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.
New research proposes Asymmetric Peak-Aware Loss (APAL) for time-series forecasting, addressing under-prediction of demand spikes by optimizing for asymmetric risks.
Research explores optimal linear combinations of binary classifiers using truth tables to partition datasets and analyze convexified empirical risk.
Research proposes "Random Logit Scaling" to defend deep neural networks against black-box score-based adversarial example attacks, aiming for scalable, low-cost defense.
Research explores reducing foundation models for dynamical systems to minimal interpretable architectures, focusing on zero-shot reconstruction.
Research explores causal inference in sequential observational settings with interference and latent confounding, using Markovian outcomes and an Ising model.
LongStraw is a new architecture-aware execution stack designed to enable long-context reinforcement learning beyond 2M tokens.
Sparse Identification of Nonlinear Dynamics (SINDy) method is introduced as a surrogate modeling technique for engineering applications with limited data.
Research proposes a hybrid neural–physics framework using ODEs for modeling dynamical systems where some equations or state variables are unknown.
Research identifies a vulnerability in World-Action Models (WAMs) where imagined futures do not reliably prevent incorrect real-world actions.
Research proposes mutable low-rank sketches using KP-trees for retrain-free, real-time updated recommendation system embeddings, tightening prediction error.
Generalized Neural Distributional Regression (GNDR) framework embeds deep neural networks into classical probability distributions using a two-step estimation.
Research proposes using LLMs to construct Bayesian Networks, blending expert judgment and data-driven learning for decision-making under uncertainty.
Research identifies 'behavioral inertia' as a core obstacle preventing LLM agents from effectively incorporating new tools and expanding toolsets.
Research explores enhancing Small Language Model (SLM) reasoning (Gemma 3, Llama 3.2) using knowledge graph grounding within a neuro-symbolic agentic framework.
Pinterest deployed a deep-learning system for causal retrieval optimization to personalize e-commerce content distribution in production.
Research uses template-constrained LLM agents for SAR ADC generation in analog design, addressing LLM hallucination and simulation failures.
Research shows LLM-synthesized code world models (CWMs) with 100% transition accuracy can still fail at planning tasks due to inadequate 'play-adequacy'.
Research evaluates six quantization methods for large code models, assessing impact on generated code quality for resource-constrained deployment.
NexForge introduces a requirement-first framework for generating executable agent tasks, overcoming limitations of substrate-first methods in scaling and domain coverage.
SIRUS is a training-free, inference-time framework for concept suppression in text-to-video models, addressing difficulties in removing target concepts.
New research suggests generative model steering is limited by training data, not model capacity, with 'examples' providing more control than 'knobs'.
Research introduces a method to calibrate segmentation models, often miscalibrated and overconfident with region-based losses, relevant for medical imaging.
NeuralChaos proposes a new method for approximating square-integrable predictable processes, relevant to stochastic control and mathematical finance.
Research introduces CatalogAgent, a supervisor-mediated self-learning system that uses LLM-based generator-evaluator frameworks for product catalog data completion.
ConFlow is a research paper introducing a constraints-guided learning method for flow matching models used in robot motion generation.
Research introduces one-shot generative design for disordered metamaterials using self-organizing neural cellular automata to overcome design complexity.
Research details SAGA, a schema-aware grounding agent for converting natural language questions into SPARQL queries over knowledge bases.
Research presents a novel system for image generation, integrating LLM-optimized negative prompts and latent-space classifier guidance to improve Stable Diffusion output quality.
ExaGEMM explores CPU-driven ML inference for very-low-bit GEMM by optimizing associative in-register computing, addressing efficiency gaps.
Research explores self-supervised learning using cross-modal data from device sensors to specialize models to specific test environments.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion