Here is today's selection of cs.CL papers exploring the intricacies of language model performance and evaluation. The papers delve into various challenges, such as handling terminology rarity in translation, evaluating dialogue with differing metrics, and probing linguistic structures across multilingual models.

Terminology Rarity Predicts Catastrophic Failure in LLM Translation of Low-Resource Ancient Languages: Evidence from Ancient Greek

This study presents the first systematic, reference-free human evaluation of large language model (LLM) machine translation (MT) for Ancient Greek (AG) technical prose. We evaluate translations by three commercial LLMs (Claude, Gemini, ChatGPT) of twenty paragraph-length passages from two works by the Greek physician Galen of Pergamum (ca. 129-216 CE): On Mixtures, which has two published English translations, and On the Composition of Drugs according to Kinds, which has never been fully translated into English. We assess translation quality using both standard automated evaluation metrics (BLEU, chrF++, METEOR, ROUGE-L, BERTScore, COMET, BLEURT) and expert human evaluation via a modified Multidimensional Quality Metrics (MQM) framework applied to all 60 translations by a team of domain specialists. On the previously translated expository text, LLMs achieved high translation quality (mean MQM score 95.2/100), with performance approaching expert level. On the untranslated pharmacological text, aggregate quality was lower (79.9/100) but with high variance driven by two passages presenting extreme terminological density; excluding these, scores converged to within 4 points of the translated text. Terminology rarity, operationalized via corpus frequency in the literary Diorisis Ancient Greek Corpus, emerged as a strong predictor of translation failure (r = -.97 for passage-level quality on the untranslated text). Automated metrics showed moderate correlation with human judgment overall on the text with a wide quality spread (Composition), but no metric discriminated among high-quality translations. We discuss implications for the use of LLMs in Classical scholarship and for the design of automated evaluation pipelines for low-resource ancient languages.

When Metrics Disagree: Automatic Similarity vs. LLM-as-a-Judge for Clinical Dialogue Evaluation

This paper details the baseline model selection, fine-tuning process, evaluation methods, and the implications of deploying more accurate LLMs in healthcare settings. As large language models (LLMs) are increasingly employed to address diverse problems, including medical queries, concerns about their reliability have surfaced. A recent study by Long Island University highlighted that LLMs often perform poorly in medical contexts, potentially leading to harmful misguidance for users. To address this, our research focuses on fine-tuning the Llama 2 7B, a transformer-based, decoder-only model, using transcripts from real patient-doctor interactions. Our objective was to enhance the model's accuracy and precision in responding to medical queries. We fine-tuned the model using a supervised approach, emphasizing domain-specific nuances captured in the training data. In the best scenario, the model results should be reviewed and evaluated by real medical experts. Due to resource constraints, the performance of the fine-tuned model was evaluated using text similarity metrics. The fine-tuned model demonstrated significant improvements across all key dimensions except GPT-4's evaluation. The evaluations of ChatGPT4 are quite different from the quantitative results; here, we not only suggest, but also propose that the result should be evaluated by human medical experts.

CLFEC: A New Task for Unified Linguistic and Factual Error Correction in paragraph-level Chinese Professional Writing

Chinese text correction has traditionally focused on spelling and grammar, while factual error correction is usually treated separately. However, in paragraph-level Chinese professional writing, linguistic (word/grammar/punctuation) and factual errors frequently co-occur and interact, making unified correction both necessary and challenging. This paper introduces CLFEC (Chinese Linguistic & Factual Error Correction), a new task for joint linguistic and factual correction. We construct a mixed, multi-domain Chinese professional writing dataset spanning current affairs, finance, law, and medicine. We then conduct a systematic study of LLM-based correction paradigms, from prompting to retrieval-augmented generation (RAG) and agentic workflows. The analysis reveals practical challenges, including limited generalization of specialized correction models, the need for evidence grounding for factual repair, the difficulty of mixed-error paragraphs, and over-correction on clean inputs. Results further show that handling linguistic and factual Error within the same context outperform decoupled processes, and that agentic workflows can be effective with suitable backbone models. Overall, our dataset and empirical findings provide guidance for building reliable, fully automatic proofreading systems in industrial settings.

Universal Conceptual Structure in Neural Translation: Probing NLLB-200's Multilingual Geometry

Do neural machine translation models learn language-universal conceptual representations, or do they merely cluster languages by surface similarity? We investigate this question by probing the representation geometry of Meta's NLLB-200, a 200-language encoder-decoder Transformer, through six experiments that bridge NLP interpretability with cognitive science theories of multilingual lexical organization. Using the Swadesh core vocabulary list embedded across 135 languages, we find that the model's embedding distances significantly correlate with phylogenetic distances from the Automated Similarity Judgment Program ($ρ= 0.13$, $p = 0.020$), demonstrating that NLLB-200 has implicitly learned the genealogical structure of human languages. We show that frequently colexified concept pairs from the CLICS database exhibit significantly higher embedding similarity than non-colexified pairs ($U = 42656$, $p = 1.33 \times 10^{-11}$, $d = 0.96$), indicating that the model has internalized universal conceptual associations. Per-language mean-centering of embeddings improves the between-concept to within-concept distance ratio by a factor of 1.19, providing geometric evidence for a language-neutral conceptual store analogous to the anterior temporal lobe hub identified in bilingual neuroimaging. Semantic offset vectors between fundamental concept pairs (e.g., man to woman, big to small) show high cross-lingual consistency (mean cosine = 0.84), suggesting that second-order relational structure is preserved across typologically diverse languages. We release InterpretCognates, an open-source interactive toolkit for exploring these phenomena, alongside a fully reproducible analysis pipeline.

GLUScope: A Tool for Analyzing GLU Neurons in Transformer Language Models

We present GLUScope, an open-source tool for analyzing neurons in Transformer-based language models, intended for interpretability researchers. We focus on more recent models than previous tools do; specifically we consider gated activation functions such as SwiGLU. This introduces a new challenge: understanding positive activations is not enough. Instead, both the gate and the in activation of a neuron can be positive or negative, leading to four different possible sign combinations that in some cases have quite different functionalities. Accordingly, for any neuron, our tool shows text examples for each of the four sign combinations, and indicates how often each combination occurs. We describe examples of how our tool can lead to novel insights. A demo is available at https: //sjgerstner.github.io/gluscope.