Mark Zuckerberg and Meta are spending billions of dollars for top talent to make up ground in the generative artificial intelligence race, sparking doubt about the wisdom of the spree. OpenAI boss Sam Altman recently lamented that Meta has offered 100 million bonuses to engineers who jump to Zuckerberg's ship, where hefty salaries await. A few OpenAI employees have reportedly taken Meta up on the offer, joining Scale AI founder and former chief executive Alexandr Wang at the Menlo Park-based tech titan.

Recent works have highlighted the significance of memory mechanisms in LLM-based agents, which enable them to store observed information and adapt to dynamic environments. However, evaluating their memory capabilities still remains challenges. Previous evaluations are commonly limited by the diversity of memory levels and interactive scenarios. They also lack comprehensive metrics to reflect the memory capabilities from multiple aspects. To address these problems, in this paper, we construct a more comprehensive dataset and benchmark to evaluate the memory capability of LLM-based agents. Our dataset incorporates factual memory and reflective memory as different levels, and proposes participation and observation as various interactive scenarios. Based on our dataset, we present a benchmark, named MemBench, to evaluate the memory capability of LLM-based agents from multiple aspects, including their effectiveness, efficiency, and capacity. To benefit the research community, we release our dataset and project at https://github.com/import-myself/Membench.

Current evaluation frameworks for multimodal generative AI struggle to establish trustworthiness, hindering enterprise adoption where reliability is paramount. We introduce a systematic, quantitative benchmarking framework to measure the trustworthiness of progressively integrating cross-modal inputs such as text, images, captions, and OCR within VisualRAG systems for enterprise document intelligence. Our approach establishes quantitative relationships between technical metrics and user-centric trust measures. Evaluation reveals that optimal modality weighting with weights of 30% text, 15% image, 25% caption, and 30% OCR improves performance by 57.3% over text-only baselines while maintaining computational efficiency. We provide comparative assessments of foundation models, demonstrating their differential impact on trustworthiness in caption generation and OCR extraction-a vital consideration for reliable enterprise AI. This work advances responsible AI deployment by providing a rigorous framework for quantifying and enhancing trustworthiness in multimodal RAG for critical enterprise applications.

We propose a novel approach for detecting personal data in structured datasets, leveraging GPT-4o, a state-of-the-art Large Language Model. A key innovation of our method is the incorporation of contextual information: in addition to a feature's name and values, we utilize information from other feature names within the dataset as well as the dataset description. We compare our approach to alternative methods, including Microsoft Presidio and CASSED, evaluating them on multiple datasets: DeSSI, a large synthetic dataset, datasets we collected from Kaggle and OpenML as well as MIMIC-Demo-Ext, a real-world dataset containing patient information from critical care units. Our findings reveal that detection performance varies significantly depending on the dataset used for evaluation. CASSED excels on DeSSI, the dataset on which it was trained. Performance on the medical dataset MIMIC-Demo-Ext is comparable across all models, with our GPT-4o-based approach clearly outperforming the others. Notably, personal data detection in the Kaggle and OpenML datasets appears to benefit from contextual information. This is evidenced by the poor performance of CASSED and Presidio (both of which do not utilize the context of the dataset) compared to the strong results of our GPT-4o-based approach. We conclude that further progress in this field would greatly benefit from the availability of more real-world datasets containing personal information.

Large language models (LLMs) have demonstrated their strong capabilities in various domains, and have been recently integrated for graph analysis as graph language models (GLMs). With LLMs as the predictor, some GLMs can interpret unseen tasks described by natural language, and learn from a few examples in the prompts without parameter tuning, known as in-context learning (ICL). Another subset of GLMs utilizes abundant training labels to enhance model performance, known as instruction tuning. However, we argue that ICL on graphs has effectiveness issues due to fixed parameters and efficiency issues due to long context. Meanwhile, the large amount of labeled data required for instruction tuning can be difficult to obtain in real-world scenarios. To this end, we aim to introduce an extra parameter adaptation stage that can efficiently tailor GLMs to an unseen graph and task with only a few labeled examples, in exchange for better prediction accuracy and faster inference speed. For implementation, in this paper we propose GraphLAMA method, with its model backbone and learning schemes specialized for efficient tuning and inference. Specifically, for model backbone, we use a graph neural network (GNN) with several well-designed components to transform nodes into the representation space of LLM tokens. Task instructions can then be represented as a mixture of node and language tokens. In the pre-training stage, model parameters except the LLM will be trained with different tasks to capture general knowledge. In the adaptation stage, only a few pre-trained parameters will be updated based on few-shot examples. Extensive experiments on few/zero-shot node classification and summary generation show that our proposed GraphLAMA achieves state-of-the-art performance with 4.91% absolution improvement in accuracy. Compared with ICL, our inference speed can be 10 times faster under 5-shot setting.

