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 explores optimizing transformer architectures for specific datasets to understand optimal task-specific inductive biases beyond current scaling methods.
Research proposes Source Node Influence Pruning (SNIP) for Unsupervised Graph Domain Adaptation (UGDA) to improve knowledge transfer by filtering irrelevant source nodes.
Research identifies encoder Jacobian conditioning as key for stable trimodal contrastive learning, preventing representation collapse or amplification.
Research explores decomposing LLM reasoning into strategy-conditioned representations, aiming to explicitly identify distinct problem-solving methods.
Research explores methods to improve LLM planning capabilities by addressing the gap between theoretical Turing-completeness and empirical performance.
Research paper proposes MXSens, a sensitivity-aware mixed-precision quantization method to improve 4-bit LLM inference accuracy and efficiency without frequent dequantization.
Research evaluates time series foundation models for zero-shot crowd forecasting in special events, addressing scarce historical data.
Research introduces a few-shot Inverse Reinforcement Learning (IRL) method for learning from limited demonstrations by decomposing rewards.
Researchers propose using 2026 FIFA World Cup as a contamination-free benchmark for LLM forecasting agents, testing four frontier models.
Research explores feature attribution methods to improve explainability of deep learning models in predictive process monitoring within business operations.
Research paper explores the concept of 'representation' in ML, differentiating engineering views from philosophical claims of 'mental representation' in scaled models.
Researchers introduced Mean Root Square Normalization (MRSNorm), a method to improve numerical stability and gradient flow in sequence models by pairing channels into 2D phasors.
New research explores the theoretical foundations of $\max$@$k$ reinforcement learning, a metric common in large reasoning models for difficult tasks.
Researchers introduced Cognitive Memory Primitive (CMP), an AI architecture that learns via local, gradient-free updates and no backpropagation.
Research explores using chain-of-thought (CoT) reasoning to stabilize value conflict resolution in LLMs, addressing challenges with scalar rewards in RLHF.
Research uses topological methods to understand TabPFN's reliability on complex tabular data, enhancing model interpretability.
DiFA proposes a training-free framework for diffusion model inference that refines data prediction by accounting for statistical uncertainty.
Research explores using LLMs to generate and optimize GPU kernel code for performance, focusing on validation and profiling for practical effectiveness.
New research introduces Random Feature Information Gain (RFIG) for reinforcement learning exploration, leveraging random Fourier features for information gain approximation.
FlashPDE is a new Triton operator library that optimizes Physics-Informed Neural Networks (PINNs) for solving PDEs by reducing memory overhead.
Adaptive Mamba Neural Operators (AMO) integrate reproducing kernels for state-space models to solve partial differential equations.
Researchers propose Similarity-based Generative Network (SGN) for synthetic data generation robust to distribution shifts, reducing the need for target-domain optimization.
Sobek proposes a new method for streaming equivariant tensor product convolutions, reducing memory and improving GPU utilization for graph neural networks.
New research from arXiv re-derives the Bellman equation from three core conditions, showing its fundamental role in sequential decision-making.
Research introduces SelectInfer, a method for selective neuron loading and computation to optimize LLMs for resource-constrained edge devices.
Research explores self-distillation to enhance rubric-based Reinforcement Learning (RL) for LLMs, addressing limitations like limited exploration.
Research identifies split conformal prediction maintains marginal coverage but fails silently on per-class coverage under distribution shift.
Research proposes repurposing edge AI inference accelerators for on-device model adaptation, addressing compute and memory limits for training.
Researchers propose Manifold-Constrained Hyper-Connections (mHC) as a parameter-efficient finetuning method by modifying residual connections.
Research explores Differentiable Logic Gate Networks (Diff-Logic) for low-latency EEG classification on edge devices, compiling models into Boolean circuits.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion