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 introduce Latent Critic, a LoRA adapter detecting and localizing LLM agent hallucinations in real time with low latency.
Research proves LLM acquisition agents require a minimum reward Signal-to-Noise Ratio to learn per-instance signal routing decisions.
Researchers created a quantitative metric framework mapping output dissimilarity and behavioral drift across 32 models from six LLM families.
Research shows latent social associations in LLMs do not reliably predict biased downstream decision-making, challenging bias test norms.
InSight-doc proposes an agentic visual framework using adaptive resolution zoom to improve long-document parsing efficiency without retrievers.
V-FiLLM introduces an evaluation framework that generates deterministic financial reasoning benchmarks from symbolic computation trees.
Research identifies 'catastrophic remembering' in agentic coding instruction files, where prompts grow unbounded due to high pruning risks.
Research reveals tool-calling evaluation pipelines for LLM agents are highly sensitive to minor prompt and pipeline variations.
Probing pre-generation activations allows LLMs to predict failure before generating tokens, enabling dynamic compute routing for reasoning.
An empirical study of 200 users reveals computational XAI correctness metrics do not reliably predict actual human understanding.
Research introduces a Multi-Group IRT framework to isolate causes of cross-lingual LLM safety guardrail degradation.
Researchers propose Entropy-Guided Supertokens to compress LLM reasoning traces, splitting tokens into structural and organic types.
Researchers propose Policy-Masked Private Experts to restrict forward-pass routing to authorized MoE parameters based on policy.
Researchers demonstrate that Rectified Flow generative models leak membership signals of training data along their interpolation paths.
Research shows LLM-as-a-judge evaluation accuracy degrades when evidence extraction is separated from final verdict generation.
Researchers introduce Leak It, a probabilistic method to extract training data from black-box LLMs by analyzing output sample distributions.
Researchers propose Living-Harness, a self-evolving agent framework that dynamically updates its own execution harness to prevent recurring failures.
Researchers propose AgentSnare, a defense framework that uses deceptive environment observations to mislead and disrupt autonomous LLM pentesting agents.
A forensic audit of a radiology VLM benchmark found inconsistencies across datasets, DICOM rendering, prompts, APIs, and statistical code artifacts.
Research explores market designs for AI model training data, aiming to compensate human content creators beyond current 'free-for-all' or 'strong IP' models.
Euclid-MCP proposes a standardized protocol server for integrating LLMs with Prolog-based symbolic reasoning, aiming for reliable logical outputs.
Researchers propose a statistically-grounded method for sparse-feature intervention in LLMs, enhancing activation steering for behavioral control without fine-tuning.
Research highlights gaps in current LLM benchmarks, arguing they fail to measure analytical knowledge work and judgment critical for white-collar tasks.
Research introduces PUPPET, a taxonomy and resource to predict human belief change in dialogues with manipulative LLMs.
New research introduces an exact method for measuring state usage in selective state-space models (SSMs) like Mamba, detailing how information flows.
Research introduces TSCoNet, a two-stage Copula CNN-LSTM model for uncertainty-aware spatio-temporal forecasting of correlated environmental variables.
Research proposes a two-rate error measurement for LLM protocols to audit correction vs. corruption, improving understanding of their impact.
Research finds automated evaluation of LLM agents is unreliable, with errors propagating through tool-use chains. Benchmarked 9 LLMs.
InfiniteScienceGym is a new procedurally generated benchmark for evaluating LLMs on scientific reasoning from empirical data, aiming to overcome biases in human-curated datasets.
Research identifies decision boundary proximity as a common cause for miscalibrated confidence and paraphrase sensitivity in medical Vision-Language Models.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion