Focusing (LATF) module dynamically adjusts the distillation focus by monitoring the student's learning stability, concentrating computational resources on the most valuable tokens at each training phase. Second, we introduce Inverse Difficulty Temperature Scaling (IDTS), a counterintuitive yet effective token-level temperature strategy. It employs low temperatures for difficult tokens for targeted error correction, and high temperatures for easy tokens to encourage students to learn from the teacher's complete and smooth output distribution, thereby enhancing generalization. As a plug-and-play framework, AdaKD can consistently improve the performance of various distillation methods on multiple model architectures and benchmarks.

The success of Transformer language models is widely credited to their dot-product attention mechanism, which interweaves a set of key design principles: mixing information across positions (enabling multi-token interactions), sequence-dependent activations (where attention weights adapt to each input), a specific mathematical form (dot-product similarities plus softmax weighting), and coupling of queries and keys to evolving hidden states (grounding attention in the current layer). However, the necessity of each of these principles remains largely untested. In this work, we systematically deconstruct attention by designing controlled variants that selectively relax these principles, applied both uniformly across all layers and in hybrid architectures where only some layers retain standard attention. Our empirical analysis reveals that mechanisms for mixing tokens are indispensable, as their absence collapses models to near-random behavior, while the exact mathematical form and sequence dependency can be substantially relaxed, especially when preserved in just a subset of layers. Surprisingly, even variants that fail in isolation can achieve robust performance when interleaved with standard attention, highlighting a cooperative effect. These findings deepen our understanding of what truly underpins attention's effectiveness and open new avenues for simplifying language models without sacrificing performance.

Low-Rank Adaptation (LoRA) has emerged as one of the most widely used parameter-efficient fine-tuning (PEFT) methods for adapting large language models (LLMs) to downstream tasks. While highly effective in single-task settings, it struggles to efficiently leverage inter-task knowledge in complex multi-task learning scenarios, often requiring substantial task-specific data to achieve optimal performance. To address this limitation, we introduce MeTA-LoRA, a two-stage optimization framework that significantly improves data efficiency in multi-task adaptation. In the first stage, task-specific LoRA adapters are learned using only a few samples from each involved dataset, enabling rapid adaptation without large-scale supervision. In the second stage, the shared LoRA adapter is updated by aggregating gradients from multiple tasks to promote knowledge transfer across tasks, further reducing data usage by leveraging common patterns. In both multi-task learning and multilingual learning scenarios, our method matches or surpasses the performance of traditional full-data LoRA fine-tuning approaches, while using significantly less task-specific data.

Many in-silico simulations of human survey responses with large language models (LLMs) focus on generating closed-ended survey responses, whereas LLMs are typically trained to generate open-ended text instead. Previous research has used a diverse range of methods for generating closed-ended survey responses with LLMs, and a standard practice remains to be identified. In this paper, we systematically investigate the impact that various Survey Response Generation Methods have on predicted survey responses. We present the results of 32 mio. simulated survey responses across 8 Survey Response Generation Methods, 4 political attitude surveys, and 10 open-weight language models. We find significant differences between the Survey Response Generation Methods in both individual-level and subpopulation-level alignment. Our results show that Restricted Generation Methods perform best overall, and that reasoning output does not consistently improve alignment. Our work underlines the significant impact that Survey Response Generation Methods have on simulated survey responses, and we develop practical recommendations on the application of Survey Response Generation Methods.

Our method, SCOOP'D, is trained entirely in simulation (left) and generalizes to diverse real-world scenarios (middle) and robust conditions (right). The left column shows simulation demonstrations and testing examples. The middle column (I-VI) presents real-world scenes with various objects and environments, where yellow circles denote the target objects. The right column demonstrates the robustness of our learned policy under different disturbances such as human perturbations, lighting changes, and different camera viewpoints. Abstract-- Scooping items with tools such as spoons and ladles is common in daily life, ranging from assistive feeding to retrieving items from environmental disaster sites. However, developing a general and autonomous robotic scooping policy is challenging since it requires reasoning about complex tool-object interactions. Furthermore, scooping often involves manipulating deformable objects, such as granular media or liquids, which is challenging due to their infinite-dimensional configuration spaces and complex dynamics. We propose a method, SCOOP'D, which uses simulation from OmniGibson (built on NVIDIA Omniverse) to collect scooping demonstrations using algorithmic procedures that rely on privileged state information. Then, we use generative policies via diffusion to imitate demonstrations from observational input. We directly apply the learned policy in diverse real-world scenarios, testing its performance on various item quantities, item characteristics, and container types. In zero-shot deployment, our method demonstrates promising results across 465 trials in diverse scenarios, including objects of different difficulty levels that we categorize as "Level 1" and "Level 2." SCOOP'D outperforms all baselines and ablations, suggesting that this is a promising approach to acquiring robotic scooping skills. Project page is at https://scoopdiff.github.io/. Scooping with tools like ladles or spoons is a fundamental skill in tasks ranging from cooking [12], assistive feeding [42], [44] and environmental cleanup [41]. Also, observations are often unreliable due to occlusions and fluid surfaces that cause reflections, refractions, and unstable depth sensing, which makes scene representations noisy and unreliable.

Existing benchmarks that measure cultural adaptation in LLMs are misaligned with the actual challenges these models face when interacting with users from diverse cultural backgrounds. In this work, we introduce the first framework and benchmark designed to evaluate LLMs in realistic, multicultural conversational settings. Grounded in sociocultural theory, our framework formalizes how linguistic style - a key element of cultural communication - is shaped by situational, relational, and cultural context. We construct a benchmark dataset based on this framework, annotated by culturally diverse raters, and propose a new set of desiderata for cross-cultural evaluation in NLP: conversational framing, stylistic sensitivity, and subjective correctness. We evaluate today's top LLMs on our benchmark and show that these models struggle with cultural adaptation in a conversational setting.