Politeness is a central pragmatic dimension of language use, and Japanese honorifics offer a well-defined testbed for studying whether pretrained encoders represent socially meaningful distinctions. Prior BERT-based work has applied supervised models to Japanese honorific data, but we are not aware of analyses that localize honorific-level information across layers or test causal influence via activation patching in Japanese BERT-style encoders. We study these questions in LineDistilBERT using the KeiCO corpus, which labels sentences with four honorific levels. To isolate pretrained representations while still defining a task predictor, we freeze all encoder parameters and train only a lightweight [CLS] classification head as a minimal readout. We then run layerwise linear probing, training multinomial L2-regularized logistic-regression probes on [CLS] vectors from each layer to quantify linear decodability across depth and to select a best layer on development data. Finally, we test causal leverage with [CLS] activation patching, transplanting donor activations into receiver sentences at selected layers and measuring prediction transitions, logit shifts, and flip rates under standard controls. Overall, honorific level is broadly decodable across layers, and [CLS] interventions can systematically steer the frozen-encoder classifier with strong depth dependence, providing complementary evidence from probing and causal intervention for Japanese politeness in practice.

Multilingual natural language inference (NLI) benchmarks are routinely constructed by translating English datasets into target languages, based on the assumption that inference relations between premise–hypothesis pairs are preserved under translation. This assumption, however, has rarely been tested empirically at scale. We investigate this by conducting a large-scale human annotation study across nine typologically diverse languages: Arabic, German, Spanish, Portuguese, Thai, Urdu, Turkish, Mandarin, and Cantonese. We translated 1,000 English premise–hypothesis pairs, sampled across four linguistically challenging phenomena, and annotated each target-language version by a native-speaker linguist. Additionally, a native English speaker annotated the source pairs under the same protocol, allowing us to distinguish source-label disagreement from translation-induced change. While agreement between translated-language annotations and the original dataset labels ranges from 55.1% to 67.4%, agreement with the native-English annotation ranges from 57.8% to 64.2%. Entailment and Contradiction frequently shift toward Neutral, but the English control shows that some apparent drift is already present in the source items. These results call into question using transfer of English labels and highlight the need for native-language annotation in multilingual NLI evaluation.

The AmericasNLP 2026 shared task challenges systems to generate culturally grounded image captions in indigenous languages of the Americas, a setting that demands both cultural awareness and linguistic accuracy for severely underresourced languages. We present IUHoosiers, Indiana University’s system for the Guaraní track. Rather than fine-tuning, our approach centers on inference-time knowledge injection: for each test image, we retrieve relevant Guaraní grammatical and cultural resources using BM25 and inject them into a large vision language model’s prompt alongside the image, enabling language-specific cultural and linguistic grounding without any parameter updates. IUHoosiers placed first for Guaraní in both automatic evaluation (24.67 chrF++) and human evaluation (3.45/5), outperforming all other participating systems.