What We Observe as LLM Behavior Can Be a Side-effect of Inference Backend

Benchmark scores are reported as properties of a model, yet the inference framework used to produce them, such as HuggingFace, vLLM, or Ollama, are considered non-influential and their names and versions are almost never disclosed. In this work we investigate how much this choice can influence the model output. In a fully-crossed study (three instruction-tuned models x five inference frameworks x six benchmarks x four generation modes) we investigate how different tools (wrappers/backend) influence benchmark scores and how their score changes is influenced by generation hyper-parameters. We find backend to be a non-negligible factor where even under greedy, sampling-noise-free decoding, changing the backend can significantly alter models performance and this effect is structural and strongly model-dependent. Decomposing the variance according to generation mode reveal that considerable portion of the variability (roughly 39\%) a practitioner sees out-of-the-box can stem from the backend, while the remaining stems from sampling noise and each framework's default generation parameters, both of which are avoidable by disclosing and matching the generation configuration. These divergences are more pronounced on factual than on social-bias benchmarks. Overall, benchmark numbers are not backend-agnostic therefore, we recommend disclosing the backend, its version, and the full generation configuration, also using deterministic decoding for cross-backend comparison.

Robustness and User-Perceived Value of Popularity Calibration in Music Recommendation: A User Study

Popularity calibration in recommender systems has been studied both as a form of user-centered personalization and as an indicator of popularity bias. Most existing work evaluates calibration through offline metrics, often assuming that users prefer recommendation lists whose popularity distribution matches their historical consumption profile. However, user studies on calibration remain limited, and existing findings suggest that calibrated recommendations do not necessarily have a strong effect on user experience. Moreover, although prior work has shown that calibration metrics can correlate with users' perceptions of recommendation lists, the robustness of this relation remains unclear under different levels of item familiarity and incomplete user-history information. In this work, we study the perceived value and measurement reliability of popularity calibration in music recommendation. We construct personalized track lists from users' recent listening histories and use a controlled naive recommender to create lists with different popularity compositions: highpop-heavy, lowpop-heavy, and calibrated. We investigate whether users perceive differences between these lists, whether calibrated lists are preferred, how robust JSD-based popularity calibration is under different familiarity and history-availability conditions, and how computational popularity labels align with users' own popularity judgments. Our results show that users perceive differences in popularity composition, but do not clearly prefer calibrated lists. We further find that the relation between JSD and perceived popularity depends on item familiarity, list composition, and available user history, while computational and user-judged popularity labels only weakly align. These findings contribute to a more critical understanding of popularity calibration as both an offline metric and a user-facing construct.

Learning Speaker Identity Beyond Language and Modality Constraints: Insights from the POLY-SIM 2026 Challenge

Multimodal speaker identification systems typically assume the availability of complete and homogeneous audio-visual modalities during both training and testing, and assume each speaker only speaks a single language. However, in real-world applications, such assumptions often do not hold. Visual or audio information may be missing due to occlusions, camera or microphone failures, or privacy constraints. Multilingual speakers introduce additional complexity due to linguistic variability across languages. These situations constitute substantial challenges for the robustness and generalization capabilities of multimodal speaker identification systems. Aim of the POLY-SIM 2026 challenge is to address these aspects of speaker identification and to provide a standardized setup for the comparison of the proposed solutions.

A Comparative Study on Affective Cues in Text Embeddings Across Psychological Emotion Theories

Text encoders are known for their utility in natural language processing, as they are able to efficiently compress inputs into dense vectors while preserving semantics. These models have been applied to affective computing, in particular to help with solving sentiment analysis and emotion recognition tasks. Nevertheless, it remains unclear to what extent the latent representations produced by modern text encoders capture well-defined psychological theories of affect. In this work, we investigate the affective capabilities of twelve recently released text encoders by probing their generated embeddings as input features for solving regression and classification tasks across three established emotion frameworks, using both word- and sentence-level data. Additionally, we apply a semantic data-leakage prevention technique to improve robustness in word-level evaluations. Our main findings show that the latent manifolds of the latest instruction-aware open-weight encoders enclose an equal or even a larger amount of affective information in comparison with proprietary counterparts when evaluated at word level. In contrast, embeddings of task-tuned and proprietary encoders reach the highest scores on sentence-level affective classification. Furthermore, a qualitative analysis of latent representations and their encoded affective cues is provided.

SB-BEVFusion: Enhancing the Robustness against Sensor Malfunction and Corruptions

Multimodal sensor fusion has demonstrated remarkable performance improvements over unimodal approaches in 3D object detection for autonomous vehicles. Typically, existing methods transform multimodal data from independent sensors, such as camera and LiDAR, into a unified bird's-eye view (BEV) representation for fusion. Although effective in ideal conditions, this strategy suffers from substantial performance deterioration when camera or LiDAR data are missing, corrupted, or noisy. To address this vulnerability, we develop a framework-agnostic fusion module for camera and LiDAR data that allows for handling cases when one of the two modalities is missing or corrupted. To demonstrate the effectiveness of our module, we instantiate it in BEVFusion [1], a well-established framework to combine camera and LiDAR data for 3D object detection. By means of quantitative experiments on the MultiCorrupt dataset, we demonstrate that our module achieves favorable performance improvements under scenarios of missing and corrupted modalities, substantially outperforming existing unified representation approaches across a wide range of sensor deterioration scenarios and reaching state-of-the-art performance in scenarios of corrupted modality due to extreme weather conditions and sensor failure.

Facet-Level Tracing of Evidence Uncertainty and Hallucination in RAG

Retrieval-Augmented Generation (RAG) aims to reduce hallucination by grounding answers in retrieved evidence, yet hallucinated answers remain common even when relevant documents are available. Existing evaluations focus on answer-level or passage-level accuracy, offering limited insight into how evidence is used during generation. In this work, we introduce a facet-level diagnostics framework for QA that decomposes each input question into atomic reasoning facets. For each facet, we assess evidence sufficiency and grounding using a structured Facet x Chunk matrix that combines retrieval relevance with natural language inference-based faithfulness scores. To diagnose evidence usage, we analyze three controlled inference modes: Strict RAG, which enforces exclusive reliance on retrieved evidence; Soft RAG, which allows integration of retrieved evidence and parametric knowledge; and LLM-only generation without retrieval. Comparing these modes enables thorough analysis of retrieval-generation misalignment, defined as cases where relevant evidence is retrieved but not correctly integrated during generation. Across medical QA and HotpotQA, we evaluate three open-source and closed-source LLMs (GPT, Gemini, and LLaMA), providing interpretable diagnostics that reveal recurring facet-level failure modes, including evidence absence, evidence misalignment, and prior-driven overrides. Our results demonstrate that hallucinations in RAG systems are driven less by retrieval accuracy and more by how retrieved evidence is integrated during generation, with facet-level analysis exposing systematic evidence override and misalignment patterns that remain hidden under answer-level evaluation.

POLY-SIM: Polyglot Speaker Identification with Missing Modality Grand Challenge 2026 Evaluation Plan

Multimodal speaker identification systems typically assume the availability of complete and homogeneous audio-visual modalities during both training and testing. However, in real-world applications, such assumptions often do not hold. Visual information may be missing due to occlusions, camera failures, or privacy constraints, while multilingual speakers introduce additional complexity due to linguistic variability across languages. These challenges significantly affect the robustness and generalization of multimodal speaker identification systems. The POLY-SIM Grand Challenge 2026 aims to advance research in multimodal speaker identification under missing-modality and cross-lingual conditions. Specifically, the Grand Challenge encourages the development of robust methods that can effectively leverage incomplete multimodal inputs while maintaining strong performance across different languages. This report presents the design and organization of the POLY-SIM Grand Challenge 2026, including the dataset, task formulation, evaluation protocol, and baseline model. By providing a standardized benchmark and evaluation framework, the challenge aims to foster progress toward more robust and practical multimodal speaker identification systems.

Adaptive Autoguidance for Item-Side Fairness in Diffusion Recommender Systems

Diffusion recommender systems achieve strong recommendation accuracy but often suffer from popularity bias, resulting in unequal item exposure. To address this shortcoming, we introduce A2G-DiffRec, a diffusion recommender that incorporates adaptive autoguidance, where the main model is guided by a less-trained version of itself. Instead of using a fixed guidance weight, A2G-DiffRec learns to adaptively weigh the outputs of the main and weak models during training, supervised by a popularity regularization that promotes balanced exposure across items with different popularity levels. Experimental results on the MovieLens-1M, Foursquare-Tokyo, and Music4All-Onion datasets show that A2G-DiffRec is effective in enhancing item-side fairness at a marginal cost of accuracy reduction compared to existing guided diffusion recommenders and other non-diffusion baselines.

Explanatory Interactive Machine Learning for Bias Mitigation in Visual Gender Classification

Explanatory interactive learning (XIL) enables users to guide model training in machine learning (ML) by providing feedback on the model's explanations, thereby helping it to focus on features that are relevant to the prediction from the user's perspective. In this study, we explore the capability of this learning paradigm to mitigate bias and spurious correlations in visual classifiers, specifically in scenarios prone to data bias, such as gender classification. We investigate two methodologically different state-of-the-art XIL strategies, i.e., CAIPI and Right for the Right Reasons (RRR), as well as a novel hybrid approach that combines both strategies. The results are evaluated quantitatively by comparing segmentation masks with explanations generated using Gradient-weighted Class Activation Mapping (GradCAM) and Bounded Logit Attention (BLA). Experimental results demonstrate the effectiveness of these methods in (i) guiding ML models to focus on relevant image features, particularly when CAIPI is used, and (ii) reducing model bias (i.e., balancing the misclassification rates between male and female predictions). Our analysis further supports the potential of XIL methods to improve fairness in gender classifiers. Overall, the increased transparency and fairness obtained by XIL leads to slight performance decreases with an exception being CAIPI, which shows potential to even improve classification accuracy.

Robust Harmful Meme Detection under Missing Modalities via Shared Representation Learning

Internet memes are powerful tools for communication, capable of spreading political, psychological, and sociocultural ideas. However, they can be harmful and can be used to disseminate hate toward targeted individuals or groups. Although previous studies have focused on designing new detection methods, these often rely on modal-complete data, such as text and images. In real-world settings, however, modalities like text may be missing due to issues like poor OCR quality, making existing methods sensitive to missing information and leading to performance deterioration. To address this gap, in this paper, we present the first-of-its-kind work to comprehensively investigate the behavior of harmful meme detection methods in the presence of modal-incomplete data. Specifically, we propose a new baseline method that learns a shared representation for multiple modalities by projecting them independently. These shared representations can then be leveraged when data is modal-incomplete. Experimental results on two benchmark datasets demonstrate that our method outperforms existing approaches when text is missing. Moreover, these results suggest that our method allows for better integration of visual features, reducing dependence on text and improving robustness in scenarios where textual information is missing. Our work represents a significant step forward in enabling the real-world application of harmful meme detection, particularly in situations where a modality is absent.