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 Holo-NN, a holographic nearest neighbor search method aiming to improve privacy in high-dimensional embedding lookups.
Researchers propose AdaPRA, a coding agent harness designed to reduce process errors in automated data science workflows.
Researchers propose GEMS-RAG, a graph-grounded retrieval-augmented generation method for hierarchical and multimodal engineering standards.
Researchers evaluate Jev 1.13.0 as a structured System 1 model for time-series classification and stock movement prediction.
Researchers introduced DualPace, a framework decoupling scheduling to enable cooperative GPU-CPU decoding for higher LLM serving throughput.
Researchers introduce AutoTavern, a self-improving LLM agent system that builds a competitive Hearthstone Battlegrounds AI via verification gates.
An industrial case study examines training a 92B-parameter coding LLM for over 100,000 developers using software engineering expertise.
Researchers examine coding agents migrating PyTorch workloads from GPUs to TPUs, analyzing success rates, cost, and behavioral shifts.
Researchers propose HIFI, a framework designed to make resampling-based statistical inference with foundation models computationally viable.
Researchers released a battery of 28 propensities with synthetic datasets and rubrics to measure AI system behavioral tendencies.
SearchHorizonBench evaluates how AI agents determine the cost-benefit trade-off of executing additional search queries.
Researchers propose a concept-based property attack method to target internal neural network logics rather than output classes.
Researchers propose Macro Mondrian Conformal Risk Control to extend conformal prediction for general monotone loss functions.
Researchers propose optimizing streaming inference by composing multiple megakernels rather than using a single monolithic kernel.
Researchers propose ShutdownMod, a method to evaluate AI agent shutdown resistance by injecting shutdown requests into agent trajectories.
Researchers introduce FeatureLoop, a benchmark evaluating coding agents on tasks requiring iterative refinement under evolving requirements.
Researchers propose pairwise Bellman validation to address off-policy evaluation model selection without auxiliary regressions.
Researchers propose adapting neural posterior estimators for simulation-based inference in complex economic agent-based models.
A new arXiv preprint explores machine unlearning for vision models via key deletion, addressing privacy and data removal requirements.
Researchers propose AutoPipe, a budget-aware framework for automating configuration selection in LLM post-training pipelines.
Researchers propose direction-informed adaptive learning to optimize test-time compute allocation for LLM agents based on signal direction.
Researchers introduce Markovian Circuit Tracing to test whether transformer activations support compact internal state abstractions.
Researchers introduced Zero-CoT Truncation to detect evasive data contamination in large language models designed to boost benchmark performance.
Research demonstrates setting temperature to zero in LLM-as-judge safety evaluators fails to guarantee deterministic results.
Academic research presents statistical audit methods for evaluating subgroup disparity changes during continuous model updating.
Researchers propose Combinatorial Data Selection to address cross-task sample interactions during LLM fine-tuning.
A recent preprint explores how announced structural breaks impact conformal prediction reliability and frequency calibration in machine learning.
Researchers introduced CAFE, an automated feature engineering framework combining causal discovery with multi-agent reinforcement learning.
Researchers examine whether simple cosmetic modifications can trick AI age estimation systems into misclassifying minors as adults.
Researchers introduced a continuous-time actor-critic framework using normalizing flows to solve jump-diffusion control problems.