The article distinguishes prediction from causal inference and explains how confounders and counterfactuals affect causal claims. It outlines a basic causal analysis process and examples of its use in several industries.
Causal-memory is a memory system with a causal core built in Rust and SQLite (Apache-2.0). It models inhibition, stores facts and temporal state, and records decision-to-outcome causal edges in one database, enabling agents to recall events, their truth, why actions worked, and what would happen if they acted differently.
DeDiCa presents a scalable, tractable framework for causal inference with discrete variables and hidden confounding, using a tensorized Bayesian network representation. It trains once on observational data with exact confounder marginalization, enabling fast interventional queries and competitive performance with lower memory and inference time than neural causal models.
CauProphet enables an LLM agent to form and update a structured Causal Belief, improving decision-making by leveraging causal knowledge that persists over tasks. In finance and real-world tests, it outperforms baselines and enhances predictive accuracy by learning causal relationships and transferring them to future tasks.
Causals is an open knowledge graph built from news reports, with more than 369,000 nodes and 1.27 million edges. The paper describes its design and retrieval properties as a resource for studying how language models use causal knowledge.
The paper argues that token-level causation remains after distribution-based measures fail, proposing Petri Nets (via the Individual Token Philosophy) as the natural representation. It presents a two-layer model—Petri Nets for structural, distribution-free causation and conformal prediction for finite-sample causal guarantees—to deliver a complete account of individual causal effects.
Translated title. Original: IntDC: KI-freie Kausalrekonstruktion aus Beobachtungsdaten ohne Experimente
IntDC presents Interventional Embedding Entropy (IEE), a method to infer causal direction from purely observational data by embedding time-series into a delay space and discounting spurious connections from common drivers; it shows robustness across stress tests and real-world biological data like neural and epidemiological dynamics.
A new foundation model, PICLE-PFN, enables dynamic policy-aware longitudinal causal effect estimation by in-context demonstrations and synthetic pretraining, predicting CAPOs in one pass with reduced deployment cost and improved uncertainty calibration compared with traditional estimators.
The paper analyzes how predictive models retain errors under interventions, showing that observational advantages over direct causes depend on input dependence and that gradient updates can distort mechanisms. It argues that interventional error has a positive lower bound within a given causal hypothesis space, and improvement requires revising causal assumptions and facilitating autonomous hypothesis testing.
The document analyzes nine simulated scenarios to compare causal ML methods versus simple models. It shows that in simple settings, basic models with proper confounder adjustment outperform complex, flexible methods, while mis-specification and confounding drive bias; correct estimation often relies on including relevant covariates, with nonlinear confounding demanding appropriate modeling choices.