Squeezing More from Less: Smarter Quantization for Large Language Models
A new method previews future data to significantly improve the compression of large language models without sacrificing accuracy.
Science
Securing the Connected Car: A New Path for Data Privacy
![A roadside unit (RSU) securely transmits data to a cloud server via a homomorphically-encrypted, iterative process-first encrypting symmetric keys, then uploading encrypted bulk sensor messages (BSMs), allowing the cloud to perform analytics on the ciphertext itself [latex] \text{without decryption} [/latex], and finally returning encrypted results for authorized decryption at the traffic management center or RSU.](https://arxiv.org/html/2602.02717v1/Figure_HHE.png)
This review explores how combining different encryption methods can dramatically improve data security and reduce communication delays in intelligent transportation systems.
Beyond Crash Tests: Securing the Future of Data Storage
As storage systems grow in complexity, traditional testing methods struggle to guarantee data integrity, demanding new approaches to uncover hidden vulnerabilities.
Speeding Up the Blockchain: Parallel Validation for Faster Rewards
New research demonstrates that leveraging multi-core processors can dramatically accelerate blockchain validation and construction, boosting validator performance and profitability.
Beyond Simplifications: Refining Heavy-Ion Collision Analysis
![The ratio of coalescence differences-[latex]\frac{\mathcal{B}-\mathcal{B}_{\mathrm{smooth}}}{\mathcal{B}_{\mathrm{smooth}}}[/latex]- exhibits a sensitivity to source size, transverse momentum, and temperature, as demonstrated under fixed conditions of [latex]\Delta\tau=1.5[/latex] fm, [latex]\beta_S=0.5[/latex], and n=2, with right-column plots specifically referencing parameter sets from Figure 1.](https://arxiv.org/html/2602.02810v1/coal-size.png)
New research quantifies the limitations of common approximations used to understand the behavior of matter created in high-energy heavy-ion collisions.
Shrinking Giants: A New Approach to Language Model Compression
Researchers have developed a novel quantization technique that dramatically reduces the size of large language models without significant performance loss.
Unlocking the Secrets of Quantum Matter: A Critical Transition Revealed
![Delayed update algorithms within probabilistic quantum Monte Carlo simulations of the Hubbard model on honeycomb lattices demonstrate performance gains-optimized by system-specific [latex]n_{d}[/latex] values-over fast update methods, revealing that strategic deferral of updates can substantially accelerate calculations of complex quantum systems.](https://arxiv.org/html/2602.03656v1/x6.png)
New high-precision simulations pinpoint the behavior of electrons at a crucial phase transition in a model system for exotic materials.
QUIC Migration and the Middlebox Challenge
The rise of QUIC’s connection migration is forcing a rethink of how network middleboxes handle stateful connections.
Graph-Based Signatures Secure the Post-Quantum Future
![The research demonstrates the generation of graphs amenable to proper [latex]k[/latex]-colorability, highlighting a construction method with implications for graph theory and algorithm design.](https://arxiv.org/html/2602.02689v1/grp.png)
A new signature scheme, Eidolon, leverages the complexity of graph coloring to offer robust security in an era threatened by quantum computing and advanced machine learning.
Building Brain-Inspired Resilience: Error Correction for Neural Networks
A new coding framework enhances the robustness of neural networks against errors in both memory and computation, mimicking the brain’s inherent fault tolerance.