BiTrajDiff uses bidirectional diffusion for trajectory generation in offline RL, addressing distribution bias through data augmentation.
ReasonCache system accelerating large reasoning model serving through KV cache sharing among concurrent requests.
Unsupervised learning approach for decomposing neural model representation spaces into interpretable subspaces.
BOOST framework for automated joint selection of kernel and acquisition functions in Bayesian optimization.
Kairos framework for adaptive time series foundation models addressing temporal heterogeneity with parameter efficiency.
Study identifying pitfalls in KV cache compression for LLMs in realistic multi-instruction scenarios with practical implications.
Analysis showing GRPO reinforcement learning algorithm for LLM post-training is equivalent to DPO with group-level baselines.
Vision-LLM approach for spatiotemporal traffic forecasting combining visual understanding of grid-based traffic data with language model capabilities.
AMiD knowledge distillation method for LLMs using alpha-mixture assistant distribution to address capacity gaps and training instability in student-teacher alignment.
ScaLoRA progressively accumulates high-rank weight updates from low-rank factors for more effective and faster LLM fine-tuning than standard LoRA.
GIFT combines GRPO group sampling, DPO-style implicit rewards, and UNA advantage standardization for on-policy LLM fine-tuning with improved efficiency.
Develops thermodynamic framework describing stationary distributions of SGD with weight decay for scale-invariant neural networks.
Analyzes how learning rate decay reduces effective use of high-quality data in curriculum-based LLM pretraining, proposing improved curriculum strategies.
BRIDGE framework for structured prompting of LLMs to generate code with formal verification in proof assistants like Lean, handling multiple coupled domains.
Investigates last-layer retraining to mitigate spurious correlations and improve minority group performance in neural networks trained with ERM.
Analyzes goal-conditioned reinforcement learning through optimal control framework, deriving optimality gaps between goal-conditioned and dense reward objectives.
Theoretical analysis showing approximate symmetry is exponentially easier to enforce than exact symmetry in ML models, with implications for inductive biases.
Rethinking output alignment for 1-bit post-training quantization of LLMs to enable efficient deployment on resource-constrained devices.
Addresses entity alignment in knowledge graphs using graph foundation models to capture long-range dependencies across sparse KG structures.
Proposes L2R, a routing mechanism for Mixture-of-Experts models using low-rank projections and Lipschitz control to improve expert specialization and routing discriminability.
TabClustPFN extends prior-fitted networks paradigm to unsupervised tabular clustering, amortizing Bayesian inference for heterogeneous tabular data.
LoRIF: Low-rank approximation for training data attribution using influence functions, scaling to large datasets while maintaining attribution quality.
PaAno: Patch-based representation learning for time-series anomaly detection that is computationally efficient compared to large foundation models.
Analysis of Hyperbolic Graph Neural Networks proposing geometry-task alignment criterion for effective hierarchical representation learning on tree-like graphs.
Embedding perturbation technique for uncertainty quantification in LLM reasoning tasks, measuring confidence in intermediate reasoning steps.
Human-inspired reward shaping for LLM reasoning via reinforcement learning, separating exploration and consolidation phases for improved performance.
Accelerated Sequential Flow Matching for real-time probabilistic inference on streaming observations using Bayesian filtering perspective on diffusion models.
Proximal Action Replacement method for offline reinforcement learning combining actor-critic and behavior cloning to mitigate suboptimal dataset actions.
Krause Attention mechanism addressing representation collapse and attention sink phenomena in transformers through principled bounded-confidence dynamics.
ArGEnT transformer for operator learning on systems with complex, varying geometries for scientific machine learning applications like design optimization.
ACC++ improves circuit-tracing method for mechanistic interpretability by identifying attention head signals in language models via low-dimensional subspaces.
CAKE method for assessing confidence in individual clustering assignments via k-partition ensembles, addressing instability in k-means algorithms.
Study of grokking phenomenon (sudden generalization) in neural networks learning finite-dimensional algebra operations, extending prior work on group operations.
MUON+ improves the Muon optimizer for LLM pre-training by addressing norm imbalance issues in polar iterations through an additional normalization step.
MPU framework for privacy-preserving machine unlearning in LLMs using perturbed copies without sharing parameters or forget sets.
Stochastic attention via Langevin dynamics on modern Hopfield energy enables temperature-controlled retrieval and generation.
Gradient Iterated Temporal-Difference Learning addresses divergence issues in TD learning with semi-gradient updates.
TERMINATOR learns optimal early stopping points for chain-of-thought reasoning to reduce overthinking compute waste.
RL framework for diffusion language models using entropy-guided step selection and stepwise advantages for training.
M²RNN proposes non-linear RNNs with matrix-valued states for language modeling with greater expressive power than Transformers.
iAmTime foundation model for time series tasks using explicit instruction-conditioned in-context learning with demonstrations.
Polaris introduces Gödel agents for small language models enabling recursive self-improvement through policy repair via experience abstraction.
SEDGE framework for generating structured data beyond training distribution with conditions for reliable extrapolation.
Research on training LLMs to explicitly express uncertainty signals within responses during reasoning or at answer time.
UniMamba combines state-space models and attention for efficient multivariate time series forecasting with linear complexity.
Integrates persistent homology and contraction operations into graph neural networks for improved representation learning.
Framework for dynamic mid-generation abstention in chain-of-thought reasoning to reduce wasted compute on incorrect long responses.
Quotient-Space Diffusion Models leverage symmetry in generative tasks for faster 3D molecule structure generation.
LayerBoost proposes layer-aware attention reduction to improve efficient LLM inference by replacing softmax attention selectively across layers.
Research on Self-Preference Bias in LLM-as-a-Judge systems showing LLMs systematically favor their own outputs during evaluation.