Recent progress in universal multilingual named entity recognition (NER) has been driven by multilingual transformer models, task-specific architectures, custom loss functions, and large-scale training datasets. However, despite substantial prior work, we find that many design decisions for such models are made without systematic justification, with individual components evaluated only in combination rather than in isolation. We argue that this impedes progress in the field by making it difficult to identify which choices improve multilingual generalization. In this work, we conduct an extensive empirical evaluation on transformer backbones, architectures, training objectives, data composition, and threshold selection for zero-shot, multilingual NER models. Building on these findings, we present Otter, a small encoder model that achieves consistent improvements over strong multilingual NER baselines, outperforming similarly sized models by 5.7 percentage points in F1. Further, it remains within 1.4 points of Qwen3-32B and 5.4 points below Gemma3-27B, while being roughly 90x smaller and one to two orders of magnitude faster at inference.

Recent multilingual named entity recognition (NER) work has shown that large language models (LLMs) can provide effective synthetic supervision, yet such datasets have mostly appeared as by-products of broader experiments rather than as systematic, reusable resources. We introduce , a dataset-creation pipeline that scales the teacher-student paradigm to 91 languages and 25 scripts. Building on FineWeb-Edu, our approach trains regression models to identify NER-relevant passages and annotates them with multilingual LLMs, resulting in about 225k passages with 235k distinct entity labels. Our experiments show that the regression model achieves more than 84 F1, and that models trained on FiNERweb obtain comparable or improved performance in zero shot transfer settings on English, Thai, and Swahili, despite being trained on 19x less data than strong baselines. In addition, we assess annotation quality using LLM-as-a-judge and observe consistently high scores for both faithfulness (3.99/5) and completeness (4.05/5), indicating reliable and informative annotations. Further, we release the dataset with both English labels and translated label sets in the respective target languages because we observe that the performance of current state-of-the-art models drops by 0.02-0.09 F1 when evaluated using target language labels instead of English ones. We release FiNERweb together with all accompanying artifacts to the research community in order to facilitate more effective student-teacher training for multilingual named entity recognition.

Linearization has emerged as a strategy for developing efficient language models (LMs). Starting from an existing Transformer-based LM, linearization replaces the attention component with computationally efficient subquadratic token mixers. However, as an increasing number of mixers are proposed, it remains unclear which inductive biases are best suited to inherit the original Transformer’s capabilities. Furthermore, it is unknown how linearization is affected by parameter and token budget scaling. To address these questions, we propose a unified setup to compare seven representative architectures, including xLSTM, GLA, and Gated DeltaNet. Our findings reveal that performance hierarchies remain stable from 140M to 1.7B parameters, with error-correcting update rules demonstrating superior scaling exponents. We show that performance gaps are established early and persist through asymptotic maturity at 10B tokens, suggesting that state resolution is a more fundamental bottleneck than the distillation budget. Finally, while most models adapt to instruction tuning, only gated delta-rule formulations maintain the precision necessary for long-context retrieval, whereas additive models suffer from irreversible state saturation. Finally, we show that a cheap mixer-only probe predicts these downstream outcomes before a full run and, used as a diagnostic, attributes the weakest student’s scaling deficit to a fixable normalization choice rather than its update rule. These results suggest that for successful linearization, architectural inductive biases remain the primary constraint that cannot be overcome by simply scaling training compute.

Recent research has shown that filtering massive English web corpora into high-quality subsets significantly improves training efficiency. However, for high-resource non-English languages like German, French, or Japanese, aggressive filtering creates a strategic dilemma: should practitioners prioritize *diversity* by training once on large amounts of lightly filtered web data, or prioritize *quality* by strictly filtering for a high-quality core and repeating it over multiple epochs?

We investigate this trade-off for German by applying hierarchical quality filters to 500M web documents, comparing multi-epoch training on filtered subsets against single-pass training on a diverse corpus. Our experiments across multiple model scales and token budgets show that repeating high-quality data consistently outperforms single-pass training on larger, less filtered sets. Notably, the performance gap persists even after 7 epochs. Our findings suggest that for non-English LLMs, semantic concentration through quality filtering offers a more viable path to efficient language modeling than simply maximizing unique data volume.

