Search papers, labs, and topics across Lattice.
100 papers published across 7 labs.
A novel decision layer achieves perfect accuracy in expert model management, eliminating false spawns and reuses in dynamic data environments.
NanoSleep achieves superior sleep stage classification accuracy while being small enough for deployment on wearable devices, striking a crucial balance between performance and efficiency.
Rethinking the VLM training pipeline allows for the creation of specialized models that outperform larger counterparts while using far fewer resources.
Optimizing feature engineering can double model accuracy and dramatically reduce error rates in ocean colour machine learning, but a one-size-fits-all approach won't work.
Training only the projector can match or exceed the performance of fully fine-tuned multimodal models while avoiding capability drift.
Rethinking the VLM training pipeline allows for the creation of specialized models that outperform larger counterparts while using far fewer resources.
Optimizing feature engineering can double model accuracy and dramatically reduce error rates in ocean colour machine learning, but a one-size-fits-all approach won't work.
Training only the projector can match or exceed the performance of fully fine-tuned multimodal models while avoiding capability drift.
A simple Gaussian counterexample reveals that existing concentration inequalities for discounted least-squares estimators are fundamentally flawed, necessitating significant corrections.
LoRA-GA$^2$ closes the performance gap with full fine-tuning by leveraging multi-step gradient dynamics without sacrificing efficiency.
Training decision trees just got faster—DICS cuts down on search time without sacrificing accuracy by using data-informed clustering to streamline split selection.
Agents struggle to significantly improve training algorithms, with the best only achieving 25% of the potential optimization gap.
By pooling data across groups, this framework achieves faster convergence rates in nonparametric regression, challenging traditional assumptions about group-specific learning.
End-to-end learning can drastically improve early classification accuracy in non-stationary environments, outperforming traditional methods that treat classification and triggering as separate tasks.
Task-CoEvolve slashes evaluation costs by 80% while maintaining performance parity with full-set validation in LLM harness optimization.
Restricting evidence visibility in language model societies can boost compositional generalization, leading to a 20-point performance increase over fully visible models.
ML-based data compression can be environmentally sustainable, but only if it surpasses a critical break-even point in carbon savings.
A novel decision layer achieves perfect accuracy in expert model management, eliminating false spawns and reuses in dynamic data environments.
Minibatch perturbations in AdamW can have delayed and significant impacts on future loss, reshaping our understanding of optimizer dynamics.
QDOS transforms offline datasets into a treasure trove of high-value skills, significantly enhancing reinforcement learning performance in challenging tasks.
GFHNNs can outperform conventional neural networks in learning Hamiltonian dynamics while requiring significantly less training data.
Negative curvature can destabilize neural network optimization, revealing critical insights into the geometry of loss landscapes.
Achieving up to $\tilde{\mathcal{O}}(T_{total}^{-1})$ convergence rates without regularization challenges conventional wisdom about the necessity of entropy in policy gradient methods.
Adapting four ImageNet-based models to new domains, the authors achieve up to 125x fewer evaluations while improving output quality by 29% on average.
4MAS leverages asymmetric hemispheric structures and sleep-like periods to significantly enhance memory retention in continual learning tasks.
Efficiently routing queries in AI systems can be achieved without the costly overhead of exhaustive value estimation, thanks to novel policies that balance accuracy and cost.
DARS achieves superior performance in instruction-based image editing by transforming outcome-level feedback into actionable, localized supervision for both planning and rendering stages.
Axon achieves up to 107% speedup on JAX, revolutionizing how LLMs can be efficiently deployed across different frameworks without sacrificing optimization.
Achieving a Micro-F1 score of 0.8085, this framework effectively balances adaptation and retention in LLMs for smart contract vulnerability detection.
Fine-tuning with sparse attention can outperform traditional exact attention models while running efficiently on modest hardware.
Achieving leading performance in image generation with only 6 billion parameters, Swift-Image redefines the efficiency frontier for compact models.
OrthoSkillVLA preserves prior skills in pretrained VLA models while seamlessly integrating new ones, outperforming traditional methods in both simulated and real-world scenarios.
Optimal learning rates for training massive Mixture-of-Experts models can be accurately predicted from small proxy models, achieving high fidelity even at trillion-token scales.
Annotations can be transformed into powerful oracle rollouts, dramatically enhancing the efficiency of reinforcement learning for video MLLMs.
A single training example can boost model performance, but its impact fades significantly within just a few training steps.
Continuous-time reinforcement learning can effectively tackle non-Markovian Hawkes processes, outperforming discrete methods in optimization tasks.
Local algorithms can outperform theoretical predictions in constrained optimization scenarios, revealing a critical gap between finite and asymptotic performance.
A modified DMRG method outperforms gradient descent in optimizing tensor networks for quantum state representation, revealing new potential in machine learning applications.
ROR enables adaptive optimizer selection during training, cutting down the time and resources needed to find the best optimizer by up to 76%.
GEAR slashes inference time by up to 2866 times while boosting AUC scores beyond conventional supervised models.
Co-observation in training data can dramatically enhance generalization in continual learning, revealing a new dimension beyond forgetting and plasticity.
GraphK can generate graphs with variable sizes while maintaining structural integrity, outperforming traditional methods in both accuracy and efficiency.
NanoSleep achieves superior sleep stage classification accuracy while being small enough for deployment on wearable devices, striking a crucial balance between performance and efficiency.
GCNO achieves superior channel reconstruction with variable-rate encoding, allowing for seamless adaptation to different antenna configurations without retraining.
Adaptive similarity margins in HN-CLIP boost retrieval accuracy by up to 4.3% while training 2.4x faster than leading methods.
Early backpropagation can boost training throughput by over 9% without sacrificing model performance.
SparsePR cuts attention-reconstruction error while speeding up video generation by over 2.5x without sacrificing quality.
Achieving competitive medical image segmentation with just 10 annotated cases could revolutionize the efficiency of clinical workflows.
RTPO eliminates critical instability in multi-turn RL training, achieving over 21% performance improvement compared to traditional methods.
Compact object detectors can achieve state-of-the-art accuracy without the bulk of larger models, thanks to a novel multi-level knowledge distillation approach.
APEX achieves ANN-equivalent accuracy with 40% energy savings, revolutionizing the efficiency of Spiking Neural Networks in practical applications.
Hybrid quantum models can achieve competitive forecasting performance with fewer parameters and distinct learning trajectories compared to classical counterparts.
DeltaMomentum accelerates convergence by adapting momentum updates to the frequency of gradient directions, outperforming traditional methods in both speed and efficiency.
Incremental learning can render retraining policies nearly irrelevant, with no significant accuracy gains over a no-retrain baseline in many scenarios.
Gradient Mirage disrupts the gradient-objective consistency that underpins gradient matching attacks, offering a robust defense without sacrificing model performance.
HyperCut slashes design space exploration time by over 80% while doubling performance, revolutionizing inter-layer scheduling for deep neural networks.
The mapping of non-maximal probabilities to GMM components significantly influences the performance of S-JEPA encoders, revealing that structure matters as much as values in representation learning.
M-OPD's capability integration gap can be closed from 35.6% to 83.4% by addressing token-level optimization imbalances, fundamentally changing how we approach multi-teacher distillation.
Achieving global optimality in model splitting for split federated learning could revolutionize resource allocation strategies in edge computing environments.
Splitting same-outcome groups by execution quality can cut rollout requirements by over half while enhancing policy performance in complex VLA tasks.
SRT recovers up to 37% of lost knowledge in language models while enhancing new information retention, challenging the limitations of traditional replay methods.
Overparameterization in online sparse regression can lead to significant improvements in computational efficiency and statistical performance, challenging traditional optimization paradigms.
Nonlocal transitions can enhance RBM learning efficiency, achieving better sampling quality with fewer steps than traditional methods.
Integrating novelty and surprise in experience replay can dramatically accelerate learning in image-based reinforcement learning environments.
CORAM achieves up to 1.35 points of performance improvement in model merging, redefining the boundaries of orthogonal transformations in AI.
Tightening the bounds on hyperparameter tuning could revolutionize how we approach model optimization in machine learning.
Understanding the dynamics of knowledge transfer reveals that a tailored curriculum can dramatically enhance performance across reasoning tasks in large language models.
ARASH slashes the resource demands of Tabular Foundation Models while maintaining accuracy, making efficient tabular prediction accessible even in low-compute settings.
Evolution strategies can optimize large language models for long-horizon tasks with minimal GPU resources, outperforming traditional reinforcement learning approaches.
A centralized model repository can enhance CSI feedback efficiency, achieving significant performance gains while slashing local training requirements.
