Sophia
Liu
Policy Learning for Tool Orchestration in Biomedical Systems
Abstract profile. Full document pending author claim.
Authors:
Sophia Liu, Nada Amin
Date Created:
2025-01-01
Course Title:
Professor:
Not specified
About Paper:
While modern large language models (LLMs) excel at text tool invocations using constructs like branching, looping, and generation, they often struggle with complex problem-solving variable binding. TOLD allows LLMs to learn complex tasks that require precision, multi-step reasoning, and grounded tool orchestration strategies without extensive manual prompt knowledge. Tool integration has emerged as a key strategy engineering. To optimize tool-use sequences, we apply in-context for addressing these limitations, enabling access to domain- policy iteration, refining demonstrations based on model rollouts. specific knowledge and serving as a crucial mechanism to mitigate The reinforcement learning process is guided by reward models hallucinations and inaccuracies. Although prior work shows fine-tuned for biomedical knowledge. We evaluate TOLD on that LLMs can sequence tool calls in game environments and drug repurposing tasks, comparing its performance, in terms of simplified settings, real-world scientific domains, such as biologythe ability to choose the correct tool and properly formulate the and medicine, pose greater challenges, requiring latent structure, tool call (i.e. using the correct parameters), to vanilla ReAct-style domain-specific reasoning, and multi-step task planning. We LLMs. Preliminary results suggest that TOLD can enhance LLM introduce TOLD (Tool-orchestration language for discovery), a capabilities and integrate the domain-specific reasoning needed to structured tool-calling language that prompts LLMs to compose empower scientific discovery. RelationshipbetweenTSPOAla147ThrPolymorphismandPROMIS-29Health Domains in Chronic Pain Patients and Healthy Controls Danika Yang, Minhae Kim, Ekim Luo, Marco Loggia Harvard College | Dunster House | Neuroscience | 2027 The 18-kDa translocator protein (TSPO) is a mitochondrial affinity groups, which were high affinity binders (HAB), membrane protein implicated in neuroimmune signaling and mixed affinity binders (MAB), and low affinity binders (LAB). neuroinflammation. There exists a common single nucleotide All participants completed the PROMIS-29 questionnaire. polymorphism, Ala147Thr (rs6971), in the TSPO gene that alters Multivariate linear models evaluated associations between binding ligand binding affinity, thereby categorizing individuals as high-,affinity and PROMIS-29 scores. mixed-, or low-affinity binders. This variation carries important implications: TSPO-PBR28 binding not only influences PET Data collection and analysis are ongoing. Based on current imaging quantification, but the protein itself is also responsible forerature linking neuroinflammatory signaling to chronic pain steroidogenesis, and thus holds potential in modulating biological and fatigue, we hypothesize, thus far, that the HAB genotype may be associated with more pain interference, catastrophizing, processes linked to the physical and psychological manifestations fatigue, and psychosocial detriment in cLBP patients. Identifying associated with chronic pain. The current study investigates such associations could highlight TSPO-related pathways whether patient TSPO genotype correlates with patient-reported behavioral outcomes measured by the PROMIS-29 Profile, a as contributors to variability in pain-related outcomes and validated instrument assessing physical function, depression, psychosocial well-being. anxiety, fatigue, sleep disturbance, satisfaction with social roles,uture directions will incorporate cross-sectional investigations and pain interference in individuals with chronic low back pain into how binding affinity genotype may differently affect various (cLBP) and healthy controls. chronic pain conditions apart from cLBP, such as carpal tunnel syndrome, Gulf War illness, fibromyalgia, and knee osteoarthritis. Participants with diagnosed cLBP and healthy controls provided Ultimately, this research may inform biomarker-driven approaches blood samples for genotyping of the Ala147Thr (rs6971) for stratifying chronic pain phenotypes and tailoring specific polymorphism. Genotypes were used to classify TSPO binding therapeutic strategies.
Abstract:
While modern large language models (LLMs) excel at text tool invocations using constructs like branching, looping, and generation, they often struggle with complex problem-solving variable binding. TOLD allows LLMs to learn complex tasks that require precision, multi-step reasoning, and grounded tool orchestration strategies without extensive manual prompt knowledge. Tool integration has emerged as a key strategy engineering. To optimize tool-use sequences, we apply in-context for addressing these limitations, enabling access to domain- policy iteration, refining demonstrations based on model rollouts. specific knowledge and serving as a crucial mechanism to mitigate The reinforcement learning process is guided by reward models hallucinations and inaccuracies. Although prior work shows fine-tuned for biomedical knowledge. We evaluate TOLD on that LLMs can sequence tool calls in game environments and drug repurposing tasks, comparing its performance, in terms of simplified settings, real-world scientific domains, such as biologythe ability to choose the correct tool and properly formulate the and medicine, pose greater challenges, requiring latent structure, tool call (i.e. using the correct parameters), to vanilla ReAct-style domain-specific reasoning, and multi-step task planning. We LLMs. Preliminary results suggest that TOLD can enhance LLM introduce TOLD (Tool-orchestration language for discovery), a capabilities and integrate the domain-specific reasoning needed to structured tool-calling language that prompts LLMs to compose empower scientific discovery. RelationshipbetweenTSPOAla147ThrPolymorphismandPROMIS-29Health Domains in Chronic Pain Patients and Healthy Controls Danika Yang, Minhae Kim, Ekim Luo, Marco Loggia Harvard College | Dunster House | Neuroscience | 2027 The 18-kDa translocator protein (TSPO) is a mitochondrial affinity groups, which were high affinity binders (HAB), membrane protein implicated in neuroimmune signaling and mixed affinity binders (MAB), and low affinity binders (LAB). neuroinflammation. There exists a common single nucleotide All participants completed the PROMIS-29 questionnaire. polymorphism, Ala147Thr (rs6971), in the TSPO gene that alters Multivariate linear models evaluated associations between binding ligand binding affinity, thereby categorizing individuals as high-,affinity and PROMIS-29 scores. mixed-, or low-affinity binders. This variation carries important implications: TSPO-PBR28 binding not only influences PET Data collection and analysis are ongoing. Based on current imaging quantification, but the protein itself is also responsible forerature linking neuroinflammatory signaling to chronic pain steroidogenesis, and thus holds potential in modulating biological and fatigue, we hypothesize, thus far, that the HAB genotype may be associated with more pain interference, catastrophizing, processes linked to the physical and psychological manifestations fatigue, and psychosocial detriment in cLBP patients. Identifying associated with chronic pain. The current study investigates such associations could highlight TSPO-related pathways whether patient TSPO genotype correlates with patient-reported behavioral outcomes measured by the PROMIS-29 Profile, a as contributors to variability in pain-related outcomes and validated instrument assessing physical function, depression, psychosocial well-being. anxiety, fatigue, sleep disturbance, satisfaction with social roles,uture directions will incorporate cross-sectional investigations and pain interference in individuals with chronic low back pain into how binding affinity genotype may differently affect various (cLBP) and healthy controls. chronic pain conditions apart from cLBP, such as carpal tunnel syndrome, Gulf War illness, fibromyalgia, and knee osteoarthritis. Participants with diagnosed cLBP and healthy controls provided Ultimately, this research may inform biomarker-driven approaches blood samples for genotyping of the Ala147Thr (rs6971) for stratifying chronic pain phenotypes and tailoring specific polymorphism. Genotypes were used to classify TSPO binding therapeutic strategies.
Source:
Harvard / Devin Morales, Jason Yu, Sam Van OOijen, Doeke Hekstra / 2025
Topics:
tool, pain, affinity, binding, chronic, tspo, llms, domain, model, told, patient, binder
