arXiv subjects

Tomas Ruiz

Publications and source records attributed to Tomas Ruiz.

4 recordsLinked to original sources

arXiv2026

Rethinking Ground Truth: A Case Study on Human Label Variation in MLLM Benchmarking

Human Label Variation (HLV), i.e. systematic differences among annotators' judgments, remains underexplored in benchmarks despite rapid progress in large language model (LLM) development. We address this gap by introducing an evaluation protocol for multimodal large language model (MLLM) benchmarking that explicitly accounts for two conditions: (1) human label agreement and (2) disagreement. We apply this protocol to two state-of-the-art MLLM families (Gemma 3, Qwen 2.5 VL) using non-aggregated human annotations from a social media content classification dataset. Across tasks, we find that larger models tend to perform best on high-agreement subsets, yet often underperform medium-sized models when human disagreement is high, indicating that parameter count alone does not determine sensitivity to ambiguity and subjectivity. These results show that benchmarks based solely on consensus labels can overstate model capabilities in such domains and that incorporating human label variation yields more realistic and robust assessments of MLLMs in content moderation pipelines.

cs.CL↗

arXiv2026

FlashSampling: Fast and Memory-Efficient Exact Sampling

Sampling from a categorical distribution is mathematically simple, but in large-vocabulary decoding, it often triggers extra memory traffic and extra kernels after the LM head. We present FlashSampling, an exact sampling primitive that fuses sampling into the LM-head matmul and never materializes the logits tensor in HBM. The method is simple: compute logits tile-by-tile on chip, add Gumbel noise, keep only one maximizer per row and per vocabulary tile, and finish with a small reduction over tiles. In tensor-parallel decoding, FlashSampling replaces the all-gather of logits with streaming peer-to-peer writes: This overlaps GPU-to-GPU communication with computation and HBM loads across up to 8 GPUs, with near-ideal scaling at large batch sizes. Our kernel is exact because argmax decomposes over partitions; grouped variants for online and tensor-parallel settings are exact by hierarchical factorization of the categorical distribution. FlashSampling demonstrates kernel-level speedups on decode workloads across 4 different datacenter GPUs (H100, H200, B200, B300), and in end-to-end vLLM experiments, it reduces time per output token by up to $10\%$ on the models we test. These results show that exact sampling, with no approximation, can be integrated into the matmul itself, consolidating the bandwidth-bound sampling step in an efficient epilogue.

cs.LG↗

arXiv2025

BoN Appetit Team at LeWiDi-2025: Best-of-N Test-time Scaling Can Not Stomach Annotation Disagreements (Yet)

Test-time scaling is a family of techniques to improve LLM outputs at inference time by performing extra computation. To the best of our knowledge, test-time scaling has been limited to domains with verifiably correct answers, like mathematics and coding. We transfer test-time scaling to the LeWiDi-2025 tasks to evaluate annotation disagreements. We experiment with three test-time scaling methods: two benchmark algorithms (Model Averaging and Majority Voting), and a Best-of-N sampling method. The two benchmark methods improve LLM performance consistently on the LeWiDi tasks, but the Best-of-N method does not. Our experiments suggest that the Best-of-N method does not currently transfer from mathematics to LeWiDi tasks, and we analyze potential reasons for this gap.

cs.CL↗

arXiv2025

PoliTok-DE: A Multimodal Dataset of Political TikToks and Deletions From Germany

We present PoliTok-DE, a large-scale multimodal dataset (video, audio, images, text) of TikTok posts from two German elections: the 2024 Saxony state election and the 2025 German federal election. The corpus contains over 930,000 posts, of which over 330,000 were later deleted from the platform (18.7% of Saxony posts, 39.7% of federal posts). In the federal-election collection, about two thirds of the deletions were creator withdrawals, and the platform-deletion rate we computed was 13.0% of all posts, more than an order of magnitude (14-19x) above the platform-wide rate TikTok reported. Posts were identified via the TikTok research API and complemented with web scraping to retrieve full multimodal media and metadata. PoliTok-DE supports social science research across substantive and methodological agendas: substantive work on intolerance and political communication, and methodological work on platform policies around deleted content and qualitative-quantitative multimodal research. To illustrate, we report a case study on intolerance and entertainment in an annotated subset of deleted posts: about one in five posts conveyed intolerance and a majority conveyed humor.

cs.SI↗