Energy-efficient model cascades can falter under data perturbations, revealing hidden vulnerabilities that could undermine their effectiveness in real-world applications.
Coiflet wavelet convolutions can outperform Haar in efficiency, slashing parameters and FLOPs while maintaining competitive accuracy.
Achieving 87.35% accuracy on CIFAR10-DVS in only 10 inference steps reveals a breakthrough in training efficiency for spiking neural networks.
Achieving over 90% accuracy retention with 50% pruning, DVBP + OB²C outperforms traditional methods by leveraging noise-filtered neuron selection and optimal weight updates.
Spectral gradient orthogonalization can boost model accuracy by over 20% in differentially private training, but only under the right conditions.
Reorganizing backward propagation around primitives slashes training time for ray-traced Gaussian rendering by up to 4x while improving output quality.
Achieving near state-of-the-art performance with a 22M encoder, DistillPath-KS16 runs over 25 times faster than its larger counterparts while retaining critical accuracy in pathology tasks.
Achieving 5-7x speedup in diffusion models without sacrificing quality, LinCa redefines the efficiency of feature caching through adaptive prediction strategies.
Sharing rollout feedback across related samples can significantly boost exploration efficiency in RLVR, leading to better reasoning capabilities in large language models.
Skill Optimizers trained through execution feedback can outperform traditional models by over 9 points, revealing a critical gap in agent learning methodologies.
A 5-step sampling schedule can deliver nearly the same quality as a 50-step one, slashing inference costs by 90%.
Repeating successful actions in real-time can dramatically boost sample efficiency in reinforcement learning, outperforming conventional methods.
Eliminating semantic redundancy in HGNNs can boost sampling performance by an order of magnitude, transforming mini-batch inference efficiency.
TileMix achieves a breakthrough in LLM inference by enabling mixed-precision attention that boosts throughput while maintaining long-context quality.
Palmyra x6 outperforms previous models in enterprise agentic tasks while maintaining a strong safety profile and low bias.
Potential theory could unlock new levels of sample efficiency in reinforcement learning algorithms.
TSFT redefines how we allocate fine-tuning resources in CRL, achieving near-oracle performance while maximizing task coverage.
QWM achieves unprecedented sample efficiency in reinforcement learning by leveraging world models without succumbing to compounding bias, outperforming prior methods on challenging benchmarks.
The form of answer labels, not just their quantity, fundamentally shapes what LLMs learn during fine-tuning, revealing a surprising causal relationship that could redefine training strategies.
Pixel-space diffusion models can now rival latent-space counterparts while achieving up to 4.75 times faster inference.
Reallocating optimization effort based on reward saturation can boost performance by up to 9.2% in complex reasoning tasks.
Achieving a new upper bound of $\omega < 2.371177 could redefine our understanding of matrix multiplication efficiency.
Analytical prior information can cut prediction errors by over 89% compared to direct learning methods when simulation data is limited.
Adam's performance can be dramatically improved with a new memory-efficient variant that retains its advantages while cutting memory usage by 50%.
Privileged Value Functions can inject crucial token-level signals into LLM reinforcement learning, leading to substantial performance gains over traditional methods.
Archive selection can dramatically shift based on unresolved components, revealing that checkpoint recovery alone is insufficient for accurate energy-oracle ranking.
Localized TabICLv2 retains nearly all the accuracy of its predecessor while slashing inference time, making it a game-changer for large-scale tabular data tasks.
FETERS achieves state-of-the-art early time-series classification with just five labeled examples, outperforming existing methods on 44 datasets.
KNNG-CS slashes selection time by up to 41.2 times while preserving model accuracy, revolutionizing coreset selection for large datasets.
Identical delay summaries can yield vastly different regret outcomes, highlighting the crucial impact of timing in bandit optimization.
A self-reinforcing instability trap in Deep Q-learning can be effectively mitigated through controlled bootstrapping and ensemble quantile estimation, leading to more stable training outcomes.
Competing dealers can increase market instability by over 3 times, revealing the critical need for a pre-deployment stability margin in machine learning models for trading.
AsyTO achieves state-of-the-art forecasting accuracy while keeping model complexity linear, challenging the notion that more parameters always lead to better performance.
Reducing the rank of gradients sent to the Transformer can increase validation loss, but surprisingly, it’s less detrimental than using a factorized forward head.
SOPD not only outperforms traditional distillation methods but also redefines how we think about trajectory corrections in model training.
