We present Team TIET’s systems for two shared tasks at #SMM4H-HeaRD 2026: Task 5 (detection of patient metadata in SARS-CoV-2 sequencing papers) and Task 1 (multilingual adverse drug event detection across six languages plus an unseen Farsi subset). For Task 5 we explore iterative LLM prompting followed by fine-tuning BiomedBERT-base with weighted cross-entropy loss and probability threshold optimization, achieving F1 = 0.760 on the official test set (above the competition mean of 0.729). For Task 1 we fine-tune XLM-RoBERTa-base with a combined language- and class-balanced sampling strategy and per-language threshold tuning, achieving macro F1 = 0.497 overall (0.608 excluding the unseen Farsi subset). We report empirical findings on BERT+LLM ensemble failure with bimodal probability distributions, the superiority of base over large model variants under limited data, and the importance of language-balanced gradient contribution in multilingual classification.

We adapt the AutoARGUE framework (Walden et al., 2026) for Task A.2 of RAG4Reports 2026, which requires ranking 57 report generation systems across 68 topics using automated evaluation. The RAGTIME-1 corpus poses a fundamental challenge: all nugget annotations use a no-reference-doc sentinel rather than ground-truth document citations, rendering the original citation-relevance gating inoperable. We address this with three adaptations: automatic sentinel detection with forced direct LLM-based nugget matching; a WEAK POSITIVE partial credit mechanism for sentences that correctly answer nuggets but lack attesting citations; and a report-level request alignment check. Our nugget_coverage_weighted metric achieves the highest topic-level Pearson correlation (r=0.599) of any non-coordinator submission, closely approaching the coordinator baseline (r=0.607).

Subword tokenizers are optimized for compression, but their boundaries also define the prediction units seen by a language model. We introduce StructPiece, an unsupervised tokenizer that enforces Unicode Extended Grapheme Cluster (EGC) boundaries and biases merge induction toward corpus-induced affix patterns without external resources. Across nine languages and five scripts, StructPiece exposes a compression–representation frontier: it produces higher-fertility segmentations with worse bits-per-character but substantially lower token-level perplexity under matched architectures, vocabulary sizes, and token budgets. Fixed-budget, granularity-matched, and component ablations show that the effect is not explained by token count alone; structural vocabulary alignment is further tested through LLaMA-3.2-1B vocabulary surgery, replacing its 128k vocabulary with 32k StructPiece tokens via mean-pooled initialization. StructPiece recovers 89.4% of baseline zero-shot reasoning performance across morphologically rich languages (X-CSQA) after 2.6B adaptation tokens, compared with 76.4% recovery for naive BPE surgery under the same checkpoint. Furthermore, structurally aligned tokens produce higher mean-pooled embedding cosine coherence, suggesting better-conditioned vocabulary projections. These results suggest that token boundaries act as a controllable structural bias influencing prediction granularity and vocabulary adaptation. StructPiece is available at https://github.com/submissionemnlp26-beep/structpiece-code.