This paper introduces PromptNCE, a method to estimate pointwise mutual information (PMI) using only LLMs and contrastive estimation prompts, circumventing the need for task-specific critics. It presents a benchmark with human-derived PMI and shows PromptNCE achieves Spearman correlation up to 0.82.
UniMamba is a new unified spatial-temporal forecasting framework that integrates efficient state-space dynamics with attention-based dependency learning to tackle multivariate time series challenges. It employs a Mamba Variate-Channel Encoding Layer and a Spatial Temporal Attention Layer to capture both global temporal dependencies and inter-variate correlations.
The user proposes creating a separate discussion thread for the ICML 2026 position track, fearing that discussions about this niche track would be submerged in the main discussion. The aim is to facilitate decisions regarding this specific track.
This research focuses on Subject-Aware Contrastive Learning, a novel AI technique developed for the effective processing and understanding of biosignals. It aims to improve the representation learning of complex biological data, offering advancements in the analysis of physiological measurements.
This content explains that matrices are the fundamental data structure for machine learning, likening them to spreadsheets. It demonstrates how to represent vectors and matrices using NumPy in Python, highlighting their importance in AI datasets.
The article, written by an experienced technical director in ML and game development, provides a deep technical look into the future of rendering in the gamedev industry. It discusses recent shifts in rendering architecture, inspired by the announcement of Nvidia's DLSS 5, moving beyond traditional hardware improvements towards new technical approaches.
This blog post introduces Gradient Descent as the fundamental optimization algorithm for neural networks, explaining how it iteratively minimizes a loss function. It uses the analogy of a blindfolded person navigating a hilly terrain to illustrate the core concept.
This paper introduces STOMP, a novel offline reinforcement learning algorithm for multi-objective optimization using smooth Tchebysheff scalarization. It addresses the limitation of linear scalarization in recovering non-convex Pareto fronts, crucial for aligning large language models and other real-world applications with conflicting rewards.
This research systematically studies and enhances the Forward-Forward (FF) algorithm by redesigning its local goodness function, which distinguishes positive from negative data. It introduces 'top-k goodness' and 'entmax-weighted energy,' demonstrating substantial accuracy improvements on benchmarks like Fashion-MNIST.
Autoencoders are neural networks that compress data into a lower-dimensional space and reconstruct the original input, learning non-linear structures unlike linear PCA. Their two-stage design features an encoder that projects input data into a latent space to extract informative features.
MLOps applies DevOps principles to machine learning systems, tackling unique challenges such as data/model versioning and experiment tracking. A mature MLOps practice ensures reproducible, reliable, and scalable ML development through versioning, automated pipelines, and continuous model monitoring in production.
This article details the construction of a deterministic-first inference router for prediction markets, designed to reduce reliance on expensive LLMs. It leverages a 6-hook system including market regime classification, anomaly detection, and confidence decay to efficiently process market questions.
This study classifies sentiment in English and Bangla reviews of Bangladeshi government mobile banking apps, using a hybrid labeling approach for 5,652 reviews. It found that traditional machine learning models like Random Forest and Linear SVM significantly outperformed fine-tuned XLM-RoBERTa for this specific task.
This article describes the creation of the first public audio dataset for coffee roasting first crack detection, addressing a significant gap in available resources. The dataset, comprising 973 annotated 10-second segments, was meticulously built from scratch and led to a model achieving 100% precision thanks to careful data splitting and loss weighting.
Discrete Tilt Matching (DTM) is a novel likelihood-free method for fine-tuning masked diffusion large language models (dLLMs), addressing the intractability of sequence-level marginal likelihoods in RL. It recasts fine-tuning as state-level matching, using a weighted cross-entropy objective with control variates for stability, and achieves strong results on various tasks like Sudoku and Countdown.
This paper introduces a practical pipeline to convert text corpora into quantitative semantic signals, employing embeddings, logprob-based evaluation, and noise reduction. The case study applies six semantic dimensions to Portuguese news articles about AI, supporting AI engineering tasks such as corpus inspection and monitoring.
This paper introduces GROW, an RL framework for open-world VLM agents, addressing limitations of existing Supervised Fine-Tuning methods. It proposes a novel approach for Group Relative Policy Optimization (GRPO) by decomposing trajectories into state-action samples rather than full entities.
The author details how their AI agent, with an Active Inference perception pipeline, learned a correction rule that led to over-correction, causing it to misclassify human speech. This incident highlights the challenge of building robust regulation mechanisms in AI systems to prevent over-generalization and suggests a need for more metabolic control.
This guide details setting up JupyterHub on a cloud GPU server to enable collaborative, multi-user environments for AI projects. It explains how JupyterHub manages individual Jupyter notebook servers, providing shared access to significant computational power.
This content explores three post-hoc methods—Platt Scaling, Isotonic Regression, and Temperature Scaling—designed to enhance the calibration of language models. These techniques aim to reduce the disparity between a model's predicted confidence and its actual accuracy.
