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 methods to certify the recall of high-recall candidate generation systems with finite-sample validity using audit labels.
Research proposes Master-Agent Proto-plan System (MAPS), a hierarchical deep reinforcement learning architecture for autonomous vehicle coordination at unsignalized intersections.
Researchers propose X$^3$-OPD, a cross-modal on-policy distillation framework to transfer reasoning from text models to audio-language models.
New research introduces TimePNS, a framework for time-series explanations that identifies both sufficient and necessary subsequences for model predictions.
Expanding Generative Flows (EFlows) are introduced, enabling flow-based generative models to handle distributions of increasing dimensionality.
Research decomposes preference-based model updates (RLHF) into spectral components, making the fine-tuning process more interpretable and controllable.
Research explores using LLMs and prediction markets to measure strategic information bias in text corpora, extracting implied beliefs from information ecosystems.
Researchers released Naver-News-KO, a 27,400-pair Korean news summarization dataset from Naver News, for open-source model fine-tuning.
Researchers released GLAN-QnA-KR, a 303K-row Korean instruction-QA corpus generated by Microsoft's Phi-3.5-MoE-instruct model.
Researchers developed a neural-symbolic pipeline combining compact transformers with a linguistically-informed mediator for gaming toxicity detection.
Domyn-Small, a 10B parameter multilingual open-weight reasoning LLM developed in Europe, achieves 32K context window via continued pre-training.
A research paper proposes a knowledge-injection framework to improve zero-shot delirium prediction using LLMs and electronic health records.
DC-Leap proposes a training-free framework to accelerate Diffusion Large Language Models (dLLMs) by addressing conservative confidence thresholds.
SonicSampler presents unified tile-aware kernels for LLM sampling and speculative verification, improving efficiency for dynamic serving workloads.
New research from arXiv proposes a novel approach to evaluate LLM performance by aligning models with diverse, real-world user expectations.
New research introduces DFAH-Bench, a replay benchmark designed to measure observable behavioral instability in financial agent decision-making processes.
Research introduces an attention-based experience replay framework for continual learning in time series forecasting models, addressing data distribution shifts.
Research finds combining regex filters with LLM alignment provides no measurable safety coverage gain against adversarial input, even when designed to bypass regex.
LeanFlow is an LLM agent system designed to translate mathematical papers into formal, buildable Lean projects, evaluating runtime mechanisms.
MiniCache is a reusable program caching framework that converts Program-of-Thought (PoT) programs into parameterized cache objects for efficient LLM inference.
ConfidenceBench evaluates 15 frontier LLMs on verbalized confidence calibration using the Brier score, addressing risks of fluent but incorrect answers.
AppWorld-UL is a new benchmark for evaluating agent-user interactions in tool-use agents, addressing diversity and environment scale limitations.
SPECTRA is a new probabilistic forecasting architecture designed for complex energy systems, integrating multi-scale decomposition and exogenous variables.
New benchmark, Frontier Financial Judgement, assesses AI agents' ability to replicate expert equity analyst judgments on stock-moving information.
Research proposes a representation learning technique using joint contrastive and classification learning to improve cross-domain generalization in optical networks.
Research demonstrates 'LeakyLMs' attacks can infer proprietary model architecture and inference optimizations from production LLMs via token timing analysis.
Research introduces Pipelined Gradient Coding, an improved method for distributed machine learning training that reduces straggler worker impact.
Research audits five diversity measures for LLM ensembles, finding they often track capability rather than true diversity when predicting majority-vote gain.
Research explores if Vision-Language Models (VLAs) with Mixture-of-Experts (MoE) action heads can learn to decompose robotic tasks into reusable primitives.
Research explores using Jungian Cognitive Functions for LLM activation steering, moving beyond static Big Five trait models for personality representation.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion