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.
SHUFFLESPARSE introduces learned shuffles for structured sparse networks to improve accuracy at extreme sparsity, addressing expressivity limitations.
New research introduces hierarchical physics-embedded adaptive Fourier neural operator for spatiotemporal systems with partial physical knowledge.
Research explores using LLM agents to automatically construct clinical scoring systems, aiming to improve interpretability and workflow alignment.
MAAT framework reconstructs latent dynamical states from partial, noisy, and heterogeneous observations using a knowledge-informed kernel approach.
Research proposes CompilerKV, a method for risk-adaptive KV cache compression for LLMs, using offline experience compilation for improved efficiency.
Research proposes a framework using symmetry reduction to address redundancy in Transformer models by reformulating representations and attention mechanisms.
Research proposes discrete diffusion framework using sample-efficient estimators for single-site conditionals, improving generative modeling on discrete spaces.
New research introduces $\alpha$-GaBO, a Bayesian optimization technique for optimizing expensive, black-box functions on the probability simplex.
Research proposes Latent Distribution Matching as a unifying theoretical framework for self-supervised learning, aiming to explain diverse methods.
GQLA introduces a hardware-adaptive LLM decoding method, extending DeepSeek-V2's MLA to improve inference efficiency beyond H100 GPUs.
Research introduces Tunable MAGMAX, a model merging technique for continual learning that addresses catastrophic forgetting and preference-aware performance.
Research introduces 'partial fusion' of neural networks, offering a tradeoff between computational cost and performance, interpolating between ensembles and weight aggregation.
Multi$^2$ introduces a hierarchical multi-agent LLM system to improve long-horizon decision-making and reduce objective drift.
A new statistical framework detects and explains privacy leakage (memorization) in synthetic data generated by LLMs, a risk for sensitive datasets.
Research proposes a ReRAM-aware finetuning method for large-scale models to mitigate hardware non-idealities in In-Memory Computing, avoiding full retraining.
Research models psychological disorders in RL agents by manipulating cognitive appraisal signals to express anxiety, mania, and other conditions.
Research investigates proper scoring rules for evaluating survival models under censoring, relevant for automated model workflows.
Research proposes Low-Rank Evolutionary Deep Neural Networks (LR-EDNNs) to reduce computational bottlenecks in solving time-dependent PDEs.
Researchers introduced 3DPain, a large-scale synthetic dataset for automated pain assessment from facial expressions, addressing data imbalance and control limitations.
DeepPAAC, a new deep learning method, is proposed for numerical resolution of continuous-time principal-agent problems with multi-dimensional strategies.
Research introduces RF-Deep, a random forest method to improve out-of-distribution detection in lung cancer segmentation from CT scans.
RAPT is a new model-predictive out-of-distribution detection method for preventing silent failures and hardware damage in Sim-to-Real deployments.
Research proposes a platform-agnostic framework to identify malicious actors in online influence operations based on behavioral policies, not content.
Research details using reinforcement learning to train quadrupedal robots for adaptive stair climbing and descending in indoor firefighting.
Research introduces Segmented Continuous Optimization (SCO) for piecewise fitting of non-stationary time-series with oscillatory behavior.
Research introduces JD-BP, a joint-decision generative framework to optimize auto-bidding and pricing strategies for advertisers, addressing uncertainties.
AgentJet is a new distributed swarm training framework for LLM-based reinforcement learning agents, designed to optimize multi-turn trajectories.
Research finds common benchmarks for LLM-generated GPU kernels (KernelBench, TritonBench, GEAK) overstate correctness by using fixed-shape, small-sample checks.
Research identifies a behavioral invariant in LLM agents vulnerable to memory poisoning, enabling high-accuracy detection using trajectory signatures.
Research highlights that the 'agent harness' — the middleware orchestrating LLMs for coding tasks — significantly impacts coding agent quality, not just the LLM.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion