New Research Explores Counterfactual Reasoning and Bias in Large Language Models

New Research Explores Counterfactual Reasoning and Bias in Large Language Models

Several research papers published this week examine how large language models handle counterfactual reasoning across different domains.

According to arxiv.org, researchers investigated GPT-3.5’s capability for counterfactual analysis in online lending contexts. The study found that prompt engineering significantly enhanced GPT’s predictions, with R-squared values increasing from 1.97% to 2.84%, approaching the 3.48% achieved by gradient-boosted regression. The paper states that “GPT exhibits logical coherence and causal reasoning in its responses” when generating counterfactual returns on investment under alternative interest rates.

Another paper introduced InvestLogicBench, a benchmark containing 201,247 documented decisions from 151 real-world investors, according to arxiv.org. Testing four leading LLMs, researchers found that “logical plausibility remains near 4/5 while event grounding is only 0.8—2.8/5,” exposing what the authors describe as “polished but weakly grounded reasoning that outcome-only evaluation hides.”

A third study presented Poli-Bias, a counterfactual framework for measuring political bias in LLMs by comparing responses to paired prompts where country identities are systematically swapped, according to arxiv.org. Testing 13 contemporary LLMs, researchers found that “country identities and user affiliations can systematically affect how equivalent actions are described, evaluated, and defended under international law.”

Finally, arxiv.org reports on research comparing three methods for grounding a frozen Qwen2.5-32B-Instruct model in a wastewater simulator for causal question answering, achieving accuracy rates ranging from 75.8% to 99.5% on a 198-question benchmark.