Reasoning Models: How Easily Can They Be Led Astray?
![Confidence-aware generation ([latex]CARG[/latex]) consistently stabilizes performance across standard large language models, yet fails to benefit-and sometimes diminishes the accuracy of-models demonstrating substantial reasoning capabilities.](https://arxiv.org/html/2602.13093v1/x2.png)
New research reveals that while large reasoning models demonstrate improved consistency, they remain surprisingly vulnerable to subtle manipulation in extended conversations.
Science
Navigating Risk in Collaborative AI
New research provides a mathematically proven method for training multi-agent systems to make stable, risk-aware decisions in complex environments.
Can Your Voice Search Handle the Noise?
A new benchmark assesses how well spoken query retrieval systems perform in real-world acoustic conditions, moving beyond ideal lab settings.
Secure Ballots: The Cryptographic Underpinnings of Electronic Voting
This review explores the essential cryptographic techniques that guarantee both the privacy of voters and the integrity of electronic elections.
Erasing Memories: Protecting Language Models After Compression
New research shows how to effectively remove specific information from large language models even after they’ve been compressed for efficient deployment.
Dancing with Protocols: A Language for Secure Choreography
A new formal language, CryptoChoreo, enables rigorous modeling and automated verification of complex, stateful cryptographic protocols.
Smarter Solvers: Boosting Accuracy in Corrupted Linear Systems
A new iterative method intelligently filters out unreliable data, accelerating and improving the robustness of solutions for linear equations affected by noise and errors.
Scaling Electronic Correlation to Nanoscale Systems
A new computational method efficiently tackles the complex problem of calculating electronic correlation energies in large, complex materials.
Fortifying Containers: A Multi-Layered Defense Against Subtle DDoS Attacks
![Figure 1:Dynamic Blacklisting A system dynamically adjusts its acceptance criteria, discarding previously viable options-a process akin to iteratively refining a solution space by strategically eliminating unproductive pathways-and prioritizing alternatives based on real-time performance feedback, effectively transforming a static rule set into a self-correcting mechanism where [latex] \text{Acceptance} = f(\text{Performance}, \text{Time}) [/latex].](https://arxiv.org/html/2602.11407v1/dynamic_blacklisting.jpg)
This review explores a robust security architecture designed to protect containerized systems from the growing threat of low-rate Distributed Denial of Service attacks.
The Price of Privacy: Memory Limits for Tracking Online Users
New research demonstrates that maintaining user privacy in streaming data requires surprisingly large memory resources, fundamentally limiting the efficiency of certain algorithms.