Framework for versioned AI component governance in embodied agent systems with compatibility checking and rollback at deployment lifecycle.
Technique for offloading KV cache to reduce memory and latency bottlenecks in long-context LLM inference while maintaining accuracy.
Lightning OPD enables efficient offline on-policy distillation for LLM post-training by precomputing teacher log-probabilities without live server.
SeaEvo uses LLM-guided evolutionary search for algorithm discovery with persistent population-level state management and structured reasoning.
Neural cellular automaton with discrete bottleneck learns compositional semantic parsing rules without hand-written rules, enabling structural generalization.
Persistent Visual Memory module strengthens visual attention in autoregressive LVLMs to prevent signal dilution during long sequence generation.
Meta's Code World Model preparedness assessment for frontier risks including code generation, reasoning, and misalignment evaluation.
Multi-agent RL approach for tactical deconfliction between heterogeneous unmanned aerial system fleets in dense urban airspace.
Cross-lingual safety alignment framework using self-distillation to transfer safeguards from high-resource to low-resource languages in LLMs.
DGPO improves RL credit assignment for LLM alignment on complex reasoning via distribution-guided policy optimization with finer-grained step isolation.
CoREB introduces a contamination-limited multitask benchmark for code search beyond first-stage retrieval, including reranking and developer-style queries.
MACS improves multimodal MoE inference efficiency using modality-aware capacity scaling to address straggler effects in expert parallelism.
Empirical study evaluating privacy awareness of Vision-Language Models when deployed as autonomous cognitive cores in physical environments.
EGA proposes a residual adapter for vector search systems to handle distribution shift with frozen vision encoders.
Active learning method optimizes communication structure in LLM-based multi-agent systems to improve performance and reduce token usage with limited budgets.
CRAFT proposes continual learning for LLMs via low-rank interventions on hidden representations to prevent catastrophic forgetting during fine-tuning.
VideoRouter addresses efficiency bottlenecks in long video understanding using query-adaptive dual routing to compress visual tokens dynamically.
Safety Anchor defends LLM safety alignment against harmful fine-tuning by introducing geometric bottlenecks in parameter space.
VISD: structured self-distillation for video reasoning combines RLVR and token-level credit assignment in VideoLLMs.
Entropy-regularized adjoint matching for offline RL with flow-matching policies reduces popularity bias in behavior distribution.
NOVA world model framework representing state as neural network weights with latent structural disentanglement for interpretability.
NavOne: top-down vision-language navigation using one-step global planning on maps for efficient path reasoning.
Asymmetric on-policy distillation improves token-level student learning by addressing variance and exploration issues.
Q-MMR framework for off-policy evaluation in MDPs using moment matching and recursive reweighting of trajectories.
AI CFD Scientist: open-source AI agent for autonomous computational fluid dynamics discovery using physics-aware reasoning.
RateQuant applies rate-distortion theory for mixed-precision KV cache quantization with per-head bit allocation.
LKV learns head-wise budgets and token selection for KV cache eviction in long-context LLM inference optimization.
Positive-and-Negative Decoding framework reducing object hallucination in vision-language models via attention rebalancing.
Analysis of strain and vorticity in velocity field Jacobians to understand integration error in flow matching models.
Hierarchical ensemble pipeline for anomaly detection in ESA satellite telemetry using shapelet and statistical features.
Toeplitz MLP Mixer architecture for sequence modeling with O(dn log n) complexity as efficient transformer alternative.
Transformer-based sequence models for wildlife species classification from GPS movement trajectories, outperforming LSTMs and CNNs.
Semantic State Abstraction Interfaces for mapping news text to interpretable coordinates in LLM-augmented portfolio decision systems.
Analysis of model-based RL training on imagined trajectories, quantifying effects of dynamics and reward model errors on policy optimization.
Gauge-aware aggregation method for federated LoRA addressing representation dependence in decentralized LLM adaptation.
Quantum-inspired fast weight programmer combining KAN and quantum circuits for scalable sequence learning on NISQ devices.
Geometric Kolmogorov-Arnold Networks using learned Riemannian metrics for geometry-aware approximation in input space warping.
Closed-form upper bound derivation for admissible learning-rate steps in belief-space dynamics using KL/Bregman geometry.
Gradient Extrapolation-Based Policy Optimization for efficient LLM reasoning improvement, reducing computational cost versus multi-step lookahead.
Sparse attention indexing method for efficient LLM inference that guarantees zero false negatives in KV cache selection during decoding.
Adam optimizer revisited for streaming reinforcement learning without replay buffers, enabling online policy updates from continuous interactions.
Calibration method for Process Reward Models using conditional optimal transport to improve inference-time scaling in LLM reasoning.
Error attribution framework for multi-agent LLM systems using conformal prediction with distribution-free coverage guarantees.
Monge Inception Distance metric for evaluating generative models using sliced Wasserstein distance, addressing FID limitations.
Model-to-Data framework improving transparency and explainability of Graph Neural Networks by shifting complexity from models to data representations.
PAC learning theory for autoregressive chain-of-thought reasoning in LLMs, extending online learning analysis to token generation mechanisms.
Self-evolving trading agent using interpretable rubric policy with bounded prompt optimization for financial markets.
Unified measure-theoretic framework showing diffusion, score-based, and flow matching as instances of vector field learning.
Shadow Mask Distillation technique compressing KV cache during RL post-training of LLMs for memory efficiency.
Analysis of benchmark-utility gap in generative AI across 28 real-world deployments, identifying evaluation failures.