Paper: arXiv 2508.18427
Authors: Fabrizio Dimino, Krati Saxena, Bhaskarjit Sarmah, Stefano Pasquali
Abstract
The growing adoption of large language models (LLMs) in finance exposes high-stakes decision-making to subtle, underexamined positional biases. The complexity and opacity of modern model architectures compound this risk. We present the first unified framework and benchmark that not only detects and quantifies positional bias in binary financial decisions but also pinpoints its mechanistic origins within open-source Qwen2.5-instruct models (1.5B-14B). Our empirical analysis covers a novel, finance-authentic dataset revealing that positional bias is pervasive, scale-sensitive, and prone to resurfacing under nuanced prompt designs and investment scenarios, with recency and primacy effects revealing new vulnerabilities in risk-laden contexts. Through transparent mechanistic interpretability, we map how and where bias emerges and propagates within the models to deliver actionable, generalizable insights across prompt types and scales. By bridging domain-specific audit with model interpretability, our work provides a new methodological standard for both rigorous bias diagnosis and practical mitigation, establishing essential guidance for responsible and trustworthy deployment of LLMs in financial systems.
Complexity vs Empirical Score
- Math Complexity: 4.0/10
- Empirical Rigor: 7.5/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper employs sophisticated but applied machine learning concepts (positional bias, mechanistic interpretability) without heavy theoretical derivations or dense mathematical formalisms. It demonstrates high empirical rigor with a novel finance-authentic dataset, cross-scale model testing (1.5B-14B parameters), multiple prompt designs, and a clear path to actionable, implementation-ready insights.
Research Flowchart
flowchart TD A["Research Goal: Trace positional bias in LLM financial decisions"] --> B["Methodology: Mechanistic interpretability<br>(Qwen2.5 1.5B-14B)"] B --> C["Input: Finance-authentic<br>binary decision dataset"] C --> D["Computational Process: Detect, quantify & trace bias"] D --> E["Key Findings:<br>- Pervasive & scale-sensitive bias<br>- Mechanistic origins mapped<br>- Recency/Primacy effects in risk contexts"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes the first unified framework and benchmark to detect, quantify, and mechanistically interpret positional bias in binary financial decisions using Qwen2.5 models.
Source sentence
We present the first unified framework and benchmark that not only detects and quantifies positional bias in binary financial decisions but also pinpoints its mechanistic origins within open-source Qwen2.5-instruct models (1.5B–14B).
Assumes
No explicit assumptions found in the text.
Data
- novel, finance-authentic dataset benchmark unknown
Source sentence
Our empirical analysis covers a novel, finance-authentic dataset revealing that positional bias is pervasive, scale-sensitive, and prone to resurfacing under nuanced prompt designs and investment scenarios, with recency and primacy effects revealing new vulnerabilities in risk-laden contexts.
Costs
Not applicable
Finds 5
- Positional bias is pervasive, scale‑sensitive, and exhibits recency and primacy effects across financial prompts. statistical
Source sentence
Our empirical analysis covers a novel, finance-authentic dataset revealing that positional bias is pervasive, scale-sensitive, and prone to resurfacing under nuanced prompt designs and investment scenarios, with recency and primacy effects revealing new vulnerabilities in risk-laden contexts.
- Increasing model size from 1.5B to 14B reduces bias overall, but the Risk category remains stubbornly biased even at large scale. statistical
Source sentence
stubbornly biased even at a large scale, suggesting domain-specific factors like volatility and exposure anchor the model to positional cues.
- Prompt ordering independently influences bias magnitude; simple reordering causes significant shifts, especially for Innovation prompts. statistical
Source sentence
Prompt ordering independently influences bias magnitude, with simple reordering causing significant shifts, especially in Innovation, emphasizing the need for deterministic prompt templates or randomized prompting with correction.
- Prompt framing affects bias: conservative strong wording amplifies primacy bias, while moderate phrasing reduces bias. statistical
Source sentence
Conservative, strong wording amplifies positional heuristics, likely reinforcing primacy bias, while Moderate phrasing promotes balanced evaluation and reduces bias.
- Mechanistic analysis shows bias consistently originates in a stable set of mid‑to‑late transformer layers and a small recurrent group of attention heads acting as core bias engines. other
Source sentence
positional bias consistently centers on a stable set of mid-to-late transformer layers and a small, recurrent group of attention heads that act as core “bias engines.”
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 28 related papers). Coverage is partial, so this is neither replication nor refutation.
