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 defines 'evaluation blindness,' where AI monitoring systems fail to detect performance degradation, showing false healthy states.
Research exposes fragility in LLM-driven causal discovery (GoT-CD) used for path-specific algorithmic fairness audits.
Researchers introduced PopFS, a method for selecting a single feature set robust to performance shifts across heterogeneous populations.
Researchers propose SAKI, a score-aware low-rank KV cache compression method to improve attention accuracy in long-context models.
Researchers proposed SAGE, a noise-aware shrinkage method improving differentially private zeroth-order fine-tuning memory efficiency and utility.
Researchers evaluated prompt-based and trained methods to shorten model reasoning traces, measuring impacts on accuracy across GPQA and MMLU-Pro.
Researchers introduce CausalOPD, a distillation method targeting "first-wrong-step" errors to improve causal reasoning in smaller, local models.
Researchers propose DiagLoop, a counterfactual reinforcement learning framework that converts causal guidelines into diagnostic training data.
Research demonstrates that LLM quantization to INT8 and INT4 alters predicted class probabilities and degrades classification reliability.
Researchers propose a closed-form linear mapping to transfer KV caches between different-sized models in the same family, bypassing prefill.
Researchers introduce AgentStream, a framework evaluating how self-evolving LLM agents perform and adapt under continuous streaming tasks.
Researchers introduce KernelBrain, an AI agent system that automates GPU kernel optimization using LLM-guided mutation and budget-aware search.
Researchers introduce DenialRAG, a technique where a single poisoned document in a RAG corpus forces an LLM to deny correct facts.
TraceCompiler compiles noisy and repetitive LLM agent execution traces into structured, mostly deterministic tool workflows.
An empirical study analyzing real-world adoption, performance trade-offs, and system designs of LLM serving frameworks.
Researchers propose a framework for causal inference on unstructured outcomes like text and images, moving beyond traditional scalar metrics.
Researchers exploit vulnerability in flow-matching vision-language-action (VLA) models using adversarial patch attacks.
Researchers propose evaluating AI scientist agents using adversarial, fast-moving real-world environments to prevent benchmark contamination.
Research demonstrates 14 common machine unlearning methods are easily reversed using a simple linear map, failing to permanently erase data.
Researchers propose PRIVEE, a privacy-preserving framework for Vertical Federated Learning to prevent feature inference attacks.
Researchers propose a unified training-serving system combining reinforcement learning with adaptive speculative decoding to speed up LLM serving.
Researchers propose a multi-level annotator modeling framework to address the reproducibility crisis in subjective human evaluations of LLMs.
Researchers introduced a leakage-aware audit framework for conformal triage models to prevent critical errors during environmental shifts.
Researchers identify a vulnerability where reintroducing privileged context to distilled student models degrades inference performance.
Researchers propose Prediction-Enhanced Monte Carlo (PEMC), using ML surrogates as control variates to accelerate complex simulations.
Researchers propose Assimilative Causal Inference (ACI), combining Bayesian data assimilation with dynamical models to trace causal links.
Researchers propose AgenticSCR, an autonomous agentic framework designed to detect early-stage, context-dependent software vulnerabilities.
Researchers propose LoBoost, a model-native local conformal prediction method that improves uncertainty quantification for gradient-boosted trees.
Researchers demonstrate that LLM agents store malicious prompt injections 97.5% of the time, but downstream execution varies independently.
Researchers introduced WebStep, a benchmark of 1,800 tasks evaluating web agents via intermediate semantic states rather than terminal success.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion