Large language models (LLMs) show promise for scientific applications but struggle when domain knowledge is encoded in complex structures. Chemistry exemplifies this challenge: molecular structures encode rich relational data that dictates chemical meaning, yet standard formats like SMILES obscure this through arbitrary linearization. In this work, we systematically characterize which structural information lost during linearization most severely hampers LLM reasoning, decomposing the problem into IUPAC-grounded dimensions: substituent positions (locants), attachment points (anchoring), and modification constraints (valency). Our diagnostic reveals non-uniform failures across dimensions, with locant identification being especially challenging. We directly address the failures by grounding the linearized representations in structural context with each component targeting its corresponding diagnostic dimension. General-purpose LLMs with explicit grounding can outperform chemistry-specialized models on structure-aware molecule manipulation, suggesting structural context can compensate for domain-specific training. Moreover, we investigate a training-free inductive protocol that nearly doubles the accuracy of weaker models. These results point toward explicit structural grounding as a principled path to more reliable molecular reasoning across model scales.

Multilingual neural machine translation (MNMT) offers the convenience of translating between multiple languages with a single model. However, MNMT often suffers from performance degradation in high-resource languages compared to bilingual counterparts. This degradation is commonly attributed to parameter interference, which occurs when parameters are fully shared across all language pairs. In this work, to tackle this issue we propose a gradient-based gradual pruning technique for MNMT. Our approach aims to identify an optimal sub-network for each language pair within the multilingual model by leveraging gradient-based information as pruning criterion and gradually increasing the pruning ratio as schedule. Our approach allows for partial parameter sharing across language pairs to alleviate interference, and each pair preserves its unique parameters to capture language-specific information. Comprehensive experiments on IWSLT and WMT datasets show that our approach yields a notable performance gain on both datasets.

This paper contains the description for the submission of Karlsruhe Institute of Technology (KIT) for the multilingual TEDx translation task in the IWSLT 2021 evaluation campaign. Our main approach is to develop both cascade and end-to-end systems and eventually combine them together to achieve the best possible results for this extremely low-resource setting. The report also confirms certain consistent architectural improvement added to the Transformer architecture, for all tasks: translation, transcription and speech translation.