pith. sign in

arxiv: 2509.21080 · v2 · submitted 2025-09-25 · 💻 cs.CL · cs.AI· cs.CY

InsideOut: Measuring and Mitigating Insider-Outsider Bias in Interview Script Generation

Yixin Wan , Xingrun Chen , Kai-Wei Chang This is my paper

Pith reviewed 2026-05-18 14:02 UTC · model grok-4.3

classification 💻 cs.CL cs.AIcs.CY
keywords insider-outsider biasLLM cultural fairnessinterview script generationbias mitigationagent-based methodsInsideOut benchmarkcultural alignment gap
0
0 comments X

The pith

Large language models default to insider perspectives for mainstream cultures and outsider stances for others when generating interview scripts, but agent-based methods can reduce this bias by over 80 percent on key measures.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper identifies a cultural positioning problem in LLMs during interview script generation, where models treat dominant cultures as their own and frame others as external. It introduces the InsideOut benchmark of 4000 prompts across ten cultures plus three metrics to track how often outputs adopt insider or outsider tones. Tests on five current models confirm the pattern is strong and consistent, with insider tones appearing in most US scripts but far less often elsewhere. The authors then show that inference-time agent frameworks can correct much of the imbalance without retraining the underlying models.

Core claim

LLMs exhibit insider-outsider bias by positioning themselves as cultural insiders for mainstream groups while externalizing less dominant cultures in interview script generation. The InsideOut benchmark quantifies this across models and cultures, and the Mitigation via Fairness Agents framework demonstrates that single-agent and hierarchical-agent pipelines can substantially lower the measured bias at inference time.

What carries the argument

The Mitigation via Fairness Agents (MFA) framework, which deploys single-agent, hierarchical-agent, or autonomous planning pipelines to intervene during script generation and adjust cultural stance.

If this is right

  • Models adopt insider tones in over 88 percent of US-contexted scripts on average.
  • Non-Western cultures receive disproportionately outsider stances in LLM-generated interview scripts.
  • MFA-SA reduces measured bias in the Llama model by 89.70 percent on the Cultural Alignment Gap metric.
  • MFA-HA reduces measured bias in the Qwen model by 82.54 percent on the Cultural Alignment Gap metric.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same positioning bias is likely to appear in other open-ended generation tasks that involve cultural content.
  • Agent-based correction could be combined with training-time methods to address the bias at multiple stages.
  • Extending the benchmark to additional cultures or real interview data would test how well the observed patterns hold outside the current prompt set.

Load-bearing premise

The three evaluation metrics accurately detect insider versus outsider positioning in the generated scripts without being overly affected by prompt wording or model-specific phrasing.

What would settle it

Re-evaluating the same model outputs with human raters from each target culture who classify insider versus outsider stance and obtaining substantially different bias percentages would indicate the automatic metrics do not reliably capture the intended phenomenon.

Figures

Figures reproduced from arXiv: 2509.21080 by Kai-Wei Chang, Xingrun Chen, Yixin Wan.

Figure 1
Figure 1. Figure 1: (L): CultureLens evaluation framework. (R): Qualitative examples of excerpts from [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of the ablation results for different bias mitigation approaches. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The proposed bias mitigation frameworks. FIP adopts a prompt-based fairness guideline [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: An overview of descriptors used in the curation of the C [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
read the original abstract

Advancements in Large language models (LLMs) have enabled a variety of downstream applications like story and interview script generation. However, recent research raised concerns about culture-related fairness issues in LLM-generated content. In this work, we identify and systematically investigate LLMs' insider-outsider bias, a phenomenon where models position themselves as "insiders" of mainstream cultures during generation while externalizing less dominant cultures. We propose the InsideOut benchmark with 4,000 generation prompts and three evaluation metrics to quantify this bias through a culturally situated interview script generation task, in which an LLM is positioned as a reporter interviewing local people across 10 diverse cultures. Empirical evaluation on 5 state-of-the-art LLMs reveals that while models adopt insider tones in over 88% US-contexted scripts on average, they disproportionately default to "outsider" stances for non-Western cultures. To mitigate these biases, we propose 2 inference-time methods: a baseline prompt-based Fairness Intervention Pillars (FIP) method, and a structured Mitigation via Fairness Agents (MFA) framework consisting of a Single-Agent (MFA-SA), a Hierarchical-Agent (MFA-HA), and an autonomous Agentic Planning (MFA-Plan) pipeline. Empirical results demonstrate that agent-based MFA methods achieve outstanding and robust performance in mitigating the insider-outsider bias: For instance, on the Cultural Alignment Gap (CAG) metric, MFA-SA reduces bias in Llama model by 89.70 % and MFA-HA mitigates bias in Qwen by 82.54%. These findings showcase the effectiveness of agent-based methods as a promising direction for mitigating biases in generative LLMs.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces the InsideOut benchmark consisting of 4,000 generation prompts and three evaluation metrics to quantify insider-outsider bias in LLMs for culturally situated interview script generation across 10 cultures. It reports that models adopt insider tones in over 88% of US-contexted scripts on average but default to outsider stances for non-Western cultures. The authors propose a baseline Fairness Intervention Pillars (FIP) method and a Mitigation via Fairness Agents (MFA) framework (including Single-Agent MFA-SA, Hierarchical-Agent MFA-HA, and MFA-Plan), with empirical results on five LLMs showing large bias reductions such as 89.70% on the Cultural Alignment Gap (CAG) metric for Llama using MFA-SA and 82.54% for Qwen using MFA-HA.

Significance. If the metrics are shown to validly and independently measure cultural positioning, the work would be significant for addressing culture-related fairness in generative applications. The introduction of a new benchmark, concrete agent-based mitigation pipelines, and reproducible quantitative results across multiple models provide practical tools and evidence for bias mitigation in LLMs.

major comments (2)
  1. [Metrics] Metrics section: The three evaluation metrics (including Cultural Alignment Gap) are defined and applied without any reported human validation, inter-rater reliability, or sensitivity analysis to prompt wording. This is load-bearing because the headline results (89.70% CAG reduction for MFA-SA on Llama; 82.54% for MFA-HA on Qwen) rest on the assumption that these metrics detect genuine insider-outsider stance rather than surface phrasing changes induced by the mitigation agents themselves.
  2. [Mitigation Methods] Mitigation evaluation: No explicit test is described for whether the CAG and other metrics remain stable when the exact agent prompts from MFA-SA/HA/Plan are used as input. Without this check, the reported robustness of the bias reductions risks circularity with the intervention design.
minor comments (2)
  1. [Benchmark Construction] Clarify the exact construction process for the 4,000 prompts to ensure balanced coverage across the 10 cultures and avoid unintended prompt artifacts.
  2. [Abstract] The abstract uses qualitative descriptors such as 'outstanding and robust performance'; replace or quantify these with reference to specific tables or statistical tests.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the detailed and constructive review of our manuscript. We appreciate the referee's focus on the validity of our evaluation metrics and the robustness of our mitigation methods. We address each major comment below and indicate the revisions we will make to strengthen the paper.

read point-by-point responses

  1. Referee: Metrics section: The three evaluation metrics (including Cultural Alignment Gap) are defined and applied without any reported human validation, inter-rater reliability, or sensitivity analysis to prompt wording. This is load-bearing because the headline results (89.70% CAG reduction for MFA-SA on Llama; 82.54% for MFA-HA on Qwen) rest on the assumption that these metrics detect genuine insider-outsider stance rather than surface phrasing changes induced by the mitigation agents themselves.

    Authors: We agree that human validation is crucial for establishing the metrics' ability to measure genuine cultural positioning. The metrics were constructed based on theoretical foundations from sociolinguistics, focusing on observable linguistic markers of insider vs. outsider perspectives. However, to directly address the concern, we will incorporate a human validation study in the revised manuscript. This will involve recruiting annotators familiar with the respective cultures to rate a representative sample of generated scripts according to insider-outsider criteria, and we will report inter-rater reliability metrics such as Cohen's kappa. We will also conduct a sensitivity analysis by varying the prompt wording slightly and observing metric consistency. These results will be added to the Metrics section to support the load-bearing claims. revision: yes

  2. Referee: Mitigation evaluation: No explicit test is described for whether the CAG and other metrics remain stable when the exact agent prompts from MFA-SA/HA/Plan are used as input. Without this check, the reported robustness of the bias reductions risks circularity with the intervention design.

    Authors: We acknowledge the potential for circularity if the metrics are sensitive to the specific phrasing in the agent prompts. In our current setup, the metrics evaluate the final output scripts generated under the influence of the MFA frameworks. To mitigate this concern, we will add a new experiment in the revised version where we directly use the agent prompts (without the full hierarchical or planning structure) as input to the base models and measure the resulting CAG and other metrics. This will allow us to isolate the effect of the agent-based coordination from prompt content alone. We expect this to show that the structured MFA approaches provide additional benefits beyond simple prompt engineering, and we will include these findings in the Mitigation Methods section. revision: yes

Circularity Check

0 steps flagged

No significant circularity: empirical benchmark and metrics are introduced independently of the mitigation results.

full rationale

The paper constructs a new InsideOut benchmark with 4000 prompts and defines three evaluation metrics (including CAG) to measure insider-outsider bias in generated interview scripts. It then applies mitigation methods (FIP, MFA variants) and reports empirical reductions on those metrics across LLMs. No equations, fitted parameters, or self-citations are shown to make the reported bias percentages (e.g., 89.70% CAG reduction) tautological with the metric definitions or prior author work. The derivation chain consists of independent artifact creation followed by measurement, satisfying the self-contained empirical case with external falsifiability via the new benchmark.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The central claims rest on the validity of the newly defined metrics and the assumption that the 10 chosen cultures plus the reporter-interview framing produce representative signals of cultural positioning. No free parameters are explicitly fitted in the abstract; the work relies on standard LLM evaluation assumptions plus the domain assumption that stance can be reliably scored from generated text.

