Beyond Short-Term Memory: Building Agents That Plan Ahead
![The study investigates how language model-based agents can be augmented with external oracles-[latex]\mathcal{O}^{\text{state}}[/latex] for summarizing state, [latex]\mathcal{O}^{\text{plan}}[/latex] for hinting waypoints, and [latex]\mathcal{O}^{\text{history}}[/latex] for rewriting task descriptions-to navigate multi-turn tasks, effectively pruning historical context and enabling agents to make decisions independent of prior steps within a Markov decision process.](https://arxiv.org/html/2601.16649v1/x1.png)
New research dissects the challenges facing conversational AI as it attempts to navigate complex, multi-step tasks requiring sustained reasoning.
Science
Simulating City Life: AI Agents Drive Realistic Traffic Patterns
Researchers are leveraging the power of artificial intelligence to create more accurate and nuanced simulations of how people move within urban environments.
Beyond AlphaGo: Injecting Human Insight into AI Game Play
Researchers have developed a new AI system that blends the power of large language models with curated human expertise to achieve expert-level performance in the complex game of Go.
Flow Control: Bridging Physics and Deep Learning
A new hybrid approach combines the strengths of finite element methods and neural networks to accurately simulate fluid dynamics in complex geometries.
Smart Data for Photonics: Learning to Predict Light
A new approach leverages intelligent data selection to dramatically accelerate the design of photonic crystals and other optical components.
Can AI Be Manipulated to Undermine Truth?
New research reveals how sophisticated language models can craft persuasive arguments that fool automated fact-checking systems, even with supporting evidence.
Mapping the Brain of the Machine
![A neural network’s internal logic is exposed through the construction of a neural graph-derived directly from its architecture-where the importance of each connection is quantified by a neural curvature [latex]\kappa\_{N}(u,v,x)[/latex] calculated across a calibration set [latex]\mathcal{D}[/latex], ultimately allowing for a ranking of connections based on their influence.](https://arxiv.org/html/2601.16366v1/img/overview.png)
New research applies the tools of differential geometry to understand how information travels within neural networks, offering a fresh perspective on network architecture and optimization.
Stochastic Activation Functions Boost Financial Forecasting
A new activation function, Brownian ReLU, leverages the principles of Brownian motion to improve the performance of long short-term memory networks.
Taming the Noise: A New Approach to High-Frequency Data Analysis
A robust subsampling method offers improved accuracy when estimating volatility from noisy, rapidly sampled financial data.
The Power of Anonymity: Pricing in Crowded Markets
New research shows that concealing price information can surprisingly optimize revenue in large-scale auctions with numerous participants.