We observe a novel phenomenon, contextual entrainment, across a wide range of language models (LMs) and prompt settings, providing a new mechanistic perspective on how LMs become distracted by “irrelevant” contextual information in the input prompt. Specifically, LMs assign significantly higher logits (or probabilities) to any tokens that have previously appeared in the context prompt, even for random tokens. This suggests that contextual entrainment is a mechanistic phenomenon, occurring independently of the relevance or semantic relation of the tokens to the question or the rest of the sentence. We find statistically significant evidence that the magnitude of contextual entrainment is influenced by semantic factors. Counterfactual prompts have a greater effect compared to factual ones, suggesting that while contextual entrainment is a mechanistic phenomenon, it is modulated by semantic factors.We hypothesise that there is a circuit of attention heads — the entrainment heads — that corresponds to the contextual entrainment phenomenon. Using a novel entrainment head discovery method based on differentiable masking, we identify these heads across various settings. When we “turn off” these heads, i.e., set their outputs to zero, the effect of contextual entrainment is significantly attenuated, causing the model to generate output that capitulates to what it would produce if no distracting context were provided. Our discovery of contextual entrainment, along with our investigation into LM distraction via the entrainment heads, marks a key step towards the mechanistic analysis and mitigation of the distraction problem.

Long-context Large Language Models (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 cache centric perspective: 1) KV cache generation, 2) KV cache compression, 3) KV cache retrieval, and 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 SCBench, we provide an extensive KV cache-centric analysis of eight categories long-context solutions, including Gated Linear RNNs (Codestal-Mamba), Mamba-Attention hybrids (Jamba-1.5-Mini), and efficient methods such as sparse attention, KV cache dropping, quantization, retrieval, loading, and prompt compression. The evaluation is conducted on six Transformer-based long-context LLMs: Llama-3.1-8B/70B, Qwen2.5-72B/32B, Llama-3-8B-262K, and GLM-4-9B. 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.

The integration of long-context capabilities with visual understanding unlocks unprecedented potential for Vision Language Models (VLMs). However, the quadratic attention complexity during the pre-filling phase remains a significant obstacle to real-world deployment. To overcome this limitation, we introduce MMInference (Multimodality Million tokens Inference), a dynamic sparse attention method that accelerates the prefilling stage for long-context multi-modal inputs. First, our analysis reveals that the temporal and spatial locality of video input leads to a unique sparse pattern, the Grid pattern. Simultaneously, VLMs exhibit markedly different sparse distributions across different modalities. We introduce a permutation-based method to leverage the unique Grid pattern and handle modality boundary issues. By offline search the optimal sparse patterns for each head, MMInference constructs the sparse distribution dynamically based on the input. We also provide optimized GPU kernels for efficient sparse computations. Notably, MMInference integrates seamlessly into existing VLM pipelines without any model modifications or fine-tuning. Experiments on multi-modal benchmarks-including Video QA, Captioning, VisionNIAH, and Mixed-Modality NIAH-with state-of-the-art long-context VLMs (LongVila, LlavaVideo, VideoChat-Flash, Qwen2.5-VL) show that MMInference accelerates the pre-filling stage by up to 8.3x at 1M tokens while maintaining accuracy. Our code is available at https://ama.ms/MMInference.

The computational challenges of Large Language Model (LLM) inference remain a significant barrier to their widespread deployment, especially as prompt lengths continue to increase. Due to the quadratic complexity of the attention computation, it takes 30 minutes for an 8B LLM to process a prompt of 1M tokens (i.e., the pre-filling stage) on a single A100 GPU. Existing methods for speeding up prefilling often fail to maintain acceptable accuracy or efficiency when applied to long-context LLMs. To address this gap, we introduce MInference (Milliontokens Inference), a sparse calculation method designed to accelerate pre-filling of long-sequence processing. Specifically, we identify three unique patterns in long-context attention matrices-the A-shape, Vertical-Slash, and Block-Sparse-that can be leveraged for efficient sparse computation on GPUs. We determine the optimal pattern for each attention head offline and dynamically build sparse indices based on the assigned pattern during inference. With the pattern and sparse indices, we perform efficient sparse attention calculations via our optimized GPU kernels to significantly reduce the latency in the pre-filling stage of longcontext LLMs. Our proposed technique can be directly applied to existing LLMs without any modifications to the pre-training setup or additional fine-tuning. By evaluating on a wide range of downstream tasks, including InfiniteBench, RULER, PG-19, and Needle In A Haystack, and models including LLaMA-3-1M, GLM-4-1M, Yi-200K, Phi-3-128K, and Qwen2-128K, we demonstrate that MInference effectively reduces inference latency by up to 10x for pre-filling on an A100, while maintaining accuracy. Our code is available at https://aka.ms/MInference.

