[Paper] Project Ariadne: A Structural Causal Framework for Auditing Faithfulness in LLM Agents

Source: arXiv - 2601.02314v1

Overview

Large Language Model (LLM) agents are increasingly being deployed to make autonomous, high‑stakes decisions—from code generation to medical triage. While “Chain‑of‑Thought” (CoT) prompting gives these agents a human‑readable reasoning trace, we still don’t know whether the trace actually drives the final answer or is just a post‑hoc justification. Project Ariadne introduces a structural‑causal framework that rigorously audits the faithfulness of those reasoning traces, exposing a systematic “faithfulness gap” in today’s state‑of‑the‑art models.

Key Contributions

• Causal Auditing Framework: Leverages Structural Causal Models (SCMs) and do‑calculus to intervene on intermediate reasoning steps, measuring how changes propagate to the final answer.
• Causal Sensitivity (φ) Metric: Quantifies the degree to which the terminal output depends on each reasoning node.
• Violation Density (ρ) & Causal Decoupling: Formal definitions for detecting when an agent’s internal logic is disconnected from its output (ρ up to 0.77 in factual/scientific tasks).
• Ariadne Score Benchmark: A new evaluation suite that scores LLM agents on the alignment between generated CoT and actual decision pathways.
• Empirical Evidence: Demonstrates that leading LLM agents (e.g., GPT‑4, Claude, Llama 2) frequently produce “Reasoning Theater”—identical answers despite contradictory internal logic.

Methodology

1. Model as an SCM – The LLM’s reasoning chain is treated as a directed graph where each node is a textual premise or inference step.
2. Hard Interventions (do‑operations) – The authors systematically flip, negate, or replace premises (e.g., change “All swans are white” to “All swans are black”).
3. Counterfactual Propagation – After each intervention, the model is asked to recompute the final answer without re‑prompting the entire chain, isolating the causal effect of the altered node.
4. Metric Computation –
   • Causal Sensitivity (φ) = |Δoutput| / |Δintervention|, measuring how much the answer changes.
   • Violation Density (ρ) = fraction of nodes where φ ≈ 0 despite contradictory content.
5. Benchmarking – A suite of factual, scientific, and reasoning tasks is used to compute the Ariadne Score for each model.

The approach is deliberately model‑agnostic: it works with any LLM that can accept CoT prompts and return deterministic outputs for a given seed.

Results & Findings

• Faithfulness Gap: Across all tested domains, the agents’ final answers were weakly sensitive to the internal reasoning, indicating that the CoT trace is often a decorative layer rather than a causal driver.
• Causal Decoupling: Flipping a premise that should logically invert the answer often left the answer unchanged, revealing reliance on latent parametric priors instead of the explicit chain.
• Ariadne Score: Provides a single‑number summary (0–1) of faithfulness; current top‑performing models score below 0.4, far from the ideal 1.0.

Practical Implications

• Safety & Compliance: For regulated sectors (finance, healthcare, autonomous systems), relying on CoT explanations alone is insufficient. Auditors can use Project Ariadne to certify that an agent’s reasoning is causally linked to its decisions.
• Debugging LLM Agents: Developers can pinpoint “dead” reasoning nodes (φ ≈ 0) and refactor prompts or fine‑tune models to make those steps influential.
• Prompt Engineering: The framework suggests that prompting strategies that force causal dependence (e.g., “You must base your answer on the following premise”) may improve faithfulness.
• Benchmarking & Competition: The Ariadne Score can become a new leaderboard metric, encouraging the community to build agents that are both accurate and explainable.
• Tooling: Open‑source libraries implementing the do‑calculus interventions could be integrated into existing LLM evaluation pipelines (e.g., 🤗 Eval, OpenAI’s Evals).

Limitations & Future Work

• Scalability: Hard interventions require multiple forward passes per reasoning node, which can be costly for long chains or large models.
• Prompt Sensitivity: The method assumes deterministic outputs; temperature‑based sampling can blur causal signals.
• Domain Coverage: Experiments focus on factual and scientific tasks; extending to creative or open‑ended generation remains open.
• Model‑Specific Optimizations: Some architectures (e.g., retrieval‑augmented models) may need adapted SCM representations.

Future Directions

• Automating intervention selection via reinforcement learning.
• Integrating causal regularization into fine‑tuning to reduce ρ.
• Exploring hybrid metrics that combine φ with traditional similarity‑based explainability scores.

Authors

• Sourena Khanzadeh

Paper Information

• arXiv ID: 2601.02314v1
• Categories: cs.AI
• Published: January 5, 2026
• PDF: Download PDF