Efficient and accurate updating of knowledge stored in Large Language Models (LLMs) is one of the most pressing research challenges today. This paper presents Larimar - a novel, brain-inspired architecture for enhancing LLMs with a distributed episodic memory. Larimar's memory allows for dynamic, one-shot updates of knowledge without the need for computationally expensive re-training or fine-tuning. Experimental results on multiple fact editing benchmarks demonstrate that Larimar attains accuracy comparable to most competitive baselines, even in the challenging sequential editing setup, but also excels in speed - yielding speed-ups of 8-10x depending on the base LLM - as well as flexibility due to the proposed architecture being simple, LLM-agnostic, and hence general. We further provide mechanisms for selective fact forgetting, information leakage prevention, and input context length generalization with Larimar and show their effectiveness. Our code is available at https://github.com/IBM/larimar

Vision Language Models (VLMs) are rapidly advancing in their capability to answer information-seeking questions. As these models are widely deployed in consumer applications, they could lead to new privacy risks due to emergent abilities to identify people in photos, geolocate images, etc. As we demonstrate, somewhat surprisingly, current open-source and proprietary VLMs are very capable image geolocators, making widespread geolocation with VLMs an immediate privacy risk, rather than merely a theoretical future concern. As a first step to address this challenge, we develop a new benchmark, GPTGeoChat, to test the ability of VLMs to moderate geolocation dialogues with users. We collect a set of 1,000 image geolocation conversations between in-house annotators and GPT-4v, which are annotated with the granularity of location information revealed at each turn. Using this new dataset, we evaluate the ability of various VLMs to moderate GPT-4v geolocation conversations by determining when too much location information has been revealed. We find that custom fine-tuned models perform on par with prompted API-based models when identifying leaked location information at the country or city level; however, fine-tuning on supervised data appears to be needed to accurately moderate finer granularities, such as the name of a restaurant or building.

The Environmental Extended Multi-Regional Input-Output analysis is the predominant framework in Ecological Economics for assessing the environmental impact of economic activities. This paper introduces ExioML, the first Machine Learning benchmark dataset designed for sustainability analysis, aimed at lowering barriers and fostering collaboration between Machine Learning and Ecological Economics research. A crucial greenhouse gas emission regression task was conducted to evaluate sectoral sustainability and demonstrate the usability of the dataset. We compared the performance of traditional shallow models with deep learning models, utilizing a diverse Factor Accounting table and incorporating various categorical and numerical features. Our findings reveal that ExioML, with its high usability, enables deep and ensemble models to achieve low mean square errors, establishing a baseline for future Machine Learning research. Through ExioML, we aim to build a foundational dataset supporting various Machine Learning applications and promote climate actions and sustainable investment decisions.

As games challenge traditional automated white-box test generators, the Neatest approach generates test suites consisting of neural networks that exercise the source code by playing the games. Neatest generates these neural networks using an evolutionary algorithm that is guided by an objective function targeting individual source code statements. This approach works well if the objective function provides sufficient guidance, but deceiving or complex fitness landscapes may inhibit the search. In this paper, we investigate whether the issue of challenging fitness landscapes can be addressed by promoting novel behaviours during the search. Our case study on two Scratch games demonstrates that rewarding novel behaviours is a promising approach for overcoming challenging fitness landscapes, thus enabling future research on how to adapt the search algorithms to best use this information.

Abstract--In this paper we show the application of the new robotic multi-platform system HSURF to a specific use case of teleoperation, aimed at monitoring and inspection. The HSURF system, consists of 3 different kinds of platforms: floater, sinker and robotic fishes. The collaborative control of the 3 platforms allows a remotely based operator to control the fish in order to visit and inspect several targets underwater following a complex trajectory. A shared autonomy solution shows to be the most suitable, in order to minimize the effect of limited bandwidth and relevant delay intrinsic to acoustic communications. The control architecture is described and preliminary results of the acoustically teleoperated visits of multiple targets in a testing pool are provided.

Treating texts as images, combining prompts with textual labels for prompt tuning, and leveraging the alignment properties of CLIP have been successfully applied in zero-shot multi-label image recognition. Nonetheless, relying solely on textual labels to store visual information is insufficient for representing the diversity of visual objects. In this paper, we propose reversing the training process of CLIP and introducing the concept of Pseudo Visual Prompts. These prompts are initialized for each object category and pre-trained on large-scale, low-cost sentence data generated by large language models. This process mines the aligned visual information in CLIP and stores it in class-specific visual prompts. We then employ contrastive learning to transfer the stored visual information to the textual labels, enhancing their visual representation capacity. Additionally, we introduce a dual-adapter module that simultaneously leverages knowledge from the original CLIP and new learning knowledge derived from downstream datasets. Benefiting from the pseudo visual prompts, our method surpasses the state-of-the-art not only on clean annotated text data but also on pseudo text data generated by large language models.

