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 a framework to convert text-based tool-calling benchmarks into audio-based evaluations for voice agents.
Researchers propose GENIE, a fine-grained evaluation metric to measure novelty in large language model outputs along task-specific features.
Researchers proposed Output Vector Editing (OVE), a post-hoc technique to mitigate training data memorization in large language models.
Researchers introduce DiaLong, a new benchmark dataset designed to quantitatively assess long-context memory and recall in LLMs.
The new VisFacor benchmark reveals multimodal LLMs lag human performance on foundational visual perception and Gestalt principles.
Researchers examine vision-language models for data visualization understanding, testing multimodal reasoning capabilities.
Researchers propose PRISM, a training-free framework combining prompting and in-context learning for financial information retrieval.
A new academic survey analyzes On-Policy Distillation methods for training large language models using student-generated trajectories.
Academic research finds depth-based dynamic inference in LLMs introduces jailbreak vulnerabilities by bypassing intermediate safety layers.
Researchers introduce Software World, a framework for simulating multi-agent collaboration to scale inference-time compute.
Researchers introduced HybridResearcher, a framework for deep research agents querying both structured databases and unstructured web sources.
Google Cloud announced Agent Compute Engine and expanded agent governance capabilities to manage secure autonomous agent workloads.
Researchers introduced ProPerMem, a benchmark evaluating memory-driven proactive assistance and long-horizon evidence mining.
Researchers introduced BASIS-Bench to evaluate whether speech LLMs distinguish beneficial personalization from demographic stereotyping.
Researchers introduced a layer-aggregating white-box probe architecture to detect LLM agent deception, achieving 98.8% AUC on a benchmark.
Researchers introduce REMORY, a neural memory network using soft tokens to compact long-horizon agent context windows.
Researchers propose SciExam for ENSO, a benchmark assessing language-model agents on building valid low-order stochastic climate models.
Research proposes aligning models using internal probes rather than output signals to prevent superficial compliance in training.
Researchers propose the Determinism-Faithfulness Assurance Harness to evaluate execution trajectories in tool-using financial LLM agents.
Researchers introduced PersonalAI 2.0, a framework improving knowledge graph retrieval and LLM agent search capabilities.
Research shows LLM decision-making conflates protected proxy attributes with predictive evidence, undermining standard demographic bias audits.
Researchers introduced FactReview, an audit framework using LLMs to extract, cross-reference, and execute code to verify technical claims.
Researchers propose Quality-Diversity Red-Teaming, an automated framework to generate diverse adversarial prompts for safety testing LLMs.
A new framework predicts vision-language model performance before training, bypassing costly compute-based scaling trial-and-error.
Academic research identifies mathematical failure modes in bilinear contrastive critics used for LLM reranking and best-of-K selection.
Research shows coding agent token consumption varies significantly across programming languages (Python, Java, Rust, OCaml), impacting inference cost.
Research explored multi-agent Q-learning coordination in a tabular predator-prey gridworld, isolating coordination structure from approximation.
New research introduces DFAH-Bench, a replay benchmark designed to measure observable behavioral instability in financial agent decision-making processes.
Research explores using Vector Symbolic Architectures (VSAs) to decode internal representations and improve interpretability of Large Language Models.
Research proposes MARS, a scalable framework for combining LLMs with knowledge graphs (KGs) using multi-hop retrieval and SPARQL generation for grounded answers.