One Example Is Enough to Pass Fairness Benchmarks: Rethinking Fairness Evaluation for Aligned LLMs

• URL: http://arxiv.org/abs/2609.14860v1
• Date: Mon, 14 Sep 2026 00:21:26 GMT
• Title: One Example Is Enough to Pass Fairness Benchmarks: Rethinking Fairness Evaluation for Aligned LLMs
• Abstract summary: Fairness benchmarks such as BBQ have become the de facto standard for fairness evaluation across major model families.<n>We argue that these benchmarks are too easy to support their role.<n>We call for evaluation suites that cover a broader spectrum of fairness alignment.
• Score: 38.07475237991597
• License: http://creativecommons.org/licenses/by-nc-sa/4.0/
• Abstract: Warning: This submission studies stereotypes and biases, and contains toxic and offensive examples, used for illustration purposes only. Fairness benchmarks such as BBQ have become the de facto standard for fairness evaluation across major model families. We argue that these benchmarks are too easy to support their role: training Qwen 2.5 7B Base with Group Relative Policy Optimization (GRPO) on a single BBQ example, or placing that example in context as a one-shot demonstration for in-context learning (ICL), lifts mean BBQ accuracy from 79.9% to 92.9% and 99.0%, respectively, closing 80% of the gap to its large-scale RLHF counterpart (96.1%) with GRPO, and surpassing it with ICL. These effects generalize across model families. A cross-conditioning analysis shows the improvement is carried by the reasoning traces generated by the model, and one example suffices to elicit a category-agnostic ``missing evidence'' reasoning pattern. We argue that BBQ-style multiple-choice abstention benchmarks measure a single structural cue, and a model that solves them does not thereby become fair. We call for evaluation suites that cover a broader spectrum of fairness alignment.

Related papers

• Procedural Fairness Failures in RLHF from Preference Averaging [0.0]
  This work defines procedural fairness in alignment and shows that standard RLHF violates this via preference averaging.<n> Preference-Aware RLHF (PA-RLHF) is introduced, separating optimization across preference modes at the reward learning stage.<n>In a controlled setting, PA-RLHF improves overall alignment accuracy from 46.9% to 67.9% and reduces the fairness gap between best and worst aligned groups from 15.9 to 9.6 percentage points.
  arXiv  Detail & Related papers  (2026-08-10T18:38:16Z)
• The Capability Frontier: Benchmarks Miss 82% of Model Performance [7.258179693117413]
  Existing benchmarks typically report accuracy for a single model on a single run.<n>This systematically understates real-world LLM capabilities, particularly under heterogeneous data distributions.
  arXiv  Detail & Related papers  (2026-06-25T10:20:47Z)
• Reasoning Quality Emerges Early: Data Curation for Reasoning Models [58.56882815783977]
  We show that diverse and challenging reasoning examples can be identified using only the initial reasoning tokens.<n>Our method outperforms existing baselines by up to 1.7% while being 91% more token efficient.
  arXiv  Detail & Related papers  (2026-06-25T09:32:58Z)
• It Takes One to Bias Them All: Breaking Bad with One-Shot GRPO [44.170225015158195]
  We show that one-shot GRPO training on a single biased example is sufficient to induce systematic bias.<n>Our results reveal a critical vulnerability in post-training: alignment can be overridden by a single example.
  arXiv  Detail & Related papers  (2026-06-09T14:44:01Z)
• Outcome Accuracy is Not Enough: Aligning the Reasoning Process of Reward Models [108.26461635308796]
  We introduce Rationale Consistency, a fine-grained metric that quantifies the alignment between the model's reasoning process and human judgment.<n>Our evaluation of frontier models reveals that rationale consistency effectively discriminates among state-of-the-art models.<n>We introduce a hybrid signal that combines rationale consistency with outcome accuracy for GenRM training.
  arXiv  Detail & Related papers  (2026-02-04T15:24:52Z)
• Beyond Correctness: Evaluating Subjective Writing Preferences Across Cultures [87.75098311090642]
  Current preference learning methods achieve high accuracy on standard benchmarks but exhibit significant performance degradation when objective quality signals are removed.<n>We introduce WritingPreferenceBench, a dataset of 1,800 human-annotated preference pairs (1,200 English, 600 Chinese) across 8 creative writing genres.
  arXiv  Detail & Related papers  (2025-10-16T12:23:13Z)
• From Harm to Help: Turning Reasoning In-Context Demos into Assets for Reasoning LMs [58.02809208460186]
  We revisit this paradox using high-quality traces from DeepSeek-R1 as demonstrations.<n>We find that adding more exemplars consistently degrades accuracy, even when demonstrations are optimal.<n>We introduce Insight-to-solve (I2S), a sequential test-time procedure that turns demonstrations into explicit, reusable insights.
  arXiv  Detail & Related papers  (2025-09-27T08:59:31Z)
• GRPO-CARE: Consistency-Aware Reinforcement Learning for Multimodal Reasoning [53.894789613838654]
  We introduce SEED-Bench-R1, a benchmark with complex real-world videos requiring balanced perception and reasoning.<n>Using SEED-Bench-R1, we find that standard GRPO, while improving answer accuracy, often reduces logical coherence between reasoning steps and answers, with only a 57.9% consistency rate.<n>We propose GRPO-CARE, a consistency-aware RL framework optimizing both answer correctness and reasoning coherence without explicit supervision.
  arXiv  Detail & Related papers  (2025-06-19T08:49:13Z)
• Towards Large Language Models that Benefit for All: Benchmarking Group Fairness in Reward Models [16.977176752570617]
  Large Language Models (LLMs) are increasingly powerful and accessible to human users.<n> Ensuring fairness across diverse demographic groups, i.e., group fairness, is a critical ethical concern.<n>This work benchmarks the group fairness of learned reward models.
  arXiv  Detail & Related papers  (2025-03-10T19:39:39Z)
• The Unbearable Weight of Massive Privilege: Revisiting Bias-Variance Trade-Offs in the Context of Fair Prediction [7.975779552420981]
  We propose a conditional-iid (ciid) model that seeks to improve on the trade-offs made by a single model. We empirically test our setup on the COMPAS and folktables datasets. Our analysis suggests that there might be principled procedures and concrete real-world use cases under which conditional models are preferred.
  arXiv  Detail & Related papers  (2023-02-17T05:34:35Z)
• On Comparing Fair Classifiers under Data Bias [42.43344286660331]
  We study the effect of varying data biases on the accuracy and fairness of fair classifiers. Our experiments show how to integrate a measure of data bias risk in the existing fairness dashboards for real-world deployments.
  arXiv  Detail & Related papers  (2023-02-12T13:04:46Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.