Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Dissecting the Ledger: Locating and Suppressing "Liar Circuits" in Financial Large Language Models

Soham Mirajkar

TL;DR

Financial LLMs suffer arithmetic hallucinations during numerical reasoning, addressed here with a mechanistic, causal-tracing framework. The authors identify a dual-stage circuit comprising distributed middle-layer computation and a late-layer aggregation bottleneck, pinpointing Layer 46 as the critical gatekeeper. Ablation of Layer 46 dramatically reduces hallucination confidence by $81.8\%$, and a linear probe trained on Layer 46 activations generalizes to unseen financial topics with $98\%$ accuracy, suggesting a universal geometry of arithmetic deception. These findings enable lightweight, topic-agnostic safety monitors that operate on internal state dynamics at aggregation layers.

Abstract

Large Language Models (LLMs) are increasingly deployed in high-stakes financial domains, yet they suffer from specific, reproducible hallucinations when performing arithmetic operations. Current mitigation strategies often treat the model as a black box. In this work, we propose a mechanistic approach to intrinsic hallucination detection. By applying Causal Tracing to the GPT-2 XL architecture on the ConvFinQA benchmark, we identify a dual-stage mechanism for arithmetic reasoning: a distributed computational scratchpad in middle layers (L12-L30) and a decisive aggregation circuit in late layers (specifically Layer 46). We verify this mechanism via an ablation study, demonstrating that suppressing Layer 46 reduces the model's confidence in hallucinatory outputs by 81.8%. Furthermore, we demonstrate that a linear probe trained on this layer generalizes to unseen financial topics with 98% accuracy, suggesting a universal geometry of arithmetic deception.

Dissecting the Ledger: Locating and Suppressing "Liar Circuits" in Financial Large Language Models

TL;DR

Financial LLMs suffer arithmetic hallucinations during numerical reasoning, addressed here with a mechanistic, causal-tracing framework. The authors identify a dual-stage circuit comprising distributed middle-layer computation and a late-layer aggregation bottleneck, pinpointing Layer 46 as the critical gatekeeper. Ablation of Layer 46 dramatically reduces hallucination confidence by , and a linear probe trained on Layer 46 activations generalizes to unseen financial topics with accuracy, suggesting a universal geometry of arithmetic deception. These findings enable lightweight, topic-agnostic safety monitors that operate on internal state dynamics at aggregation layers.

Abstract

Large Language Models (LLMs) are increasingly deployed in high-stakes financial domains, yet they suffer from specific, reproducible hallucinations when performing arithmetic operations. Current mitigation strategies often treat the model as a black box. In this work, we propose a mechanistic approach to intrinsic hallucination detection. By applying Causal Tracing to the GPT-2 XL architecture on the ConvFinQA benchmark, we identify a dual-stage mechanism for arithmetic reasoning: a distributed computational scratchpad in middle layers (L12-L30) and a decisive aggregation circuit in late layers (specifically Layer 46). We verify this mechanism via an ablation study, demonstrating that suppressing Layer 46 reduces the model's confidence in hallucinatory outputs by 81.8%. Furthermore, we demonstrate that a linear probe trained on this layer generalizes to unseen financial topics with 98% accuracy, suggesting a universal geometry of arithmetic deception.

Paper Structure

This paper contains 11 sections, 1 equation, 3 figures.

Table of Contents

  1. Introduction
  2. Methodology
  3. Task Definition
  4. Causal Tracing Setup
  5. Implementation
  6. Core Findings
  7. The Dual-Stage Mechanism
  8. Validation via Causal Suppression
  9. Robustness Verification
  10. Application: A Universal Hallucination Detector
  11. Discussion and Conclusion

Figures (3)

  • Figure 1: Causal Tracing Heatmap. The x-axis represents tokens in the financial prompt. The y-axis represents Transformer layers (0-48). Note the distributed impact in the middle layers (L12-L30) at the operand tokens and the massive peak at Layer 46 at the final token.
  • Figure 2: Robustness Analysis. Averaged Causal Impact across 5 diverse financial prompts. The high-impact region at Layer 46 remains consistent, confirming it as a structural bottleneck for arithmetic output.
  • Figure 3: Geometry of Deception. PCA projection of Layer 46 activations. Even on unseen topics (triangles), the model's internal state separates clearly into "Truth" (Green) and "Hallucination" (Red) clusters along a shared linear direction.