Search papers, labs, and topics across Lattice.
88 papers published across 7 labs.
Unlimited OCR can transcribe dozens of pages in one go, overcoming the memory bottlenecks that plague traditional OCR models.
GRINQH redefines the efficiency of LLM generation by enabling dynamic quantization that adapts to computational needs, achieving state-of-the-art performance even at 2-bit precision.
Achieving up to 3.9x compression with near-lossless quality, HyperQuant sets a new standard for quantization in large models.
Achieving a 20% boost in performance while slashing emissions by over 60% reveals the untapped potential of Deep Shift Neural Networks in sustainable AI.
Halving multi-hop accuracy loss while maintaining single-hop recall, Kamera redefines how multimodal agents can efficiently reuse cached information without retraining.
Unlimited OCR can transcribe dozens of pages in one go, overcoming the memory bottlenecks that plague traditional OCR models.
GRINQH redefines the efficiency of LLM generation by enabling dynamic quantization that adapts to computational needs, achieving state-of-the-art performance even at 2-bit precision.
Achieving up to 3.9x compression with near-lossless quality, HyperQuant sets a new standard for quantization in large models.
Achieving a 20% boost in performance while slashing emissions by over 60% reveals the untapped potential of Deep Shift Neural Networks in sustainable AI.
Halving multi-hop accuracy loss while maintaining single-hop recall, Kamera redefines how multimodal agents can efficiently reuse cached information without retraining.
Smaller models can achieve empathetic dialogue performance that rivals larger counterparts by leveraging privileged information during training.
KD can significantly enhance model performance in low-data settings, but its effectiveness hinges on the quality of the teacher model.
MOCAP slashes LLM inference latency by over 76% while boosting throughput more than threefold, redefining efficiency for long-context processing on wafer-scale chips.
CAAD achieves an 8% performance boost in speech language models while slashing inference latency and linguistic bias.
Concordia achieves fault tolerance for LLM inference by seamlessly integrating persistent kernel checkpointing, enabling rapid recovery without CPU bottlenecks.
Achieving substantial model compression with negligible accuracy loss could redefine deployment strategies for neural networks on edge devices.
ARIA reallocates training focus to areas of persistent misalignment, leading to significant performance improvements in unseen conditions.
Nexus Sampling retains crucial tokens during KV cache eviction, achieving near-dense attention performance with dramatically reduced memory usage.
SVD-Surgeon achieves superior compression of large language models without retraining, enhancing performance while reducing resource demands.
FlexServe achieves over 10X faster secure LLM inference on mobile devices without compromising privacy or performance.
Achieving up to 65% energy savings on mobile devices without sacrificing quality of experience could redefine on-device LLM deployment.
Combining advanced model compression with hardware-aware architecture search can drastically enhance real-time GNSS interference monitoring on low-resource devices.
The performance gap in LLM serving under GPU-CC is primarily due to a costly serialized VM-GPU bridge, not the GPU's computational capabilities.
Context-aware distillation not only boosts the size of the PolkitBench corpus but also proves that structured context is crucial for maintaining model performance under challenging conditions.
The Apple Neural Engine's undocumented capabilities reveal significant performance insights that could optimize machine learning applications on Apple devices.
Achieving oracle-level Bayesian predictions with a multi-task framework that adapts seamlessly to new priors, all while being orders of magnitude faster.
Achieving up to 27x speedup in execution-state restoration could revolutionize low-latency AI applications, from interactive agents to robotics.
UltraQuant slashes time-to-first-token by over 3x in cache-pressured scenarios, revolutionizing how context-heavy agents manage KV caching.
StreamKL slashes memory usage from quadratic to constant, enabling efficient long-context attention distillation on a single GPU.
Tsetlin Machines can outperform Binarized Neural Networks in accuracy while slashing execution time and energy consumption, making them a game-changer for edge AI applications.
AIR achieves over 18% better perplexity than previous methods while retaining 60% of the parameters, revolutionizing LLM compression efficiency.
Token-oriented inference optimizations can cut production costs and boost efficiency, transforming large model services from merely callable to fully operable.
Efficient evidence ordering can cut response times by up to a third without sacrificing answer quality in RAG systems.
SafeSpec achieves a remarkable 15% reduction in attack success rates while accelerating inference by over 2x, effectively marrying safety and speed in LLMs.
Achieving a 1.33x speedup in FHE NTT computations on standard AI hardware could revolutionize the efficiency of privacy-preserving applications.
Accelerating autoregressive image generation by up to 13.3x, SSD reveals that respecting the 2D structure of images can unlock unprecedented computational efficiencies.
Achieving 134 FPS with under 50 ms latency, ViCoStream redefines the capabilities of VideoLLMs for real-time streaming applications.
Whisper models can achieve unprecedented energy efficiency on the IMAX architecture, slashing power consumption by over 10x compared to top GPUs.
Shrinkage Bias in E2M1 formats could be the hidden culprit behind training instability in LLMs, but uniform grids like E1M2/INT4 offer a robust solution.
QVec reveals that the weight shifts during quantization can be leveraged to neutralize backdoor threats without retraining or additional computational burden.
SAC achieves a staggering 2.1x throughput increase by fetching only the essential KV entries for sparse attention models, revolutionizing memory efficiency in LLM inference.
SparseStack's embedding quality remains stable across different FP16 rounding methods, challenging assumptions about low-precision impacts on performance.
DO-ALL achieves long-term robustness in continual adaptation by distilling source information into compact anchors, sidestepping privacy issues without sacrificing performance.
CAHP reveals that pruning attention heads through a graph-theoretical approach can enhance model efficiency without sacrificing performance, outperforming traditional methods in high-compression settings.
BNN satisfiability is NP-complete, but robustness verification can be achieved in polynomial time through clever structural insights.
SwitchBraidNet achieves hybrid BCI performance with an INT8 model size of only 3.03 KB, making high-dimensional neural decoding feasible on low-power devices.
Compact models can surpass larger instruction-tuned models in financial sentiment analysis by leveraging intelligently generated synthetic data.
Visual-OPSD achieves a remarkable 14.3x speedup while boosting accuracy by over 3 percentage points, revealing the untapped potential of reasoning in visual thought generation.
Anti-relevance in token selection can enhance contextual fidelity, allowing for up to 94% reduction in visual tokens without sacrificing performance.
Moebius achieves high-fidelity image inpainting with less than 2% of the parameters of leading models, setting a new benchmark for efficiency in the field.
EfficientRollout slashes RL rollout latency by nearly 20% while maintaining model performance, revolutionizing how we approach decoding in reinforcement learning.
KLD may mislead researchers by appearing to correlate with model performance while failing to predict outcomes in critical silent zones.
Cumulative bit-flips can drastically alter predictions in quantized neural networks, but SPINE reveals how to strategically harden models against these vulnerabilities.
MonaVec achieves 0.960 Recall@10 with just 27 MB of storage, outperforming traditional vector-search systems while eliminating the need for training data.
Coresets often outperform state-of-the-art dataset distillation methods, revealing that less complex approaches can be more effective and efficient in data-centric learning.
Ternary SSMs can achieve competitive performance without the costly overhead of from-scratch training, slashing token requirements by 1,000x.
Robust on-device inference can be achieved through meticulous engineering choices that significantly enhance performance in microcontroller-class devices.
A small trade-off in utility can lead to significantly enhanced safety alignment in LLMs, protecting against jailbreak attacks.
AoiZora cuts video diffusion denoising latency by 1.42x on TPU v5e sub-slices by smartly aligning logical sharding with physical hardware topology.
A novel software framework transforms high-level transformer models into efficient, deployable components for low-latency jet tagging on advanced AI hardware.
Resource-level supervision boosts Website Fingerprinting accuracy by over 8% even under significant temporal drift, challenging the limitations of traditional low-level feature reliance.
Fixed-budget evaluations can significantly underestimate the capabilities of advanced language models, with larger token budgets unlocking their true potential across diverse tasks.
MIVE achieves unprecedented hardware efficiency by unifying the execution of critical normalization operations, slashing resource overhead in LLM inference.
