Search papers, labs, and topics across Lattice.
97 papers published across 11 labs.
Staged knowledge distillation allows quantum agents to learn complex visual tasks without the pitfalls of direct pixel-based training, achieving near-optimal performance with significantly smaller models.
RMMD not only accelerates model inference by 7.5x but also outperforms its teacher model on nearly all target weather variables, showcasing a breakthrough in distillation techniques.
PRR slashes decoding latency by up to 40% in long-context LLMs while maintaining accuracy, revolutionizing the efficiency of dynamic sparse attention.
Relaxed and tree-based acceptance criteria can dramatically increase the certified acceptance region in speculative decoding, challenging traditional views on distribution preservation.
Off-policy distillation fails in multi-task settings, but a two-phase approach combining it with on-policy refinement can achieve single-task expert performance across multiple tasks.
Staged knowledge distillation allows quantum agents to learn complex visual tasks without the pitfalls of direct pixel-based training, achieving near-optimal performance with significantly smaller models.
RMMD not only accelerates model inference by 7.5x but also outperforms its teacher model on nearly all target weather variables, showcasing a breakthrough in distillation techniques.
PRR slashes decoding latency by up to 40% in long-context LLMs while maintaining accuracy, revolutionizing the efficiency of dynamic sparse attention.
Relaxed and tree-based acceptance criteria can dramatically increase the certified acceptance region in speculative decoding, challenging traditional views on distribution preservation.
Off-policy distillation fails in multi-task settings, but a two-phase approach combining it with on-policy refinement can achieve single-task expert performance across multiple tasks.
Balancing forward and reverse KL divergence in knowledge distillation can lead to significant improvements in text generation quality, outperforming traditional methods by up to 0.6 points.
Achieving a 46.2% reduction in computational cost for CNNs without retraining could revolutionize how we deploy deep learning on edge devices.
COSM achieves a remarkable 2.8x improvement in PIM throughput while keeping CPU performance degradation under 2.0%.
Only a subset of design interactions in heterogeneous LLM inference are binding constraints, revealing critical insights for optimizing deployment strategies.
Achieving a 20.58x reduction in search cost for optimal neural network configurations on FPGAs could revolutionize resource-efficient AI deployment at the edge.
DOPD reveals that intelligently routing supervision based on advantage gaps can significantly enhance capability transfer in distillation, outperforming conventional methods.
Transforming probabilistic programs into dynamic graphs can drastically cut down on computation time, enabling faster and more precise MCMC inference.
HMA-Serve achieves 3.2x higher goodput and 4.8x better cost-efficiency by leveraging memory-heterogeneous accelerators, challenging the status quo of single-vendor LLM serving.
Budget-adaptive routing can outperform strong models while cutting per-frame latency by nearly 30%—a game changer for edge-cloud inference efficiency.
CAEE slashes inference latency in MoE models by intelligently pruning low-value experts, achieving efficiency gains without sacrificing accuracy.
A compact 4B model using DuoMem achieves a staggering 77.9% task success rate, rivaling a 72B teacher model while being over 3x faster in execution.
Prioritizing token coverage in KV-cache eviction can boost LLM performance by over 10 points without increasing memory usage.
QuantGuard effectively neutralizes quantization-conditioned backdoor attacks, achieving attack success rates comparable to clean models without sacrificing performance.
Unused AI computation at the edge can be harnessed to boost performance in traditional tasks without sacrificing the efficiency of primary workloads.
Tri-serve redefines energy efficiency in multimodal inference by addressing hidden power inefficiencies, achieving a 22% boost without latency trade-offs.
A new scale-invariant scaling formula for the Ozaki scheme II fast mode achieves high accuracy without sacrificing throughput, revolutionizing matrix multiplication efficiency.
Achieving near-FP16 quality with 4-bit quantization, this method enhances motion smoothness while maintaining performance on large models.
MBD-LMs achieve a 2x increase in decoding throughput while maintaining high accuracy, revolutionizing the efficiency of diffusion-based text generation.
Threshold-sensitive KV cache pruning is out; ReFreeKV's adaptive approach achieves robust memory efficiency without predefined limits.
Quantization is not just a compression technique; it fundamentally reshapes the performance and robustness of Federated Learning systems.
Achieving up to 5.87× acceleration in diffusion models while maintaining performance with just 3-5 training samples could redefine efficiency benchmarks in generative AI.
SharQ recovers up to 63% of the accuracy gap in low-bit quantization while achieving over 2x latency reduction in LLM inference.
Pruning 77.8% of visual tokens without losing performance could revolutionize the efficiency of multimodal large language models.
Cost-aware selective inference slashes unsafe false negatives in driver monitoring from 17.37% to around 5%, revolutionizing safety in automated vehicles.
ResilPhase achieves high-fidelity diffusion model acceleration by eliminating derivative instability, setting a new standard for inference speed without sacrificing quality.
Achieving 13.8x compression on LLMs while improving accuracy by 3.70 percentage points could redefine on-device inference for IIoT applications.
EpiKV achieves 72% accuracy on MATH-500 with a 4096-token cache, rivaling the best attention-based methods while dramatically improving inference speed.
Quantization can slash memory usage by 85%, but it may also double inference time and energy costs, complicating the trade-offs in LLM deployment for software repair tasks.
LogicIR achieves high-quality image restoration with a fraction of the computational cost by leveraging the efficiency of logic gate networks.
Rebinding visual cache positions can boost multimodal reasoning accuracy by 5% while slashing computational costs dramatically.
Tokens with high predictive uncertainty can dramatically enhance long-context reasoning, outperforming traditional attention-based methods in KV cache compression.
