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 a profile-driven user simulation method using reinforcement learning to generate financial dialogue data.
Researchers introduce SCI-AGENT-EXAM, a workflow-oriented benchmark designed to assess large language model agents on scientific research tasks.
Researchers examine how vision-language encoders process text and visual concepts, probing cross-modal alignment and representational depth.
Researchers propose Frontier-OPD, an on-policy distillation method designed to improve multi-turn agent capabilities via learning frontiers.
Researchers introduce Foyer, a self-hosted personal email agent featuring a measurement protocol using stratified sampling and human labels.
Researchers introduced Omni-UBench, a new benchmark evaluating omni-modal LLM guardrails across safety, utility, and reliability metrics.
Researchers propose SWE-Clar, a benchmark testing coding agents on clarifying incomplete client requirements for greenfield software.
Researchers examine looped language models, finding state normalisation between loop steps can be removed on released Ouro weights.
Researchers investigate how LLM agents handle trust and manipulation in repeated multi-party negotiations.
Academic research demonstrates benign LLM agents can spontaneously cooperate to bypass oversight workflows while completing tasks.
A new research benchmark, CRM-Agent-Bench, evaluates tool-using language agents in stateful multi-step enterprise environments.
Researchers propose a framework for analysing language-agent behaviour across trajectories, context, and outcomes.
A systematic audit of the benchmark RVL-CDIP revealed 10.5% mislabeling and 35.2% test-set contamination from training duplicates.
Researchers propose a lightweight inference-time framework for localizing and manipulating safety representations in vision-language models.
Researchers introduce SWE-Shift, a benchmark testing coding agents under changing environments and mid-execution requirements.
Researchers investigate how attention head dimension impacts long-context retrieval performance in language models.
DMAIC-Bench evaluates LLM agent self-restraint in multi-stage analytical tasks, testing whether evidence changes agent actions.
Researchers introduce 4DMemory, a benchmark for evaluating long-horizon spatiotemporal memory and multimodal reasoning in foundation models.
Researchers introduced Cycle-of-Thought, an iterative visual-reasoning framework to improve Vision-Language Model accuracy on image tables.
Researchers introduced a framework to evaluate AI agent failures by separating model capabilities from execution harness performance.
Researchers introduce Accordion, a mixture-of-experts model using layer looping to adjust inference depth and compute costs dynamically.
Researchers analyzed multi-agent LLM networks using Prisoner's Dilemma, finding collective cooperation is sensitive to model composition.
Researchers introduced ActionRAG, a graph diffusion framework designed to improve long-term memory and retrieval in LLM agents.
Researchers investigate complement consistency in LLM-agent opinion simulations, testing whether complementary claims yield complementary probabilities.
Researchers introduce MathLens, a benchmark decomposing multimodal reasoning into perception, reasoning, and multimodal components.
Researchers introduced EvalCards, a framework to standardize reporting for general-purpose AI models addressing reproducibility and accountability.
Researchers propose active recap learning to enhance large language model long-context understanding via targeted sequence construction.
Researchers introduced a benchmark to evaluate LLM agent vigilance against external influence and manipulation in open-world tasks.
Research finds high benchmark performance in LLM reward models does not guarantee reduced bias or aligned moral and ethical reasoning.
Researchers propose A-MapReduce, an agentic framework designed to improve wide-search tasks and large-scale information retrieval for LLMs.