Spiking neural network language model with softmax-free attention and spike-aware distillation for energy-efficient LLM alternatives.
Dataset of autonomous AI agent-only social media activity from Moltbook platform, including posts, profiles, and community metadata.
Lagrangian relaxation approach for coordinating large-scale multi-agent systems with shared resource constraints and dynamic population composition.
Adaptive importance sampling method to address rollout-training mismatch in low-precision RL for LLMs, preventing policy gradient bias and training collapse.
Survey of data-dependent worst-case generalization bounds for overparameterized neural networks using non-vacuous guarantees.
Technique eliminating dequantization bottleneck in quantized LLM inference via activation decomposition on AI accelerators.
AI agent using call stack information to automate game testing with improved code coverage beyond RL baselines.
Study of vector merging methods for multilingual knowledge editing in LLMs to reduce language-specific interference.
Explainable AI and statistical analysis comparing tree ensembles, DNNs, and hybrid models for UAV intrusion detection.
Quantum-inspired classical algorithm for finding sparse subnetworks in neural networks via ridgelet transform.
Security vulnerability in speculative decoding for LLM inference where draft tokens can be manipulated to collapse performance.
Study of masked autoencoders for bioacoustic recognition with weak annotations using self-supervised learning.
Sheaf-theoretic framework for detecting theory-shift in AI agents to identify when representational frameworks become obstructed.
Principled reasoning method for LLMs to improve content trustworthiness beyond smooth prose generation.
Deep RKHS network unifying Kolmogorov superposition with RKHS regularization for prediction with finite-sample guarantees.
Large-scale autoformalization benchmark with 52k graduate-level math theorems, definitions, and proofs from 103 textbooks.
Replication study evaluating toxicity in LLMs trained on web-scale data and testing mitigation strategies while maintaining utility.
Conformal procedure for aggregating multiple chain-of-thought reasoning paths with uncertainty quantification to improve reasoning reliability.
Edge-cloud cascaded architecture for automated diabetic retinopathy screening in rural areas with limited bandwidth using deep learning.
Agentic workflow framework combining interpretable prototype networks with LLMs for clinical diagnostics while avoiding hallucination through privacy-aware design.
ASH system that learns embodied policies for long-horizon tasks from unlabeled internet video through self-improvement loops and inverse dynamics models.
Reinforcement learning framework for genetic circuit design using code generation and hierarchical verification rewards for synthetic biology.
Transformer-based real-time catheter tip tracking system for autonomous robotic navigation in mechanical thrombectomy procedures.
Training-free generative sampling method using moment-matched score-smoothed overdamped Langevin dynamics without requiring neural network training.
Proposes MMGuard to prevent unauthorized fine-tuning of vision-language models through proactive data protection mechanisms before training occurs.
Formal verification approach for transformers using ReLU-catalyzed abstraction refinement to verify complex attention mechanisms in safety-critical applications.
Method for concept-level machine unlearning in vision-language models via interpretable concept decomposition to remove target knowledge precisely.
Studies reasoning necessity in chain-of-thought traces by analyzing minimal cores within overcomplete reasoning to understand how much intermediate reasoning preserves model predictions.
Proposes semantic-space alignment with reinforcement learning to extend LLMs to low-resource languages while avoiding performance degradation in general capabilities.
Nexus agentic framework for time series forecasting integrating LLMs with specialized foundation models using contextual textual signals.
MIRAGE framework discovers semantic attacks bypassing adversarial defenses to degrade autonomous vehicle HD map construction.
RL framework for training prompting policies that optimize multi-step reasoning and tool-use in frozen LLMs via iterative distillation.
Intelligence Impact Quotient metric framework quantifying organizational AI integration depth and impact beyond usage metrics.
PRISE framework for learning scenario reduction in two-stage robust optimization with discrete uncertainty.
Deep reinforcement learning method for dynamic rebalancing of dockless bike-sharing systems using graph-based simulation.
ArcGate activation function with learnable parameters offering adaptive non-linear transformations for deep networks.
Theoretical analysis of how scaling laws emerge from feature learning in hierarchical multi-layer neural networks.
ML classifiers and LLM-based approach to automatically identify and rank refactoring opportunities in BDD test suites.
Research showing existing defenses against malicious fine-tuning of foundation models fail under adaptive adversaries.
SepsisAgent augments LLMs with clinical world models to simulate patient dynamics for sepsis treatment recommendations in ICU settings.
Video2GUI: Automated framework extracting interaction trajectories from videos to generate large-scale training data for GUI agents, addressing data scarcity in multimodal LLM-based automation systems.
Method extending LLM intervention techniques beyond linear representations to non-linear manifolds for model interpretability.
Study showing LLM browser agents can be fingerprinted via UI interaction traces and timings across 14 models, enabling targeted attacks.
Non-monotone preconditioned trust-region method for parallel neural network training with domain decomposition.
Fairness metrics for multi-agent resource competition using rotational periodicity and sliding-window measures.
Analysis of vision-language model embedding spaces showing 164 noise dimensions in CLIP through spectral covariance decomposition.
Research on pluralistic AI alignment arguing aggregation methods produce sycophantic consensus rather than genuine value pluralism representation.
SpeakerLLM audio-LLM specialized for speaker understanding and verification reasoning in conversational AI applications.
AsyncFC framework enables concurrent function calling in LLM agents through asynchronous execution, decoupling decoding from function calls.
Theoretical analysis of kernel regression with gradient outer product for recovering low-dimensional structure in multi-index models.