2026· Proceedings of the 15th International Conference on Data Science, Technology and Applications· 0 citations· 14 references
TL;DR
A fundamental disconnect is suggested between a model's capacity for factual accuracy and its ability to maintain social fairness, highlighting the need for multi-dimensional evaluation frameworks for small-scale systems.
Abstract
: As small-scale, open-source Large Language Models (LLMs) proliferate for on-device and privacy-centric applications, understanding the trade-offs between their utility and behavioural reliability becomes critical. This study evaluates a suite of instruction-tuned LLMs, Gemma, Llama, and Qwen ( ≤4 B parameters), treating the model family, and scale as the primary units of analysis. Retrieval-Augmented Generation (RAG) is employed as a controlled experimental condition to assess utility gains on the Natural Questions (NQ) benchmark, while utilizing native, non-augmented configurations to establish a fairness baseline via the Bias Benchmark for QA (BBQ). The findings reveal that while RAG significantly enhances utility, often doubling Exact Match (EM) scores, these gains are non-uniform and architecture-dependent, with certain families exhibiting greater "retrieval-readiness" than others. Paradoxically, the fairness analysis shows that providing explicit context in disambiguated settings can increase stereotype engagement rather than suppressing it. These results suggest a fundamental disconnect between a model's capacity for factual accuracy and its ability to maintain social fairness, highlighting the need for multi-dimensional evaluation frameworks for small-scale systems.
Large Language Models (LLMs) often produce outputs that reflect social biases, toxicity, or unfair treatment of demographic groups, undermining trust and fairness. While prior mitigation strategies frequently rely on complex architectures, access to model internals, or costly fine-tuning, we argue that simplicity can be a strength. We introduce StarDTox, a lightweight, critique-and-revise multi-agent framework that leverages the LLM's own internal knowledge, via a small number of coordinated prompts, to self-correct harmful outputs. Dedicated agents independently assess bias and overall output quality, and their feedback is integrated to guide prompt-based revision. Without modifying model weights or requiring any extra finetuning, StarDTox offers strong bias mitigation and high-quality outputs across both open-ended text generation and structured tasks, outperforming other baselines. For the text generation task, on the RealToxicityPrompt dataset, it reduces toxicity by over 50% compared to other baselines, while maintaining over 90% fluency. In addition, in structured tasks, on the BBQ benchmark, it achieves the lowest bias scores across both ambiguous and disambiguated examples, without sacrificing accuracy.
Shirin Tahmasebi, Narjes Nikzad, A. H. Payberah et al.· Annual International Compute...· 0 citations
Benevolence bias is identified and measure, a small but consistent tendency for aligned LLMs to lean toward the kinder, safer, more socially approved answer on value-laden survey questions, and is easy to diagnose and straightforward to fix.
Yuanzi Li, Jun-Hao Wang, Minghui Liu et al.· 0 citations
ConfidenceBench, a calibration benchmark that evaluates verbalized confidence estimates in 15 frontier LLMs using the Brier score, a proper scoring rule that incentivises truthful probability reporting shows that verbalized confidence calibration is a distinct and practically important axis of LLM reliability, complementary to standard accuracy-based evaluation.
M. ffrench-Constant, Daniel Yang, Xinmeng Huang et al.· 1 citation· ⚡1
An integrated conceptual frame-work that couples attention- and perturbation-based explainability with lightweight hallucination-detection signals and token-efficient inference strategies is presented, and a set of cross-cutting consistency metrics are instrumented with a set of cross-cutting consistency metrics.
Sakshi Parate, Shreyans Sanyal· Advanced International Journ...· 0 citations
DiverValue-Bench is introduced, a population-aware benchmark for evaluating multi-dimensional value alignment across 74 countries/regions and it is shown that lightweight preference-based fine-tuning with Low-Rank Adaptation and Direct Preference Optimization substantially improves in-domain value alignment while yielding consistent out-of-domain gains.
Yao Liang, Dongcheng Zhao, Feifei Zhao et al.· 0 citations
LLM evaluators (trained reward models and prompted LLM-as-a-Judge) are routinely validated via pairwise accuracy. In a multilingual setting, this operates under the premise that high pairwise accuracy implies reliable, language-neutral scoring. We show that this assumption does not hold. We conduct experiments with semantically identical instruction-response pairs across 23 languages, and find that multilingual evaluators assign significantly different scores to different evaluation languages. The bias is statistically significant and consistent across eight open-weight evaluators of different architectures and training paradigms, persists in frontier judges, and is strongly correlated with language resource level: lower-resource languages are scored more generously. Meanwhile, these biases are invisible to pairwise accuracy: evaluators achieve above 90% pairwise accuracy, yet have up to 43% difference in acceptance rate across languages under a global decision threshold, meaning, for instance, that harmful content in lower-resource languages is more likely to pass safety filters. Per-language thresholds would require language identification, which can be defeated by code-switched prompts. We then investigate why lower-resource languages receive higher rather than lower scores, and we find that model uncertainty is linked with the effect: models tend to give higher scores when less confident, both under negative log-likelihood and under token-free uncertainty measures; however, language identity remains a significant predictor after controlling for uncertainty, and the bias cannot be explained away by content difficulty alone, but is a structural, language-level misalignment.
Yijie Zhou, Lucas Resck, Zheng Hui et al.· 0 citations