Sesgo en LLMs

Benchmarks, datasets y métodos de mitigación de sesgo en modelos de lenguaje.

← Literature Review

---

Benchmarks y Datasets

Estado	Año	Título	Tipo de método	Resumen
	2020	RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models	Benchmark / Dataset	Ver
	2020	Social Bias Frames: Reasoning about Social and Power Implications of Language	Benchmark / Dataset	Ver
	2021	TruthfulQA: Measuring How Models Mimic Human Falsehoods	Benchmark / Dataset	Ver
	2021	Bot-Adversarial Dialogue for Safe Conversational Agents	Benchmark / Dataset	Ver
	2021	StereoSet: Measuring stereotypical bias in pretrained language models	Benchmark / Dataset	Ver
	2021	BBQ: A hand-built bias benchmark for question answering	Benchmark / Dataset	Ver
	2022	ToxiGen: A Large-Scale Machine-Generated Dataset for Adversarial and Implicit Hate Speech Detection	Benchmark / Dataset	Ver
	2022	Understanding Dataset Difficulty with V-Usable Information	Benchmark / Dataset	Ver
	2022	“I’m sorry to hear that”: Finding New Biases in Language Models with a Holistic Descriptor Dataset	Benchmark / Dataset	Ver
	2023	BeaverTails: Towards Improved Safety Alignment of LLM via a Human-Preference Dataset	Benchmark / Dataset	Ver
	2023	HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models	Benchmark / Dataset	Ver
	2023	Nationality Bias in Text Generation	Benchmark / Dataset	Ver
	2025	BiasFreeBench: a Benchmark for Mitigating Bias in Large Language Model Responses	Benchmark / Dataset	Ver
	2025	Do Bias Benchmarks Generalise? Evidence from Voice-based Evaluation of Gender Bias in SpeechLLMs	Benchmark / Dataset	Ver

---

Métodos de Mitigación

Estado	Año	Título	Tipo de método	Resumen
	2021	An Empirical Survey of the Effectiveness of Debiasing Techniques for Pre-trained Language Models	Evaluación / análisis	Ver
	2021	Sustainable Modular Debiasing of Language Models	Adapters / PEFT	Ver
	2021	FairFil: Contrastive Neural Debiasing Method for Pretrained Text Encoders	Fine-tuning / data augmentation	Ver
	2022	Debiasing Pre-Trained Language Models via Efficient Fine-Tuning	Fine-tuning / data augmentation	Ver
	2022	MABEL: Attenuating Gender Bias using Textual Entailment Data	Fine-tuning / data augmentation	Ver
	2022	Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned	Evaluación / análisis	Ver
	2022	Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback	Fine-tuning / data augmentation	Ver
	2023	Language Models Get a Gender Makeover: Mitigating Gender Bias with Few-Shot Data Interventions	Fine-tuning / data augmentation	Ver
	2023	An Empirical Analysis of Parameter-Efficient Methods for Debiasing Pre-Trained Language Models	Adapters / PEFT	Ver
	2023	Mitigating Biases for Instruction-following Language Models via Bias Neurons Elimination	Edición de pesos / neuronas	Ver
	2023	Causal-Debias: Unifying Debiasing in Pretrained Language Models via Causal Invariant Learning	Causal / invariante	Ver
	2023	D-CALM: A Dynamic Clustering-based Active Learning Approach for Mitigating Bias	Fine-tuning / data augmentation	Ver
	2024	Self-Debiasing Large Language Models: Zero-Shot Recognition and Reduction of Stereotypes	Tiempo de inferencia	Ver
	2024	ChatGPT Based Data Augmentation for Improved Parameter-Efficient Debiasing of LLMs	Fine-tuning / data augmentation	Ver
	2025	BiasEdit: Debiasing Stereotyped Language Models via Model Editing	Edición de pesos / neuronas	Ver
	2025	Debiasing the Fine-Grained Classification Task in LLMs with Bias-Aware PEFT	Adapters / PEFT	Ver
	2025	BiasFilter: An Inference-Time Debiasing Framework for Large Language Models	Tiempo de inferencia	Ver
	2025	FairSteer: Inference Time Debiasing for LLMs with Dynamic Activation Steering	Tiempo de inferencia	Ver
	2025	LLM Bias Detection and Mitigation through the Lens of Desired Distributions	Fine-tuning / data augmentation	Ver
	2025	BiasGym: Fantastic LLM Biases and How to Find (and Remove) Them	Evaluación / análisis	Ver
	2025	Dissecting Bias in LLMs: A Mechanistic Interpretability Perspective	Evaluación / análisis	Ver
	2025	Aligned but Stereotypical? The Hidden Influence of System Prompts on Social Bias	Evaluación / análisis	Ver
	2026	KnowBias: Mitigating Social Bias in LLMs via Know-Bias Neuron Enhancement	Edición de pesos / neuronas	Ver
	2026	No Free Lunch in Language Model Bias Mitigation?	Evaluación / análisis	Ver

---

Estadísticas

Por tipo de método

Tipo de método	N° de papers
Benchmark / Dataset	14
Fine-tuning / data augmentation	8
Evaluación / análisis	6
Adapters / PEFT	3
Edición de pesos / neuronas	3
Tiempo de inferencia	3
Causal / invariante	1
Total	38

---

Frecuencia de datasets en papers de métodos

Número de papers de mitigación (sobre 24) que utilizan cada dataset.

Dataset	Papers que lo usan
StereoSet	18
WinoBias	15
BBQ	11
CrowS-Pairs	10
GLUE	8
SEAT	5
BOLD	3
MMLU	2
BiasFreeBench, STS-B, WNC, SentiBias, SNLI/MultiNLI, TruthfulQA, WEAT, HH-RLHF, FairFace, CUB-200, Stanford Cars, Food-101	1 cada uno

---

Métodos que miden calidad general del modelo

De los 18 papers que proponen un método de mitigación activo (excluidos los 6 de evaluación/análisis pura).

Mide calidad general	N° de papers	Papers
Sí	11	FairFil, Gira, MABEL, RLHF-Asst., Lauscher, Xie, Yang, Causal-Debias, BiasEdit, Zhao, KnowBias
No	7	Thakur, D-CALM, Gallegos, Han, BiasFilter, FairSteer, Shrestha