We present a comprehensive evaluation and extension of the LLM-Assisted Rule-Based Machine Translation (LLM-RBMT) paradigm, an approach that combines the strengths of rule-based methods and Large Language Models (LLMs) to support translation in no-resource settings. We present a robust new implementation (the Pipeline Translator) that generalizes the LLM-RBMT approach and enables flexible adaptation to novel constructions. We benchmark it against four alternatives (Builder, Instructions, RAG, and Fine-tuned translators) on a curated dataset of 150 English sentences, and compare them across translation quality and runtime. The Pipeline Translator consistently achieves the best overall performance. The LLM-RBMT methods (Pipeline and Builder) also offer an important advantage: they naturally align with evaluation strategies that prioritize grammaticality and semantic fidelity over surface-form overlap, which is critical for endangered languages where mistranslation carries high risk.

KV-cache compression reduces LLM inference memory by evicting context tokens, but when the evicted tokens contain answer-bearing evidence, the model may hallucinate instead of recognizing that the compressed context is insufficient. We address this failure from a behavioral perspective: to our knowledge, this is the first work to formulate compression-aware abstention as a learning problem, in which a model learns to answer when supporting evidence survives compression and abstain when it does not. We construct supervision from compressor survival masks and tight answer-bearing spans, labeling examples as Confident when evidence survives and Abstain when it is removed. A 10.1M-parameter LoRA adapter trained on ∼2.6K MuSiQue 2-hop QA examples reduces base-model hallucinations by 97% under prompt-style truncation while preserving correct answering on evidence-retaining examples. Unlike prompt-only abstention baselines, which over-abstain on many answerable high-retention examples, the trained adapter learns a conditional policy. We also evaluate the method under actual compressed-cache decoding, where multi-compressor training yields a 6–22× relative lift over the unaided base on evidence-retaining examples. Controlled-deletion experiments show that the learned behavior is driven by evidence content rather than input length alone.

We present an end-to-end open-weight pipeline for LLM-Assisted Rule-Based Machine Translation (LLM-RBMT) into extremely low-resource languages, with no proprietary-API dependency at any stage and no per-language linguistic effort beyond registering a language with the framework. On Owens Valley Paiute (OVP), a critically endangered Indigenous language of California, we synthesize training data entirely from the OVP language specification, fine-tune two Qwen2.5-3B LoRA adapters (one for each LLM-driven stage of the translation system), and serve both locally. The fine-tuned forward adapter matches or exceeds gpt-4o-mini on the 150-sentence forward evaluation of Coleman et al. (2026) (COMET-c 0.472 vs. 0.448, same evaluator), and a separately fine-tuned backward adapter approaches it on a 40-item hand-validated backward reference set, both running locally. The full recipe runs on a single consumer GPU.

We describe our submission to all five tracks of the AmericasNLP 2026 Shared Task on Cultural Image Captioning: Bribri, Guaraní, Yucatec Maya, Orizaba Nahuatl, and Wixárika. Our system is an LLM-assisted rule-based machine translation (LLM-RBMT) captioner. For each language, a coding agent reads the small development split and open-web linguistic references and writes a complete Pydantic grammar package with a closed vocabulary. At inference time, a vision–language model sees the image and the schema, emits a structured SentenceList under constrained decoding, and a deterministic Python renderer produces the surface string. The model never generates target-language tokens. The same architecture handles all five languages with no fine-tuning, no parallel corpora, and no human edits to the generated packages. On the official test set, the system placed first on human evaluation in Bribri and Orizaba Nahuatl, third on Yucatec Maya, and first on ChrF++ in Yucatec Maya. We suggest that a strength of the approach is that outputs are restricted to simple sentences that are grammatically correct by construction, modulo the correctness of the generated grammar itself.

We propose a new paradigm for machine translation that is particularly useful for no-resource languages (those without any publicly available bilingual or monolingual corpora): LLM-RBMT (LLM-Assisted Rule Based Machine Translation). Using the LLM-RBMT paradigm, we design the first language education/revitalization-oriented machine translator for Owens Valley Paiute (OVP), a critically endangered Indigenous American language for which there is virtually no publicly available data. We present a detailed evaluation of the translator’s components: a rule-based sentence builder, an OVP to English translator, and an English to OVP translator. We also discuss the potential of the paradigm, its limitations, and the many avenues for future research that it opens up.