Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
CoNN-AL framework for data-efficient inverse design in automotive engineering using stream-based active learning and uncertainty estimation.
Method for optimizing area under insertion/deletion curves to improve faithfulness assessment of XAI attribution methods and feature importance rankings.
ORCA identifies compositional failures in text-to-image diffusion models (attribute binding, spatial relations, object counting) stemming from misaligned vision-language representations.
Framework for building 2D-constrained minimal transformers enabling full visualization of internal representations and inference of learned algorithms from geometry.
Benchmark of 7,599 multicomponent material configurations evaluating 11 pretrained ML force-fields against DFT, identifying error sources in high-entropy compositional generalization.
Analysis reveals catastrophic forgetting in LLM continual learning concentrates in output embeddings of rare tokens; proposes data-free mitigation via parameter freezing.
SkillForge: RL method for co-evolving LLM agent skills and policy via dynamic skill lifecycle stages (trial, active, stable, retired) to prevent skill accumulation.
Dual-QK: Rotation-based 2-bit KV cache quantization with query-channel pruning to reduce memory cost in long-context LLM inference while preserving query-key invariance.
Study of homogenization failure mode in multi-agent systems via conformity, polarization, and inertia metrics; evaluated on code generation, hiring, and peer review tasks.
UltraText Bench evaluates dense visual text rendering in image generators across 432 bilingual prompts with four-twelve text regions per scene.
FAB attack demonstrates backdoor vulnerabilities in acoustic foundation models via physically realizable triggers without pre-training data access.
BoT-GRPO improves process-reward RL for reasoning via token-level rewards and bag-of-tokens aggregation, eliminating value network overhead.
DisParQ learns spatially grounded, discrete concept representations from frozen self-supervised vision backbones without labels or language supervision.
MIRROR framework shifts LLM personalization from imitation to preference internalization via reference-revealed self-distillation.
Weakly supervised labeling framework for removing EMG artifacts from multichannel EEG using frequency-aware high-dimensional representations.
LatentGRM performs efficient reward modeling via semantic compression and latent reasoning, reducing inference cost versus token-level generation.
Stratechery analysis of Apple's smart home strategy via partner integration and speculation on Amazon's agent positioning.
AdaPS-LiNGAM addresses small-sample causal discovery in linear non-Gaussian acyclic models by adaptive predecessor selection.
Sequential Isolation Methodology reduces LLM inference benchmark variance via controlled regression testing on vLLM across GPU concurrency levels.
Decoupling Architecture (AO-DA) separates logic from persona in edge LLM agents via dual inference paths on INT4 base model with LoRA.
Deep learning autoregressive models now match dynamical systems in weather forecasting; review explores extension to climate prediction under altered forcings.
Bolzano multi-agent system with verifier solves 8 expert-selected mathematical proofs; demonstrates LLM-based proof search on 3,800+ open problems without human guidance.
Mean field theory characterizes fluctuations of nonlinear observables in wide shallow networks trained via SGD.
Offline RL for autonomous driving addresses distribution shift by constraining candidate ego trajectories under interaction constraints beyond policy action space.
PARC-Loc uses partial assignment and relational consistency for text-to-point-cloud city-scale localization across submap boundaries.
Smaller transformer architecture executes Lloyd's k-means clustering in-context with reduced embedding size versus prior work.
Flow matching models with modified training objectives improve multi-step reasoning; proves two-layer Transformer solves graph reachability with steps proportional to target distance.
EntroPrefill uses Renyi entropy to prune mid-prefill context in transformers for RAG with stability guarantees on attention mass.
Shaer framework generates Classical Arabic poetry jointly conditioned on semantic descriptions, meter subforms, and hemistich counts using enriched 116k-poem corpus.
SoftSEEPS differentiable SEEPS approximation trains ML precipitation forecasting; joint RMSE loss improves IMERG dataset at 0.1° resolution.