Pruned LLMs can ace multiple-choice tests while failing to generate correct answers, revealing a critical evaluation blind spot in model assessment.
Bifrost achieves up to 53.4x latency reduction in LLM inference by strategically combining TEE and FHE, redefining privacy in cloud AI services.
LUMEN slashes recovery times in distributed LLM serving by intelligently coordinating load-aware recovery strategies during worker failures.
A two-loop configuration in LoopCoder-v2 boosts code generation performance by over 50% compared to a non-looped baseline, while more loops actually hinder results.
d-OPSD enables dLLMs to learn from their own future outputs, drastically improving sample efficiency and performance in reasoning tasks.
JetSpec shatters the speed ceiling of speculative decoding, achieving up to 9.64x acceleration on complex tasks while maintaining high acceptance rates.
CoreMem achieves a remarkable +4.51 percentage point improvement in open-domain reasoning accuracy while operating within a strict 8 GB VRAM budget.
Learned attention allocation patterns reveal that SWA is best positioned in lower layers, challenging conventional wisdom on attention distribution in LLMs.
LENS achieves an impressive 2.15% mean prediction error in latency estimation for LLM inference on NPUs, even without access to microarchitecture details.
Pruning up to 70% of operators in SSMs can maintain performance while slashing inference latency, unlocking their potential for real-world applications.
Tail latency in LLM serving can be cut by up to 50% without relying on length predictions, reshaping how we optimize inference performance.
Shift-and-sum quantization reduces reconstruction errors in visual autoregressive models, achieving state-of-the-art performance in post-training quantization.
ActiveSAM accelerates open-vocabulary segmentation by up to 5.5x while boosting accuracy, setting a new benchmark for efficiency in semantic segmentation tasks.
Orthogonal obfuscation slashes token recovery rates from LLM inference to just 1.3%, safeguarding sensitive data without compromising performance.
Ultra-low token pruning can achieve 92.1% of peak performance with only 16 visual tokens, thanks to a novel approach that preserves distribution consistency.
Achieving up to 88x efficiency gains, Taylor-Calibrate transforms the way we initialize hybrid linear attention models, drastically reducing the training burden.
QK normalization can be effectively integrated into MLA without the overhead of full key caching, leading to improved performance and efficiency.
FEnc2 achieves up to 228.83x speedup in private inference by revolutionizing how data is packed and processed in Fully Homomorphic Encryption.
J4D achieves up to 11.60% higher accuracy than standard JPEG at the same compression rate, revolutionizing how we think about image compression for DNNs.
Tropical achieves a remarkable balance in LLM serving, outperforming both disaggregated and non-disaggregated systems by enhancing request throughput while maintaining low latency.
SMEPilot achieves up to 3.94× faster LLM inference by intelligently balancing workloads between CPU and SME units based on operator characteristics.
Memory clock changes can increase edge inference latency by up to 48%, revealing critical gaps in traditional latency estimators.
Coding agents can achieve up to 3.5x faster completion times with a new KVCache management strategy that understands their unique workload patterns.
SwiftCache slashes latency by 69% and boosts context length nearly fourfold by enabling cross-model KV cache sharing on GPUs.
TokenPilot slashes inference costs by up to 87% without sacrificing performance, tackling the critical trade-off between context management and cache efficiency in LLM agents.
KVEraser achieves a 3-4x speedup over full recomputation while maintaining high performance in long-context tasks, revolutionizing how we handle context updates in LLMs.
Achieving real-time, energy-efficient inference on microcontrollers, this work shows that advanced AI can thrive even in the most resource-constrained environments.
TreeGRNG achieves a 3.7× reduction in energy usage while enhancing distribution accuracy, revolutionizing the implementation of Bayesian Neural Networks in ultra-low power hardware.
LLMs can autonomously optimize AI models for embedded devices, achieving 250x compression with minimal accuracy loss—something human experts struggle to match.
Fixed-reference delta compression achieves nearly 50% weight reduction while only sacrificing 8.3% accuracy, revolutionizing DNN deployment on FPGAs.
