Universal multimodal embedders enable retrieval across text, image, and combined queries, but their dense representations incur high memory and inference costs. Post-hoc sparsification could reduce these costs but remains underexplored for multimodal retrieval. We introduce PUMA, a sparse autoencoder recipe that maps universal multimodal embeddings to compact sparse codes without retraining the backbone: a pretraining stage preserves dense dot-product geometry, after which the sparse encoder is fine-tuned for retrieval. We evaluate on five benchmarks covering text-to-image and composed image retrieval. On Qwen3-VL-Embedding-2B, PUMA is statistically indistinguishable from or improves over dense retrieval on four of five datasets. We further identify two failure modes of post-hoc sparsification: insufficient pre-TopK support and retrieval-misaligned active support. PUMA reduces per-vector storage by 8-16x (FP32) and runs up to 25x faster over exact dense scoring on larger candidate pools, offering a practical approach to storage- and latency-efficient multimodal retrieval.

Large Language Models (LLMs) are increasingly used as evaluators for code generation, but existing judge benchmarks mainly target general-purpose programming. Constraint Programming (CP) requires a stricter notion of correctness: a model is correct only if it faithfully represents the intended solution space, not merely if it executes, preserves satisfiability, or returns one valid solution. We introduce CPJudgeBench, a benchmark for meta-evaluating LLM judges of CP modeling correctness through solver-supported solution-space semantics. It contains 522 solver-validated candidate models from 82 CP problem families in CPMpy and MiniZinc, labeled as equivalent, unsound, incomplete, mixed, or non-executable. Across our experiments, we meta-evaluate a total of 15 LLM judges through semantic-label, binary-correctness, and score-based protocols. Results show that fine-grained semantic judging remains difficult, especially for reference-free MiniZinc, while stronger reasoning- and code-oriented models perform more reliably. Score-based LLM judgments also align better with solution-space similarity than standard text or code metrics. The benchmark system is available at https://github.com/Nastaran95/CPJudgeBench.

Inductive link prediction is emerging as a key paradigm for real-world knowledge graphs (KGs), where new entities frequently appear and models must generalize to them without retraining. Predicting links in a KG faces the challenge of guessing previously unseen entities by leveraging generalizable node features such as subgraph structure, type annotations, and ontological constraints. However, explicit type information is often lacking or incomplete. Even when available, type information in most KGs is often coarse-grained, sparse, and prone to errors due to human annotation. In this work, we explore the potential of pre-trained language models (PLMs) to enrich node representations with implicit type signals. We introduce TyleR, a Type-less yet type-awaRe approach for subgraph-based inductive link prediction that leverages PLMs for semantic enrichment. Experiments on standard benchmarks demonstrate that TyleR outperforms state-of-the-art baselines in scenarios with scarce type annotations and sparse graph connectivity. To ensure reproducibility, we share our code at https://github.com/sisinflab/tyler .

Large Language Models (LLMs) have rapidly become central to NLP, demonstrating their ability to adapt to various tasks through prompting techniques, including sentiment analysis. However, we still have a limited understanding of how these models capture sentiment-related information. This study probes the hidden layers of LLaMA models to pinpoint where sentiment features are most represented and to assess how this affects sentiment analysis.Using probe classifiers, we analyze sentiment encoding across layers and scales, identifying the layers and pooling methods that best capture sentiment signals. Our results show that sentiment information is most concentrated in mid-layers for binary polarity tasks, with detection accuracy increasing up to 14% over prompting techniques. Additionally, we find that in decoder-only models, the last token is not consistently the most informative for sentiment encoding. Finally, this approach enables sentiment tasks to be performed with memory requirements reduced by an average of 57%.These insights contribute to a broader understanding of sentiment in LLMs, suggesting layer-specific probing as an effective approach for sentiment tasks beyond prompting, with potential to enhance model utility and reduce memory requirements.

Factual hallucinations are a major challenge for Large Language Models (LLMs). They undermine reliability and user trust by generating inaccurate or fabricated content. Recent studies suggest that when generating false statements, the internal states of LLMs encode information about truthfulness. However, these studies often rely on synthetic datasets that lack realism, which limits generalization when evaluating the factual accuracy of text generated by the model itself. In this paper, we challenge the findings of previous work by investigating truthfulness encoding capabilities, leading to the generation of a more realistic and challenging dataset. Specifically, we extend previous work by introducing: (1) a strategy for sampling plausible true-false factoid sentences from tabular data and (2) a procedure for generating realistic, LLM-dependent true-false datasets from Question Answering collections. Our analysis of two open-source LLMs reveals that while the findings from previous studies are partially validated, generalization to LLM-generated datasets remains challenging. This study lays the groundwork for future research on factuality in LLMs and offers practical guidelines for more effective evaluation.