Generating realistic 3D indoor scenes from user inputs remains a challenging problem in computer vision and graphics, requiring careful balance of geometric consistency, spatial relationships, and visual realism. While neural generation methods often produce repetitive elements due to limited global spatial reasoning, procedural approaches can leverage constraints for controllable generation but struggle with multi-constraint scenarios. When constraints become numerous, object collisions frequently occur, forcing the removal of furniture items and compromising layout completeness. To address these limitations, we propose RoomCraft, a multi-stage pipeline that converts real images, sketches, or text descriptions into coherent 3D indoor scenes. Our approach combines a scene generation pipeline with a constraint-driven optimization framework. The pipeline first extracts high-level scene information from user inputs and organizes it into a structured format containing room type, furniture items, and spatial relations. It then constructs a spatial relationship network to represent furniture arrangements and generates an optimized placement sequence using a heuristic-based depth-first search (HDFS) algorithm to ensure layout coherence. To handle complex multi-constraint scenarios, we introduce a unified constraint representation that processes both formal specifications and natural language inputs, enabling flexible constraint-oriented adjustments through a comprehensive action space design. Additionally, we propose a Conflict-Aware Positioning Strategy (CAPS) that dynamically adjusts placement weights to minimize furniture collisions and ensure layout completeness. Extensive experiments demonstrate that RoomCraft significantly outperforms existing methods in generating realistic, semantically coherent, and visually appealing room layouts across diverse input modalities.

Recent Multimodal Large Language Models (MLLMs) have demonstrated strong performance in visual grounding, establishing themselves as a general interface for various vision-language applications. This progress has driven the development of token pruning methods to mitigate the high computational costs associated with processing numerous visual tokens. However, we observe that pruning significantly weakens the model's grounding ability, leading to incorrect predictions and drastic performance degradation. In Referring Expression Comprehension (REC), for instance, pruning causes the accuracy of LLaVA on the RefCOCO validation set to drop from 56.14% to 15.34%. Our analysis identifies misaligned position IDs after pruning as the primary cause of this degradation, as both the order and value of these IDs are crucial for maintaining performance in grounding tasks. To address this issue, we propose Grounding-Aware Token Pruning (GAP), a simple yet effective adjustment to position IDs that recovers REC accuracy back to 51.42%, which is 90% of the original performance in the without pruning setting, all while requiring no additional training, memory, or computational overhead. Applied to models such as Shikra, MiniGPTv2, and the LLaVA series, our method consistently improves performance across various token pruning strategies.

A growing body of work has been querying LLMs with political questions to evaluate their potential biases. However, this probing method has limited stability, making comparisons between models unreliable. In this paper, we argue that LLMs need more context. We propose a new probing task, Questionnaire Modeling (QM), that uses human survey data as in-context examples. We show that QM improves the stability of question-based bias evaluation, and demonstrate that it may be used to compare instruction-tuned models to their base versions. Experiments with LLMs of various sizes indicate that instruction tuning can indeed change the direction of bias. Furthermore, we observe a trend that larger models are able to leverage in-context examples more effectively, and generally exhibit smaller bias scores in QM. Data and code are publicly available.

The AI for Nuclear Energy workshop at Oak Ridge National Laboratory evaluated the potential of Large Language Models (LLMs) to accelerate fusion and fission research. Fourteen interdisciplinary teams explored diverse nuclear science challenges using ChatGPT, Gemini, Claude, and other AI models over a single day. Applications ranged from developing foundation models for fusion reactor control to automating Monte Carlo simulations, predicting material degradation, and designing experimental programs for advanced reactors. Teams employed structured workflows combining prompt engineering, deep research capabilities, and iterative refinement to generate hypotheses, prototype code, and research strategies. Key findings demonstrate that LLMs excel at early-stage exploration, literature synthesis, and workflow design, successfully identifying research gaps and generating plausible experimental frameworks. However, significant limitations emerged, including difficulties with novel materials designs, advanced code generation for modeling and simulation, and domain-specific details requiring expert validation. The successful outcomes resulted from expert-driven prompt engineering and treating AI as a complementary tool rather than a replacement for physics-based methods. The workshop validated AI's potential to accelerate nuclear energy research through rapid iteration and cross-disciplinary synthesis while highlighting the need for curated nuclear-specific datasets, workflow automation, and specialized model development. These results provide a roadmap for integrating AI tools into nuclear science workflows, potentially reducing development cycles for safer, more efficient nuclear energy systems while maintaining rigorous scientific standards.

Inference accounts for the majority of latency and energy consumption in large language model (LLM) deployments, often exceeding 90% of total cost. While training-time efficiency has seen extensive progress, runtime optimization remains a key bottleneck, particularly under autoregressive decoding. Existing approaches -- such as pruning, quantization, early exits, and speculative decoding -- often require retraining, architectural changes, or disrupt decoding compatibility. We introduce QuickSilver, a modular, token-level framework that enables semantic adaptivity at inference time without altering model weights or structure. QuickSilver integrates four synergistic mechanisms: (i) Dynamic Token Halting, which halts computation for tokens with converged representations; (ii) KV Cache Skipping, which selectively suppresses memory writes to reduce attention overhead; and (iii) Contextual Token Fusion, which collapses redundant tokens into shared paths to shrink sequence length. Unlike speculative decoding or MoE routing, QuickSilver operates entirely on frozen, dense models and requires no auxiliary networks. Applied to GPT-2 and Llama-2 across WikiText-103 and C4, QuickSilver achieves up to 39.6% FLOP reduction with negligible perplexity degradation (<=0.2).