The LARK Lab from CU Anschutz will be sharing research updates from their ongoing projects.

 

Presentation 1.  

Speaker: He Cheng, PhD (CU Anschutz LARK Lab’s postdoc. He graduated with a PhD. in CS from Utah State). 
Title: LogosKG: Hardware-Optimized Scalable and Interpretable Knowledge Graph Retrieval
Abstract: This talk will present LogosKG, a hardware-optimized framework for scalable and interpretable knowledge graph (KG) retrieval. Even though there are already many systems or libraries that can achieve retrieval, they struggle to balance efficiency, scalability, and interpretability. To overcome this issue, LogosKG decomposes KGs into subject, object, and relation sparse matrices and reformulates KG retrieval as sparse matrix operations. LogosKG integrates novel graph-scaling techniques including degree-aware partitioning, cross-graph routing, and on-demand caching, each designed for efficient parallel computation across distributed graph workloads. Experimental results show substantial performance gains over CPU and GPU baselines. A case study will also be included to demonstrate how LogosKG can be utilized to facilitate graph analysis, LLM diagnostic predictions, and govern the alignment between them, paving a path toward next-generation KG–LLM integration.  

Presentation 2. 

Speaker:  Saksham Khatwani (CU Boulder CS MS student)

Title: Brittleness and Promise: Knowledge-Graph-Based Reward Modeling for Diagnostic Reasoning. 

Abstract:  Even though Large language models (LLMs) show promise in diagnostic reasoning, they suffer from hallucination and often lack reliable, knowledge-grounded inference. Biomedical knowledge graphs (KGs) provide structured medical knowledge that may enable more structural reasoning. This talk will focus on our exploration of a reward-style approach, inspired by “reward-model-as-reasoning”, that trains an LLM to verify whether a candidate KG path plausibly connects a patient’s findings to a diagnosis. Clinically, this parallels how physicians evaluate and refine candidate explanations rather than exhaustively search possibilities. We will present our findings on KG-based reasoning, comparing five graph-based task formulations and seven training paradigms across two medical KGs and three open-source LLMs. While reward-style and distillation methods show promise in path-judging accuracy, they transfer weakly to downstream tasks such as diagnosis summarization and medical QA, underscoring the limited generalization of current KG-based reasoning approaches.

Presentation 3. 

Speaker: Shiyue Hu (CU Boulder CLASIC MS student)

Title: Investigating Racial Bias in LLMs Through Mechanistic Interpretability

Abstract: This ongoing research investigates racial bias in Large Language Models (LLMs) through mechanistic interpretability. We employ linear probing on LLMs, using race conditioned medical prompts to extract and analyze layer specific MLP neuron activations. Our goal is to identify how racial demographic information is encoded and noticeably represented within the model’s intermediate layers. By analyzing probe weights and neuron activation calculations, we aim to identify bias related neurons that show different activity when processing racial identity information. The project is developing a sign-aware Logit Lens to semantically decode these neurons’ contributions, thereby mapping internal states to interpretable concepts. This work aims to deepen the understanding of bias propagation and inform mitigation strategies in LLMs.