CoT Average Causal Effect (CACE)

• CoT Average Causal Effect (CACE) is a metric that quantifies causal effects of interventions on parent reasoning steps in chain-of-thought outputs using a structural causal model framework.
• It integrates answer-level and logical content shifts through a convex weighting parameter to assess and repair non-causal reasoning in large language models.
• Empirical studies show that higher CACE correlates with improved accuracy on math reasoning tasks and helps identify and correct flawed reasoning sequences.

The CoT Average Causal Effect (CACE) formalizes the causal relationship between individual reasoning steps in Chain-of-Thought (CoT) outputs from LLMs. Defined within the structural causal model (SCM) framework, CoT CACE quantifies the mean causal effect of an intervention on the parent steps in the CoT, measuring the resulting changes in both the downstream logical inference and the final answer. This SCM-grounded metric is inspired by classical causal effect estimands yet adapted to the context of machine reasoning, enabling rigorous assessment and “causalization” (i.e., repair) of model-generated reasoning sequences (Fu et al., 25 Feb 2025).

1. Formal SCM Framework and CoT Causal Graph

CoT Average Causal Effect is grounded in explicit SCM notation, consistent with the causal modeling traditions of Pearl (2009). Consider:

• Q: question input (exogenous)
• IS: system-level instruction or prompt (exogenous)
• C = [c₁, ..., cₙ]: intermediate reasoning steps (endogenous)
• A = [a₁, ..., aₖ]: answer tokens

The SCM imposes structural equations of the type

$$c_i \leftarrow f_i(Q, IS, c^{pa}_i), \qquad a_j \leftarrow g_j(Q, IS, C, C^{pa}, a_1,...,a_{j-1}),$$

where each $c_i$ depends only on designated parent steps $c^{pa}_i$, and possibly $Q$ and $IS$. This causal graph treats each reasoning step as a downstream node, with edges connecting parent steps to their children (Fu et al., 25 Feb 2025).

2. Definition and Computation of CoT CACE

CoT Average Causal Effect is explicitly defined using interventions (“do-operations”) on parent steps:

• The effect on the logical content:

$$\gamma_l(c^{pa}_i;Q,IS) = E[c_i \mid do(c^{pa}_i), Q, IS] - E[c_i \mid c^{pa}_i, Q, IS]$$

• The effect on the answer:

$$\gamma_a(c^{pa}_i;Q,IS) = E[a_i \mid do(c^{pa}_i), Q, IS] - E[a_i \mid c^{pa}_i, Q, IS]$$

• The CoT Average Causal Effect:

$$\gamma_{CoT}(c^{pa}_i;Q,IS) = \alpha \cdot \gamma_a + (1-\alpha) \cdot \gamma_l, \quad \alpha \in [0,1]$$

The weighting parameter $\alpha$ allows joint consideration of answer-level and step-level shifts. For the first step ($i=1$), a specialized effect $\gamma_{fs}$ is defined on the opening answer token (Fu et al., 25 Feb 2025).

3. Identification Assumptions

Interpretation of CoT CACE as a causal (as opposed to merely associational) effect necessitates three critical assumptions:

• SUTVA (consistency): No interference; intervention on $c^{pa}_i$ affects only downstream nodes as dictated by SCM.
• Unconfoundedness: Given $(Q,IS)$, no unmeasured variable simultaneously affects the chosen $c^{pa}_i$ and the outcomes $(c_i, a_i)$.
• Overlap: Every possible $c^{pa}_i$ configuration occurs with positive probability, ensuring well-posed do-interventions.

Under these assumptions, $\gamma_l$ and $\gamma_a$ are empirically identifiable from interventional runs or controlled LLM prompting (Fu et al., 25 Feb 2025).

4. Algorithmic Estimation in Practice

Expectation-based CoT CACE cannot be evaluated analytically for high-dimensional, language-based $c^{pa}_i$. The practical estimation pipeline leverages LLMs both to regenerate candidate steps under interventions and to score impact:

• For each CoT instance and step $i$:
  • Score the shift in logical content ($\gamma_l$) and answer-level shift ($\gamma_a$) by LLM-based evaluation of outputs under factual and interventional prompts.
  • Aggregate into $\gamma_{CoT}$ via the defined convex sum.
• Apply a threshold $\sigma$ (“causal confidence”) to decide if the step is adequately justified; if not, invoke “causalization,” prompting the LLM to produce new candidate $c_i$ with higher causal support.
• Iteratively refine deficient steps until all steps exceed the CACE threshold.

The procedure is formalized in pseudocode (Algorithm 1, (Fu et al., 25 Feb 2025)).

5. Empirical and Theoretical Properties

Large-scale empirical analysis demonstrates that higher average CACE across CoT steps correlates with better Exact Match (EM) accuracy on math reasoning datasets (GSM8K, Math, OlympiadBench, Omni-MATH). The metric can localize non-causal or vacuously justified steps; for instance, in arithmetic errors, both $\gamma_l$ and $\gamma_a$ can be near zero until causalization repairs the reasoning. Additional metrics, such as heterogeneous effect (HE) and factual average treatment effect (ATE), are also employed for comprehensive causal analysis of the stepwise reasoning process (Fu et al., 25 Feb 2025).

6. Relation to Classical Causal Effect Estimands

While inspired by the potential outcomes and principal stratification literature (e.g., CACE in randomized experiments), the CoT CACE operates on language and logic objects within a model-generated reasoning trajectory rather than treatment assignment or compliance status in human subjects. Identifiability, interpretable interventions, and SUTVA analogs are preserved via careful design of LLM prompts and structural equations. This suggests a bridge between algorithmic interpretability and formal causal inference (Fu et al., 25 Feb 2025).

7. Applications, Limitations, and Future Directions

Applications of CoT CACE include:

• Quantitative evaluation of the sensitivity and necessity of each reasoning step.
• Automatic repair and improvement of LLM reasoning by enforcing causal validity.
• Diagnostic tools for debugging and surfacing vacuous logic in complex model outputs.

Limitations stem from the reliance on accurate LLM-generated judgments for both interventional responses and causalization; hints of unmeasured confounding or insufficient support for certain $c^{pa}_i$ configurations may limit identifiability. Promising directions include refinement of prompt-based intervention strategies, formal guarantees under distribution shift, and integration with other causal metrics for multi-stage reasoning assessment (Fu et al., 25 Feb 2025).