Search signals, briefings, company results, benchmarks and glossary terms.
Search signals, briefings, company results, benchmarks and glossary terms.
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.
Research audits hyperbolic vision-language models, finding that claimed geometric benefits often don't manifest at their operating points.
Research introduces Gaussian invariant MCMC sampling methods (RWM, MALA, Hessian MALA) for improved statistical efficiency via analytical tractability.
Hyperflux introduces an L0 pruning method modeling weight removal as an evolving system, improving understanding and efficiency of neural network pruning.
Research proves diffusion model samplers (DDIM, DDPM) achieve $k/\varepsilon$ iteration complexity for low-dimensional data, accelerating sampling.
Research explores belief-contingent interventions like 'common identity' to counter machine-assisted statistical discrimination, beyond belief-free methods.
Research explores robust Bayesian decision making under adversarial uncertainty, improving decision stability when outcome models are perturbed.
Researchers propose Eigenbasis-Independent Learnable Spectral Positional Encodings for directed graphs to overcome $O(n^3)$ complexity and gauge ambiguity.
Researchers propose x-prediction, a training-free method to accelerate generation in diffusion and flow matching models by reducing neural function evaluations.
Research proposes a layer-parallel inference method to reduce encrypted nonlinear depth in Transformers, addressing a key bottleneck in FHE for AI.
Research establishes new theoretical bounds for minimum block width in residual neural networks with inner width one, improving universal approximation.
Research identifies a low-cost method using Armijo backtracking to estimate loss sharpness, improving Adam optimizer stability without complex Hessian computations.
Research introduces DemoPSD, a self-distillation method improving LLM reasoning by mitigating overfitting from teacher supervision in cross-domain tasks.
Research introduces WildChat benchmark and evaluation protocol for multi-agent routing from natural language prompts, focusing on set-valued prediction and cost-awareness.
New research proposes the Vigilant Evaluator of Representations (VER) framework to detect explanatory insufficiency in learned ML representations beyond traditional metrics.
New research introduces the Bootstrap Theory of Representational Emergence (TBER) to explain how AI systems develop new representations.
New research addresses KL-regularized reinforcement learning and contextual bandits with model misspecification using function approximation.
TimeSAE is a new research method for generating causal sparse explanations of black-box time series model predictions, addressing out-of-distribution generalization.
Research presents methods to estimate the amplification factor of generative networks in simulations, crucial for understanding statistical precision.
Research introduces a robust group Lasso regularized rank regression method to handle heavy-tailed noise and outliers in high-dimensional data.
Research on monitoring probability forecast calibration with an application to concept drift detection in image classification models.
Research paper details a framework for analyzing learning dynamics of high-dimensional, multi-class stochastic gradient descent (SGD) problems.
Research proposes ACES, a novel method using Leave-One-Out AUC Consistency to evaluate LLM-generated code tests, addressing the circular dependency challenge.
Research identifies a critical feature norm threshold in neural networks that predicts the onset of neural collapse, invariant to training conditions.
Research explores challenges and opportunities of using machine learning emulators to reduce computational costs in climate modeling.
Research finds chain-of-thought (CoT) reasoning in vision-language models (VLMs) can degrade uncertainty quantification reliability, inducing overconfidence.
Research indicates Vision-Language-Action (VLA) models can be effective continual learners via sequential fine-tuning in reinforcement learning without catastrophic forgetting.
Researchers introduce GenGNN, a message-passing graph neural network for discrete graph generation, potentially moving beyond Graph Transformers.
New research proposes a Softmax-Weighted Switching Gradient method for distributed stochastic minimax optimization with stochastic constraints for federated learning.
Research on multi-platform learning finds that user choice can lead to model overspecialization and arbitrary loss increases, even with optimal algorithms.
Research introduces BRIDGE, a method for distilling Chain-of-Thought reasoning from large to smaller LLMs by using structure-aware masking to address capacity mismatch.
Researchers introduced SynthSAEBench, a new benchmark and toolkit using large-scale synthetic data to evaluate Sparse Autoencoders more precisely.
Research explores probabilistic wind power forecasting using gradient boosting trees and weather ensembles to improve renewable energy grid integration.
Research introduces INFORM, an interpretability analysis for multi-expert LLM systems, aiming to decouple expert interaction and sequencing.
CoGenCast is a new generative AI framework for time series forecasting combining autoregressive LLMs and diffusion-like models for improved accuracy.
Research proposes HyperNet-Adaptation for diffusion-based generative test case creation, improving functional evaluation for deep learning systems.
Research introduces BalDRO, a framework using distributionally robust optimization to improve LLM unlearning by addressing sample imbalance in forget sets.
CUDA-L2, a system combining LLMs and RL, automates and optimizes Half-precision General Matrix Multiply (HGEMM) CUDA kernels, surpassing cuBLAS performance.
Research explores methods to control the trade-off between LLM inference costs and in-context recall, moving beyond fixed training-time settings.
Research introduces CANDI, a hybrid discrete-continuous diffusion model, and a framework for understanding noise corruption in discrete data.
Research proposes a new reinforcement learning algorithm to improve reasoning in diffusion large language models (dLLMs), aiming for higher inference throughput.
New research suggests multi-hop reasoning failures in LLMs stem from missing supervision on "bridge entities," addressable with "identity bridge" supervision.
New research proposes an 'attribution-guided continual learning' method to mitigate catastrophic forgetting in LLMs by identifying and protecting critical parameters.
COSMOS is a new federated learning framework addressing architectural and statistical heterogeneity by using clustered server models and pseudo-labels.
Researchers introduced CTFusion, a new CTF-based benchmark for evaluating LLM agents, addressing data contamination issues in existing CTF challenges.
Research introduces a novel retrieval method using multiple query vectors simultaneously for complex reasoning tasks, moving beyond single-vector queries.
Researchers propose Multi-Scale Separable Fourier Neural Networks (MS-SFNN) to solve high-frequency Partial Differential Equations (PDEs), addressing spectral bias.
Research systematically studies knowledge distillation effectiveness with ResNet teacher-student pairs on CIFAR-10, focusing on student capacity.
Research introduces 'Prune, Update and Trim' (PUT), a post-training pruning method to reduce LLM inference costs and resource requirements.
Research explores Constrained Reinforcement Learning for safe and energy-efficient control of heat pump systems, balancing thermal comfort and optimization.
Research proposes using cross-model consensus from independently trained LLMs to improve reasoning accuracy, outperforming self-consistency and reward models.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion