Causal LLMs (2026) | Causal Reasoning, Causal Discovery, Causal Agents & Benchmarks for Agentic AI

Why Causal Reasoning Is the Next Frontier

8 entries

Language models are trained on correlations in text, yet agents must predict what will happen if they act: which tool call will fix the error, which change caused the regression, what would have happened otherwise. That is causal reasoning, which Judea Pearl describes as a ladder from association to intervention to counterfactuals. Research on causal LLMs follows three threads. First, measuring whether LLMs genuinely reason causally or only repeat causal facts from training data. Second, combining LLM knowledge with formal causal tools for discovery and inference. Third, building agents whose planning, memory and safety rely on causal structure. Evidence is mixed: several studies find strong performance on commonsense causal questions but poor results on formal inference, so treat this as an active research area.

01Causal Reasoning and Large Language Models: Opening a New Frontier for CausalityKıcıman et al. on LLM strengths and failures across causal tasks, a key starting paper. 02Causal Parrots: Large Language Models May Talk Causality But Are Not CausalArgues that LLMs recite causal facts without reliable causal reasoning. 03Causality for Large Language ModelsSurvey of how causality can improve LLM reasoning, fairness, safety and robustness. 04Large Language Models and Causal Inference in CollaborationSurvey of how LLMs and causal inference can work together. 05The Book of Why (Pearl)Companion site for the book that popularised the ladder of causation. 06Elements of Causal InferenceOpen-access text on causal models and learning. 07Introduction to Causal Inference (Brady Neal)Free course with lecture notes. 08Causal Inference for the Brave and TrueFree book with Python examples.

Do LLMs Reason Causally? Benchmarks & Evaluations

11 entries

Measuring causal ability beyond memorised facts.

09CLadder: Assessing Causal Reasoning in Language ModelsBenchmark built on Pearl's ladder with formal causal queries. 10CLadder code and dataBenchmark generation code. 11CLadder on Hugging FaceDataset download. 12Can Large Language Models Infer Causation from Correlation? (Corr2Cause)Tests inferring causal relations from statistical correlations. 13Corr2Cause code and dataData and code for the benchmark. 14A Critical Review of Causal Reasoning Benchmarks for Large Language ModelsReviews what current benchmarks do and do not measure. 15CounterBench: Evaluating and Improving Counterfactual Reasoning in LLMsBenchmark focused on counterfactual questions. 16Counterfactual reasoning: Testing language models' understanding of hypothetical scenariosEarly study of counterfactual understanding. 17CaT-Bench: Causal and Temporal Dependencies in PlansTests understanding of step dependencies in plans, relevant to agents. 18Com2: A Causal-Guided Benchmark for Complex Commonsense ReasoningCausal-guided commonsense benchmark. 19Unveiling Causal Reasoning in Large Language Models: Reality or Mirage?Asks whether LLM causal reasoning is genuine and proposes methods to improve it.

LLM-Assisted Causal Discovery & Inference

9 entries

LLMs supply domain knowledge while statistics test it.

20Causal Discovery with Language Models as Imperfect ExpertsUses LLM suggestions as noisy priors for structure learning. 21Integrating Large Language Model for Improved Causal DiscoveryCombines LLM knowledge with data-driven discovery. 22Efficient Causal Graph Discovery Using Large Language ModelsReduces the number of LLM queries needed to build a causal graph. 23Large Language Models for Constrained-Based Causal DiscoveryLLMs as conditional-independence oracles. 24On the Reliability of Large Language Models for Causal DiscoveryExamines consistency and memorisation risks. 25Multi-Agent Causal Discovery Using Large Language ModelsSeveral LLM agents collaborate on discovery. 26Causal Discovery from Data Assisted by Large Language ModelsCombines data with LLM assistance. 27Can Large Language Models Help Experimental Design for Causal Discovery?LLMs choose which interventions to run. 28Sequential Causal Discovery with Noisy Language Model PriorsHandles unreliable LLM priors.

Causal Analysis Agents & Toolkits

12 entries

Agents and libraries that run causal analysis end to end.

