AAAI2026 LLM Safety Causal Discovery LLM-guided Fairness Auditing Active Learning Dynamic Scoring Bias Paths

Uncovering Bias Paths with LLM-guided Causal Discovery: An Active Learning and Dynamic Scoring Approach¶

Conference: AAAI2026 arXiv: 2506.12227
Authors: Khadija Zanna, Akane Sano (Rice University) Code: Not released
Area: AI Safety Keywords: Causal Discovery, LLM-guided, Fairness Auditing, Active Learning, Dynamic Scoring, Bias Paths

TL;DR¶

This paper proposes a hybrid causal discovery framework that integrates LLM semantic priors with statistical signals. Through an active learning strategy and a dynamic scoring mechanism, the framework prioritizes querying the most informative variable pairs, effectively recovering fairness-critical causal paths (e.g., sex→education→income) under noise and confounding conditions, substantially outperforming classical CD methods and naïve LLM-based approaches.

Background & Motivation¶

A Causal Perspective on Fairness Auditing¶

Bias in machine learning models deployed in high-stakes domains such as hiring, lending, education, and healthcare has attracted growing concern. Such bias often arises not through the direct effect of sensitive attributes (e.g., gender, race) on outcomes, but through indirect structural paths—where sensitive attributes influence final decisions via proxy variables or confounded relationships. Conventional fairness auditing relies on statistical disparity metrics (e.g., demographic parity), which cannot reveal the causal propagation mechanisms underlying bias, potentially rendering interventions ineffective or misleading. Causal Discovery (CD) offers tools to identify such paths, enabling the distinction between genuine causal effects and spurious associations introduced by confounding or measurement artifacts.

Limitations of Prior Work¶

Classical CD methods (e.g., PC algorithm, GES) rely on conditional independence tests and strong assumptions (e.g., faithfulness), and tend to fail under noisy data, latent confounding, or incomplete metadata. Optimization-based methods (NOTEARS, DAGMA) are computationally expensive and highly sensitive to hyperparameters. Recent work has explored LLMs as auxiliary tools for CD, leveraging semantic knowledge to infer causal directions from variable metadata. However, two critical issues persist: (1) naïve use of LLMs may over- or under-attribute causality, particularly when sensitive attributes are involved; (2) BFS-based LLM query strategies (Jiralerspong et al. 2024) treat all variable pairs uniformly, wasting query budgets on non-informative relationships, while early errors propagate through the discovery process.

Core Motivation¶

In fairness-sensitive settings, real-world datasets lack known ground-truth causal structures, making systematic evaluation of CD methods extremely challenging. This paper is motivated by two goals: (1) to design an adaptive query strategy that efficiently recovers fairness-relevant causal paths by balancing semantic and statistical signals; and (2) to construct a semi-synthetic benchmark with known bias paths to support rigorous evaluation.

Core Problem¶

How can causal paths from sensitive attributes (e.g., sex, race) to outcomes (e.g., income) be recovered efficiently and accurately under noise, confounding, and data corruption? Classical CD methods exhibit insufficient precision in this setting, while naïve LLM approaches tend to over-predict. The core challenge is to design a hybrid framework that achieves both accuracy and query efficiency.

Method¶

Overall Architecture¶

The proposed framework augments the BFS-based CD pipeline with two key components: active learning and dynamic scoring. The overall procedure is: (1) identify independent root variables; (2) compute dynamic scores for each unqueried variable pair; (3) prioritize querying the highest-scoring pair via the LLM; (4) if the LLM returns "Yes" and adding the edge does not create a cycle, insert the directed edge into the causal graph; (5) repeat until the maximum number of iterations is reached or all scores fall below a threshold.

Key Design 1: Dynamic Scoring Mechanism¶

Each unqueried variable pair \((x, y)\) receives a composite score integrating statistical signals, model confidence, and query history:

\[S(x,y) = w_{\text{stat}} \cdot \text{StatScore}(x,y) + w_{\text{conf}} \cdot \text{LLMConf}(x,y) + w_{\text{hist}} \cdot \text{HistScore}(x,y)\]

The individual components are defined as follows:

Statistical Score: Combines mutual information (MI) and partial correlation (PCorr) to capture both linear and nonlinear dependencies:

\[\text{StatScore}(x,y) = \frac{\text{MI}(x,y) + \text{PCorr}(x,y)}{2}\]

PCorr is computed as a conditional partial correlation conditioned on the currently discovered parent set, following the iterative conditioning procedure of the PC algorithm.

LLM Confidence Score: A sigmoid transformation of token-level log-probabilities, reflecting the LLM's certainty regarding the causal relationship of the variable pair:

\[\text{LLMConf}(x,y) = \frac{1}{1 + e^{-\text{confidence}}}\]

Query History Score: Penalizes repeated queries and encourages broader exploration:

\[\text{HistScore}(x,y) = \frac{1.5}{1 + \text{query\_count}(x,y)}\]

The weights \(w_{\text{stat}}, w_{\text{conf}}, w_{\text{hist}}\) are tuned via Bayesian optimization (Gaussian Process surrogate + gp_minimize).

Key Design 2: Active Learning Query Strategy¶

At each step, the highest-scoring variable pair is selected for querying:

\[(x^*, y^*) = \arg\max_{(x,y) \in \text{Unqueried}} S(x,y)\]

The LLM operates in a multi-turn dialogue mode, maintaining context awareness over prior reasoning. If the response is "Yes" and no cycle is introduced, a directed edge is added, constructing the adjacency matrix:

\[A(i,j) = \begin{cases} 1 & \text{if } X_i \rightarrow X_j \text{ is predicted} \\ 0 & \text{otherwise} \end{cases}\]

Key Design 3: Fairness Path Analysis¶

After learning the causal graph, all directed paths from each sensitive attribute \(S\) to the outcome \(Y\) are enumerated and classified as:

Direct paths: \(S \rightarrow Y\)
Indirect paths: \(S \rightarrow \cdots \rightarrow Y\) (mediated through intermediate variables)
Spurious paths: involving \(S\) but not reaching \(Y\)

Causal effect decomposition quantifies bias as \(TE = DE + IE\), and a normalized bias coefficient is computed:

\[C_{\text{bias}} = \frac{TE}{\text{Var}(Y)}\]

Key Design 4: Semi-Synthetic Benchmark Construction¶

A semi-synthetic causal graph with 15 variables is constructed based on the UCI Adult dataset. Direct edges and indirect paths from race and sex to income (e.g., sex→education→income) are injected, along with noise, data corruption, and a latent confounding variable \(U \sim \mathcal{N}(0,1)\), providing a controlled ground-truth evaluation environment.

Key Experimental Results¶

Main Results: Structural Recovery Performance Across Three Datasets¶

Method	Dataset	Accuracy ↑	F1 ↑	Precision	Recall	NHD ↓	Predicted Edges
PC	Adult (15n, 28e)	0.239	0.382	0.352	0.420	0.193	33
GES	Adult	0.296	0.473	0.368	0.580	0.203	44
NOTEARS	Adult	0.021	0.039	0.035	0.045	0.260	27
DAGMA	Adult	0.099	0.180	0.141	0.250	0.283	50
LLM Pairwise	Adult	0.307	0.470	0.331	0.813	0.253	69
LLM BFS	Adult	0.299	0.456	0.332	0.750	0.305	64
Proposed	Adult	0.413	0.585	0.792	0.464	0.109	17
PC	Child (20n, 25e)	0.146	0.255	0.273	0.239	0.097	22
NOTEARS	Child	0.216	0.356	0.403	0.319	0.080	20
Proposed	Child	0.364	0.533	0.601	0.479	0.082	20
PC	Neuropathic (221n)	0.041	0.078	0.092	0.068	0.025	563
LLM BFS	Neuropathic	0.000	0.000	0.000	0.000	0.903	43
Proposed	Neuropathic	0.073	0.136	0.690	0.075	0.109	84

Fairness Path Recovery¶

Ground-truth: 2 direct paths (sex→income, race→income), 25 indirect paths, TE=4.89, \(C_{\text{bias}}=28.46\).

Method	Direct Paths	Indirect Paths	Spurious Paths	\(C_{\text{bias}}\)
PC	0	Few	Many	Low
GES	0	Moderate	Many	Low
LLM Pairwise	3 (includes spurious age path)	Excessive	Many	Inflated
Proposed	2 (sex and race only)	Partially recovered	Very few	Near ground-truth

The proposed method is the only approach that correctly recovers both true direct paths (sex→income, race→income) while excluding the spurious age→income path.

Hyperparameter Sensitivity Analysis¶

A Random Forest regression analysis of Bayesian optimization trials reveals that the maximum number of iterations has the largest impact on F1, followed by the scoring threshold and LLM temperature. Anti-correlations among scoring weights indicate competitive trade-offs—semantic guidance dominates in small graphs, while MI/PCorr statistical signals take precedence in larger graphs (e.g., Neuropathic).

Highlights & Insights¶

Adaptive Query Strategy: The dynamic scoring mechanism adaptively balances semantic versus statistical signals according to graph structure and data quality—relying on LLM semantic judgments during early exploration and shifting toward statistical signal verification in later stages, reflecting a progressive transition from semantic exploration to empirical refinement.
Precise Fairness Path Recovery: The proposed method is the only approach to correctly identify the sex→income and race→income direct paths while avoiding the spurious age→income path, demonstrating strong fairness diagnostic capability.
High Precision with Low False Positives: On the Adult dataset, Precision reaches 0.792 (far exceeding all baselines), with only 17 edges predicted (ground-truth: 28), reflecting a conservative yet reliable strategy.
Cross-Scale Generalization: The method maintains optimal F1 across graphs ranging from 15 to 221 nodes, remaining effective on the large-scale Neuropathic network where LLM BFS fails entirely.
Semi-Synthetic Benchmark: A reproducible benchmark with controlled ground-truth is provided for fairness-sensitive CD evaluation.

Limitations & Future Work¶

Low Recall: The conservative strategy yields a Recall of only 0.464 on the Adult dataset, potentially missing certain fairness-relevant indirect paths.
LLM Dependency and Reproducibility: GPT-4-based inference introduces stochasticity and version sensitivity; inconsistencies with results reported by Jiralerspong et al. highlight the absence of standardized evaluation protocols for LLM-based CD.
Semi-Synthetic Data Limitations: The causal graph treats income as a terminal node, ignoring downstream feedback effects and temporal dynamics.
Computational Cost: Bayesian optimization combined with LLM queries incurs high overhead on large graphs; the Neuropathic dataset is significantly affected by token-limit constraints.
Risk of LLM Social Bias: Social biases encoded in LLMs may lead to inferences that reflect stereotypes rather than true causal mechanisms.
Observational CD Only: Interventional or experimental causal discovery methods (e.g., do-calculus, instrumental variables) are not considered.

PC / GES: Classical constraint- and score-based methods achieve F1 of only 0.38–0.47 under noise and fail to recover direct fairness paths.
NOTEARS / DAGMA: Continuous optimization methods perform worst on fairness-relevant structures (F1 < 0.18) and are highly sensitive to hyperparameters.
LLM Pairwise (Kıcıman et al. 2023): Achieves high Recall (0.813) but severely over-predicts (69 edges vs. 28 ground-truth), introducing spurious fairness paths that inflate \(C_{\text{bias}}\).
LLM BFS (Jiralerspong et al. 2024): Fixed-order querying is sensitive to early errors and fails completely on large-scale graphs (Neuropathic F1=0).
Takayama et al. 2024: LLM-augmented statistical method, but with a single prior injection step, lacking the dynamic adaptivity of the proposed approach.
The core contribution lies in the dynamic scoring + active learning query prioritization mechanism, which substantially reduces NHD while maintaining high precision.

Broader Implications:

Path-Level Interpretability over Aggregate Statistics: Summary metrics such as \(C_{\text{bias}}\) can be inflated by spurious paths; path-level analysis provides more reliable fairness diagnoses, with important implications for real-world auditing.
Complementarity of Semantic Priors and Statistical Signals: LLM semantic guidance is especially valuable in the early stages of graph discovery when statistical signals are weak; as evidence accumulates, statistical signals progressively take over. This adaptive balancing strategy generalizes to other LLM-assisted scientific discovery tasks.
Fairness Auditing as a Tool: The framework can serve as an algorithmic auditing tool for high-risk domains such as hiring and lending, enabling non-technical stakeholders to trace bias propagation paths.

Rating¶

Novelty: ⭐⭐⭐⭐ — Introducing active learning and dynamic scoring into LLM-guided CD is a meaningful contribution, though the overall framework represents an incremental improvement over BFS-based methods.
Experimental Thoroughness: ⭐⭐⭐⭐ — Three datasets, multiple baselines, hyperparameter sensitivity analysis, and fairness path evaluation provide comprehensive coverage, though the low-Recall issue is not deeply analyzed.
Writing Quality: ⭐⭐⭐⭐ — Well-structured with clear motivation and rigorous formalism; some discussions are somewhat verbose.
Value: ⭐⭐⭐⭐ — The intersection of fairness, causal discovery, and LLMs addresses a practically significant problem, and the semi-synthetic benchmark contributes to the community; however, LLM reproducibility concerns limit broader adoption.