We release our German language models (called **Boldt**), as well as our cleaned evaluation benchmarks to the research community. Our experiments indicate that they achieve state-of-the-art results despite training on 10-360x fewer tokens than comparable models.

BabyLM is a pressure test for language-modeling assumptions: which ideas still work when the data budget is closer to a child’s linguistic input than to web-scale pretraining? We use the 2026 English tracks to ask this question in a controlled way, holding the corpus, tokenizer, backbone size family, training schedule, and public evaluation wrapper fixed while varying two token mixers and six objectives. Across 432 completed scale-grid runs on the 10M-word Strict-Small and 100M-word Strict tracks, the strongest result is also the most conservative one. Autoregressive Transformers are the best public-evaluation models at every tested scale, reaching peak fine-tuning macro scores of 64.88 and 66.94 and peak BLiMP scores of 64.70 and 75.13. Masked language modeling is the strongest non-autoregressive baseline. The negative results are informative: Gated DeltaNet improves Entity Tracking but loses grammar and broad transfer, while diffusion and flow objectives show local advantages in probes that rarely survive the official scoring interface. The lesson for BabyLM is not that newer objectives or recurrent mixers are unhelpful, but that small-data gains depend on matching the objective, architecture, sampler, and scorer.

In the small-data regime of the BabyLM Challenge, a model’s architecture still shapes what it learns. We ask how much an inductive bias toward recurrence helps, and whether current benchmarks can detect that help. Our model adapts the Hierarchical Reasoning Model, a small recurrent puzzle-solver that reuses one shared block many times per token instead of stacking more layers. We make three changes to turn it into a language model: a sequence-mixing layer that can carry order-dependent state, a gate on one of the two connections between its modules, and a tunable number of recurrent cycles. The result matches a parameter-matched GPT-2, and we use it to study three questions. Q1: recurrence buys depth without adding parameters. On a controlled state-tracking task, running more cycles of the same weights tracks longer sequences, even when the extra cycles are added only at inference. Q2: the advantage is largest where data is scarcest. leads GPT-2 on BLiMP at words, ties at , and, trained to convergence, still leads at tokens (out of budget, where data and tokenizer also change). Q3: the benchmark meant to reward state-tracking cannot see it. A null model that never reads the puzzle outscores every trained model on entity_tracking, because 28.5% of its items reveal the answer in the option list and every trained model is at chance on the rest. Gains on that task measure how well a model has learned the frequency of one answer. Because the suite misses a capability we can demonstrate directly, we recommend pairing benchmarks with controlled architectural probes.

Zero-shot named entity recognition (NER) is the task of detecting named entities of specific types (such as Person or Medicine) without any training examples. Current research increasingly relies on large synthetic datasets, automatically generated to cover tens of thousands of distinct entity types, to train zero-shot NER models. However, in this paper, we find that these synthetic datasets often contain entity types that are semantically highly similar to (or even the same as) those in standard evaluation benchmarks. Because of this overlap, we argue that reported F1 scores for zero-shot NER overestimate the true capabilities of these approaches. Further, we argue that current evaluation setups provide an incomplete picture of zero-shot abilities since they do not quantify the label shift (i.e., the similarity of labels) between training and evaluation datasets. To address these issues, we propose Familarity, a novel metric that captures both the semantic similarity between entity types in training and evaluation, as well as their frequency in the training data, to provide an estimate of label shift. It allows researchers to contextualize reported zero-shot NER scores when using custom synthetic training datasets. Further, it enables researchers to generate evaluation setups of various transfer difficulties for fine-grained analysis of zero-shot NER.

Sample efficiency is a crucial property of language models with practical implications for training efficiency. In real-world text, information follows a long-tailed distribution. Yet, we expect models to learn and recall frequent and infrequent facts. Sample efficient models are better equipped to handle this challenge of learning and retaining rare information without requiring excessive exposure. This study analyzes multiple models of varying architectures and sizes, all trained on the same pre-training data. By annotating relational facts with their frequencies in the training corpus, we examine how model performance varies with fact frequency. Our findings show that most models perform similarly on high-frequency facts but differ notably on low-frequency facts. This analysis provides new insights into the relationship between model architecture, size, and factual learning efficiency.

