Measure-theoretic analysis of contrastive learning mechanisms beyond alignment-uniformity decomposition, revealing explicit energy landscapes.
Studies embedding condensation phenomenon in small language models and proposes dispersion loss to improve generalization and match larger model representations.
Technique using 2-bit KV-cache quantization to reduce memory bottlenecks in autoregressive video generation models from 30GB to deployable sizes.
Proposes rationality measurement framework and theory for evaluating reinforcement learning agent behavior in deployment settings.
Unified system integrating reinforcement learning with adaptive speculative decoding to accelerate LLM serving while enabling continuous speculator training.
arXiv paper on binary flow matching for discrete generative modeling, extending flow matching framework to binary manifolds.
arXiv paper on Control Reinforcement Learning for interpretable token-level LLM steering using sparse autoencoder features.
arXiv paper on black-box detection of LLM API changes using low-cost token-level observation via border input approach.
arXiv paper on CrispEdit: scalable non-destructive LLM editing using second-order optimization to preserve model capabilities.
arXiv paper on extreme LLM compression via sub-1-bit quantization, identifying spectral energy gains through latent geometry alignment.
Task-driven subspace decomposition method for LoRA-based continual learning balancing knowledge sharing and task isolation.
Mechanistic interpretability study of grokking in Transformers through architectural topology modifications to understand delayed generalization.
Visual representation framework encoding signals as functions with low-rank adaptations to diffusion foundation models for compact storage and reuse.
Novel multi-treatment generalization bound and optimal balancing weight estimator for individual treatment effect estimation in causal inference.
FastDSAC framework scales maximum entropy reinforcement learning to high-dimensional humanoid control by addressing exploration inefficiency in stochastic policy gradients.
Visualization approach for understanding critic network loss landscapes in actor-critic reinforcement learning algorithms.
Dimension-wise structured pruning method for LLMs merging task-specific importance scores for efficient deployment.
Physics-informed reinforcement learning for adaptive sensing in high-dimensional low-sample-size datasets.
Gradient-boosted attention mechanism applying boosting principles within single transformer layer for error correction.
Reinforcement learning method addressing delayed feedback using homomorphic approach to avoid state-space explosion.
Empirical study analyzing how reasoning evolves from supervised fine-tuning to RL in LLMs using chess as evaluation domain.
Master Key Hypothesis demonstrating capability transfer across model scales through linear subspace alignment without retraining.
Method leveraging LLMs for schema-adaptive tabular representation learning to improve generalization across varying EHR schemas.
Theoretical analysis of signal propagation at initialization in normalization-free transformers using averaged partial Jacobian norms.
Uses LLMs to enrich medical knowledge graphs for improved clinical concept representation learning from EHR data.
Single-layer Mamba variant optimized for time series classification with minimal architectural modifications.
Framework for dynamic mid-generation abstention in LLM chain-of-thought reasoning to reduce wasted compute on incorrect outputs.
Linear attention mechanism using delta rule with fine-grained gating control for improved associative recall in long-context transformers.
Analysis reveals attention heads in LLMs detect false claims but agree anyway; identifies shared sycophancy-lying circuit across models.
Switching-system analysis framework for Q-learning convergence and stability guarantees via Lyapunov theory.
Universal Transformers with learned memory tokens as computational scratchpad for adaptive reasoning on combinatorial tasks.
Agentic system fusing atomic-scale models and LLMs for accelerated superconductor discovery via multi-dimensional reasoning.
Federated learning framework combining differential privacy and homomorphic encryption for cardiovascular disease modeling.
Study of GNN expressive power for solving semidefinite programming problems as ML surrogates.
CastFlow enables LLM-based agents with specialized roles for time series forecasting via adaptive workflows.
Neural network approach using diverse object features for open set recognition of novel classes.
Flow matching method for non-autoregressive text generation using conditional sampling-hybrid inference.
Heima framework compresses chain-of-thought reasoning in multimodal LLMs into abstract tokens for efficiency gains.
Multi-agent reinforcement learning method using optimistic epsilon-greedy exploration to address value underestimation in CTDE paradigm.
Introduces TF1-EN-3M, open dataset of three million synthetic moral fables generated by small LLMs for training aligned language models.
Develops techniques to measure memorization of copyrighted books in open-weight LLMs, examining relationship between memorization and copyright.
Introduces AVA-Bench for systematic evaluation of vision foundation models, identifying limitations in VQA-based evaluation protocols.
Proposes method for unsupervised discovery of interpretable features in LLM activation spaces using neuron group composition.
Proposes methods for enforcing tail calibration in probabilistic forecasting models when model class is misspecified.
Introduces logit-gap steering metric to quantify safety margins in RLHF-aligned language models through single scalar measurement at decoding.
Presents NaviMaster, a unified agent for GUI and embodied navigation tasks using MDP formulation, enabling multi-domain learning with shared policy.
Proposes CorrSteer for steering LLM generation using sparse autoencoder features without contrastive datasets, enabling efficient control at inference time.
Uses LLMs to construct nomological networks in psychological measurement, automating validity assessment for clinical and policy applications.
Proposes ParamInter tool for analyzing high-dimensional parameter spaces using guided visual interpolations and t-SNE representations.
Evaluates quantization's impact on Vision-Language Models beyond accuracy, examining effects on reliability metrics like OOD detection for efficient deployment.