axioms (2)
  • domain assumption The three evaluation metrics accurately quantify insider versus outsider cultural stance in generated interview scripts
    The paper's bias measurements and mitigation success rates depend directly on these metrics performing as intended.
  • domain assumption The 4000 prompts and 10 cultures provide a sufficient and balanced sample for detecting systematic bias patterns
    All reported percentages and reduction figures are computed over this constructed test set.
invented entities (1)
  • insider-outsider bias no independent evidence
    purpose: To name and frame the observed tendency of LLMs to adopt insider tones for mainstream cultures and outsider tones for others
    New descriptive term introduced to organize the empirical observations in the interview generation task.

pith-pipeline@v0.9.0 · 5843 in / 1701 out tokens · 62465 ms · 2026-05-18T14:02:05.943882+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

  1. [2]

    Emg-transnn-mha: A transformer-based model for enhanced motor intent recognition in assistive robotics,

    ISBN 0313297703. Siddhesh Pawar, Arnav Arora, Lucie-Aim´ee Kaffee, and Isabelle Augenstein. Presumed cultural identity: How names shape llm responses.arXiv preprint arXiv:2502.11995, 2025. Richard Peet. From eurocentrism to americentrism.Antipode, 37(5), 2005. Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Da...

    work page doi:10.1109/bigdata62323.2024.10825082 2025

  2. [3]

    Large language models help humans verify truthfulness – except when they are convincingly wrong

    Association for Computational Linguistics. ISBN 979-8-89176-251-0. doi: 10.18653/v1/ 2025.acl-long.445. URLhttps://aclanthology.org/2025.acl-long.445/. Yixin Wan, George Pu, Jiao Sun, Aparna Garimella, Kai-Wei Chang, and Nanyun Peng. ”kelly is a warm person, joseph is a role model”: Gender biases in llm-generated reference letters, 2023a. URLhttps://arxiv...

    work page doi:10.18653/v1/ 2025

  3. [4]

    Can you describe a tradition or practice that holds significance in your community?

    Cultural Neutrality * Guideline: Avoid assumptions or stereotypes about the culture’s values, behaviors, or beliefs. Use open-ended, factual prompts instead of leading questions. * Template: “Can you describe a tradition or practice that holds significance in your community?” (Avoid: “Why is your culture so focused on family values?”)

    work page

  4. [5]

    What are some key elements of this ceremony, and what do they represent?

    Contextual Awareness Without Exoticism * Guideline: Frame cultural practices in a neutral tone without sensationalizing or framing them as ”unusual.” * Template: “What are some key elements of this ceremony, and what do they represent?” (Avoid: “Why is this ceremony so strange compared to Western weddings?”)

    work page

  5. [6]

    How does your community typically celebrate important life events?

    Balanced Language Use * Guideline: Use precise and respectful language. Avoid comparative language that implicitly ranks cultures. * Template: “How does your community typically celebrate important life events?” (Avoid: “How does your culture differ from American norms in celebrating birthdays?”)

    work page

  6. [7]

    How would you explain this custom to someone unfamiliar with it?

    Insider V oice Empowerment * Guideline: Let the interviewee frame their own cultural narrative. Avoid inserting interpretations or explanations for them. * Template: “How would you explain this custom to someone unfamiliar with it?” (Avoid: “So this custom is kind of like Thanksgiving, right?”)

    work page

  7. [8]

    What historical or social factors have shaped this practice?

    Equal Depth and Curiosity * Guideline: Ask equally detailed and curious questions across all cultures to prevent showing over-familiarity or superficiality. * Template: “What historical or social factors have shaped this practice?” (Avoid: asking only factual surface-level questions to certain groups and deep philosophical ones to others)

    work page

  8. [9]

    Is this tradition still widely practiced today, or is it more associated with older generations or specific regions?

    Temporal and Regional Specificity * Guideline: Clarify if a cultural trait is regional, contemporary, or historical to avoid overgeneralization. * Template: “Is this tradition still widely practiced today, or is it more associated with older generations or specific regions?” (Avoid: “So all people from this culture do this?”)

    work page

  9. [10]

    Are there different perspectives or interpretations of this tradition within your community?

    Recognition of Cultural Dynamism * Guideline: Acknowledge that cultures evolve and contain internal diversity. * Template: “Are there different perspectives or interpretations of this tradition within your community?” (Avoid: “Is this the only correct way this is done?”)

    work page

  10. [11]

    What are some values or principles that guide daily life in your culture?

    Avoidance of Deficit Framing * Guideline: Do not frame cultural differences as problems or limitations. * Template: “What are some values or principles that guide daily life in your culture?” (Avoid: “What challenges does your culture face in adapting to modernity?”)

    work page

  11. [12]

    We’re hoping to understand how cultural practices shape community life. Would you feel comfortable sharing examples from your experience?

    Transparent Intent * Guideline: Share the purpose of the interview in a way that respects the cultural knowledge being shared. * Template: “We’re hoping to understand how cultural practices shape community life. Would you feel comfortable sharing examples from your experience?”

    work page

  12. [13]

    Consider involving cultural consultants in the review process

    Reflection and Review * Guideline: Before finalizing, review the script for imbalance, jargon, or assumptions. Consider involving cultural consultants in the review process. Table 10: Input prompt and full generated FIP guidelines for interview generation. 20

    work page