Despite recent advances in Reasoning Language Models (RLMs), most research focuses solely on English, even though many models are pretrained on multilingual data. In this work, we investigate: Is English the most token-efficient language for reasoning? We evaluate three open-source RLMs: DeepSeek R1, Qwen 2.5, and Qwen 3, across four math datasets and seven typologically diverse languages. We find that reasoning in non-English languages not only reduces token usage, but also preserves accuracy. These gains persist even after translating the reasoning traces into English, suggesting genuine shifts in reasoning behavior rather than surface-level linguistic effects. The extent of improvement, however, depends on the model’s multilingual strength. Our findings motivate a broader view of reasoning in language models, highlighting the potential of multilingual reasoning and the importance of strong multilingual foundations. The code for our work can be found: https://github.com/microsoft/EfficientXLang.
论文检索
输入标题、作者或关键词，从 7,591 篇学术成果中精准定位
Cognitive science offers rich theories of learning and communication, yet these are often difficult to operationalize at scale. We demonstrate how natural language processing can bridge this gap by applying psycholinguistic theories of discourse to real-world educational data. We investigate linguistic alignment – the convergence of conversational partners’ word choice, grammar, and meaning – in a longitudinal dataset of real-world tutoring interactions and associated student test scores. We examine (1) the extent of alignment, (2) role-based patterns among tutors and students, and (3) the relationship between alignment and learning outcomes. We find that both tutors and students exhibit lexical, syntactic, and semantic alignment, with tutors aligning more strongly to students. Crucially, tutor lexical alignment predicts student learning gains, while student lexical alignment negatively predicts them. As a lightweight, interpretable metric, linguistic alignment offers practical applications in intelligent tutoring systems, educator dashboards, and tutor training.
Large language models (LLMs) are increasingly expected to function as collaborative partners, engaging in back-and-forth dialogue to solve complex, ambiguous problems. However, current LLMs often falter in real-world settings, defaulting to passive responses or narrow clarifications when faced with incomplete or under-specified prompts—falling short of proactively gathering the missing information that is crucial for high-quality solutions. In this work, we introduce a new task paradigm: proactive information gathering, where LLMs must identify gaps in the provided context and strategically elicit implicit user knowledge through targeted questions. To systematically study and train this capability, we design a scalable framework that generates partially specified, real-world tasks, masking key information and simulating authentic ambiguity. Within this setup, our core innovation is a reinforcement finetuning strategy rewards questions that elicit genuinely new, implicit user information—such as hidden domain expertise or fine-grained requirements—that would otherwise remain unspoken. Experiments demonstrate that our trained Qwen-2.5-7B model significantly outperforms o3-mini by 18% on automatic evaluation metrics. More importantly, human evaluation reveals that clarification questions and final outlines generated by our model are favored by human annotators by 42% and 28% respectively. Together, these results highlight the value of proactive clarification in elevating LLMs from passive text generators to genuinely collaborative thought partners.
Language models built from various sources are the foundation of today’s NLP progress. However, for many low-resource languages, the diversity of domains is often limited, more biased to a religious domain, which impacts their performance when evaluated on distant and rapidly evolving domains such as social media. Domain adaptive pre-training (DAPT) and task-adaptive pre-training (TAPT) are popular techniques to reduce this bias through continual pre-training for BERT-based models, but they have not been explored for African multilingual encoders. In this paper, we explore DAPT and TAPT continual pre-training approaches for African languages social media domain. We introduce AfriSocial, a large-scale social media and news domain corpus for continual pre-training on several African languages. Leveraging AfriSocial, we show that DAPT consistently improves performance (from 1% to 30% F1 score) on three subjective tasks: sentiment analysis, multi-label emotion, and hate speech classification, covering 19 languages. Similarly, leveraging TAPT on the data from one task enhances performance on other related tasks. For example, training with unlabeled sentiment data (source) for a fine-grained emotion classification task (target) improves the baseline results by an F1 score ranging from 0.55% to 15.11%. Combining these two methods (i.e. DAPT + TAPT) further improves the overall performance. The data and model resources are available at HuggingFace.
