Research on budgeted attention allocation for transformers, enabling multiple cost-quality trade-offs in deployed systems via head-gating mechanisms.
int4 KV cache quantization technique for Apple Silicon that runs faster than fp16 via fused Metal kernels, implemented as HuggingFace Cache subclass.
Empirical study of proactive LLM-based coding assistants that infer developer intent from IDE interactions and repository context to reduce interaction overhead.
Active learning approach to optimize communication structure in multi-agent LLM systems, reducing token usage and improving performance with limited training data.
SafeHarbor: hierarchical memory-augmented safety guardrail for LLM agents that balances security with reducing over-refusal in tool-use scenarios.
Auto research system using specialist AI agents to autonomously develop training recipes through closed empirical loops with hypothesis testing and code generation.
CRAFT: continual learning framework for LLMs that avoids catastrophic forgetting by learning low-rank interventions on hidden representations instead of updating weights.
Framework for behavioral evaluation of agentic stock prediction systems using LLM judges and closed-loop RL feedback. Addresses hidden decision quality in aggregated metrics.
Methods for nonlinear causal discovery using Fourier feature approximations. Scales Gaussian process marginal likelihood scoring and kernel tests.
Theoretical research proving transformers can implement policy-improvement reinforcement learning in-context without parameter updates via explicit constructions.
Research on adaptive LoRA component selection in federated learning with differential privacy. Addresses aggregation error and stability in distributed fine-tuning.
Research on unified multimodal models bridging understanding and generation with explicit synergy supervision for visual tasks.
Research on gradient noise imbalance in LLM training. Proposes Adam optimizer calibration via signal-to-noise ratio for modular improvements.
Research introducing LCC-LLM framework for malware attribution using code-centric LLMs. Benchmark dataset for static analysis tasks.
LeakDojo framework evaluates and benchmarks RAG system vulnerabilities to database leakage attacks.
LoopTrap demonstrates termination poisoning attacks where adversaries manipulate LLM agent loop termination judgment.
VideoRouter uses query-adaptive dual routing to compress long video tokens for efficient multimodal LLM inference.
CITE provides anytime-valid statistical inference for LLM self-consistency with data-dependent stopping rules.
LOVER is an unsupervised verifier regularized by logical rules to enhance LLM reasoning without labeled data.
LLM-driven design space exploration automates FPGA accelerator design for AI workloads using language models.
PCNET uses probabilistic circuits to detect and correct hallucinations in LLMs without corrupting correct generations.
SIREN addresses selection bias in LLM evaluation benchmarks through repeated-split reporting to correct winner's curse.
PragLocker protects LLM agent prompts from theft in untrusted deployments using non-portable prompt techniques.
Safety Anchor defense against harmful LLM fine-tuning using geometric bottlenecks to block orthogonal attack trajectories.
Theoretical analysis of uniform convergence for halfspaces showing VC bounds are tight for inhomogeneous case.
PersonaKit platform for testing diverse personas in full-duplex spoken dialogue systems with realistic turn-taking.
Research proving randomized Hadamard transforms preserve orthogonality in quantization for compression and acceleration.
Large-scale empirical study evaluating synthetic data augmentation for time series forecasting across architectures and datasets.
SelectiveRM framework using optimal transport to handle noisy preferences in LLM reward model training for RLHF.
TFM-Retouche introduces input-space adapter for tabular foundation models enabling efficient task adaptation without full fine-tuning.
MTG-Causal-RL benchmark for reinforcement learning combining sequential decision-making, hidden information, and causal structure using Magic: The Gathering.
nGPT architecture with normalized weights and activations on unit hypersphere enables stable 4-bit precision training without random transforms or scaling tricks.
Milestone-Guided Policy Learning for long-horizon language agents addresses credit misattribution and sample inefficiency through intermediate milestone supervision.
VISD enhances VideoLLMs for complex reasoning combining RL with verifiable rewards and structured self-distillation for fine-grained credit assignment.
Conditioning via injection in Decision Transformer augments RTG with external information outside sequential modeling to improve offline RL sample efficiency.
Schedule-and-Calibrate method for utility-guided multi-task RL in code LLMs, addressing non-uniform task difficulty in unified post-training approach.
BUILD-AND-FIND protocol for evaluating AI coding agents on repository-level engineering tasks, assessing generated code as both task solution and artifact for future work.
Prologue approach for autoregressive image generation prepending learnable prologue tokens to bridge reconstruction-generation gap in visual token sequences.
Listwise Policy Optimization reveals geometric structure in group-based RL with verifiable rewards for LLM post-training, connecting to target projection on response simplex.
SymDrift approach for one-shot generative modeling of physical systems using equivariant diffusion models that respect global symmetries like rotations.
Suite of detectors for AI-generated images with explainability focus, trained on large-scale photorealistic fake image dataset to support disinformation detection.
Theoretical framework unifying goal-conditioned RL and mutual information skill learning through control-maximization principle for unsupervised skill discovery.
Method using Wasserstein gradient flow to jointly train discrete image tokenizers and autoregressive prior models, addressing mismatch between tokenization and generation.
AdaGamma proposes state-dependent discount factors for deep actor-critic RL methods, adapting planning horizon and bootstrapping strength dynamically per state.
Entropy-Regularized Adjoint Matching for offline RL integrating flow-matching generative policies to address popularity bias in low-density action regions.
Hard Negative Captions dataset and method to improve fine-grained visual-linguistic understanding in image-text matching models through hard negative mining.
In-context learning approach for black-box optimization that handles unreliable feedback from multiple sources (experts, simulators, heuristics).
Event-aware generative world model combining video diffusion models with structured kinematic-to-visual action fields for robotic control.
Variable codebook size quantization technique for autoregressive visual generation, addressing information-theoretic limits of fixed-size codebooks.
Statistical method for direct optimization of minimum-volume prediction regions in multivariate conditional quantile regression.