The rapid advancement of Large Language Models (LLMs) and Large Multimodal Models (LMMs) has heightened the demand for AI-based scientific assistants capable of understanding scientific articles and figures. Despite progress, there remains a significant gap in evaluating models' comprehension of professional, graduate-level, and even PhD-level scientific content. Current datasets and benchmarks primarily focus on relatively simple scientific tasks and figures, lacking comprehensive assessments across diverse advanced scientific disciplines. To bridge this gap, we collected a multimodal, multidisciplinary dataset from open-access scientific articles published in Nature Communications journals. This dataset spans 72 scientific disciplines, ensuring both diversity and quality. We created benchmarks with various tasks and settings to comprehensively evaluate LMMs' capabilities in understanding scientific figures and content. Our evaluation revealed that these tasks are highly challenging: many open-source models struggled significantly, and even GPT-4V and GPT-4o faced difficulties. We also explored using our dataset as training resources by constructing visual instruction-following data, enabling the 7B LLaVA model to achieve performance comparable to GPT-4V/o on our benchmark. Additionally, we investigated the use of our interleaved article texts and figure images for pre-training LMMs, resulting in improvements on the material generation task. The source dataset, including articles, figures, constructed benchmarks, and visual instruction-following data, is open-sourced.

While Large Language Models (LLMs) excel in question-answering (QA) tasks, their multi-step reasoning abilities on multiple evidence integration on Multi-hop QA tasks remain underexplored. LLMs sometimes generate answers that rely on internal memory rather than reasoning given context, which brings concerns about the evaluation quality of real reasoning abilities. The counterfactual QA task can separate internal memory from reasoning abilities, but focusing solely on final-QA performance without evaluating the multi-step reasoning process is insufficient for reporting LLMs' real reasoning abilities. Current Multi-hop QA (MHQA) benchmarks are factual and annotated on open-source corpora such as Wikipedia, although useful for multi-step reasoning evaluation, showing limitations due to potential data contamination in LLMs pre-training stage. To address this issue, we introduce the Inherent Reasoning Evaluation (IRE) method, a novel evaluation way that jointly evaluates the LLMs' chain-of-reasoning performance based on the first knowledge-edited counterfactual multi-hop QA data which involves editing the original Wikipedia passages, reducing data contamination risks. The IRE comprehensively assesses reasoning chains through sub-QA and final-QA evaluations. Our comparisons reveal significant performance gaps for several LLMs between Wikipedia-based benchmarks and IRE, deeming data contamination issues in existing benchmarks. We believe that the IRE benchmark will enhance and facilitate trustworthy LLM evaluations.

Detecting and understanding out-of-distribution (OOD) samples is crucial in machine learning (ML) to ensure reliable model performance. Current OOD studies, in general, and in the context of ML, in particular, primarily focus on extrapolatory OOD (outside), neglecting potential cases of interpolatory OOD (inside). This study introduces a novel perspective on OOD by suggesting OOD can be divided into inside and outside cases. In addition, following this framework, we examine the inside-outside OOD profiles of datasets and their impact on ML model performance. Our analysis shows that different inside-outside OOD profiles lead to nuanced declines in ML model performance, highlighting the importance of distinguishing between these two cases for developing effective counter-OOD methods.

The emergence of large language models (LLMs) has transformed research and practice in a wide range of domains. Within the computing education research (CER) domain, LLMs have received plenty of attention especially in the context of learning programming. Much of the work on LLMs in CER has however focused on applying and evaluating proprietary models. In this article, we evaluate the efficiency of open-source LLMs in generating high-quality feedback for programming assignments, and in judging the quality of the programming feedback, contrasting the results against proprietary models. Our evaluations on a dataset of students' submissions to Python introductory programming exercises suggest that the state-of-the-art open-source LLMs (Meta's Llama3) are almost on-par with proprietary models (GPT-4o) in both the generation and assessment of programming feedback. We further demonstrate the efficiency of smaller LLMs in the tasks, and highlight that there are a wide range of LLMs that are accessible even for free for educators and practitioners.