Large language models (LLMs) have delivered significant breakthroughs across diverse domains but can still produce unreliable or misleading outputs, posing critical challenges for real-world applications. While many recent studies focus on quantifying model uncertainty, relatively little work has been devoted to diagnosing the source of uncertainty. In this study, we show that, when an LLM is uncertain, the patterns of disagreement among its multiple generated responses contain rich clues about the underlying cause of uncertainty. To illustrate this point, we collect multiple responses from a target LLM and employ an auxiliary LLM to analyze their patterns of disagreement. The auxiliary model is tasked to reason about the likely source of uncertainty, such as whether it stems from ambiguity in the input question, a lack of relevant knowledge, or both. In cases involving knowledge gaps, the auxiliary model also identifies the specific missing facts or concepts contributing to the uncertainty. In our experiment, we validate our framework on AmbigQA, OpenBookQA, and MMLU-Pro, confirming its generality in diagnosing distinct uncertainty sources. Such diagnosis shows the potential for relevant manual interventions that improve LLM performance and reliability.
Instruction fine-tuning enhances the alignment of autoregressive language models (ArLMs) with human intent but relies on large-scale annotated datasets prone to label and text noise. In this paper, we show that existing noise detection techniques designed for autoencoder models (AeLMs) do not directly generalize to ArLMs due to differences in learning dynamics. We propose TDRanker, a novel approach leveraging training dynamics to rank datapoints from easy-to-learn to hard-to-learn, effectively identifying noisy instances. Our method demonstrates robustness across multiple model architectures covering both autoencoder and autoregressive language models (GPT-2, BERT, LaMini-Cerebras-256M) and across various dataset noise levels, achieving at least 2x faster denoising than previous techniques. Applied to real-world classification and generative tasks, TDRanker significantly improves data quality and model performance. These findings suggest that TDRanker provides a scalable solution for refining instruction-tuning datasets, enhancing the reliability of fine-tuned ArLMs in practical applications.
Understanding and solving complex reasoning tasks is vital for addressing the information needs of a user. Although dense neural models learn contextualised embeddings, they underperform on queries containing negation. To understand this phenomenon, we study negation in traditional neural information retrieval and LLM-based models. We (1) introduce a taxonomy of negation that derives from philosophical, linguistic, and logical definitions; (2) generate two benchmark datasets that can be used to evaluate the performance of neural information retrieval models and to fine-tune models for a more robust performance on negation; and (3) propose a logic-based classification mechanism that can be used to analyze the performance of retrieval models on existing datasets. Our taxonomy produces a balanced data distribution over negation types, providing a better training setup that leads to faster convergence on the NevIR dataset. Moreover, we propose a classification schema that reveals the coverage of negation types in existing datasets, offering insights into the factors that might affect the generalization of fine-tuned models on negation. Our code is publicly available on GitHub, and the datasets are available on HuggingFace.
Large language models (LLMs) have recently been proposed as general-purpose agents for experimental design, with claims that they can perform in-context experimental design. We evaluate this hypothesis using open-source instruction-tuned LLMs applied to genetic perturbation and molecular property discovery tasks. We find that LLM-based agents show no sensitivity to experimental feedback: replacing true outcomes with randomly permuted labels has no impact on performance. Across benchmarks, classical methods such as linear bandits and Gaussian process optimization consistently outperform LLM agents. We further propose a simple hybrid method, LLM-guided Nearest Neighbour (LLMNN) sampling, that combines LLM prior knowledge with nearest-neighbor sampling to guide the design of experiments. LLMNN achieves competitive or superior performance across domains without requiring significant in-context adaptation. These results suggest that current open-source LLMs do not perform in-context experimental design in practice and highlight the need for hybrid frameworks that decouple prior-based reasoning from batch acquisition with updated posteriors.
The frequency distribution of words in human-written texts roughly follows a simple mathematical form known as Zipf’s law. Somewhat less well known is the related Heaps’ law, which describes a sublinear power-law growth of vocabulary size with document size. We study the applicability of Zipf’s and Heaps’ laws to texts generated by Large Language Models (LLMs). We empirically show that Heaps’ and Zipf’s laws only hold for LLM-generated texts in a narrow model-dependent temperature range. These temperatures have an optimal value close to t=1 for all the base models except the large Llama models, are higher for instruction-finetuned models and do not depend on the model size or prompting. This independently confirms the recent discovery of sampling temperature dependent phase transitions in LLM-generated texts.
Instruction-tuned large language models (LLMs) employ structured templates, such as role markers and special tokens, to enforce format consistency during inference. However, we identify a critical limitation of such formatting: it induces a phenomenon we term diversity collapse, where the model generates semantically similar outputs for open-ended inputs, undermining creativity and variability. We systematically evaluate this effect across tasks like story completion and free-form generation, finding that (1) diversity collapse persists even under high-temperature sampling, and (2) structural tokens in templates significantly constrain the model’s output space. To contextualize these findings, we fine-tune using a range of structured prompts and then evaluate them across three axes: downstream task performance, alignment behavior, and output diversity. Our analysis shows that format consistency between fine-tuning and inference is crucial for structure-sensitive tasks (e.g., GSM8K, IFEval), but has marginal influence on knowledge-heavy tasks (e.g., MMLU, WebQuestions). In contrast, output diversity is primarily governed by the presence or absence of structural tokens, with minimal formatting yielding the most diverse outputs. These findings reveal that current prompting conventions, while beneficial for alignment, may inadvertently suppress output diversity, underscoring the need for diversity-aware prompt design and instruction tuning.
Fine-grained control is essential for precise and customizable text generation, yet existing benchmarks evaluate models on only a few attributes, typically fewer than five. We introduce UltraBench, a new benchmark for extremely fine-grained controllable generation (EFCG), which evaluates large language models (LLMs) under dense, multi-attribute constraints. Each sample includes approximately 70 attributes, combining LLM-extracted soft constraints (e.g., style and tone) with programmatically enforced hard constraints (e.g., word count). Using UltraBench, we conduct a comprehensive evaluation of state-of-the-art LLMs and prompting strategies. Models such as GPT-4.1 and Qwen3-8B perform well on individual constraints, achieving instruction-level accuracy above 70%, but consistently fail to satisfy all constraints simultaneously. To understand this limitation, we analyze model behavior across different dimensions. First, we observe a clear position bias: models tend to prioritize constraints presented later in the prompt while neglecting those that appear earlier. Second, we find that structural and formatting-related constraints are significantly more difficult to satisfy than content- or style-based ones, suggesting that current models struggle to coordinate global structure with token-level control. Finally, our error analysis reveals distinct failure modes: GPT-4.1 often attempts to follow constraints but falls short in precision, whereas LLaMA frequently omits constraints, particularly in multi-turn settings. These findings highlight fundamental limitations in EFCG and underscore the need for new methods that support dense, instruction-sensitive generation.
Time series Forecasting with large language models (LLMs) requires bridging numerical patterns and natural language. Effective forecasting on LLM often relies on extensive pre-processing and fine-tuning. Recent studies show that a frozen LLM can rival specialized forecasters when supplied with a carefully engineered natural-language prompt, but crafting such a prompt for each task is itself onerous and ad-hoc. We introduce FLAIRR-TS, a test-time prompt optimization framework that utilizes an agentic system: a Forecaster-agent generates forecasts using an initial prompt, which is then refined by a refiner agent, informed by past outputs and retrieved analogs. This adaptive prompting generalizes across domains using creative prompt templates and generates high-quality forecasts without intermediate code generation. Experiments on benchmark datasets show FLAIRR-TS improves forecasting over static prompting and retrieval-augmented baselines, approaching the performance of specialized prompts.FLAIRR-TS provides a practical alternative to fine-tuning, achieving strong performance via its agentic approach to adaptive prompt refinement and retrieval.
Context Minimization for Resource-Constrained Text Classification: Optimizing Performance-Efficiency Trade-offs through Linguistic Features
PDF ↗Pretrained language models have transformed text classification, yet their computational demands often render them impractical for resource-constrained settings. We propose a linguistically-grounded framework for context minimization that leverages theme-rheme structure to preserve critical classification signals while reducing input complexity. Our approach integrates positional, syntactic, semantic, and statistical features, guided by functional linguistics, to identify optimal low-context configurations. We present a methodical iterative feature exploration protocol across 6 benchmarks, including our novel CMLA11 dataset. Results demonstrate substantial efficiency gains: 69-75% reduction in GPU memory, 81-87% decrease in training time, and 82-88% faster inference. Despite these resource savings, our configurations maintain near-parity with full-length inputs, with F1 (macro) reductions averaging just 1.39-3.10%. Statistical significance testing confirms minimal practical impact, with some configurations outperforming the baseline. SHAP analysis reveals specific feature subsets contribute most significantly across datasets, and these recurring configurations offer transferable insights, reducing the need for exhaustive feature exploration. Our method also yields remarkable data compression (72.57% average reduction, reaching 92.63% for longer documents). Ablation studies confirm synergistic feature contributions, establishing our context minimization as an effective solution for resource-efficient text classification with minimal performance trade-offs.
Legal research depends on headnotes: concise summaries that help lawyers quickly identify relevant cases. Yet, many court decisions lack them due to the high cost of manual annotation. To address this gap, we introduce the Swiss Landmark Decisions Summarization (SLDS) dataset containing 20K rulings from the Swiss Federal Supreme Court, each with headnotes in German, French, and Italian. SLDS has the potential to significantly improve access to legal information and transform legal research in Switzerland. We fine-tune open models (Qwen2.5, Llama 3.2, Phi-3.5) and compare them to larger general-purpose and reasoning-tuned LLMs, including GPT-4o, Claude 3.5 Sonnet, and the open-source DeepSeek R1. Using an LLM-as-a-Judge framework, we find that fine-tuned models perform well in terms of lexical similarity, while larger models generate more legally accurate and coherent summaries. Interestingly, reasoning-focused models show no consistent benefit, suggesting that factual precision is more important than deep reasoning in this task. We release SLDS under a CC BY 4.0 license to support future research in cross-lingual legal summarization.
The growing emotional stress in modern society has increased the demand for Emotional Support Conversations (ESC). While Large Language Models (LLMs) show promise for ESC, they face two key challenges: (1) low strategy selection accuracy, and (2) preference bias, limiting their adaptability to users’ emotional needs. Existing supervised fine-tuning (SFT) struggles to address these issues, as it rigidly trains models on single gold-standard responses without modeling nuanced strategy trade-offs. To overcome these limitations, we propose a novel two-stage framework that optimizes strategy selection preferences at each dialogue turn. We first leverage Monte Carlo Tree Search to construct ESC-Pro, a high-quality preference dataset with turn-level strategy-response pairs. Then training on ESC-Pro with Chain-of-Strategy Optimization (CSO) improves both strategy accuracy and bias mitigation, enabling LLMs to generate more empathetic and contextually appropriate responses. Experiments on LLaMA-3.1-8B, Gemma-2-9B, and Qwen2.5-7B demonstrate that CSO outperforms standard SFT, highlighting the efficacy of fine-grained, turn-level preference modeling in ESC.
ConText-LE: Cross-Distribution Generalization for Longitudinal Experiential Data via Narrative-Based LLM Representations
PDF ↗Longitudinal experiential data offers rich insights into dynamic human states, yet building models that generalize across diverse contexts remains challenging. We propose ConText-LE, a framework that systematically investigates text representation strategies and output formulations to maximize large language model cross-distribution generalization for behavioral forecasting. Our novel Meta-Narrative representation synthesizes complex temporal patterns into semantically rich narratives, while Prospective Narrative Generation reframes prediction as a generative task aligned with LLMs’ contextual understanding capabilities. Through comprehensive experiments on three diverse longitudinal datasets addressing the underexplored challenge of cross-distribution generalization in mental health and educational forecasting, we show that combining Meta-Narrative input with Prospective Narrative Generation significantly outperforms existing approaches. Our method achieves up to 12.28% improvement in out-of-distribution accuracy and up to 11.99% improvement in F1 scores over binary classification methods. Bidirectional evaluation and architectural ablation studies confirm the robustness of our approach, establishing ConText-LE as an effective framework for reliable behavioral forecasting across temporal and contextual shifts.
Large Language Models (LLMs) have shown extraordinary success across various text generation tasks; however, their potential for simple yet essential text classification remains underexplored, as LLM pre-training tends to emphasize generation over classification. While LLMs with instruction tuning can transform classification into a generation task, they often struggle to categorize nuanced texts. One such example is text revision, which involves nuanced edits between pairs of texts. Although simply fine-tuning LLMs for revision classification seems plausible, it requires a large amount of revision annotations, which are exceptionally expensive and scarce in the community. To address this issue, we introduce a plug-and-play layer-wise parameter-efficient fine-tuning (PEFT) framework, i.e., IR-Tuning, which fine-tunes a subset of important LLM layers that are dynamically selected based on their gradient norm distribution, while freezing those of redundant layers. Extensive experiments suggest that IR-Tuning surpasses several layer-wise PEFT baselines over diverse text revisions, while achieving fast convergence, low GPU memory consumption, and effectiveness on small revision corpora.
Large language models are increasingly used for teaching and self-learning foreign languages. However, their capability to meet specific linguistic constraints is still underexplored. This study compares the effectiveness of prompt engineering in guiding ChatGPT (4o and 4o-mini), and Llama 3 to rephrase general-domain texts to meet CEFR A1-level constraints in English and Italian, making them suitable for beginner learners. It compares 4 prompt engineering approaches, built upon iterative paraphrasing method that gradually refines original texts for CEFR compliance. The approaches compared include paraphrasing with or without Chain-of-Thought, as well as grammar and vocabulary simplification performed either simultaneously or as separate steps. The findings suggest that for English the best approach is combining COT with separate grammar and vocabulary simplification, while for Italian one-step strategies have better effect on grammar, and two-step strategies work better for covering the vocabulary. The paraphrasing approach can approve compliance, although at this point it is not cost-effective. We release a dataset of pairs original sentence-beginner level paraphrase (both in Italian and in English) on which further work could be based.
Annotated datasets for supervised learning tasks often contain incorrect gold annotations, i.e. label noise. To address this issue, many noisy label learning approaches incorporate metrics to filter out unreliable samples, for example using heuristics such as high loss or low confidence. However, when these metrics are integrated into larger pipelines, it becomes difficult to compare their effectiveness, and understand their individual contribution to reducing label noise. This paper directly compares popular sample metrics for detecting incorrect annotations in named entity recognition (NER). NER is commonly approached as token classification, so the metrics are calculated for each training token and we flag the incorrect ones by defining metrics thresholds. We compare the metrics based on (i) their accuracy in detecting the incorrect labels and (ii) the test scores when retraining a model using the cleaned dataset. We show that training dynamics metrics work the best overall. The best metrics effectively reduce the label noise across different noise types. The errors that the model has not yet memorized are more feasible to detect, and relabeling these tokens is a more effective strategy than excluding them from training.
Users of Augmentative and Alternative Communication (AAC) may write letter-by-letter via an interface that uses a character language model. However, most state-of-the-art large pretrained language models predict subword tokens of variable length. We investigate how to practically use such models to make accurate and efficient character predictions. Our algorithm for producing character predictions from a subword large language model (LLM) provides more accurate predictions than using a classification layer, a byte-level LLM, or an n-gram model. Additionally, we investigate a domain adaptation procedure based on a large dataset of sentences we curated based on scoring how useful each sentence might be for spoken or written AAC communication. We find our procedure further improves model performance on simple, conversational text.