Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
Delta Tokenizer reduces time-to-first-token for agentic LLM serving by tokenizing only incremental session deltas on CPU.
AGORA framework investigates how team composition and social structure shape self-evolving multi-agent LLM systems.
Group Relative Policy Optimization trains Qwen2.5 and Llama3 on perception-to-action rules via reinforcement learning for Theory-of-Mind reasoning.
StateCraft decouples abstract state from visual state in video world models to maintain consistency across long-horizon interactions.
JetSpec++ co-optimizes speculative decoding via intra-block self-refinement and inter-block foresight drafting to hide draft model cost.
Decentralized active-inference agents per HVAC zone minimize expected free energy without inter-agent communication for multi-zone control.
Benchmark comparing 10 self-distillation methods in agentic RL under unified conditions, revealing training stability and computational cost trade-offs often obscured in prior work.
LiFi framework generates LiDAR point clouds from multi-view RGB images using geometric and semantic guidance for simulation and perception tasks.
Hybrid RL method combining supervised learning stability with policy optimization performance, addressing hyperparameter sensitivity in standard RL.
PROMETHEUS framework applies causal multi-treatment ranking to population health stratification for resource-constrained clinical care prioritization.
SNACK linearizes graphs as sequences for generative models with structural constraint enforcement via invalid logit masking, enabling GFlowNet training.
Anthropic opens Singapore office and signs MOU with Singapore government on safe AI adoption.
LabOPBench multimodal benchmark: 756 expert-authored questions in organic synthesis and materials science with evidence-grounded decision-making evaluation.
Convergence analysis of α-stable Lévy score-based generative models with non-asymptotic bounds on sampling and heavy-tailed data modeling.
Semi-supervised graph learning for code-to-architecture mapping using dependency graphs and self-training to recover modular decomposition from source files.
Analysis of looped transformers reveals failure modes beyond computation: recurrent reasoners lose retention or state reusability when iterating past training depth.
SPECIFICITY PADDING exploit shows prompt-prediction measures of human contribution in AI-assisted writing are unreliable and easily gamed.
Demonstrates backdoor injection into flow-based IDS models via 0.001% data poisoning; proposes diagnostic methodology to localize triggers.
PRISMATIC multi-agent architecture coordinates specialized medical agents on blackboard system to reduce prescribing errors for multimorbid patients.
CAST history-adaptive spectral clipping improves gradient conditioning for MoE pretraining; 19.08B params, 83.89B tokens, lower train/validation loss.
Qwen-MT-Lumen open-source model for WMT2026 translation task; four-stage pipeline combining pretraining, SFT, quality-aware training, on-policy distillation.
ReSemReason framework addresses reachable-but-invalid problem in biomedical multi-hop QA by validating relation-semantic reasoning over knowledge graphs.
DSF method extracts generative priors from pretrained discriminative model loss gradients to enable zero-task-guidance image generation.
Wenyan Thinking rewrites chain-of-thought reasoning in Classical Chinese to reduce token budget for Qwen3-4B-Thinking-2507 inference.
ProtModernBERT applies modern architectural refinements to encoder-only transformers for low-compute protein language modeling on amino acid sequences.
MakSafe spatiotemporal semantic framework formally specifies and verifies road vehicle safety rules via semantic zones and trace-tree reasoning.
Theoretical result: proportional cake cutting under equal-length constraint requires exactly 2n-2 cuts; improves prior quadratic bound.
Empirical study identifies three distinct failure modes in goal misgeneralization across Procgen and robotics tasks: functional substitution, policy shortcuts, and cue competition.
ARP-CD training-free framework reduces hallucinations in video LLMs via attention-guided perturbations in contrastive decoding without model retraining.
TIMERSI deploys time-series foundation models as self-improving agentic systems using coding agents to refine pipelines on historical data without retraining.