Real-time AI-curated news from 94,959+ articles across 50+ sources. Sentiment analysis, importance scoring, and key takeaways — updated every 15 minutes.
Researchers introduce DPPrefSyn, an algorithm for generating differentially private synthetic preference data to train large language models while protecting user privacy. The method combines the Bradley-Terry preference model with DP-PCA to create synthetic training data from private datasets, achieving competitive alignment performance with formal privacy guarantees.
GaMi is a multimodal material identification system that combines mmWave and acoustic sensing to accurately identify materials regardless of geometric variations like shape, orientation, and distance. Using cross-modal subtractive disentanglement and contrastive learning, the system achieves 95.2% accuracy on 20 materials and demonstrates few-shot generalization across different devices.
Researchers propose a constrained optimization framework for unlearning in diffusion models that balances removing undesirable data while preserving model utility. Using KL divergence and likelihood constraints with primal-dual algorithms, the approach achieves superior performance in concept and data unlearning compared to existing weight-based methods.
Researchers introduce BioConCal, a supervised scoring system that evaluates biomedical entity candidates surfaced by multiple LLMs across five public datasets. The tool improves candidate verification from 75.3% to 91% AUROC by leveraging agreement patterns and document features, enabling more efficient curator review workflows rather than recovering missed entities.
Researchers identify Supervision Fidelity Decay (SFD) as a critical limitation in on-policy distillation where teacher model confidence deteriorates as student-generated reasoning chains lengthen. They propose Lookahead Group Reward (LGR) with entropy-triggered tree-attention to strengthen supervision signals, achieving 2.57-point improvements on math and code benchmarks, with gains reaching 4.92 points on AIME-26.
Researchers propose Hide-and-Seek, a machine learning framework that detects failures in Vision-Language-Action (VLA) models during robot execution by identifying failure-indicative actions from trajectory-level data alone. The method achieves state-of-the-art performance across multiple VLA policies and robotic platforms without requiring expensive step-level annotations or external models.
Researchers propose Canopy Entropy (CE*), a new metric that reveals fine-tuning reorganizes uncertainty in language models rather than simply reducing it. The measure shows that fine-tuned models convert token-level uncertainty into more semantically meaningful and informative outputs, fundamentally changing how we understand model alignment and information generation.
Researchers propose Safe Equilibrium Policy Optimization (SEPO), a training method that prevents language model agents from exploiting weaker opponents, colluding on harmful outcomes, or externalizing costs during multi-agent interactions. The technique augments standard reward optimization with penalties for exploitability and collusion risk, demonstrated across strategic domains including Prisoner's Dilemma, auctions, and poker.
Researchers introduce Sophrosyne, a system that improves Text2SQL agents by moderating their exploration of database APIs. The solution addresses over-exploration by fine-grained APIs, reducing unnecessary schema queries by 4.6x while improving SQL generation accuracy by up to 12.4 percentage points.
Researchers propose FedVPA-GP, a federated learning framework that enables privacy-preserving alignment of large language models while preserving diverse user preferences instead of averaging them into a single monolithic reward model. The approach uses a Gumbel-Softmax prior and orthogonal loss to prevent posterior collapse and successfully disentangles conflicting user intents in decentralized settings.
Researchers introduce PatchWorld, a gradient-free framework that converts offline trajectories into executable Python world models for AI agents operating in partially observable environments. The method achieves 76.4% success on planning tasks without requiring LLM calls during prediction, while revealing a fundamental tradeoff between observation accuracy and decision-making utility in executable world models.
Researchers introduce SURE, a unified experimentation framework that standardizes evaluation metrics and training pipelines for speech understanding models, addressing reproducibility challenges that have hindered fair comparison of speech foundation models and Speech LLMs across different deployment scenarios.
Researchers present a novel inverse reinforcement learning framework that handles multiple imperfect demonstrators with varying suboptimality levels, using a feasible-reward-set approach with linear constraints. The method includes theoretical guarantees for reward recovery and practical algorithms tested on grid-worlds and LLM fine-tuning, addressing a significant gap in real-world IRL applications.
BlueFin is a new benchmark dataset that evaluates how well large language model agents perform on real-world financial spreadsheet tasks, revealing that even frontier LLMs struggle significantly with complex spreadsheet manipulation and analysis despite their advanced capabilities.
Researchers demonstrate that toxic language in prompts significantly degrades the factual accuracy of large language models, even when semantic content remains identical. By analyzing internal model activations, they identify that toxicity amplifies perturbation-sensitive nodes while leaving core reasoning pathways relatively stable, revealing a critical vulnerability in LLM reliability.
Researchers propose DareU, a novel LLM unlearning framework that uses data attribution rewards and reinforcement learning to remove training data influence from large language models. Unlike existing approaches that maximize loss on forget sets, this method reduces attribution scores to forgotten data owners, addressing critical issues of over-forgetting and model utility degradation.
Researchers introduced the Tacit Understanding Index (TUX), a new framework for measuring how well AI language models align with human values and reasoning without explicit instructions. Testing across 241 humans and 200 LLM profiles, they found that AI-human pairs with similar personality traits achieved significantly higher alignment, suggesting tacit understanding is structured and measurable rather than random.
Researchers tested whether large language models inherit moral reasoning patterns from the institutional environments of the languages they were trained on. Across nine languages and six frontier LLMs, moral divergence emerged specifically in institutionally ambiguous scenarios and correlated with real-world institutional quality differences, suggesting language encodes institutional experience that influences AI decision-making.
Researchers introduce AMix-2, a protein-text foundation model that treats protein sequences as a native modality in large language models alongside natural language. The model uses a novel block-wise diffusion approach instead of traditional left-to-right generation, paired with a new ProteinArena benchmark for evaluating protein AI systems.
Researchers introduce ImmersiveTTS, an AI model that generates natural speech integrated within environmental audio contexts using multimodal diffusion transformers and domain-specific representation alignment. The advancement addresses a key challenge in audio generation: seamlessly combining speech with background environmental sounds while maintaining acoustic quality and intelligibility.
Researchers propose a 'claim network' framework that transforms flat citation graphs into typed, stance-labeled networks for scientific literature. By reifying each cross-document reference as a typed claim with source, target, text, and stance classification, the approach enables richer document understanding than traditional knowledge graphs and demonstrates improvements in retrieval-augmented generation tasks.
Researchers introduce VACSR, a variational adapter method that improves cross-modal similarity representation in vision-language models by treating annotation limitations as a variational inference problem. The approach addresses the problem of binary classification boundaries compressing continuous similarity spaces, reducing false negatives and improving generalization across image-text retrieval and domain adaptation tasks.
This paper analyzes why reinforcement learning methods that update policies based on reward signals without explicitly tracking uncertainty can still be effective. Researchers prove that annealed softmax policies achieve near-optimal regret rates in many-armed Bayesian bandit settings when many near-optimal actions exist, providing theoretical justification for uncertainty-agnostic approaches used in modern language model training.
Researchers introduce a structured visual perturbation framework to analyze how Vision-Language-Action (VLA) models ground their autonomous driving decisions in visual information. The study reveals uneven visual dependency across different abstraction levels, highlighting the need for better diagnostic tools to ensure safer, more robust autonomous driving systems.
Researchers propose dynamic Stiefel routing, a novel machine learning approach using expert projection filters on the Stiefel manifold to improve cross-domain EEG decoding without requiring target-domain calibration data. The method addresses a fundamental degeneracy problem where naive routing collapses to ensemble averaging, introducing three structural properties that enable genuine domain-specialized routing with significant accuracy improvements across datasets.
