Research

TL;DR: We steer LLMs away from hallucinations by dynamically amplifying context-sensitive attention heads.

Large language models (LLMs) often produce fluent but factually incorrect statements, even when relevant evidence is available, due to misallocation of attention between contextual inputs and parametric knowledge. Ensuring that models actively reason over context and retrieve relevant information is critical for trustworthy and interpretable AI. We introduce COMPASS (Context-Modulated PID Attention Steering System), a lightweight, interpretable framework that dynamically steers attention to retrieved context during generation. Using the Context Reliance Score (CRS), COMPASS identifies which attention heads are underutilizing context, and a PID controller adjusts them in real time to improve evidence grounding and factual consistency. This mechanism enables the model to demonstrate advanced reasoning by actively returning to context and retrieving supporting information when needed, without retraining or multi-pass decoding. Across benchmarks including HotpotQA, XSum, HaluEval, and RAGTruth, COMPASS reduces hallucinations by 2.8–5.8% absolute while revealing how attention heads contribute to context-aligned reasoning. These results show that feedback-driven, interpretable control can enhance reasoning, retrieval, and evidence based generation in LLMs.

TL;DR: Using Colley's matrix (a traditional sports-ranking system), we propose an urban sustainability index/ranking list to rival Yale EPI.

Using different multipliers and b-vectors as applications of Colley's Matrix, we have produced many different viable rankings for different countries based on their environmental metrics: percent renewable electricity, pollution AQI, recycling EPI, CO2 emissions, waste, and wastewater generated. We found that all the different rankings produce similar results and trends as Yale EPI, even with some variation.

TL;DR: A comparative analysis of PID, PI, PD, LQR, and hybrid controllers on a 6-DOF robotic arm, measuring accuracy via MAE and efficiency via motor output.

Robotic systems rely heavily on feedback control loops for stability, control, and precision. This study evaluates and compares PID, PI, PD, LQR, and hybrid PD+LQR controllers using an assembled SO101 6-degree-of-freedom Arduino robotic arm. The study measures both accuracy and efficiency through programmed, predefined three-dimensional paths (circle, square, heart). Accuracy was evaluated using Mean Average Error (MAE), and efficiency was assessed using motor output magnitude. Results showed that PD controllers achieved the highest accuracy with statistical significance (p = 0.046), while PI controllers demonstrated the best energy efficiency with 43x less motor output than LQR. Contrary to expectations, LQR-based systems showed the least efficiency, though hybrid PD+LQR approaches outperformed standalone LQR controllers.

TL;DR: Investigating Pd islands on Ag(111) substrates as cost-effective heterogeneous catalysts for CO oxidation using LT-STM and TPD techniques.

The development of high-performance, cost-effective catalysts is essential to address rising global energy demands. This study evaluated the structure and functionality of Pd islands on Ag(111) substrates as catalysts for CO oxidation. Pd islands were synthesized using physical vapor deposition (PVD) and verified via low-temperature scanning tunneling microscopy (LT-STM). Temperature programmed desorption (TPD) assessed catalytic activity. Results showed Pd/Ag samples were more effective at oxidizing CO than pure Ag(111), with peak desorption at 86.1°C and a 250% increase in initial coverage. STM imaging confirmed that island geometry expands with increased Pd coverage, supporting the theory that threefold Pd site geometry leads to greater desorption efficiency.