HyperDFlash achieves remarkable decoding speedups and accuracy improvements by aligning residual streams with the MHC architecture, outperforming existing methods.
Adaptive page-aware scheduling can boost LLM serving throughput by over 1.3x on commodity GPUs, challenging existing optimization paradigms.
CAT-Q quantizes large language models with unprecedented efficiency, achieving superior performance while slashing training token requirements by 100,000X.
Compressing state-of-the-art image generation models by up to 75% without sacrificing quality could revolutionize resource efficiency in AI image synthesis.
ViQ achieves a groundbreaking balance between semantic richness and detail in visual representations, enabling efficient multimodal training without sacrificing quality.
Achieving near-optimal accuracy while slashing costs, this cascaded framework ensures only the toughest queries hit the expensive models.
Self-distillation may boost accuracy but comes at the hidden cost of significantly reduced output diversity, risking performance in diverse scenarios.
Achieving up to 6.72× compression with only minor accuracy trade-offs, HiReLC revolutionizes how we approach neural network optimization.
Small models can achieve competitive performance with innovative data generation and distillation techniques, challenging the notion that bigger is always better in model training.
Student models can guide the creation of complex teacher models, yielding unexpected computational efficiencies and improved performance.
Low-bit quantization can inflate reasoning length, leading to hidden compute costs that traditional accuracy metrics overlook.
IF-Beta allows student models to achieve superior performance with significantly less data and compute, challenging the traditional reliance on full datasets in knowledge distillation.
Curation of training data for brevity can yield a staggering 35x improvement in inference efficiency without sacrificing accuracy in VLMs.
A single elastic model can adapt to multiple sparsity configurations without re-optimization, revolutionizing LLM deployment efficiency.
Achieving comparable performance to full-precision models, BITEMBED slashes storage costs and enhances embedding efficiency with extreme low-bit quantization.
Key-switching overhead in secure Transformer inference can be drastically reduced by offloading computations to a preprocessing phase, enabling efficient deployment even under resource constraints.
PGL-Net achieves state-of-the-art dehazing quality with over 10x faster inference, making it a game-changer for real-time applications.
Achieving a 75% reduction in inference latency while maintaining high reconstruction quality could revolutionize how we deploy Implicit Neural Representations in real-time applications.
Achieving linear-time efficiency in remote sensing instance segmentation could revolutionize how we deploy vision models in resource-constrained environments.
EmuGEMM achieves up to 5.5x speedup over cuBLAS ZGEMM while maintaining accuracy, revolutionizing low-precision matrix multiplication on modern GPUs.
Achieving a staggering 220x speedup in MaxSim scoring while preserving exact retrieval quality could revolutionize the efficiency of multi-vector retrieval systems.
Achieving 51% higher throughput with low-precision arithmetic while maintaining accuracy could redefine efficiency benchmarks in semiconductor simulations.
SOLAR achieves zero observed violations in Speed-of-Light performance analysis, transforming how we approach deep learning optimizations.
Achieving up to 13.9x speedup in LLM inference by leveraging cache-resident execution could redefine performance benchmarks for large models on standard CPUs.
Lightweight transformers can match traditional ML accuracy in fault detection but at a staggering cost of latency and model size, challenging their practicality in real-time applications.
Blockwise policy-drift gating boosts on-policy distillation performance, increasing solve rates by effectively managing policy drift without changing teacher signals.
Compression strategies can preserve general knowledge in LLMs better than previously thought, thanks to innovative supervision techniques.
CompressKV achieves over 97% performance retention with just 3% of the KV cache, revolutionizing resource efficiency in long-context LLMs.
LAS2 redefines the balance between speed and accuracy in stereo matching, proving that efficient models can achieve state-of-the-art results without heavy computational demands.
Memory efficiency in 3DGS-SLAM can be dramatically improved by over 60% without sacrificing accuracy, making it viable for large-scale autonomous driving applications.
Achieving state-of-the-art accuracy in knowledge distillation while maintaining a lightweight framework could redefine efficiency benchmarks in model training.
Achieving a 5x speedup in document parsing without sacrificing accuracy could redefine efficiency benchmarks in Vision-Language Models.
Spectral analysis reveals that tokens with dynamic cross-layer evolution are crucial for preserving semantic integrity, leading to more efficient MLLMs.
PriorTR reveals that ignoring model-induced priors can lead to the loss of critical task-specific information, enhancing MLLM efficiency without sacrificing accuracy.
Achieving bit-perfect random access through dual compression layers in just 0.334ms could revolutionize data retrieval in compressed formats.
Output compression can slash inference costs by up to 3x, but input compression leads to higher costs and accuracy collapse—an unexpected trade-off for LLMs.
Block-GTQ slashes quantization error by up to 80%, transforming how we optimize key-cache performance in long-context models.
Traditional pruning methods compromise prediction accuracy, but a new rationale-informed strategy ensures both accuracy and evidence alignment in VLMs.
Co-located speculative decoding consistently beats distributed variants in latency, challenging the assumption that edge-cloud setups always yield faster inference.
CrossPool achieves a staggering 10.4x reduction in tail latency for bursty long-context requests by decoupling model weights from KV-cache in GPU memory.
The effectiveness of antidistillation defenses hinges on the threat model used, revealing that current defenses may offer a false sense of security.
Speculative decoding at temperature zero shows no significant safety divergence, challenging assumptions about draft model outputs contaminating safety-scored results.
Asynchronous OPD can boost training throughput significantly while managing the challenges of stale data, transforming the efficiency of large language model fine-tuning.
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.
