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 improved Gaussian approximation bounds for the privacy analysis of differentially private stochastic gradient descent.
Researchers propose COPE, a composable performance emulation method for mixture-of-experts model offloading across memory and compute tiers.
An arXiv research paper explores the theoretical foundations and construction of autonomous AI agents capable of long-term planning.
Researchers propose OPAL, a reactive optimizer using a language-model agent to dynamically rewrite training code during a single run.
Researchers propose Knowledge Evolution to address error propagation and memory corruption in continual autonomous agent learning.
Researchers propose recovering missing enterprise configuration rules from historical transactions to automate field predictions in order systems.
Researchers introduce ParetoRSI-Gym, a framework for evaluating multi-domain recursive self-improvement in language model post-training.
Researchers propose dynamic multi-step decoding for time-series foundation models to balance error accumulation and temporal dependencies.
A research paper presents a framework for actionable transparency and valid explanations in algorithmic insurance pricing models.
Researchers analyze finite-sample sensitivity of kernel metrics like KID and KAD used for evaluating generative models.
Researchers propose an approximate unlearning technique for prototype-based classifiers to address data removal requirements.
Researchers propose AT-AT, a method for automatically optimizing agent tools using execution traces from environment interactions.
Researchers evaluated 21 open-weight decision models to test whether direct predictions can replace RAG rerankers with lower latency.
Researchers propose streaming recurrent memory architectures to process continuous high-bandwidth data without large offline replay buffers.
Researchers propose a certified robustness framework using f-divergences to defend machine learning models against backdoor poisoning attacks.
Researchers propose MINT, a graph neural network method using attentive message passing to detect illicit financial transaction accounts.
Researchers propose localized CBFI to recover latent feature interaction structures in complex tabular machine learning models.
Researchers propose Adaptive Memory Index Evolution, a lightweight layer to improve long-horizon agent memory access.
Researchers propose BellowsKV, a new KV-cache compression method sharing latent representations across layers to reduce GPU memory bottlenecks.
Researchers argue machine unlearning should prioritize risk-aware output controls over parameter-space retrain indistinguishability.
A preprint examines multi-objective reinforcement learning for AI data centre siting, power and cooling decisions over long asset lifespans.
Researchers introduce SkillLifeBench to evaluate security risks across the lifecycle of web agent skills and version changes.
Researchers propose InterQ-RL to address execution mismatch in low-precision quantized rollouts for multi-turn agent reinforcement learning.
Researchers introduced StressSim, an adversarial user simulation framework designed to stress-test multi-turn AI agents.
Researchers introduce ExCal, an optimization-free multivariate time series anomaly detector combining raw state retrieval and robust statistics.
Researchers study optimal allocation of strong-verifier budgets in GRPO to mitigate reward hacking during reinforcement learning.
Researchers propose FOCUS, a reinforcement learning method using selective branching to improve long-horizon language-model agent credit assignment.
A new research paper proposes a work-centric, system-level perspective for evaluating and optimizing digital work agents across tools.
Researchers propose Organon, a recursive self-improving agent framework for adaptive scientific discovery across diverse domains.
An arXiv survey examines uncertainty quantification techniques for foundation models across financial QA, forecasting, and decision-making.