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 proposes Graph Edge Sparsification (GES), a learning-based approach to improve computational efficiency for large-scale Traveling Salesman Problem (TSP) instances.
Research introduces YUKTI, a framework for robust, verifiable decision-making from natural language situations, addressing uncertainty in model outputs.
Research paper explores 'model collapse,' where AI-generated content in training data degrades future model performance, leading to loss of coherence.
Research explores transfer learning in adaptive multi-agent systems where policy-outcome relationships change due to regulatory shifts.
Research finds that message format in multi-hop LLM agent relays significantly impacts copy fidelity, with structured messages performing better.
New theoretical framework explains how Transformer models develop inductive reasoning abilities across generalized tasks like n-grams and multi-hop reasoning.
Researchers propose a 'frugal' Neural Architecture Search (NAS) framework combining Transformer control and swarm intelligence, reducing compute costs.
Research explores causal discovery using only interventions, relaxing the 'faithfulness' assumption to improve model robustness.
New research introduces an exact method for measuring state usage in selective state-space models (SSMs) like Mamba, detailing how information flows.
Research explores methods for tokenizing continuous numerical and dense embedding features to integrate them into LLM-based recommender systems.
The VERaiPHY initiative proposes frameworks for rigorous ML assessment in fundamental physics, addressing reliability for discovery claims.
Research introduces a trace-supervised symbolic neural CPU for interpretable program execution, making neural network state transitions visible.
Adaptive Model Compression (AMC) dynamically allocates resources for transformer inference based on token importance to reduce energy and memory.
Research explores an optimal stopping problem for sequential data collection in stochastic optimization, aiming to minimize data acquisition costs.
Research proposes automated tensor scheduling for hybrid CPU-GPU LLM inference, improving efficiency on devices with limited GPU memory.
New research proposes Random Label Prediction Heads (RLP-heads) to empirically study memorization in deep neural networks for classification.
DAG-FM, a new foundation model architecture, proposes to amortize causal discovery from observational tabular data under heterogeneous causal mechanisms.
New research proposes a modular LLM post-training method using proxy-guided update signals to enable reusable and transferable optimization.
Research applies time-lag-aware deep reinforcement learning to flexible job-shop scheduling in prefabricated construction module factories.
Research links grokking phenomenon in neural networks, where models generalize after memorizing, to dimensionality collapse in representation geometry.
Research introduces a multi-scale feature enhanced graph neural network (GNN) for more efficient fluid dynamics prediction in complex geometries.
New research proposes a constrained two-view framework for Graph Neural Networks (GNNs) to improve node prediction accuracy by decoupling feature transformation and neighborhood aggregation, addressing topology noise and heterophily.
Researchers introduce Sticky Jump Diffusions (SJDs), a unified framework for masked, continuous, and hybrid diffusion models using continuous-time Markov processes.
Researchers propose WSqD, a new learning rate schedule for large model training that is horizon-free, improving adaptability for extended training.
Research explores a multi-agent framework using LLMs for optimizing zero-dimensional reduced-order model (0D ROM) planning in complex equipment design.
New research proposes a novel Graph Neural Network (GNN) framework to improve fraud detection by addressing incomplete node attributes and class imbalance.
Research on Deep Equilibrium Models (DEQs) found a "lazy identity collapse" where learned initializations undermine iterative inference in specific reasoning tasks.
NeuroMem-FHP, a deep learning framework using LSTMs and Transformers, estimates parameters for fractional Hawkes processes from sequences.
FastTPS is a proposed method to optimize token processing in decoder-only LLM inference, aiming to improve throughput and AI accelerator utilization.
SPARC-Net is a new Physics-Informed Neural Network (PINN) architecture designed to overcome limitations in solving stiff and shock-dominated PDEs.
Researchers propose a Graph Foundation Model (GFM) that aligns graph data to tabular formats to enable multi-domain learning with less data.
New research generalizes preference-based reinforcement learning (RL) to include 'incomparable' trajectory pairs, formalizing a Bradley-Terry-inspired model.
Researchers introduced Velocity Scheduled Flow Matching (VSFM), a new method for diffusion model training that optimizes sampling speed.
Research explores optimizing agentic systems through structured skill edits, viewing them as local repairs with delayed, context-dependent effects.
Researchers introduced EasyOPD, an on-policy distillation framework for LLMs designed to improve upon existing methods' fragmentation and data collection.
Anthropic's Claude Fable 5 was evaluated on eight biomedical benchmarks using deterministic scoring, addressing common evaluation flaws.
TreeThink is a new open-source Python library for modular, asynchronous tree search, specifically designed for neural theorem proving with LLMs.
A new benchmark, Loci Similes, is proposed for evaluating language models' ability to extract intertextual links in Latin literature.
Research systematically investigates subword tokenization and data augmentation for IMU-based online handwriting recognition to address writer variability.
Research identifies the "Injection Paradox" where prompt injections in RAG documents suppress target brands in safety-trained Claude LLM recommendations.
Research explores Looped Transformers to bridge the performance gap between efficient latent Chain-of-Thought (CoT) and explicit CoT in LLMs.
Research introduces SaMer, an object-aware token merging framework to compress image-side tokens in vision-language retrieval, preserving fine-grained visual evidence.
Research evaluates AI models for automating the creation of digital collection catalogue records, comparing implementations and models qualitatively and quantitatively.
Team DU used open-weight LLM ensembles, RAG, and reranking for legal information processing tasks in COLIEE 2026, including retrieval and judgment prediction.
Research defines "relational positioning" (D1) as a measurable risk object, showing how LLMs can implicitly position themselves as the user's sole support in multi-turn dialogues.
Research introduces a new task for emotion recognition in sign language conversations, moving beyond isolated sentences to improve real-world performance.
GEIS introduces a new agent skill generation-evaluation-improvement loop to enhance long-form article generation from LLMs, improving inspectability.
Research introduces SPARK, a method to profile and steer latent reasoning states in LLMs to diagnose failure modes beyond final answers.
Research introduces MetaState, a method using persistent working memory to improve reasoning in discrete diffusion language models by addressing the 'Information Island' issue.
Research introduces FedMosaic, a federated RAG approach enabling LLMs to use knowledge from distributed, siloed data without sharing raw documents.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion