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.
Researchers propose a direction-aware loss function for time series forecasting to improve upward/downward move prediction accuracy.
Researchers introduce an observability ladder assessing how well model-written summaries of hidden reasoning traces reflect actual accuracy.
Researchers propose Gecko, a framework for fast private inference that secures public encoder offloading to protect user data and model weights.
Academic research analyzes why frontier LLMs structurally fail at predictive analytics over tabular data compared to specialized models.
Researchers introduce PRECOG, a retrieval mechanism for State-Space Models that reduces prefill costs to O(1) by injecting pre-computed states.
A research paper analyzes 761 million GitHub Copilot production traces, revealing distinctive workload properties of agentic coding systems.
Researchers propose a novel framework for causal inference using unstructured data treatments like text, images, or clinical decisions.
Research introduces a framework to evaluate enterprise AI agents by replaying temporal, permission-aware data states rather than static snapshots.
Researchers demonstrate a novel backdoor vulnerability in multi-agent systems triggered collectively by peer communication thresholds.
Researchers used repeated game theory to show frontier LLMs adopt extreme, rigid safety-risk policies in multi-agent competitive races.
Researchers propose Climate-Dyna, a model-based reinforcement learning framework to optimize residual climate hedging valuation adjustment (HVA) and hedge discovery.
Researchers propose a projection-consistent neural operator to speed up local-stochastic volatility calibration by avoiding iterative loops.
Researchers prove that runtime safety monitors using linear temporal logic fail to block attacks due to LLM response entropy variations.
Researchers identify structural flaws in using reference-based text evaluation metrics as optimization targets, as models learn to game them.
Researchers identify that LLM-based agent replanning creates heavy-tailed latency bottlenecks as accumulated context windows grow.
Researchers identify a fundamental performance ceiling in ML models that pair long-term aggregate labels with short-window observation inputs.
Researchers introduced semantic self-segmentation, a method where coding agents declare their own trajectory boundaries to improve training data.
Researchers propose a cooperative coevolution strategy to make full-parameter post-training of agentic LLMs viable on limited GPU hardware.
Researchers propose a low-cost telemetry-based monitoring system to detect and repair LLM agent failures without expensive LLM-as-a-judge evaluators.
Researchers propose a privacy-aware sparsity tuning method to protect over-parameterized models against membership inference attacks.
Researchers propose MUSS, an algorithmic approach to select relevant and diverse subsets, optimizing retrieval-augmented generation (RAG).
An academic survey identifies intersectional bias and fairness vulnerabilities in Augmented Graph Learning architectures.
Researchers propose Quality-Diversity Red-Teaming, an automated framework to generate diverse adversarial prompts for safety testing LLMs.
Researchers propose a low-cost, black-box method using 'Border Inputs' to detect silent under-the-hood changes in hosted LLM APIs.
Researchers propose a safety filtering framework using constricting barrier functions to guarantee hard safety constraints on generative models.
Researchers introduce OTora, a framework targeting Reasoning-Level Denial-of-Service (R-DoS) in LLM agents by inflating reasoning depth.
Researchers propose a method to dynamically scale inference compute by recursively re-running top decoder layers during high-uncertainty tokens.
Researchers propose a new anomaly detection method designed specifically for high-dimensional, heterogeneous relational databases.
Researchers proposed HERALD, a system optimizing CPU-GPU KV cache offloading to improve serving throughput for long-context LLMs.
Researchers developed a history-aware financial path generator using Denoising Levy Probabilistic Models for exotic derivatives trading.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion