Causal Resilience Benchmarks: Dataset Construction for Systematic Evaluation of LLM Reasoning Robustness

TL;DR: Let’s make a new benchmark where we systematically remove, replace, or perturb key reasoning steps and see how well models can “bounce back” (or not!) by resampling. This would let us quantify reasoning resilience across models.

Research Question: How robust are LLMs to targeted perturbations of their reasoning chains, and can systematic resampling-based benchmarks quantify and compare model resilience across tasks and architectures?

Hypothesis: Some LLMs and prompt strategies exhibit higher resilience—maintaining answer accuracy and coherent reasoning even after key steps are removed or edited—than others; this resilience can be benchmarked and used as a model selection criterion.

Experiment Plan: Construct a benchmark dataset spanning math, logic, ethics, and causal inference, with annotated “critical” reasoning steps per question. For each sample, systematically perturb (remove/replace/scramble) key steps and resample downstream CoTs. Measure resilience (e.g., probability of answer recovery, reasoning coherence, causal impact) across models and prompt types. Release the dataset and metrics for the community. Expected outcome: Reveals model- and prompt-specific differences in resilience, identifies brittle areas, and guides robustness improvements.

References:

• 1. Debjit Paul, West, R., Bosselut, A., & Faltings, B. (2024). Making Reasoning Matter: Measuring and Improving Faithfulness of Chain-of-Thought Reasoning. Conference on Empirical Methods in Natural Language Processing. 2. Lei Wang & Shen, Y. (2024). Evaluating Causal Reasoning Capabilities of Large Language Models: A Systematic Analysis Across Three Scenarios. Electronics. 3. Tinghui Zhu, Zhang, K., Xie, J., & Su, Y. (2024). Deductive Beam Search: Decoding Deducible Rationale for Chain-of-Thought Reasoning. arXiv.org.