Architecture Design (Transformers, SSMs, MoE)

RhyMix outperforms existing models by dynamically adapting its forecasting strategy to capture both rhythmic and local patterns without sacrificing efficiency.

Achieving 205× latency reduction in FPGA-based neural networks could redefine the benchmarks for ultra-fast inference in latency-critical applications.

Layer patching can dramatically enhance model performance in size interpolation, revealing that simple strategies often outperform complex methods.

Multi-deformation modeling can be achieved more effectively by choosing the right integration strategy, impacting the robustness of dynamic scene reconstruction.

EV-MoE not only enhances feature representation but also introduces a large-scale benchmark that redefines multi-query vehicle ReID evaluation in complex environments.

Fiber Memory can slash weight-delivery energy by over 70% while eliminating redundant storage across thousands of AI accelerators.

Migrating large-model inference to non-GPU accelerators like Huawei Ascend reveals eight critical limitations that could derail performance and reliability.

Transforming LLM prototypes into auditable agents can ensure compliance and safety without sacrificing performance, achieving full utility in complex enterprise applications.

MAESTRO prunes MoE models more effectively by leveraging the interdependencies of expert routing, achieving superior performance retention and consistency across tasks.

Delta-style networks outperform traditional methods by leveraging key-dependent projections, revealing a new path to efficient transformer inference.

Targeted layer insertion based on rigorous error estimation leads to superior generalization in neural networks, outperforming traditional architecture adaptation techniques.

Moderately expressive neural networks outperform more complex models in recovering mechanistic operators from sparse data, revealing the critical balance needed in architecture and optimization.

FMMVCC achieves unprecedented clustering performance for univariate time series by efficiently capturing long-range dependencies with linear complexity.

Deep ReLU networks can create intricate piecewise linear partitions of input space, fundamentally altering our understanding of their training dynamics.

Modality-aware graph encoding in generative models can dramatically enhance the fidelity and efficiency of neuroimaging feature analysis.

Structural designers thrive on friction, and interactive AI can enhance creativity by preserving the reflective challenges of the design process.

Tailoring sparsity to layer importance can slash perplexity by over 1.9 points, challenging the one-size-fits-all approach in transformer pruning.

Spectral preprocessing of query-key projections can reduce attention computation costs by up to 79% while preserving performance in character-level language modeling.

TF-Engram achieves a notable performance boost in LLMs by integrating scalable, train-free semantic memory without the typical overhead of retraining.

eBIM redefines blockchain infrastructure management by integrating RISC-V to reclaim trust and flexibility lost in traditional BaaS models.

Achieving robust video semantic communication under extreme conditions, MamVSC maintains high quality even with 90% packet loss and -8 dB SNR.

ATLAS automates the transition from high-level deep learning models to FPGA implementations, drastically reducing the manual effort required for custom hardware acceleration.

Achieving a 2.52x speedup in quantum circuit execution could redefine efficiency benchmarks in scalable quantum computing.

Achieving a 3.5x boost in Energy-Delay-Inverse-Yield, ThermoDSE redefines the optimization landscape for chiplet-based DNN accelerators under thermal constraints.

Sparse state vector simulations can drastically cut down computational costs while accurately predicting outputs of peaked quantum circuits.

Kimi Delta Attention with Muon outperforms other architectures in validation loss, while introducing Cross-Layer Value Routing reveals new avenues for optimizing memory management in linear attention models.

Scaling linear RNNs with Sparse Delta Memory leads to dramatic gains in long-context recall without increasing computational overhead.

LingBot-Video bridges the gap between digital creativity and physical actuation, achieving unprecedented efficiency in video pretraining for embodied intelligence.

Switching between autoregressive and diffusion modes allows Nemotron-Labs-Diffusion to achieve unprecedented throughput and efficiency in language modeling.

Evolving LLMs into a heterogeneous intelligent ecology could fundamentally alter how we approach AI alignment and governance.

A single minimal infinite-state tool can elevate finite-precision models to Turing completeness, while finite-state tools add virtually no expressivity.

xDECAF transforms data flow analysis in information security with a robust framework that has already gained traction across several research lines.

Unlocking over 14,000 unique neural architectures, LEMUR 2 sets a new standard for cross-domain evaluation and deployment in AI design.

Achieving a 5 to 6 orders of magnitude speedup in effective resistance analysis could revolutionize the design of reliable 3D IC power delivery networks.

Direct BRAM-DSP connections can boost FPGA performance by 25% without the need for extensive architectural overhauls.

Scattering networks can achieve optimal separation capacity by strategically tuning filter frequencies and ensuring well-conditioned geometric couplings.

Linear attention fails to capture spectral variations in graphs, but Graph Convolutional Attention achieves superior denoising by directly utilizing the graph spectrum.

NTK regression can require exponentially more samples than necessary for compositional tasks, revealing a critical gap in our understanding of neural network performance.

ELSA3D outperforms existing unified 3D models by halving computational costs while enhancing cross-modal reasoning precision.

Finite-width neural networks converge to their Gaussian-process limits with error bounds that shrink as the network width increases, revealing a surprising robustness across architectures.

Models that excel in static conditions may falter dramatically as data evolves, revealing a trade-off between initial accuracy and long-term reliability.

Achieving a 99.81% logical accuracy in quantum error correction by intelligently routing only 3.3%-6.2% of syndromes to a refinement stage reveals a transformative approach to real-time decoding.

A novel error-correction method for QCNNs reduces qubit overhead while significantly improving learning rates in noisy environments.

RoME achieves superior robustness against multiple adversarial threats by intelligently routing them through specialized expert pathways, outperforming existing methods in both accuracy and resilience.

UBEP slashes All-to-All latency by over 52%, unlocking the full potential of Mixture-of-Experts models on superpod architectures.

Reducing picture buffer access from 152 to 1 could revolutionize the efficiency of video encoding in VVC.

Achieving over 4600x speedup in OTA design without sacrificing accuracy could revolutionize circuit modeling practices.

Achieving a 94.54% reduction in analysis time while maintaining just 1.63% average accuracy error could revolutionize reliability assessments in digital circuit design.

CAP achieves up to 86% higher throughput in MoE models while preserving accuracy, revolutionizing how we optimize expert placement and pruning.

Achieving up to 30.9x speedup in ML inference by optimizing workload partitioning between CPUs and CIM accelerators reveals the untapped potential of heterogeneous computing.