TLDR: CARENLI is a new framework that improves how large language models (LLMs) reason in clinical settings. It breaks down complex clinical inference into four specific reasoning types (causal, compositional, epistemic, risk) and uses specialized AI agents (planner, solver, verifier, refiner) for each. This structured approach significantly boosts accuracy, especially in causal attribution and risk abstraction, by guiding LLMs to apply principled reasoning rules rather than generic heuristics. The main challenge identified is accurately classifying the reasoning type, but the framework offers a path to safer and more auditable AI in healthcare.
Large Language Models (LLMs) have shown impressive capabilities across many natural language processing tasks, often outperforming humans on standard benchmarks. However, in critical domains like clinical research and practice, these successes can hide significant limitations, especially when tasks require structured, domain-specific reasoning rather than just recognizing surface-level patterns. Errors in such contexts can have serious consequences, as models might possess relevant medical knowledge but fail to apply it according to established inferential rules, leading to incorrect or unsafe conclusions.
Clinical Natural Language Inference (NLI) is a prime example of this challenge. Unlike general NLI, which can sometimes be solved by simple word matching, clinical inference demands sophisticated reasoning. This includes understanding causal relationships, complex interactions between multiple factors, resolving conflicting evidence, and assessing latent health risks. Mistakes here could mean confusing cause with correlation, overlooking dangerous drug interactions, or ignoring critical risks where severity outweighs how often they occur.
To address these issues, researchers have introduced a new framework called CARENLI (Compartmentalised Agentic Reasoning for Clinical NLI). This innovative approach aims to improve the reliability and auditability of LLMs in clinical NLI by separating knowledge access from the process of principled inference. The details of this work can be found in the research paper: Compartmentalised Agentic Reasoning for Clinical NLI.
Understanding Clinical Reasoning Families
The CARENLI framework is built upon a benchmark that breaks down clinical inference into four distinct reasoning families, each with its own specific rules and criteria:
- Causal Attribution: This involves distinguishing between simple observations and true cause-and-effect claims, requiring careful consideration of timing, control groups, and potential confounding factors.
- Compositional Grounding: This evaluates whether a statement’s clinical validity can be inferred from complex configurations involving multiple interacting variables, such as drug dose, co-existing conditions, or trial phases.
- Epistemic Verification: This tests the ability to determine the truth of a claim based on supporting evidence, rather than simply accepting it due to the authority of the source. It involves evaluating evidence hierarchies.
- Risk State Abstraction: This focuses on reasoning about potential or hidden clinical risks and probable outcomes, especially when these are not explicitly stated in the text. It integrates severity and likelihood.
Previous studies have shown that LLMs often know the relevant clinical facts but struggle to apply them correctly within these specific reasoning frameworks, frequently defaulting to generic heuristics.
How CARENLI Works: A Multi-Agent Approach
CARENLI tackles this “knowledge-reasoning dissociation” by adopting a compartmentalized agentic framework. It routes each premise-statement pair to a family-specific “solver” and enforces auditable procedures through a series of specialized agents:
- Planner: This agent is responsible for identifying the dominant reasoning signature of a given problem and routing it to the correct reasoning family (Causal, Compositional, Epistemic, or Risk).
- Solver: Once routed, the solver applies reasoning type-specific rules and decision procedures to the task. For example, a Causal Attribution solver would require evidence of interventional contrast, while an Epistemic Verification solver would prioritize objective measurements over interpretations.
- Verifier: This agent audits the solver’s output for both factual accuracy (ensuring all knowledge is premise-grounded or clinically admissible) and adherence to the family-specific reasoning schemas. It flags violations.
- Refiner: The final stage, the refiner, applies minimal corrections prescribed by the verifier. It doesn’t generate new inferences but edits the solver’s output to ensure factual validity and alignment with principled reasoning.
This modular design prevents LLMs from taking “heuristic shortcuts” and instead guides them to follow principled decision rules, making the reasoning process more transparent and reliable.
Significant Performance Gains and Key Insights
The researchers evaluated CARENLI across four contemporary LLMs (GPT-4o, GPT-4o-mini, Gemini 2.5 Pro, and DeepSeek R1) and found substantial improvements in reasoning fidelity. CARENLI consistently achieved accuracies exceeding 70% in three of the four reasoning types, with peak performance reaching 98.0% in Causal Attribution and 81.2% in Risk State Abstraction. These gains were often more than 30 percentage points higher than traditional “reasoning-agnostic” prompting methods.
A crucial finding was that while the solvers themselves were highly capable of following the expected inferential schemas, the main bottleneck was often the initial classification of the reasoning family by the Planner. When the correct reasoning family was provided directly (an “Oracle Planner” condition), accuracy in Compositional Grounding, for instance, dramatically rose from 18–48% to 90%, and in Risk State Abstraction, it reached 100%. This indicates that LLMs possess the underlying capacity for structured reasoning but need explicit guidance to apply it correctly.
The Verifier stage also proved highly effective, often exceeding the solver’s initial accuracy in correctly identifying invalid reasoning. This suggests that LLMs are better at recognizing errors than at producing perfectly valid reasoning from the start. The Refiner, while making minimal edits, successfully corrected a substantial share of epistemic errors, particularly those related to factual inaccuracies.
Also Read:
- Pinpointing Errors in AI Teams: A New Approach to Debugging Multi-Agent Systems
- Small Language Models Show Promise in Formal Logic Reasoning for Ontology Engineering
Implications for Safer AI in Healthcare
The CARENLI framework demonstrates that structuring LLM inference into specialized, auditable agents can significantly improve their reliability in safety-critical applications like clinical NLI. By enforcing adherence to domain-grounded inferential schemas, CARENLI offers a pathway towards more trustworthy and transparent AI systems in healthcare. The ability to diagnose reasoning failures and enforce principled decision rules is vital for deploying LLMs in contexts where erroneous conclusions can have direct implications for patient safety.