29Causal Agent based on Large Language ModelAgent that uses tools to answer causal questions. 30Causal-Copilot: An Autonomous Causal Analysis AgentLLM agent that automates causal discovery, inference and interpretation. 31Causal-Copilot on GitHubCode for the Causal-Copilot agent. 32ORCA: ORchestrating Causal AgentOrchestrates causal analysis steps with agents. 33CausalSteward: An Agentic Divide-Conquer-Combine Copilot for Causal Discovery2026 preprint on agentic copilots for discovery. 34Causal Discovery in the Era of Agents2026 preprint on how agents change causal discovery practice. 35PyWhyOpen-source ecosystem for causal machine learning. 36PyWhy-LLMExperimental library integrating LLMs into causal analysis. 37DoWhyCausal inference library with explicit assumptions and refutation tests. 38causal-learnPython package for causal discovery. 39EconMLHeterogeneous treatment effect estimation. 40dodiscoverExperimental causal discovery library.

Causal Reasoning Inside Agentic Systems

11 entries

Using causality for planning, memory, world models and safety.

41Chain-of-Thought Prompting Elicits Reasoning in Large Language ModelsStep-by-step prompting underlying most agent reasoning. 42ReAct: Synergizing Reasoning and Acting in Language ModelsInterleaves reasoning with actions. 43Tree of ThoughtsDeliberate search over reasoning paths. 44Reflexion: Language Agents with Verbal Reinforcement LearningAgents learn from feedback on failures. 45ToolformerLanguage models that learn to call tools. 46When Do Causal World Models Help Modular LLM Agents2026 preprint on when causal world models benefit agents. 47Causal Intervention-Based Memory Selection for Long-Horizon LLM Agents2026 preprint on choosing memories by causal effect. 48Active Causal Discovery Benchmark: Evaluating LLM Agents Under Budgeted Interventions2026 benchmark where agents must intervene to learn structure. 49CausalGame: Benchmarking Causal Thinking of LLM Agents in Games2026 benchmark of agent causal thinking. 50Enhancing LLM Agent Safety via Causal Influence PromptingPrompting agents to reason about the causal effects of actions. 51Causality Laundering: Denial-Feedback Leakage in Tool-Calling LLM Agents2026 preprint on a security issue in tool-calling agents.

10 entries

Neighbouring pages on this site.

52Causal AI ModelsCausal AI methods and libraries. 53Graph & Causal Representation LearningGraph and causal representation learning. 54Agentic AIAgentic AI hub. 55Top Agentic AI FrameworksFrameworks for building agents. 56Loop vs Graph EngineeringAgent control-flow design. 57LLM-as-a-Judge SolutionsEvaluating model outputs with LLMs. 58Large Language ModelsLLM hub. 59Guardrails for Agentic AIGuardrails for agents. 60World Action ModelsWorld models for agents. 61Reinforcement Learning AlgorithmsReinforcement learning.

Frequently Asked Questions

What are causal LLMs?

Causal LLMs is a research direction on how large language models reason about cause and effect and how causal methods can make them and the agents built on them more reliable. It covers evaluating causal reasoning, LLM-assisted causal discovery, causal analysis agents and causal designs for agent planning, memory and safety.

Can LLMs actually reason causally?

Evidence is mixed. LLMs often answer commonsense causal questions well, but benchmarks such as CLadder and Corr2Cause show weaker results on formal causal inference, and some papers argue they mostly repeat causal facts from training. Results depend heavily on the benchmark, model and prompting, so validate on your own task.

How are causal methods used with LLM agents?

Agents can call tools such as DoWhy or causal-learn, use LLM knowledge as priors for causal discovery that data then tests, and use causal structure to select memories, plan interventions or assess whether an action will cause harm. Many of these ideas are recent preprints.

Where should I start?

Read Kıcıman et al. and the CLadder and Corr2Cause papers for the evaluation picture, then try DoWhy with PyWhy-LLM or Causal-Copilot. Pearl's Book of Why and Brady Neal's course cover the underlying theory.

How current is this page?

Links were checked in October 2026. Many entries, especially those from 2026, are recent arXiv preprints that have not been peer reviewed, and results on LLM causal ability vary widely by benchmark and model. Sites that block automated checks are kept and may need a manual click.

Causal LLMs: The Next Frontier of Agentic Reasoning