New publication by Prof. Edward Szczerbicki on causal reasoning in large language models
An international team of scientists, including Prof. Edward Szczerbicki from the Faculty of Management and Economics at Gdańsk Tech, has developed a neurosymbolic framework SACR (Structure-Aware Causal Reasoning) aimed at enhancing the ability of large language models to perform counterfactual reasoning. The research results were published in the journal “International Journal of Approximate Reasoning.”
An international team of scientists, including Prof. Edward Szczerbicki from the Faculty of Management and Economics at Gdańsk Tech, has developed a neurosymbolic framework SACR (Structure-Aware Causal Reasoning) aimed at enhancing the ability of large language models to perform counterfactual reasoning. The research results were published in the journal “International Journal of Approximate Reasoning.”
The authors of the article “Structure-Aware Causal Reasoning: A Neurosymbolic Framework for Counterfactual Inference in Large Language Models” are Maiqi Wang and Haoxi Zhang from Chengdu University of Information Technology in China, and Prof. Edward Szczerbicki from Gdańsk Tech.
What if?
Counterfactual reasoning involves analyzing alternative scenarios to events that have already occurred – in other words, answering questions like “what if?”. It is one of the important human reasoning abilities, particularly valuable in analyzing causal relationships and decision-making.
Large language models excel at natural language processing and recognizing semantic relationships, but counterfactual reasoning remains a challenge for them. The problem arises especially in complex structures with many interrelated causal pathways.
– The critical mismatch in current approaches is that standard models process causal graphs as flattened textual context, failing to distinguish proper causal mechanisms from spurious correlations in the presence of colliders and multi-path dependencies. This leads to frequent hallucinations when answering “what if?” questions – explains Prof. Edward Szczerbicki.
SACR organizes causal dependencies
The team’s response to this problem is SACR (Structure-Aware Causal Reasoning) – a neurosymbolic framework combining the capabilities of large language models with formal methods of analyzing causal structures.
Unlike approaches primarily based on crafting suitable prompts, SACR uses, among others, d-separation theory and Markov boundary extraction. This allows isolating from the graph those dependencies that are truly relevant to a given problem and limits the impact of accidental or apparent correlations on the reasoning process.
– We propose and rigorously verify the neurosymbolic SACR framework, which integrates d-separation logic and an extension of Markov boundaries in interpreting causal graphs modeling LLM-based reasoning. SACR significantly outperforms state-of-the-art models in complex logical scenarios, achieving higher accuracy in multi-path settings and thus also in “what if” scenarios – says Prof. Szczerbicki.
In conducted experiments, SACR achieved results in complex multi-path scenarios that were 22% better than GPT-4, while maintaining competitive accuracy in benchmarks related to counterfactual causal inference.
The next step: greater model realism
The authors note that the developed solution still has limitations. Some of its elements rely on heuristic strategies, and the graphs used in the studies do not always fully reflect the complexity of real-world systems, where dependencies may be temporal, cyclic, and multi-level in nature.
– Our future research will address, among other things, path validity models based on information-theoretic metrics and collider interpretations based on classifiers, which should improve the system’s precision and adaptability – announces Prof. Szczerbicki. – We also plan to increase graph realism, broaden the domain scope, and introduce more diverse counterfactual queries to more accurately evaluate causal reasoning capabilities.
The research is aligned with the development of methods intended to allow artificial intelligence systems not only to recognize data patterns but also to better understand causal relationships and use them in reasoning processes.
International Journal of Approximate Reasoning is a peer-reviewed scientific journal dedicated, among others, to artificial intelligence, reasoning methods, and issues related to uncertainty and data imprecision. Publications in this journal are highly scored in the ministerial list of scientific journals.
The article “Structure-Aware Causal Reasoning: A Neurosymbolic Framework for Counterfactual Inference in Large Language Models” is available online.