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 introduced EpiSEM-Bench, a benchmark evaluating tool-using LLM agent safety and execution when evidence is missing.
An arXiv paper introduces a selective auditing methodology for online monitoring of AI decisions using verifier agents and confidence scores.
Researchers propose a fragility-aware classification framework to measure and mitigate the risk of confident misjudgments in risk assessment models.
Research paper examines how increased computational resources enable transferable adversarial attacks and defenses in classification tasks.
Researchers propose MarkTune, a model-embedded LLM watermarking method designed to balance quality and detectability without inference overhead.
Researchers introduced Brittlebench, a framework measuring LLM performance degradation caused by noisy or rephrased user prompts.
Researchers propose 'invariance auditing', a diagnostic tool for machine learning feature extractors to check invariant sets.
Researchers proposed an attack-agnostic neural network robustness metric based on the Fisher Information Matrix spectral norm.
Researchers introduced PROACT-Agent, a framework synthesizing synthetic trajectory data to train real-time guardrails for AI agents.
Researchers introduced MAVOS-DD, a 350-hour multilingual audio-video deepfake detection benchmark covering eight languages.
Research investigates pretraining contamination in medical vision-language benchmarks, showing that public evaluation sets may overlap with training data.
Researchers propose a decentralized learning framework addressing both privacy attacks and Byzantine poisoning via robust aggregation.
A new research paper introduces ATTUNE, an automated method for tuning LLM training throughput across varying GPUs and frameworks.
Researchers introduced WhatIf-Bench to evaluate how time series foundation models handle scenario analysis and intervention-aware forecasting.
Researchers introduced TauIndianBankBench, a benchmark evaluating whether conversational AI agents accurately follow Indian banking procedures.
Researchers introduce FusionEval, a benchmark assessing multi-expert model fusion across long chains of thought and tool use.
Researchers introduce CiteResolve, a benchmark of 423 hallucinated citations for testing source recoverability in text generation.
Researchers propose Self-Assessment Reinforcement Learning to improve search-state awareness in retrieval-augmented large language models.
Researchers introduce JSFactCheckBench, highlighting how evaluation scores for LLM fact-checkers depend on design choices.
Researchers present Protegrity-PII, a PII detection model claiming 0.915 typed recall and low-cost GPU inference on a new benchmark.
Academic research reveals flaws in coding-agent benchmarks where test grading relies on undocumented project details.
Researchers propose compiling autoregressive LLM control into policy-exact parallel inference for hierarchical RAG systems.
Researchers propose ULME, a unified model approach coupling text embedding and language generation to address deployment inefficiencies.
Researchers introduced UniGuard, a framework to train AI guardrails that maintain safety consistency across complex agentic workflows.
Researchers introduce TRIAD, a new benchmark designed to evaluate how omni-modal large language models resolve ambiguity across text, vision, and audio.
Research demonstrates LLM persona sampling matches survey answer distributions but fails to recreate cohesive individual respondents.
ReflectPO introduces reflection-guided policy optimization to improve long-horizon task execution for LLM agents.
Researchers introduced IQuest-Q1, a 320B parameter sparse Mixture-of-Experts agentic foundation model optimized for command-line tasks.
Researchers evaluate compact language model encoders from 17M to 1B parameters for closed-set decision tasks like tool selection.
Research on inference-time context augmentation reveals limitations in how models utilize provided input, impacting reliability.