The Archive

Shows one-step gradient delay in asynchronous pipeline parallelism does not harm LLM pretraining convergence, enabling efficient GPU utilization at scale.

GROW² enables robots to ground open-world tool use by hierarchically decomposing affordance grounding into object-part semantics and action localization.

Conservative offline training amplifies reward hacking in Qwen3-14B during online RL adaptation, contradicting conventional safety wisdom about policy constraints.

DOPD mitigates privilege illusion in on-policy distillation by decoupling transferable knowledge from non-replicable information asymmetries via dual supervision.

Theoretically explains why embedding norms in scale-invariant contrastive models encode semantic specificity despite being ignored by cosine similarity metrics.

Agents-A1 (35B MoE) achieves trillion-parameter-scale performance by scaling agent horizon to 45K-token trajectories and cross-domain abilities without parameter scaling.

C²R regularization eliminates feature splitting and absorption in large Sparse Autoencoders via cross-sample consistency constraints for improved LLM interpretability.

MESA framework identifies critical communication channels in multi-agent systems to prioritize security defenses against inter-agent attacks.

The startup, which runs a popular free AI leaderboard, launched its commercial service just last September.

Paper analyzes backdoor attacks and defenses in semantic communication systems over wireless multiple access channels.

Study shows cognitive heuristics bias LLM vulnerability detection in code, revealing systematic vulnerabilities in AI-assisted security tools.

Hybrid detection framework for crypto-ransomware targeting enterprise shared storage and network drives.

Uncertainty-aware decision-making algorithms for LLMs based on Bayesian theory to improve reliability in ambiguous tasks.

Single-stage geo-localization framework with 220k+ ground-satellite image pairs for cross-view object matching.

Theoretical framework establishing sample complexity bounds for valid transport map estimation in generative models.

SWE-Interact benchmark simulates realistic multi-turn developer workflows with vague requirements and iterative feedback for coding agents.

Mixture of Experts architecture for unified malware classification, packing detection, and family attribution across diverse binaries.

Study finds LLM conversations exhibit model-specific attractor states across self-play and mixed-play debates on controversial topics.

Detects memory-poisoning attacks on LLM agents via behavioral invariants in memory-tool call sequences, achieving 0.9563 AUC.

Cursor has launched a new mobile app for remote oversight over coding agents.

Proposes joint optimization for hybrid attention layer selection in Transformers, treating layer importance as interdependent for long-context efficiency.

Human Creativity Benchmark separates evaluator convergence and divergence in creative AI evaluation, preserving taste-based disagreement as signal.

TraceLab releases 4,300 real coding-agent sessions with 350K LLM steps for analyzing serving workloads across agents and models.

Analyzes convergence of continual classification in homogeneous deep networks via projection theory.

Poller uses LLMs as poetry evaluators by role-playing authors for Chinese poetry understanding tasks.

Linguistic Firewall improves multi-agent routing by using geometric profiling to verify agent competence empirically rather than self-description.

Hybrid quantum-classical workflow for protein-ligand binding energy via generative-ML-assisted quantum selected CI on Fujitsu FX700.

Clover tool tracks student engagement with AI code completion via behavioral metrics and attention checks to measure critical evaluation.

OTF decomposes visual transitions into sparse primitives to learn robust latent action representations in multi-object scenes.

TRACE framework detects emotional entrainment in dyadic speech via temporal modeling; introduces DyadEE dataset with synthetic disruption controls.