EMERGE Benchmark Catches AI Failing to Keep Knowledge Graphs Current
A new benchmark pairs 233K Wikipedia passages with 1.45 million Wikidata edits to test whether models can actually keep knowledge graphs current.
20 stories
A new benchmark pairs 233K Wikipedia passages with 1.45 million Wikidata edits to test whether models can actually keep knowledge graphs current.
Researchers use small language models trained on child-directed speech as simulated learners to generate testable hypotheses about how children acquire dative verbs.
A new PyTorch library turns B-spline and Legendre curves into trainable layers you can drop into KANs, embeddings, and functional regression.
A new theoretical analysis shows REINFORCE-based methods like PPO and GRPO can silently kill the base model's best responses, and offers a fix.
A new framework from Georgia Tech uses Bayesian uncertainty to tell SAM exactly where to ask for its next click, beating human annotators on 14 of 16 datasets.
A new evolution-strategy approach ditches random subspace projections for black-box prompt tuning and beats BBT across seven GLUE tasks.
An EPFL team gave frontier LLMs trained EEG and ECG network weights as text and asked them to debug, repair, and even write networks from scratch, without any gradient descent.
A new splatting method extends 3D Gaussian Splatting from points to connected curves, producing strands that work for hair, wires, and physics simulation.
A new ACL 2026 method called CLaRE predicts which facts will break when you edit an LLM, using forward activations instead of gradients.
A new method discovers compact, interpretable polynomial learning rules per layer, offering a symbolic alternative to backpropagation for small networks.
A research prototype streams a live 3D Gaussian Splatting model instead of video frames, letting thin clients render their own viewpoints from cloud scenes.
A new benchmark tests whether multimodal models can use eye-gaze to follow attention, infer intent, and predict actions in live egocentric video.
Google DeepMind researchers show a coordinated swarm of attacker agents still cracks frontier models that human red-teamers struggle to break.
Stanford and ETH Zürich researchers introduce a per-token mixed-policy self-distillation method that cuts supervised tokens by 38 to 69 percent while preserving general skills.
A new research paper from TU Munich shows that cloud rendering can skip path tracing entirely and still match reference quality in about 10 milliseconds.
Stanford researchers show AI scientists invent more original ideas when they get less task information up front, then learn from past experiments to pick winners.
LAION releases a 365M-clip audio dataset with 707M captions and a family of open CLAP2 models that set new state of the art on MAEB-30.
A new survey maps how TinyML is reshaping farm sensing on ESP32, STM32, and UAV hardware, exposing serious gaps in energy and memory reporting.
Deletion Tuning learns a binary keep-or-delete mask over gate projections, matching full fine-tuning accuracy while zeroing under 0.05% of weights.
A new 500-question long-video benchmark forces models to not just answer correctly but prove where in the video the evidence lives. Frontier systems collapse.