Previous work finds that recent long-context language models fail to make equal use of information in the middle of their inputs, preferring pieces of information located at the tail ends which creates an undue bias in situations where we would like models to be equally capable of using different parts of the input. Thus far, the problem has mainly only been considered in settings with single pieces of critical information, leading us to question what happens when multiple necessary pieces of information are spread out over the inputs. Here, we demonstrate the effects of the "lost in the middle" problem in the multi-hop question answering setting -- in which multiple reasoning "hops" over disconnected documents are required -- and show that performance degrades not only with respect to the distance of information from the edges of the context, but also between pieces of information. Additionally, we experiment with means of alleviating the problem by reducing superfluous document contents through knowledge graph triple extraction and summarization, and prompting models to reason more thoroughly using chain-of-thought prompting.

Long-context LLMs have enabled numerous downstream applications but also introduced significant challenges related to computational and memory efficiency. To address these challenges, optimizations for long-context inference have been developed, centered around the KV cache. However, existing benchmarks often evaluate in single-request, neglecting the full lifecycle of the KV cache in real-world use. This oversight is particularly critical, as KV cache reuse has become widely adopted in LLMs inference frameworks, such as vLLM and SGLang, as well as by LLM providers, including OpenAI, Microsoft, Google, and Anthropic. To address this gap, we introduce SCBench(SharedContextBench), a comprehensive benchmark for evaluating long-context methods from a KV cachecentric perspective: 1) KV cache generation, 2) KV cache compression, 3) KV cache retrieval, 4) KV cache loading. Specifically, SCBench uses test examples with shared context, ranging 12 tasks with two shared context modes, covering four categories of long-context capabilities: string retrieval, semantic retrieval, global information, and multi-task. With it, we provide an extensive KV cache-centric analysis of eight categories long-context solutions, including Gated Linear RNNs, Mamba-Attention hybrids, and efficient methods such as sparse attention, KV cache dropping, quantization, retrieval, loading, and prompt compression. The evaluation is conducted on 8 long-context LLMs. Our findings show that sub-O(n) memory methods suffer in multi-turn scenarios, while sparse encoding with O(n) memory and sub-O(n^2) pre-filling computation perform robustly. Dynamic sparsity yields more expressive KV caches than static patterns, and layer-level sparsity in hybrid architectures reduces memory usage with strong performance. Additionally, we identify attention distribution shift issues in long-generation scenarios. https://aka.ms/SCBench.

We introduce Olympus, a new approach that transforms Multimodal Large Language Models (MLLMs) into a unified framework capable of handling a wide array of computer vision tasks. Utilizing a controller MLLM, Olympus delegates over 20 specialized tasks across images, videos, and 3D objects to dedicated modules. This instruction-based routing enables complex workflows through chained actions without the need for training heavy generative models. Olympus easily integrates with existing MLLMs, expanding their capabilities with comparable performance. Experimental results demonstrate that Olympus achieves an average routing accuracy of 94.75% across 20 tasks and precision of 91.82% in chained action scenarios, showcasing its effectiveness as a universal task router that can solve a diverse range of computer vision tasks. Project page: http://yuanze-lin.me/Olympus_page/

Large Language Models (LLMs) are commonly evaluated using human-crafted benchmarks, under the premise that higher scores implicitly reflect stronger human-like performance. However, there is growing concern that LLMs may ``game" these benchmarks due to data leakage, achieving high scores while struggling with tasks simple for humans. To substantively address the problem, we create GAOKAO-Eval, a comprehensive benchmark based on China's National College Entrance Examination (Gaokao), and conduct ``closed-book" evaluations for representative models released prior to Gaokao. Contrary to prevailing consensus, even after addressing data leakage and comprehensiveness, GAOKAO-Eval reveals that high scores still fail to truly reflect human-aligned capabilities. To better understand this mismatch, We introduce the Rasch model from cognitive psychology to analyze LLM scoring patterns and identify two key discrepancies: 1) anomalous consistent performance across various question difficulties, and 2) high variance in performance on questions of similar difficulty. In addition, We identified inconsistent grading of LLM-generated answers among teachers and recurring mistake patterns. we find that the phenomenons are well-grounded in the motivations behind OpenAI o1, and o1's reasoning-as-difficulties can mitigate the mismatch. These results show that GAOKAO-Eval can reveal limitations in LLM capabilities not captured by current benchmarks and highlight the need for more LLM-aligned difficulty analysis.

The introduction of Transformers architecture has brought about significant breakthroughs in Deep Learning (DL), particularly within Natural Language Processing (NLP). Since their inception, Transformers have outperformed many traditional neural network architectures due to their "self-attention" mechanism and their scalability across various applications. In this paper, we cover the use of Transformers in Robotics. We go through recent advances and trends in Transformer architectures and examine their integration into robotic perception, planning, and control for autonomous systems. Furthermore, we review past work and recent research on use of Transformers in Robotics as pre-trained foundation models and integration of Transformers with Deep Reinforcement Learning (DRL) for autonomous systems. We discuss how different Transformer variants are being adapted in robotics for reliable planning and perception, increasing human-robot interaction, long-horizon decision-making, and generalization. Finally, we address limitations and challenges, offering insight and suggestions for future research directions.