We propose TD-GEN, a graph generation framework based on tree decomposition, and introduce a reduced upper bound on the maximum number of decisions needed for graph generation. The framework includes a permutation invariant tree generation model which forms the backbone of graph generation. Tree nodes are supernodes, each representing a cluster of nodes in the graph. Graph nodes and edges are incrementally generated inside the clusters by traversing the tree supernodes, respecting the structure of the tree decomposition, and following node sharing decisions between the clusters. Further, we discuss the shortcomings of the standard evaluation criteria based on statistical properties of the generated graphs. We propose to compare the generalizability of models based on expected likelihood. Empirical results on a variety of standard graph generation datasets demonstrate the superior performance of our method.

Uncertainty of labels in clinical data resulting from intra-observer variability can have direct impact on the reliability of assessments made by deep neural networks. In this paper, we propose a method for modelling such uncertainty in the context of 2D echocardiography (echo), which is a routine procedure for detecting cardiovascular disease at point-of-care. Echo imaging quality and acquisition time is highly dependent on the operator's experience level. Recent developments have shown the possibility of automating echo image quality quantification by mapping an expert's assessment of quality to the echo image via deep learning techniques. Nevertheless, the observer variability in the expert's assessment can impact the quality quantification accuracy. Here, we aim to model the intra-observer variability in echo quality assessment as an aleatoric uncertainty modelling regression problem with the introduction of a novel method that handles the regression problem with categorical labels. A key feature of our design is that only a single forward pass is sufficient to estimate the level of uncertainty for the network output. Compared to the 0.11 ± 0.09 absolute error (in a scale from 0 to 1) archived by the conventional regression method, the proposed method brings the error down to 0.09 ± 0.08, where the improvement is statistically significant and equivalents to 5.7% test accuracy improvement. The simplicity of the proposed approach means that it could be generalized to other applications of deep learning in medical imaging, where there is often uncertainty in clinical labels.

Accurate detection of end-systolic (ES) and end-diastolic (ED) frames in an echocardiographic cine series can be difficult but necessary pre-processing step for the development of automatic systems to measure cardiac parameters. The detection task is challenging due to variations in cardiac anatomy and heart rate often associated with pathological conditions. We formulate this problem as a regression problem and propose several deep learning-based architectures that minimize a novel global extrema structured loss function to localize the ED and ES frames. The proposed architectures integrate convolution neural networks (CNNs)-based image feature extraction model and recurrent neural networks (RNNs) to model temporal dependencies between each frame in a sequence. We explore two CNN architectures: DenseNet and ResNet, and four RNN architectures: long short-term memory, bi-directional LSTM, gated recurrent unit (GRU), and Bi-GRU, and compare the performance of these models. The optimal deep learning model consists of a DenseNet and GRU trained with the proposed loss function. On average, we achieved 0.20 and 1.43 frame mismatch for the ED and ES frames, respectively, which are within reported inter-observer variability for the manual detection of these frames.

Echocardiography (echo) is a skilled technical procedure that depends on the experience of the operator. The aim of this paper is to reduce user variability in data acquisition by automatically computing a score of echo quality for operator feedback. To do this, a deep convolutional neural network model, trained on a large set of samples, was developed for scoring apical four-chamber (A4C) echo. In this paper, 6,916 end-systolic echo images were manually studied by an expert cardiologist and were assigned a score between 0 (not acceptable) and 5 (excellent). The images were divided into two independent training-validation and test sets. The network architecture and its parameters were based on the stochastic approach of the particle swarm optimization on the training-validation data. The mean absolute error between the scores from the ultimately trained model and the expert's manual scores was 0.71 ± 0.58. The reported error was comparable to the measured intra-rater reliability. The learned features of the network were visually interpretable and could be mapped to the anatomy of the heart in the A4C echo, giving confidence in the training result. The computation time for the proposed network architecture, running on a graphics processing unit, was less than 10 ms per frame, sufficient for real-time deployment. The proposed approach has the potential to facilitate the widespread use of echo at the point-of-care and enable early and timely diagnosis and treatment. Finally, the approach did not use any specific assumptions about the A4C echo, so it could be generalizable to other standard echo views.

In the above-title paper [ibid., vol. 36, no. 6, pp. 1221-1230, Jun. 2017], the first footnote should have indicated the following information: A. H. Abdi and C. Luong are joint first authors.