Subjective NLP tasks such as sentiment analysis and hate speech classification often involve inherent annotator disagreement, reflecting diverse perspectives shaped by annotators’ lived experiences. Although conventional approaches resolve disagreement through majority voting or aggregation, these methods risk erasing valuable nuances and minority viewpoints. Recent embedding-based and multitask models have advanced the modeling of annotator-specific judgments, yet their robustness to annotation noise remains underexplored. In this work, we systematically investigate how state-of-the-art multi-annotator learning models perform in the presence of noisy labels, and find that embedding-based approaches are more robust than aggregation and multitask baselines on datasets with many sparse annotators. Furthermore, we propose a new methodology, Noise-Robust Annotator Embedding (NRA-Embed), which integrates Robust InfoNCE (RINCE) contrastive loss to enhance models’ robustness under noisy annotation conditions. Through extensive experiments, we demonstrate that NRA-Embed effectively captures subjective disagreement while maintaining robustness to noise, achieving performance that is comparable to state-of-the-art in clean settings while providing numerically greater robustness under noise.

Prior research has demonstrated that the choice of prompt template during Supervised Fine-Tuning (SFT) significantly impacts the robustness of safety alignment afterwards. However, the influence of template selection during Knowledge Distillation (KD) from teacher to student remains largely unexplored. Thus, we fill this gap by analyzing how different template configurations influence the pre-existing safety alignment of the student. We observe a significant degradation of safety alignment present in the aligned base instruct-tuned model. Specifically, we find that utilizing chat templates renders the model more compliant with harmful queries compared to a non-chat template. These findings are consistent across three models: LLaMA, Gemma and Qwen model families and are evaluated across multiple safety benchmarks. We further show that using a non-chat template during distillation better preserves the base student’s internal representations, while chat template distillation induces a larger representational shift.

Code: https://github.com/anjilab/role-of-prompt-template-in-kd

Large Language Models (LLMs) are increasingly used as conversational search engines for academic literature, yet whether they judge papers on content or on authority signals has not been tested causally. We investigate authority bias: systematic preference for papers based on author prestige, venue, and citations rather than content. Holding title and abstract constant, we vary authority metadata across three counterfactual conditions (original, flipped, boosted) over eight LLMs (five open-weight and three frontier closed-weight) in an in-context, single-turn, top-1 recommendation setting. Our experiments show that authority bias is substantial and directional, varies markedly across models, and is only partially addressable through prompt-level debiasing. We further document a say-do gap: debiasing instructions suppress authority mentions far faster than authority-driven flips, so surface auditing systematically underestimates behavioral bias.

Code: https://github.com/jinaduuthman/Authority-Bias-In-Conversational-Search-Engine

Data: https://huggingface.co/datasets/uthmanjinadu/authority-bias-paper-recommendation

To explore a more scalable path for adding multimodal capabilities to existing LLMs, this paper addresses a fundamental question: Can a unimodal LLM, relying solely on text, reason about its own informational needs and provide effective feedback to optimize a multimodal model? To answer this, we propose a method that enables a language agent to give feedback to a vision-language model (VLM) to adapt text generation to the agent’s preferences. Our results from different experiments affirm this hypothesis, showing that LLM preference feedback significantly enhances VLM descriptions. Using our proposed method, we find that the VLM can supply multimodal scene descriptions to help the LLM better understand multimodal context, leading to improvements of maximum 13% in absolute accuracy compared to the baseline multimodal approach. Furthermore, a human study validated our AI-driven feedback, showing a 64.6% preference alignment rate between the LLM’s choices and human judgments. Extensive experiments provide insights on how and why the method works and its limitations.