Amazon is still seen as a bit of a laggard in the race to develop advanced artificial intelligence, but it has quietly created a lab that is now setting records when it comes to AI performance. Amazon's AGI SF Lab, which is located in San Francisco and dedicated to building artificial general intelligence, or AI that surpasses the capabilities of humans, revealed the first fruits of its work today: A new AI model capable of powering some of the most advanced AI agents available anywhere. The new model, called Amazon Nova Act, outperforms ones from OpenAI and Anthropic on several benchmarks designed to gauge the intelligence and aptitude of AI agents, Amazon says. On the benchmarks GroundUI Web and ScreenSpot, Amazon Nova Act performs better than Claude 3.7 Sonnet and OpenAI Computer Use Agent. A major part of Amazon's plan to compete in the AI market is to focus on building agents, and the new model's abilities reflect its efforts to build a generation of tools that can measure up to the very best available.

OpenAI isn't slowing down when it comes to building extra functions and add-ons into its ChatGPT AI bot, and one of the newest features to roll out--exclusive to paying users, for now--is ChatGPT Projects. This is a major step forward for keeping conversations and data organized in ChatGPT: It gives you the ability to put your discussions with ChatGPT in separate spaces, like folders in a filing cabinet, complete with uploaded documents, web searches, custom instructions, and whatever else you've added. You can have one project for researching birthday present ideas, for example, and one for analyzing the current state of the movie industry. It's up to you how you use them, but Projects can make a genuine difference to workflows in ChatGPT. Projects can include conversations, files, and instructions.

We study the problem of building a visual concept library for visual recognition. Building effective visual concept libraries is challenging, as manual definition is labor-intensive, while relying solely on LLMs for concept generation can result in concepts that lack discriminative power or fail to account for the complex interactions between them. Our approach, ESCHER, takes a library learning perspective to iteratively discover and improve visual concepts. ESCHER uses a vision-language model (VLM) as a critic to iteratively refine the concept library, including accounting for interactions between concepts and how they affect downstream classifiers. By leveraging the in-context learning abilities of LLMs and the history of performance using various concepts, ESCHER dynamically improves its concept generation strategy based on the VLM critic's feedback. Finally, ESCHER does not require any human annotations, and is thus an automated plug-and-play framework. We empirically demonstrate the ability of ESCHER to learn a concept library for zero-shot, few-shot, and fine-tuning visual classification tasks. This work represents, to our knowledge, the first application of concept library learning to real-world visual tasks.

This study presents an empirical investigation into the energy consumption of Discriminative and Generative AI models within real-world MLOps pipelines. For Discriminative models, we examine various architectures and hyperparameters during training and inference and identify energy-efficient practices. For Generative AI, Large Language Models (LLMs) are assessed, focusing primarily on energy consumption across different model sizes and varying service requests. Our study employs software-based power measurements, ensuring ease of replication across diverse configurations, models, and datasets. We analyse multiple models and hardware setups to uncover correlations among various metrics, identifying key contributors to energy consumption. The results indicate that for Discriminative models, optimising architectures, hyperparameters, and hardware can significantly reduce energy consumption without sacrificing performance. For LLMs, energy efficiency depends on balancing model size, reasoning complexity, and request-handling capacity, as larger models do not necessarily consume more energy when utilisation remains low. This analysis provides practical guidelines for designing green and sustainable ML operations, emphasising energy consumption and carbon footprint reductions while maintaining performance. This paper can serve as a benchmark for accurately estimating total energy use across different types of AI models.

Most discussions about Large Language Model (LLM) safety have focused on single-agent settings but multi-agent LLM systems now create novel adversarial risks because their behavior depends on communication between agents and decentralized reasoning. In this work, we innovatively focus on attacking pragmatic systems that have constrains such as limited token bandwidth, latency between message delivery, and defense mechanisms. We design a $\textit{permutation-invariant adversarial attack}$ that optimizes prompt distribution across latency and bandwidth-constraint network topologies to bypass distributed safety mechanisms within the system. Formulating the attack path as a problem of $\textit{maximum-flow minimum-cost}$, coupled with the novel $\textit{Permutation-Invariant Evasion Loss (PIEL)}$, we leverage graph-based optimization to maximize attack success rate while minimizing detection risk. Evaluating across models including $\texttt{Llama}$, $\texttt{Mistral}$, $\texttt{Gemma}$, $\texttt{DeepSeek}$ and other variants on various datasets like $\texttt{JailBreakBench}$ and $\texttt{AdversarialBench}$, our method outperforms conventional attacks by up to $7\times$, exposing critical vulnerabilities in multi-agent systems. Moreover, we demonstrate that existing defenses, including variants of $\texttt{Llama-Guard}$ and $\texttt{PromptGuard}$, fail to prohibit our attack, emphasizing the urgent need for multi-agent specific safety mechanisms.

This paper introduces TaxoMatic, a framework that leverages large language models to automate definition extraction from academic literature. Focusing on the media bias domain, the framework encompasses data collection, LLM-based relevance classification, and extraction of conceptual definitions. Evaluated on a dataset of 2,398 manually rated articles, the study demonstrates the frameworks effectiveness, with Claude-3-sonnet achieving the best results in both relevance classification and definition extraction. Future directions include expanding datasets and applying TaxoMatic to additional domains.

