Can LLMs Identify Event Causality More Accurately through Debate? A Systematic Assessment of LLMs’ Factuality and Reasoning
Bibliographic record
Abstract
Event Causality Identification (ECI) aims to determine both the existence and direction of a causal relation between two events. Given the remarkable capabilities of large language models (LLMs) in various natural language processing (NLP) tasks, several studies have explored their performance in the ECI task. However, applying LLMs to the ECI task carries risks, as their generations often contain factual and reasoning errors. Due to the powerful capabilities of debate, some studies have leveraged it to mitigate these issues. However, debate can sometimes lead LLMs to generate even more errors, suggesting that debate may not effectively enhance the quality of LLMs’ generations. Therefore, we propose a novel debate framework for LLMs in the ECI task to validate whether LLMs can identify event causality more accurately through debate. As LLMs’ generations often mix correct information with factual and reasoning errors, binary judgments on their generations may not fully reveal their potential in real-world use cases. And factual errors reflect flaws in the factuality of LLMs’ generations, while reasoning errors reveal weaknesses in their reasoning abilities. Therefore, we propose a Fine-Grained Factuality Score (FGFS) to better evaluate the factuality of LLMs’ generations and a Fine-Grained Reasoning Score (FGRS) to assess their reasoning abilities more effectively. Additionally, we redefine factuality and correct reasoning for LLMs’ generations at a fine-grained level to ensure more precise evaluation results for our metrics. Finally, considering the lack of automated evaluation for FGFS and FGRS and the time-consuming and costly nature of human evaluation, we propose automated methods for both proposed metrics and validate their superior performance by human expertise.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.016 | 0.181 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.002 |
| Bibliometrics | 0.006 | 0.002 |
| Science and technology studies | 0.001 | 0.002 |
| Scholarly communication | 0.006 | 0.020 |
| Open science | 0.002 | 0.006 |
| Research integrity | 0.003 | 0.003 |
| Insufficient payload (model declined to judge) | 0.009 | 0.002 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".