We study architectural and optimization techniques for sample-efficient language modeling under the constraints of the BabyLM 2025 shared task. Our model, BLaLM, replaces self-attention with a linear-time mLSTM token mixer and explores lightweight enhancements, including short convolutions, sliding window attention with dynamic modulation, and Hedgehog feature maps. To support training in low-resource settings, we curate a high-quality corpus emphasizing readability and pedagogical structure. Experiments across both strict and strict-small tracks show that (1) linear attention combined with sliding window attention consistently improves zero-shot performance, and (2) the Muon optimizer stabilizes convergence and reduces perplexity over AdamW. These results highlight effective strategies for efficient language modeling without relying on scale.

Instruction-tuned Large Language Models (LLMs) have recently showcased remarkable ability to generate fitting responses to natural language instructions. However, an open research question concerns the inherent biases of trained models and their responses. For instance, if the data used to tune an LLM is dominantly written by persons with a specific political bias, we might expect generated answers to share this bias. Current research work seeks to de-bias such models, or suppress potentially biased answers.With this demonstration, we take a different view on biases in instruction-tuning: Rather than aiming to suppress them, we aim to make them explicit and transparent. To this end, we present OpinionGPT, a web demo in which users can ask questions and select all biases they wish to investigate. The demo will answer this question using a model fine-tuned on text representing each of the selected biases, allowing side-by-side comparison. To train the underlying model, we identified 11 different biases (political, geographic, gender, age) and derived an instruction-tuning corpus in which each answer was written by members of one of these demographics. This paper presents OpinionGPT, illustrates how we trained the bias-aware model and showcases the web application (available at https://opiniongpt.informatik.hu-berlin.de).

Recent advancements in large language models (LLMs) have showcased their exceptional abilities across various tasks, such as code generation, problem-solving and reasoning. Existing benchmarks evaluate tasks in isolation, yet the extent to which LLMs can understand prose-style tasks, identify the underlying problems, and then generate appropriate code solutions is still unexplored. Addressing this gap, we introduce PECC, a novel benchmark derived from Advent Of Code (AoC) challenges and Project Euler, including 2396 problems. Unlike conventional benchmarks, PECC requires LLMs to interpret narrative-embedded problems, extract requirements, and generate executable code. A key feature of our dataset is the complexity added by natural language prompting in chat-based evaluations, mirroring real-world instruction ambiguities. Results show varying model performance between narrative and neutral problems, with specific challenges in the Euler math-based subset with GPT-3.5-Turbo passing 50% of the AoC challenges and only 8% on the Euler problems. By probing the limits of LLMs’ capabilities, our benchmark provides a framework to monitor and assess the subsequent progress of LLMs as a universal problem solver.

This paper explores the potential of recurrent neural networks (RNNs) and other subquadratic architectures as competitive alternatives to transformer-based models in low-resource language modeling scenarios. We utilize HGRN2 (Qin et al., 2024), a recently proposed RNN-based architecture, and comparatively evaluate its effectiveness against transformer-based baselines and other subquadratic architectures (LSTM, xLSTM, Mamba). Our experimental results show that, our HGRN2 language model, outperforms transformer-based models in both the 10M and 100M word tracks of the challenge, as measured by their performance on the BLiMP, EWoK, GLUE and BEAR benchmarks. Further, we show the positive impact of knowledge distillation. Our findings challenge the prevailing focus on transformer architectures and indicate the viability of RNN-based models, particularly in resource-constrained environments.

Most NLP tasks are modeled as supervised learning and thus require labeled training data to train effective models. However, manually producing such data at sufficient quality and quantity is known to be costly and time-intensive. Current research addresses this bottleneck by exploring a novel paradigm called zero-shot learning via dataset generation. Here, a powerful LLM is prompted with a task description to generate labeled data that can be used to train a downstream NLP model. For instance, an LLM might be prompted to “generate 500 movie reviews with positive overall sentiment, and another 500 with negative sentiment.” The generated data could then be used to train a binary sentiment classifier, effectively leveraging an LLM as a teacher to a smaller student model. With this demo, we introduce Fabricator, an open-source Python toolkit for dataset generation. Fabricator implements common dataset generation workflows, supports a wide range of downstream NLP tasks (such as text classification, question answering, and entity recognition), and is integrated with well-known libraries to facilitate quick experimentation. With Fabricator, we aim to support researchers in conducting reproducible dataset generation experiments using LLMs and help practitioners apply this approach to train models for downstream tasks.