Zero-shot named entity recognition (NER) is the task of detecting named entities of specific types (such as 'Person' or 'Medicine') without any training examples. Current research increasingly relies on large synthetic datasets, automatically generated to cover tens of thousands of distinct entity types, to train zero-shot NER models. However, in this paper, we find that these synthetic datasets often contain entity types that are semantically highly similar to (or even the same as) those in standard evaluation benchmarks. Because of this overlap, we argue that reported F1 scores for zero-shot NER overestimate the true capabilities of these approaches. Further, we argue that current evaluation setups provide an incomplete picture of zero-shot abilities since they do not quantify the label shift (i.e., the similarity of labels) between training and evaluation datasets. To address these issues, we propose Familiarity, a novel metric that captures both the semantic similarity between entity types in training and evaluation, as well as their frequency in the training data, to provide an estimate of label shift. It allows researchers to contextualize reported zero-shot NER scores when using custom synthetic training datasets. Further, it enables researchers to generate evaluation setups of various transfer difficulties for fine-grained analysis of zero-shot NER.

Scaling large language models (LLMs) significantly improves performance but comes with prohibitive computational costs. Mixture-of-Experts (MoE) models offer an efficient alternative, increasing capacity without a proportional rise in compute requirements. However, training MoE models from scratch poses challenges like overfitting and routing instability. We present an efficient training recipe leveraging pre-trained dense checkpoints, training an 8-Expert Top-2 MoE model from Llama 3-8B with less than $1\%$ of typical pre-training compute. Our approach enhances downstream performance on academic benchmarks, achieving a $\textbf{2%}$ improvement in 0-shot accuracy on MMLU, while reaching a Model FLOPs Utilization (MFU) of $\textbf{46.8%}$ during training using our framework. We also integrate online upcycling in NeMo for seamless use of pre-trained weights, enabling cost-effective development of high-capacity MoE models.

The centralization of Artificial Intelligence (AI) poses significant challenges, including single points of failure, inherent biases, data privacy concerns, and scalability issues. These problems are especially prevalent in closed-source large language models (LLMs), where user data is collected and used without transparency. To mitigate these issues, blockchain-based decentralized AI (DeAI) has emerged as a promising solution. DeAI combines the strengths of both blockchain and AI technologies to enhance the transparency, security, decentralization, and trustworthiness of AI systems. However, a comprehensive understanding of state-of-the-art DeAI development, particularly for active industry solutions, is still lacking. In this work, we present a Systematization of Knowledge (SoK) for blockchain-based DeAI solutions. We propose a taxonomy to classify existing DeAI protocols based on the model lifecycle. Based on this taxonomy, we provide a structured way to clarify the landscape of DeAI protocols and identify their similarities and differences. We analyze the functionalities of blockchain in DeAI, investigating how blockchain features contribute to enhancing the security, transparency, and trustworthiness of AI processes, while also ensuring fair incentives for AI data and model contributors. In addition, we identify key insights and research gaps in developing DeAI protocols, highlighting several critical avenues for future research.

Stance detection is crucial for fostering a human-centric Web by analyzing user-generated content to identify biases and harmful narratives that undermine trust. With the development of Large Language Models (LLMs), existing approaches treat stance detection as a classification problem, providing robust methodologies for modeling complex group interactions and advancing capabilities in natural language tasks. However, these methods often lack interpretability, limiting their ability to offer transparent and understandable justifications for predictions. This study adopts a generative approach, where stance predictions include explicit, interpretable rationales, and integrates them into smaller language models through single-task and multitask learning. We find that incorporating reasoning into stance detection enables the smaller model (FlanT5) to outperform GPT-3.5's zero-shot performance, achieving an improvement of up to 9.57%. Moreover, our results show that reasoning capabilities enhance multitask learning performance but may reduce effectiveness in single-task settings. Crucially, we demonstrate that faithful rationales improve rationale distillation into SLMs, advancing efforts to build interpretable, trustworthy systems for addressing discrimination, fostering trust, and promoting equitable engagement on social media.

Large language models (LLMs) often exhibit subtle yet distinctive characteristics in their outputs that users intuitively recognize, but struggle to quantify. These "vibes" -- such as tone, formatting, or writing style -- influence user preferences, yet traditional evaluations focus primarily on the singular axis of correctness. We introduce VibeCheck, a system for automatically comparing a pair of LLMs by discovering identifying traits of a model (vibes) that are well-defined, differentiating, and user-aligned. VibeCheck iteratively discovers vibes from model outputs and then utilizes a panel of LLM judges to quantitatively measure the utility of each vibe. We validate that the vibes generated by VibeCheck align with those found in human discovery and run VibeCheck on pairwise preference data from real-world user conversations with Llama-3-70b vs GPT-4. VibeCheck reveals that Llama has a friendly, funny, and somewhat controversial vibe. These vibes predict model identity with 80% accuracy and human preference with 61% accuracy. Lastly, we run VibeCheck on a variety of models and tasks including summarization, math, and captioning to provide insight into differences in model behavior. VibeCheck discovers vibes like Command X prefers to add concrete intros and conclusions when summarizing in comparison to TNGL, Llama-405b often overexplains its thought process on math problems compared to GPT-4o, and GPT-4 prefers to focus on the mood and emotions of the scene when captioning compared to Gemini-1.5-Flash. Code can be found at https://github.com/lisadunlap/VibeCheck