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 propose minimal-edit on-policy distillation to improve visual mathematical reasoning in smaller vision-language models.
Researchers introduce SWE-Verifier, a lightweight generative model designed to selectively verify coding agent actions and reduce API costs.
Researchers examine portfolio allocation using reinforcement learning and LLMs to process aging signals from SEC disclosures.
A new academic survey on time series anomaly detection reviews recent architectures including diffusion models, state space models, and LLMs.
Researchers introduced Federated GraphRAG to enable Graph Retrieval-Augmented Generation across distributed, privacy-constrained data.
Researchers propose Gandalf, a method to extend flash storage lifetime for KV cache offloading in LLM serving workloads.
Researchers introduce Trace2World, a method for separating behavior reconstruction from code implementation to scale GUI agent training environments.
Research demonstrates that foundation models fail to maintain consistent knowledge retrieval across different molecular identifiers like SMILES and names.
A new research paper examines how hard latency and energy constraints alter time-series model selection in resource-limited settings.
Researchers propose XALIGN, a visual analytics workflow to compare rule-based and other explanations for high-stakes domains.
Researchers propose a personalized federated learning protocol combining geometric consensus and support-adaptive statistical personalization.
New research evaluates LLM inference optimizations like quantization and speculative decoding on Apple Silicon architecture.
Researchers propose a method for small language models to accurately select tools from large repositories exceeding 10,000 APIs.
Researchers propose a shift-aware score transport method for empirical privacy auditing via membership inference without model retraining.
Researchers propose CRANE, a post-hoc counterfactual explanation method designed for classifiers handling multimodal inputs.
Researchers analyzed OLMo-3 checkpoints, finding reinforcement learning with verifiable rewards may not induce power-seeking behaviors.
Researchers propose KV-sharing to mitigate linear memory growth in looped transformer architectures during inference.
Researchers introduced CausalCoT, a method to prevent attention leakage and ensure chain-of-thought rationales causally drive model output.
An empirical study measures handshake overhead in AI agent protocols, detailing setup latency across frameworks like MCP.
Researchers introduced IRAH, combining tree search with meta-learning to optimize RAG retrieval pipelines under limited evaluation budgets.
Researchers propose a Viterbi tracking method to detect synchronous AI model training via facility-level power consumption patterns.
Researchers propose agentic verified optimization, combining coding agents with formal verification to prevent test exploitation.
New research from arXiv proposes testing frameworks for continuously evolving AI analysts to control cumulative Type I error risks.
Researchers introduced LLVMhammer, showing Rowhammer DRAM attacks can degrade GPU LLM inference speed and cost without altering output.
A new arXiv paper introduces a partition theory of AI evaluation, demonstrating fundamental mathematical limits on model oversight.
Researchers introduced ReDence, an open-source hybrid framework combining agents and decision models to detect social engineering scams.
Research shows offline verification tests have mathematical limits and can fail to detect safety risks in deployed autonomous agents.
Researchers study attribution problems in multi-agent LLM pipelines for requirements engineering, separating prompt quality from typed contracts.
Researchers propose AgenticEvolve, an evolutionary program search framework using persistent LLM agents to design and run experiments.
Researchers propose Conservative Meta-Agent Search (CMAS) to manage switching costs and risk constraints in non-stationary environments.