Stereotypes as Heuristics for Resolving Ambiguous Contexts

This direction is split into other directions: ambiguity-induced bias amplification. Its last version stays readable and citable here.

Version 1 of 171 papers101 conceptsfrozen 2026-10-04

Large Language Models (LLMs) are fundamentally designed to be "completion engines." When presented with a prompt, their objective is to minimize surprise by predicting the most statistically likely continuation. However, this objective creates a critical failure mode when the model encounters ambiguity—scenarios where the provided information is insufficient to reach a factual conclusion.

Instead of acknowledging a lack of information, models often treat ambiguity as a "tie-breaking" problem. To resolve the uncertainty, they fall back on the strongest statistical associations in their training data: social stereotypes. This research direction shifts the focus from viewing bias as a purely "toxic" trait to understanding it as a flawed reasoning heuristic used to fill information gaps.

Ambiguity as a Revelatory Mechanism

Traditional bias benchmarks often look for overt toxicity. However, researchers like Andrew Mao and Jordan Lee Boyd-Graber argue in Eliciting Bias in Question Answering Models through Ambiguity that ambiguity serves as a "revelatory mechanism." By presenting underspecified queries where no single correct answer exists, researchers can force a model to rely on its internal priors, making latent biases observable.

The most prominent tool for this is the Bias Benchmark for QA (BBQ), introduced by Alicia Parrish, Nikita Nangia, and colleagues in BBQ: A Hand-Built Bias Benchmark for Question Answering. BBQ specifically tests two conditions:

  1. Ambiguous Contexts: Where the answer is not present (e.g., "A doctor and a nurse were at the hospital. Who was tired?").
  2. Disambiguated Contexts: Where the answer is explicitly stated (e.g., "...The doctor was tired.").

The "intuition" behind BBQ is that a fair model should choose an "unknown" or "not enough information" option in the ambiguous case. Instead, state-of-the-art models frequently default to stereotypical associations (e.g., assuming the doctor is male or the nurse is female) to resolve the uncertainty.

The Fallback Mechanism

Why does this happen? Maarten Sap, Yejin Choi, and others suggest in Social Bias Frames: Reasoning about Social and Power Implications of Language that models default to high-frequency, generic stereotypes because they lack a structured way to model the "implied" logic of social dynamics.

This "fallback mechanism" is not limited to text. In the realm of vision, Alberto Testoni and Raquel Fernández introduced RACQUET in RACQUET: Unveiling the Dangers of Overlooked Referential Ambiguity in Visual LLMs. They found that when a Vision-Language Model is asked a question about an image with multiple potential subjects (referential ambiguity), it doesn't ask for clarification. Instead, it "picks a winner" based on visual saliency or harmful social stereotypes, mapping subjective adjectives to specific individuals based on gender or ethnicity.

Biased or Just Flawed?

A provocative framing emerges from Akshita Jha and Sanchit M. Kabra in Biased or Flawed? Mitigating Stereotypes in Generative Language Models by Addressing Task-Specific Flaws. They argue that what we perceive as "bias" is often a fundamental failure in reading comprehension—specifically, the inability to recognize when a context is under-informative.

By instruction-tuning models to recognize information gaps and explicitly report when a question cannot be answered, they reduced stereotypical responses by over 60%. This suggests that bias mitigation doesn't always require "debiasing" specific identity groups; rather, it requires improving the model's ability to abstain under uncertainty.

The Cost of Stereotypical Heuristics

The reliance on stereotypes isn't just a social problem; it's an accuracy problem. Alicia Parrish notes that models suffer an "accuracy cost" when factual information explicitly contradicts a stereotype. In these cases, the internal "gut feeling" of the model (its inductive bias) can be so strong that it overrides the provided context, leading the model to provide a biased answer even when the correct one is right in front of it.

This is further complicated by language. Arianna Bisazza and colleagues found in MBBQ: A Dataset for Cross-Lingual Comparison of Stereotypes in Generative LLMs that a model's reliance on stereotypes increases as its linguistic proficiency decreases. When a model struggles to reason through an ambiguous context in a lower-resource language, it leans even harder on biased heuristics as a crutch.

Toward Uncertainty-Aware Fairness

The consensus across these 64 papers is that a "fair" model is not one that has been scrubbed of all stereotypes, but one that is uncertainty-aware.

As Yinong Oliver Wang and Katherine Metcalf propose in Is Your Model Fairly Certain? Uncertainty-Aware Fairness Evaluation for LLMs, we should penalize models for being "confidently wrong" about stereotypes. A truly robust system should reflect the inherent ambiguity of the world by increasing its uncertainty or asking for more data, rather than using societal prejudices as a convenient shortcut for completion.

Papers in this version

Showing 60 of 71 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_stereotypes_as_heuristics_for_resolving__55dd8d99,
  title = {Stereotypes as Heuristics for Resolving Ambiguous Contexts},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/stereotypes-as-heuristics-for-resolving-ambiguous-contexts/rdr_55dd8d99979b468db7af1f59410c7c86},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_stereotypes_as_heuristics_for_resolving__55dd8d99_v1,
  title = {Stereotypes as Heuristics for Resolving Ambiguous Contexts},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/stereotypes-as-heuristics-for-resolving-ambiguous-contexts/rdr_55dd8d99979b468db7af1f59410c7c86/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}