Recent advancements in Chain of Thought (COT) generation have significantly improved the reasoning capabilities of Large Language Models (LLMs), with reinforcement learning (RL) emerging as an effective post-training approach. Multimodal Large Language Models (MLLMs) inherit this reasoning potential but remain underexplored in tasks requiring both perception and logical reasoning. To address this, we introduce SEED-Bench-R1, a benchmark designed to systematically evaluate post-training methods for MLLMs in video understanding. It includes intricate real-world videos and complex everyday planning tasks in the format of multiple-choice questions, requiring sophisticated perception and reasoning. SEED-Bench-R1 assesses generalization through a three-level hierarchy: in-distribution, cross-environment, and cross-environment-task scenarios, equipped with a large-scale training dataset with easily verifiable ground-truth answers. Using Qwen2-VL-Instruct-7B as a base model, we compare RL with supervised fine-tuning (SFT), demonstrating RL's data efficiency and superior performance on both in-distribution and out-of-distribution tasks, even outperforming SFT on general video understanding benchmarks like LongVideoBench. Our detailed analysis reveals that RL enhances visual perception but often produces less logically coherent reasoning chains. We identify key limitations such as inconsistent reasoning and overlooked visual cues, and suggest future improvements in base model reasoning, reward modeling, and RL robustness against noisy signals.

In e-commerce, behavioral data is collected for decision making which can be costly and slow. Simulation with LLM powered agents is emerging as a promising alternative for representing human population behavior. However, LLMs are known to exhibit certain biases, such as brand bias, review rating bias and limited representation of certain groups in the population, hence they need to be carefully benchmarked and aligned to user behavior. Ultimately, our goal is to synthesise an agent population and verify that it collectively approximates a real sample of humans. To this end, we propose a framework that: (i) creates synthetic shopping agents by automatically mining personas from anonymised historical shopping data, (ii) equips agents with retail-specific tools to synthesise shopping sessions and (iii) introduces a novel alignment suite measuring distributional differences between humans and shopping agents at the group (i.e. population) level rather than the traditional "individual" level. Experimental results demonstrate that using personas improves performance on the alignment suite, though a gap remains to human behaviour. We showcase an initial application of our framework for automated agentic A/B testing and compare the findings to human results. Finally, we discuss applications, limitations and challenges setting the stage for impactful future work.

Sparse autoencoders (SAEs) have emerged as a workhorse of modern mechanistic interpretability, but leading SAE approaches with top-$k$ style activation functions lack theoretical grounding for selecting the hyperparameter $k$. SAEs are based on the linear representation hypothesis (LRH), which assumes that the representations of large language models (LLMs) are linearly encoded, and the superposition hypothesis (SH), which states that there can be more features in the model than its dimensionality. We show that, based on the formal definitions of the LRH and SH, the magnitude of sparse feature vectors (the latent representations learned by SAEs of the dense embeddings of LLMs) can be approximated using their corresponding dense vector with a closed-form error bound. To visualize this, we propose the ZF plot, which reveals a previously unknown relationship between LLM hidden embeddings and SAE feature vectors, allowing us to make the first empirical measurement of the extent to which feature vectors of pre-trained SAEs are over- or under-activated for a given input. Correspondingly, we introduce Approximate Feature Activation (AFA), which approximates the magnitude of the ground-truth sparse feature vector, and propose a new evaluation metric derived from AFA to assess the alignment between inputs and activations. We also leverage AFA to introduce a novel SAE architecture, the top-AFA SAE, leading to SAEs that: (a) are more in line with theoretical justifications; and (b) obviate the need to tune SAE sparsity hyperparameters. Finally, we empirically demonstrate that top-AFA SAEs achieve reconstruction loss comparable to that of state-of-the-art top-k SAEs, without requiring the hyperparameter $k$ to be tuned. Our code is available at: https://github.com/SewoongLee/top-afa-sae.

The ACPBench dataset provides atomic reasoning tasks required for efficient planning. The dataset is aimed at distilling the complex plan generation task into separate atomic reasoning tasks in their easiest possible form, boolean or multiple-choice questions, where the model has to choose the right answer from the provided options. While the aim of ACP-Bench is to test the simplest form of reasoning about action and change, when tasked with planning, a model does not typically have options to choose from and thus the reasoning required for planning dictates an open-ended, generative form for these tasks. To that end, we introduce ACPBench Hard, a generative version of ACPBench, with open-ended questions which the model needs to answer. Models that perform well on these tasks could in principle be integrated into a planner or be used directly as a policy. We discuss the complexity of these tasks as well as the complexity of validating the correctness of their answers and present validation algorithms for each task. Equipped with these validators, we test the performance of a variety of models on our tasks and find that for most of these tasks the performance of even the largest models is still subpar. Our experiments show that no model outperforms another in these tasks and with a few exceptions all tested language models score below 65%, indicating that even the current frontier language models have a long way to go before they can reliably reason about planning. ACPBench Hard collection is available at the following link: https://ibm.github.io/ACPBench. Introduction The ability to reason and plan is the cornerstone of artificial intelligence. With the introduction of large language models, a major focus in the field is on testing their abilities in these two fields, reasoning and planning. For reasoning, the majority of work focuses on the mathematical reasoning (Cobbe et al. 2021) and logical inference (Saparov and He 2023). For planning, most work focused on the ability to produce or validate a plan (V almeekam et al. 2023a; Stein et al. 2024). To tackle this gap, recent work introduced an ACPBench dataset (Kokel et al. 2025), a benchmark for testing the reasoning abilities about action, change, and planning, separating the planning process into the atomic reasoning tasks performed by planners.