This paper describes a submission to the constrained English-to-Simplified-Chinese track of the WMT26 Model Compression shared task. Starting from Gemma~3 12B, the system applies native BitsAndBytes NF4 with double quantization to the 144 gate, up, and down projections in the model’s 48 text-transformer MLP blocks. Attention projections, embeddings, normalization layers, the tied language-model head, and vision modules remain in BF16. The method requires no calibration or fine-tuning data and produces a mixed checkpoint whose quantized weights remain low-bit at inference. On the official 945-segment blind set, the system scores 0.7669 CometKiwi-XXL and 2.7836 MetricX-24-XXL, slightly improving on the organizer global-q4 baseline’s 0.7665 and 2.8147, respectively. Its 11.8~GB artifact is 48.4\% of BF16 size, and H100 throughput is essentially tied with global q4 (866.9 versus 867.9 source characters/s). On 332 reference-backed WMT25 segments, its chrF and BLEU gains over global q4 are significant under paired bootstrap resampling. A transferred selected-layer q3+Zstd codec reaches statistically indistinguishable local quality but expands to BF16-like memory after decoding.

This paper describes a selected-layer codec compression approach submitted to the IWSLT 2026 Model Compression Shared Task for constrained English-to-Chinese speech translation. The approach is compared against standard quantization, global codec compression, and a pruning-plus-codec variant. The results indicate that translation quality after compression depends strongly on where compression is applied. In these experiments, selected-layer compression preserves translation quality better than uniform global compression, with one variant achieving the highest COMET score among compressed systems and another providing the strongest overall quality-compression trade-off among the custom codec methods. These results suggest that simple layer-aware post-hoc compression is a viable approach for model compression in constrained English-to-Chinese speech translation.

Pretraining optimizes a language model with self-supervised objectives over raw text, shaping the representations later measured by diagnostic evaluations. Masked language modeling (MLM) predicts hidden tokens from context, combining the distributional hypothesis with the cloze task. Less is known about alternative objectives under small data budgets. Using BabyLM 2026’s 10M-token Strict-Small setting, this paper asks whether auxiliary objectives jointly improve grammar, world-knowledge plausibility, entity tracking, and reading prediction, or instead benefit some areas at the expense of others. The experiments compare an MLM-only baseline with several multi-task objective mixtures while holding the corpus, architecture, tokenizer, training budget, and evaluation pipeline fixed. Results are selective: MLM-only is strongest on grammatical minimal-pair judgments, while auxiliary objectives improve entity-state tracking and reading prediction. Within the controlled fast evaluation, single-mask semantic cloze gives the clearest directional world-knowledge plausibility gain; grammar repair does not recover grammatical performance, and broad plausibility classification produces no positive EWoK delta. The central finding is objective-interface alignment: auxiliary objectives transfer most clearly when their loss format resembles the diagnostic interface used at evaluation time. Auxiliary objectives therefore act as targeted steering signals rather than uniform improvements over MLM.

Predicting the difficulty of exam questions or items is essential to effectively assembling and calibrating exams. While item response theory (IRT) models can estimate item difficulty, they require student responses that are costly and rarely available at scale. Natural language processing methods offer a text-only alternative; however, due to the scarcity of real-world labeled data, prior work often relies on synthetic or domain-specific corpora, limiting generalizability and overlooking the nuanced challenges of real-world text-based item difficulty estimation. Addressing this gap, we benchmark 122 classifiers on 935 German Vocational Education and Training (VET) items labeled via previous IRT analysis to assess feasibility under real-world conditions. In our setup, a stacked ensemble that combines linguistic features, pre-trained embeddings, and external semantic resources outperforms both transformer-based models and few-shot large language models, achieving moderate performance. We report findings and discuss limitations in the context of German VET.

In education, high-quality exams must cover broad specifications across diverse difficulty levels during the assembly and calibration of test items to effectively measure examinees’ competence. However, balancing the trade-off of selecting relevant test items while fulfilling exam specifications without bias is challenging, particularly when manual item selection and exam assembly rely on a pre-validated item base. To address this limitation, we propose a new mixed-integer programming re-ranking approach to improve relevance, while mitigating bias on an industry-grade exam assembly platform. We evaluate our approach by comparing it against nine bias mitigation re-ranking methods in 225 experiments on a real-world benchmark data set from vocational education services. Experimental results demonstrate a 17% relevance improvement with a 9% bias reduction when integrating sequential optimization techniques with improved contextual relevance augmentation and scoring using a large language model. Our approach bridges information retrieval and exam assembly, enhancing the human-in-the-loop exam assembly process while promoting unbiased exam design

A major bottleneck in exam construction involves designing test items (i.e., questions) that accurately reflect key content from domain-aligned curricular materials. For instance, during formative assessments in vocational education and training (VET), exam designers must generate updated test items that assess student learning progress while covering the full breadth of topics in the curriculum. Large language models (LLMs) can partially support this process, but effective use requires careful prompting and task-specific understanding. We propose a new key point extraction method for retrieval-augmented item generation that enhances the process of generating test items with LLMs. We exhaustively evaluated our method using a TREC-RAG approach, finding that prompting LLMs with key content rather than directly using full curricular text passages significantly improves item quality regarding key information coverage by 8%. To demonstrate these findings, we release EdTec-ItemGen, a retrieval-augmented item generation demo tool to support item generation in education.

Selecting and assembling test items from a validated item database into comprehensive exam forms is an under-researched but significant challenge in education. Search and retrieval methods provide a robust framework to assist educators when filtering and assembling relevant test items. In this work, we present EdTec-QBuilder, a semantic search tool developed to assist vocational educators in assembling exam forms. To implement EdTec-QBuilder’s core search functionality, we evaluated eight retrieval strategies and twenty-five popular pre-trained sentence similarity models. Our evaluation revealed that employing cross-encoders to re-rank an initial list of relevant items is best for assisting vocational trainers in assembling examination forms. Beyond topic-based exam assembly, EdTec-QBuilder aims to provide a crowdsourcing infrastructure enabling manual exam assembly data collection, which is critical for future research and development in assisted and automatic exam assembly models.

Most research on natural language processing treats bias as an absolute concept: Based on a (probably complex) algorithmic analysis, a sentence, an article, or a text is classified as biased or not. Given the fact that for humans the question of whether a text is biased can be difficult to answer or is answered contradictory, we ask whether an “absolute bias classification” is a promising goal at all. We see the problem not in the complexity of interpreting language phenomena but in the diversity of sociocultural backgrounds of the readers, which cannot be handled uniformly: To decide whether a text has crossed the proverbial line between non-biased and biased is subjective. By asking “Is text X more [less, equally] biased than text Y?” we propose to analyze a simpler problem, which, by its construction, is rather independent of standpoints, views, or sociocultural aspects. In such a model, bias becomes a preference relation that induces a partial ordering from least biased to most biased texts without requiring a decision on where to draw the line. A prerequisite for this kind of bias model is the ability of humans to perceive relative bias differences in the first place. In our research, we selected a specific type of bias in argumentation, the stance bias, and designed a crowdsourcing study showing that differences in stance bias are perceptible when (light) support is provided through training or visual aid.