5 papers in the archive : Eliciting ESG Preferences for Reinforcement Learning-Based …
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (63)
- Large language models in finance: Opportunities and challenges (2023)
- Predictably biased: The influence of position bias on language model predictions (2020)
- Large language models exhibit systematic position bias in multiple-choice questions (2023)
- Calibrating factual knowledge in pretrained language models (2021)
- Fairness in financial AI: Detecting and mitigating algorithmic bias (2023)
- Judgment in managerial decision making (2013)
- Global Sustainable Investment Review 2020 (2021)
- Social Group Bias in AI Finance (2025) — link
and 55 more
- Grammar prompting for domain-specific language generation with large language models (2023)
- Grammar-constrained decoding for structured NLP tasks without finetuning (2023) — link
- Simulating the Survey of Professional Forecasters (2024)
- Large language models as simulated economic agents: What can we learn from homo silicus? (2023)
- AI and Operational Losses: Evidence from US Bank Holding Companies (2024)
- Analysing the generalisation and reliability of steering vectors (2024)
- Fragile Preferences: A Deep Dive Into Order Effects in Large Language Models (2025) — link
- Humans or llms as the judge? a study on judgement biases, 2024 (2024)
- Are LLMs Rational Investors? A Study on Detecting and Reducing the Financial Bias in LLMs (2024) — link
- Evaluating Company-specific Biases in Financial Sentiment Analysis using Large Language Models — → 2411.00420
- Assessing look-ahead bias in stock return predictions generated by gpt sentiment analysis — → 2309.17322
- Individual comparisons by ranking methods (1945)
- Portfolio Selection (1952)
- Prospect Theory: An Analysis of Decision under Risk (1979)
- Are large language models strategic decision makers? a study of performance and bias in two-player non-zero-sum games (2024) — link
- Mechanistic interpretability of large language models with applications to the financial services industry (2024)
- Eliminating position bias of language models: A mechanistic approach (2024) — link
- Serial position effects of large language models (2024) — link
- Unveiling selection biases: Exploring order and token sensitivity in large language models (2024) — link
- Fairer preferences elicit improved human-aligned large language model judgments (2024) — link
- Can many-shot in-context learning help long-context llm judges? see more, judge better! (2024)
- Attention Speaks Volumes: Localizing and Mitigating Bias in Language Models (2024) — link
- Small Language Models are the Future of Agentic AI (2025) — link
- On positional bias of faithfulness for long-form summarization (2024) — link
- How susceptible are llms to influence in prompts? (2024) — link
- Decision-making behavior evaluation framework for llms under uncertain context (2024)
- Prompt position really matters in few-shot and zero-shot NLU tasks (2023) — link
- Estimates of Location Based on Rank Tests (1963) — link
- Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing (1995) — link
- Bootstrap Methods and Their Application (1997) — link
- Systematic evaluation of llm-as-a-judge in llm alignment tasks: Explainable metrics and diverse prompt templates (2024) — link
- Dissecting Disposition Effect of Large Language Models in Financial Decisions (2024)
- How robust are LLMs to in-context majority label bias? (2023) — link
- Social bias evaluation for large language models requires prompt variations (2024) — link
- Judging the judges: A systematic investigation of position bias in pairwise comparative assessments by llms (2024)
- Mitigate position bias in large language models via scaling a single dimension (2024) — link
- Position engineering: Boosting large language models through positional information manipulation (2024) — link
- Do RAG Systems Suffer From Positional Bias? (2025) — link
- Eliminating Position Bias of Language Models: A Mechanistic Approach (2025) — link
- Position of Uncertainty: A Cross-Linguistic Study of Positional Bias in LLMs (2025) — link
- Order Matters: Investigate the Position Bias in Multi-constraint Instruction Following (2025) — link
- Wait, that’s not an option: LLMs Robustness with Incorrect Multiple-Choice Options (2024) — link
- Qwen2.5 Technical Report (2024) — link
- Self-Debiasing Models by Conditioning on Their Own Biases (2023)
- Interpretability in the Wild: A Circuit for Indirect Object Identification in GPT-2 Small (2022)
- Reducing Bias in Language Models via Domain Adapted LoRA Fine-Tuning (2023)
- Mitigating Positional Bias in Multiple-Choice QA with Permutation-Based Calibration (2022)
- LangFair: A Python Package for Assessing Bias and Fairness in Large Language Model Use Cases (2025) — link
- Interpretable LLMs for Credit Risk: A Systematic Review and Taxonomy — → 2506.04290
- Dissecting Bias in LLMs: A Mechanistic Interpretability Perspective (2025) — link
- Beyond the black box: Interpretability of llms in finance — → 2505.24650
- Understanding and mitigating gender bias in llms via interpretable neuron editing (2025) — link
- Interpreting bias in large language models: a feature-based approach (2024) — link
- Evaluate Bias without Manual Test Sets: A Concept Representation Perspective for LLMs (2025) — link
- Unibias: Unveiling and mitigating llm bias through internal attention and ffn manipulation (2024)
4 of 63 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.