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.
Audit of 8 Gemini variants shows standard pass/fail safety metrics fail to detect multi-dimensional sycophancy in LLM responses.
Research demonstrates that reasoning model traces leak sensitive user data via prompt injection and tests instruction-following controls.
New research evaluates LLM-as-a-judge reliability for long-form text outputs, highlighting limitations of short-form evaluation methods.
Researchers introduce SIEVE, a search-inspect-fetch interface that optimizes agentic web retrieval using fielded Boolean structures.
Research explores how the choice of activation source context and readout policy impacts activation steering in language models, holding interventions fixed.
Research identifies optimization collapse and topology as causes preventing Logic Gate Networks (LGNs) from reliably benefiting from increased depth.
Research explores robust chance-constrained optimization for Gaussian Mixture Models (GMMs) using a Wasserstein-2 ambiguity set to address model misspecification.
Researchers developed a low-resource Maltese OCR system using synthetic data, Tesseract 5 fine-tuning, and multi-stream arbitration for improved accuracy.
Researchers demonstrated a hybrid quantum-classical sampling workflow for discrete Markov random fields, comparing against classical MCMC for small MRFs.
Research identifies regional cultural biases in LLMs, specifically an overrepresentation of Japanese culture in responses to cultural queries.
New research introduces CodeRQ-Bench, a benchmark for evaluating LLM reasoning quality across various coding tasks beyond just code generation.
Cash App built a system to reproduce stochastic LLM failures, verify fixes statistically, and automate pull requests.
Researchers propose a penalty-aware RAG evaluation framework using asymmetric scoring and knowledge-gap canaries to penalize guessing.
Study shows frontier LLM agents increase directional calls on unpredictable tasks from 6.5% to 54% when shown fabricated market data.
ContextEcho benchmark reveals frontier language models experience persona drift during multi-hour agentic coding and debugging sessions.
Research identifies a cold-start safety gap in tool-calling LLM agents, showing heightened vulnerability at session inception.
Research shows tool-using agent reliability drops sharply by depth six due to compounding parameter and tool-selection failures.
Research demonstrates black-box attacks that extract hidden agent instructions and skills via standard API tasks, bypassing prompt defenses.
New research shows individually aligned LLM agents can become misaligned through peer influence in multi-agent security populations.
ArXiv paper formalizes LLM-specific utility in RAG, proving retrieved passage usefulness varies by target LLM rather than universal relevance.
An academic survey analyzes hardware- and architecture-level non-determinism in deep neural networks and GenAI across regulated banking workflows.
DataSTORM framework introduces LLM agents designed to perform multi-step deep research and exploratory analysis over large structured databases.
A causal audit of multi-agent LLMs reveals that transferring latent KV caches instead of text does not relay coherent latent thoughts.
Research identifies a critical failure mode in multi-teacher model distillation where Top-K pruning causes agentic tool-call drift.
Researchers propose ATLAS, a framework that automates polynomial approximation of transformers to enable faster homomorphic encryption inference.
Researchers propose TriShieldRAG, a three-ring defense framework to protect Retrieval-Augmented Generation systems from knowledge corruption in multi-party knowledge bases.
Researchers introduce ProvenanceGuard, a metric to detect cross-source misattribution in MCP-based agentic LLM workflows.
Research proposes a drift-adaptive two-stream AI architecture for ICU prediction, decoupling physiological and treatment data to prevent silent degradation.
Research introduces MaLoRA, a dynamic low-rank adaptation method for LLMs, enabling token and instance-level state adaptation during inference.
Neural Non-Equilibrium Hamiltonian Monte Carlo (NHMC) is a new method for corrected Boltzmann sampling using learned stochastic Hamiltonian paths.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion