We investigate the abilities of 28 Large language Models (LLMs) to reason about cardinal directions (CDs) using a benchmark generated from a set of templates, extensively testing an LLM's ability to determine the correct CD given a particular scenario. The templates allow for a number of degrees of variation such as means of locomotion of the agent involved, and whether set in the first, second or third person. Even the newer Large Reasoning Models are unable to reliably determine the correct CD for all questions. This paper summarises and extends earlier work presented at COSIT-24.

The following paper introduces a general linguistic creativity test for humans and Large Language Models (LLMs). The test consists of various tasks aimed at assessing their ability to generate new original words and phrases based on word formation processes (derivation and compounding) and on metaphorical language use. We administered the test to 24 humans and to an equal number of LLMs, and we automatically evaluated their answers using OCSAI tool for three criteria: Originality, Elaboration, and Flexibility. The results show that LLMs not only outperformed humans in all the assessed criteria, but did better in six out of the eight test tasks. We then computed the uniqueness of the individual answers, which showed some minor differences between humans and LLMs. Finally, we performed a short manual analysis of the dataset, which revealed that humans are more inclined towards E(extending)-creativity, while LLMs favor F(ixed)-creativity.

The proliferation of software vulnerabilities presents a significant challenge to cybersecurity, necessitating more effective detection methodologies. We introduce White-Basilisk, a novel approach to vulnerability detection that demonstrates superior performance while challenging prevailing assumptions in AI model scaling. Utilizing an innovative architecture that integrates Mamba layers, linear self-attention, and a Mixture of Experts framework, White-Basilisk achieves state-of-the-art results in vulnerability detection tasks with a parameter count of only 200M. The model's capacity to process sequences of unprecedented length enables comprehensive analysis of extensive codebases in a single pass, surpassing the context limitations of current Large Language Models (LLMs). White-Basilisk exhibits robust performance on imbalanced, real-world datasets, while maintaining computational efficiency that facilitates deployment across diverse organizational scales. This research not only establishes new benchmarks in code security but also provides empirical evidence that compact, efficiently designed models can outperform larger counterparts in specialized tasks, potentially redefining optimization strategies in AI development for domain-specific applications.

In this paper, we introduce ECom-Bench, the first benchmark framework for evaluating LLM agent with multimodal capabilities in the e-commerce customer support domain. ECom-Bench features dynamic user simulation based on persona information collected from real e-commerce customer interactions and a realistic task dataset derived from authentic e-commerce dialogues. These tasks, covering a wide range of business scenarios, are designed to reflect real-world complexities, making ECom-Bench highly challenging. For instance, even advanced models like GPT-4o achieve only a 10-20% pass^3 metric in our benchmark, highlighting the substantial difficulties posed by complex e-commerce scenarios. The code and data have been made publicly available at https://github.com/XiaoduoAILab/ECom-Bench to facilitate further research and development in this domain.

Prompt injection attack has been listed as the top-1 security threat to LLM-integrated applications, which interact with external environment data for complex tasks. The untrusted data may contain an injected prompt trying to arbitrarily manipulate the system. Model-level prompt injection defenses have shown strong effectiveness, but are currently deployed into commercial-grade models in a closed-source manner. We believe open-source secure models are needed by the AI security community, where co-development of attacks and defenses through open research drives scientific progress in mitigating prompt injection attacks. To this end, we develop Meta SecAlign, the first fully open-source LLM with built-in model-level defense that achieves commercial-grade performance, powerful enough for complex agentic tasks. We provide complete details of our training recipe, an improved version of the SOTA SecAlign defense. We perform the most comprehensive evaluation to date on 9 utility benchmarks and 7 security benchmarks on general knowledge, instruction following, and agentic workflows. Results show that Meta SecAlign, despite being trained on generic instruction-tuning samples only, surprisingly confers security in unseen downstream tasks, including tool-calling and web-navigation, in addition to general instruction-following. Our best model -- Meta-SecAlign-70B -- establishes a new frontier of utility-security trade-off for open-source LLMs. Even compared to closed-course commercial models such as GPT-5, our model is much securer than most of them. Below are links for the code (https://github.com/facebookresearch/Meta_SecAlign), Meta-SecAlign-70B(https://huggingface.co/facebook/Meta-SecAlign-70B), and Meta-SecAlign-8B(https://huggingface.co/facebook/Meta-SecAlign-8B) models.

The task of reconstructing unknown textual inputs to language models is a fundamental auditing primitive that allows us to assess the model's vulnerability to a range of security issues, including stealing hidden system prompts, detecting backdoors, and leaking private data. Existing inversion works assume access to differing levels of information (e.g. requiring input-output examples, the model parameters, intermediate activations or output logits) but oftentimes fail to fully reconstruct the desired input. In this paper, we present the Sparse One-hot Discrete Adam (SODA) algorithm, a search-based inversion method that can accurately reconstruct the input text, given white-box access to the language model and its output. Our experiments demonstrate for the first time that exact language model inversion is possible on both natural language and random inputs. Indeed, SODA achieves respectively 98% and 79% reconstruction rates on inputs with lengths up to 10 tokens. Furthermore, we show that input length and vocabulary size have a far greater impact on the probability of a successful reconstruction than the size of the language model itself, thus allowing us to scale to models from 33M to 3B parameters.

Despite progress in Large Vision-Language Models (LVLMs), their capacity for visual reasoning is often limited by the binding problem: the failure to reliably associate perceptual features with their correct visual referents. This limitation underlies persistent errors in tasks such as counting, visual search, scene description, and spatial relationship understanding. A key factor is that current LVLMs process visual features largely in parallel, lacking mechanisms for spatially grounded, serial attention. This paper introduces Visual Input Structure for Enhanced Reasoning (VISER), a simple, effective method that augments visual inputs with low-level spatial structures and pairs them with a textual prompt that encourages sequential, spatially-aware parsing. We empirically demonstrate substantial performance improvements across core visual reasoning tasks, using only a single-query inference. Specifically, VISER improves GPT-4o performance on visual search, counting, and spatial relationship tasks by 25.0%, 26.8%, and 9.5%, respectively, and reduces edit distance error in scene description by 0.32 on 2D datasets. Furthermore, we find that the visual modification is essential for these gains; purely textual strategies, including Chain-of-Thought prompting, are insufficient and can even degrade performance. VISER underscores the importance of visual input design over purely linguistically based reasoning strategies and suggests that visual structuring is a powerful and general approach for enhancing compositional and spatial reasoning in LVLMs.

In many real-world applications, evaluating the quality of instances is costly and time-consuming, e.g., human feedback and physics simulations, in contrast to proposing new instances. In particular, this is even more critical in reinforcement learning, since it relies on interactions with the environment (i.e., new instances) that must be evaluated to provide a reward signal for learning. At the same time, performing sufficient exploration is crucial in reinforcement learning to find high-rewarding solutions, meaning that the agent should observe and learn from a diverse set of experiences to find different solutions. Thus, we argue that learning from a diverse mini-batch of experiences can have a large impact on the exploration and help mitigate mode collapse. In this paper, we introduce mini-batch diversification for reinforcement learning and study this framework in the context of a real-world problem, namely, drug discovery. We extensively evaluate how our proposed framework can enhance the effectiveness of chemical exploration in de novo drug design, where finding diverse and high-quality solutions is crucial. Our experiments demonstrate that our proposed diverse mini-batch selection framework can substantially enhance the diversity of solutions while maintaining high-quality solutions. In drug discovery, such an outcome can potentially lead to fulfilling unmet medical needs faster.

Time series forecasting aims to model temporal dependencies among variables for future state inference, holding significant importance and widespread applications in real-world scenarios. Although deep learning-based methods have achieved remarkable progress, they still exhibit suboptimal performance in long-term forecasting. Recent research demonstrates that large language models (LLMs) achieve promising performance in time series forecasting, but this progress is still met with skepticism about whether LLMs are truly useful for this task. To address this, we propose Time-Prompt, a framework for activating LLMs for time series forecasting. Specifically, we first construct a unified prompt paradigm with learnable soft prompts to guide the LLM's behavior and textualized hard prompts to enhance the time series representations. Second, to enhance LLM' comprehensive understanding of the forecasting task, we design a semantic space embedding and cross-modal alignment module to achieve fusion of temporal and textual data. Finally, we efficiently fine-tune the LLM's parameters using time series data. Furthermore, we focus on carbon emissions, aiming to provide a modest contribution to global carbon neutrality. Comprehensive evaluations on 6 public datasets and 3 carbon emission datasets demonstrate that Time-Prompt is a powerful framework for time series forecasting.

Critical Learning Periods comprehend an important phenomenon involving deep learning, where early epochs play a decisive role in the success of many training recipes, such as data augmentation. Existing works confirm the existence of this phenomenon and provide useful insights. However, the literature lacks efforts to precisely identify when critical periods occur. In this work, we fill this gap by introducing a systematic approach for identifying critical periods during the training of deep neural networks, focusing on eliminating computationally intensive regularization techniques and effectively applying mechanisms for reducing computational costs, such as data pruning. Our method leverages generalization prediction mechanisms to pinpoint critical phases where training recipes yield maximum benefits to the predictive ability of models. By halting resource-intensive recipes beyond these periods, we significantly accelerate the learning phase and achieve reductions in training time, energy consumption, and CO$_2$ emissions. Experiments on standard architectures and benchmarks confirm the effectiveness of our method. Specifically, we achieve significant milestones by reducing the training time of popular architectures by up to 59.67%, leading to a 59.47% decrease in CO$_2$ emissions and a 60% reduction in financial costs, without compromising performance. Our work enhances understanding of training dynamics and paves the way for more sustainable and efficient deep learning practices, particularly in resource-constrained environments. In the era of the race for foundation models, we believe our method emerges as a valuable framework. The repository is available at https://github.com/baunilhamarga/critical-periods