Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
QuantWAMs introduces a post-training quantization framework to efficiently deploy World Action Models (WAMs) by addressing calibration and execution challenges.
A new open-source framework, READII-2-ROQC, uses volume-preserving negative controls to assess if imaging AI features capture independent spatial signals.
Research explores Graph Neural Networks for multilevel preconditioning to improve iterative solvers for large, sparse linear systems in scientific computing.
Researchers propose ScaFE, a system using LLM-generated clinical feature programs to classify pathological scars from images, addressing data scarcity and governance.
Research proposes Doubly Robust Functional Representation Learning for longitudinal causal inference, handling irregular time-series data in studies.
MixFrag proposes a fragility-guided mixed-precision post-training quantization (PTQ) framework for Vision Transformers (ViTs) to optimize deployment on resource-constrained devices.
ReToken introduces a single learnable embedding to improve visual retrieval in vision-language models by selecting relevant visual tokens, addressing long visual context challenges.
Researchers propose AdamG, a new parameter-free optimizer designed to automatically adapt to diverse optimization problems without manual learning rate tuning.
Researchers propose Attention Heads and MLP Pruning (AMP) to significantly reduce LLM computational costs and accelerate inference for resource-limited deployments.
Research explores necessary disclosures for fairness audit reports when demographic data used for protected labels is incomplete.
Research introduces importance-aware aggregation for federated graph learning (FGL) to improve global model performance under domain shifts across clients.
Research paper explores using category-wise influence functions to analyze classifier performance ceilings and improve data quality for model training.
Researchers identify attention rank-collapse in long-context models and evaluate polylogarithmic rescaling to preserve retrieval quality.
Research proposes 'soft-label training' using a distribution of annotator judgments instead of single majority labels for ambiguous tasks.
Researchers introduced Dynamically Scaled Activation Steering (DSAS), a method-agnostic framework for guiding generative model behavior more efficiently.
Research suggests that training future AI systems on content generated by a diversity of models, rather than a single model, mitigates knowledge collapse.
Research introduces concurrency enhancements for training Kolmogorov-Arnold Networks (KANs) using a Newton-Kaczmarz method, claiming better accuracy and speed than MLPs.
Researchers introduced Theseus, a training-free method to transport task-specific parameter updates across large language models with different architectures.
New research shows previous claims of large language models learning effectively from 100% noisy data using RLVR are invalid due to data contamination.
Research proposes Quadratic Objective Perturbation (QOP) for differentially private empirical risk minimization, addressing limitations of Linear Objective Perturbation (LOP) with unbounded gradients.
Research proposes a neurosymbolic approach to imitation learning combining neural networks for high-dimensional data and symbolic methods for generalization.
SNAC-Pack 2.0 introduces Scaled-Out Surrogate Neural Architecture Codesign, optimizing model architecture for hardware cost, especially FPGAs, beyond just accuracy.
New research introduces strategic classification models where agents' responses to classifiers induce genuine improvements, not just cosmetic feature changes.
Research shows averaging evaluation metrics masks significant capability trade-offs in LLM pruning, particularly in code retention.
Research characterizes the sparsity and geometry of parameter updates in on-policy distillation, a post-training recipe for language and vision-language models.
FORGE is a new LLM training technique that eliminates the need to materialize gradients, reducing memory usage by fusing optimizer steps.
Research introduces Conformal Cascade, a multi-tier LLM inference method offering distribution-free accuracy guarantees, addressing miscalibrated confidence scores.
Research proposes a neural network method for anomaly detection in cybersecurity, addressing novel attack evasion without prior anomaly data or distributions.
Researchers developed MMC+, an enhanced framework for scalable drift monitoring in medical imaging AI, building on the CheXstray framework.
CLIP-Guided Backdoor Defense (CGD) is a new method for mitigating backdoor attacks in DNNs by efficiently separating poisoned training data.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion