arxiv arXiv cs.CL · 8h ago

Tatoxa: A Novel Text Detoxification System for Low-Resource Tatar

The paper introduces Tatoxa, a state-of-the-art system designed for automated text detoxification in the low-resource language of Tatar. This work addresses the lack of research attention given to abusive content mitigation in languages with limited digital resources. The authors present a new dataset specifically created for fine-tuning and evaluating detoxification models within these constrained settings. Comparative experiments demonstrate that Tatoxa outperforms both existing open-source and proprietary commercial large language models on key quality metrics. Furthermore, the study investigates cross-lingual transfer capabilities to assess the viability of using data from other languages. Results indicate that training on native Tatar data is significantly more effective than transferring knowledge from culturally close languages like Russian. Even when a large Russian corpus is available, cross-lingual approaches perform worse than models trained exclusively on native Tatar text.

arxiv arXiv cs.CL · 8h ago

Detect, Unlearn, Restore: Defending Text Summarization Models Against Data Poisoning

The study addresses the threat of training-time data poisoning during fine-tuning for abstractive text summarization models. Adversaries manipulate small task-specific datasets to induce persistent summarization failures while maintaining standard evaluation metrics. A unified post-hoc defense framework is proposed to detect and remediate poisoning across the machine learning supply chain. In white-box settings, detection relies on influence-function analysis identifying abnormally high training influence in poisoned pairs. Black-box defenses utilize behavioral auditing based on increased sensitivity to semantics-preserving perturbations. The authors introduce novel attacks targeting factual distortion and representational bias that evade conventional alarms. Experiments across nine architectures and six benchmarks show 85-92% detection precision for the proposed defenses. Gradient-ascent unlearning restores up to 96% of original behavior with less than 0.6% ROUGE degradation.

arxiv arXiv cs.CL · 9h ago

Natural Ungrokking: Asymmetric Control of Which Rules Survive Pretraining

A study identifies 'natural ungrokking,' a phenomenon where small language models lose learned grammatical rules midway through pretraining despite the evidence remaining in the data. Researchers observed that a model learning pronoun-gender agreement with Sue collapsed from 0.94 accuracy to near zero by step 3,500 without any corresponding spike in the loss curve. The survival of these rules is determined by support frequency within the training stream, while the data-to-parameter ratio only modulates the depth of the collapse. This emergence-then-collapse dynamic was replicated across multiple corpora, budgets, and seeds, and confirmed in public Pythia checkpoints where collapse depth correlated with model scale. The forgetting process acts as a displacement mechanism where a competing surface pattern out-competes the rule, causing the log-probability margin to cross zero within 100 steps of behavioral failure. Control over this fate is asymmetric; while injecting counter-evidence can destroy rules via a monotone dose-response, restoring support even at 450 times the sustaining level fails to recover them.

arxiv arXiv cs.CL · 9h ago

Keyword Lexicon Blindness Distorts Rhetorical Stance Measurement

A study analyzing 85 interviews with four public intellectuals reveals that keyword-based scoring can produce statistical artifacts regarding rhetorical stance. Initial analysis showed a robust negative-affect and emphatic-certainty co-occurrence pattern with high correlation coefficients ranging from r = 0.72 to 0.93. However, replacing this method with LLM-based zero-shot semantic classification on the full diarized corpus of 32,625 sentences significantly reduced these correlations. For instance, Dalio's correlation dropped from 0.851 to 0.206, while other speakers exhibited negative or null relationships between negativity and certainty. In contrast, the LLM analysis revealed a strong coupling between negative sentiment and hedging language, aligning with conventional expectations of pessimistic discourse. The discrepancy stems from three structural failures in keyword lexicons: syntactic blindness, polysemy blindness, and categorical absence. These flaws can invert semantic meaning, such as scoring 'never absolutely totally confident' as high certainty. The authors argue that keyword counts measure lexical co-occurrence tendencies rather than epistemic certainty, constituting a category error.

arxiv arXiv cs.CL · 10h ago

Structuring an Arabic-English Machine-Readable Dictionary Using Parsing Expression Grammars

This paper presents a method for structuring a machine-readable version of the Arabic-English Al-Mawrid dictionary, addressing the lack of standardization in printed formats. The approach converts unstructured streams of words and punctuation into explicit hierarchical structures that define entry components such as subentries, domain labels, and translation equivalences. Parsing serves as the central step within a cascaded design, implemented using the parsing expression grammars formalism. This technique allows for the automatic or semi-automatic organization of dictionary entries despite the absence of microstructure standardization in Arabic dictionaries. The study demonstrates that inducing microstructure enables plausible accuracy in structuring these complex lexical resources. By transforming raw text into defined formats, the work supports downstream natural language processing applications requiring machine-readable lexical data.

arxiv arXiv cs.CL · 10h ago

Automatic Generation of Highlights for Academic Paper Via Prompt-based Learning

This study investigates prompt-based learning for the automatic generation of academic paper highlights to address the lack of labeled training data in existing supervised methods. The researchers designed task-specific prompt templates combined with paper abstracts as inputs for several language models, including locally deployed GPT-2 and T5, as well as ChatGPT accessed via API. Experiments conducted on three datasets demonstrated that ChatGPT with prompt templates achieved performance comparable to previous supervised methods without requiring task-specific training samples. When a small number of examples were added to the prompts, the model significantly outperformed state-of-the-art methods on two of the datasets. The analysis revealed that while ChatGPT possesses strong language modeling capabilities, its performance is highly sensitive to the specific information provided within the prompt. Case studies indicated that the generated highlights are generally coherent, informative, and closely resemble those written by authors. This approach does not rely on domain-specific training corpora, supporting downstream text mining and bibliometric research for papers lacking existing highlights.

arxiv arXiv cs.CL · 10h ago

Measuring Research Difficulty in NLP: An Inverted U-Shaped Relationship with Academic Impact

This study proposes a comprehensive evaluation system for measuring the difficulty of academic research, focusing on Natural Language Processing as a case study. The authors extract internal and external features from papers, including collaboration, content, and references, to compute multiple difficulty indicators. These indicators are weighted using the entropy weight method and summed to generate a final research difficulty score. Academic impact is quantified by citation frequency, while expert assessments validate the reliability of the measurement approach. Empirical results indicate that page count, reference count, and high-level institutional participation significantly correlate with academic impact. Crucially, the analysis reveals an inverted U-shaped relationship between research difficulty and impact. This suggests that moderately difficult research tends to achieve the highest level of academic influence.

arxiv arXiv cs.CL · 10h ago

Data-Driven Evolution of Library and Information Science Research Methods (1990-2022)

This study analyzes the influence of data-centric research on Library and Information Science by examining methodological evolution from 1990 to 2022. Researchers automatically extracted four key categories of data-driven entities from academic papers: algorithms and models, data resources, software and tools, and metrics. The analysis evaluates trends across three dimensions, including temporal characteristics, topic-specific evolution, and cross-method features. Findings identify data resources as the primary driver of methodological changes within the discipline. The research reveals a cyclical pattern characterized by emergence followed by stability or practical application in LIS methods. This perspective highlights how big data advancements have reshaped the field's technical landscape over three decades.

arxiv arXiv cs.CL · 11h ago

iLLaDA: An 8B Masked Diffusion Language Model with Fully Bidirectional Attention

The authors introduce iLLaDA, an 8B parameter masked diffusion language model trained from scratch using fully bidirectional attention. This approach contrasts with the predominant autoregressive factorization and causal attention used in modern large language models. The model's pre-training scaled to 12 trillion tokens, followed by supervised fine-tuning on a 25 billion-token instruction corpus for 12 epochs. iLLaDA maintains the masked diffusion objective throughout both training phases and employs variable-length generation for efficiency. It also introduces confidence-based scoring to enhance performance on multiple-choice evaluation tasks. Benchmark results show significant improvements over its predecessor, LLaDA, including gains of 21.6 points on BBH and 14.9 points on ARC-Challenge for the base model. The instruction-tuned variant achieved increases of 14.5 points on MATH and 16.5 points on HumanEval. Despite its non-autoregressive nature, iLLaDA remains competitive with Qwen2.5 7B across several metrics.

arxiv arXiv cs.CL · 11h ago

Local Branch Routing: Efficient Trainable Test-Time Scaling for Language Models

The authors introduce Local Branch Routing (LBR), a token-level framework designed to improve language model reasoning through efficient test-time scaling. LBR expands a small local lookahead tree and forwards all sampled branches through the model, using a lightweight router to select the depth-1 subtree for commitment. This approach allows each token decision to utilize evidence from candidate local futures without incurring the computational costs of full solution-level search. The method employs a prune-shift-grow decoding process that preserves discrete branch identities and defines a tractable tree-trajectory likelihood. Consequently, LBR enables end-to-end reinforcement learning with verifiable rewards, jointly optimizing the base model and router under the same likelihood-ratio principle as discrete-token RLVR. Experimental results on synthetic hierarchical-planning tasks demonstrate that post-candidate hidden states provide useful routing evidence. Furthermore, benchmarks in mathematical reasoning show that LBR improves both Pass@1 and Pass@32 metrics compared to discrete chain-of-thought and other baselines.

arxiv arXiv cs.CL · 11h ago

Neural Machine Translation for Low-Resource Tangkhul-English

This study addresses low-resource machine translation for the Tangkhul-English language pair, focusing on a severely under-resourced Tibeto-Burman language with minimal prior NLP infrastructure. The authors present two systems: a primary model based on ByT5-large and a contrastive system using mT5-small, both fine-tuned on 38,336 parallel sentence pairs. Evaluation on a held-out test set of 3,856 sentences shows the ByT5-large system achieving a corpus BLEU score of 39.97 and a chrF++ score of 58.07. Additional metrics include a BERTScore F1 of 0.8104 and a COMET score of 0.7302 using the wmt22-comet-da model. The research highlights orthographic challenges related to Tangkhul's Latin-script diacritics as a specific technical hurdle. Furthermore, the training corpus exhibits domain bias, consisting primarily of biblical texts, stories, and conversational data. Future work aims to improve performance through data diversification and domain adaptation strategies.

arxiv arXiv cs.CL · 11h ago

Computational Stylometry of English Pali Canon Translations Across Pitakas

This study presents a computational stylometric analysis of the Tipitaka across all three Pitakas in English translation, extending previous work on the Sutta Pitaka. The corpus comprises 134,831 segments from Bhikkhu Sujato's Sutta Pitaka, Bhikkhu Brahmali's Vinaya Pitaka, I.B. Horner's 1938 Vinaya translation, three English translations of the Abhidhammattha Sangaha, and cross-tradition Vinaya texts. The authors compute Zipf rank-frequency distributions, MATTR-500 lexical diversity, numeral-word density, and vocabulary overlap metrics. Main findings indicate that all corpora show Zipf-consistent distributions with R-squared values above 0.989. The Sutta and Theravada Vinaya exhibit nearly identical lexical diversity scores of 0.399 and 0.400, while the Sangaha corpus is more diverse at 0.560. The Sangaha corpus also displays the highest numeral-word density at 3.26%, reflecting its systematic enumeration of categories. Additionally, the Mulasarvastivada Vinaya shares significant vocabulary overlap with the Theravada Vinaya, whereas two English translations of the same source share only 24.2% of their vocabulary.

arxiv arXiv cs.CL · 11h ago

Story Operators: Decomposing the Original to Sequel Transformation in Embedding Space

This study models literary transformations as geometric operations within a sentence-embedding space using all-mpnet-base-v2 vectors from the PG19 corpus. By calculating displacement vectors between original novels and their sequels, the author decomposes these changes along a content basis derived via PCA. Analysis of thirteen verified author pairs reveals a taxonomy of sequel types: formulaic, concentrated, and compositional. Formulaic transformations involve minimal rank changes, such as Doyle's Holmes collections with a norm of 0.12. Concentrated shifts are dominated by a single axis, exemplified by Alcott's Little Women to Little Men where 75% of the change occurs on one move. Compositional transformations involve many small axes, seen in works by Twain, Burroughs, and Nesbit. For Tom Sawyer to Huckleberry Finn, the dominant axis is structural, reflecting a shift from domesticity to picaresque adventure rather than surface themes like vernacular voice. The geometric findings are corroborated against Mark Twain's documented authorial intent in letters to Howells.

arxiv arXiv cs.CL · 11h ago

Introducing corpora Hlava Cor and Hlava AD: Human Label Variation in Coreference and Discourse Relations

Researchers have created two new corpora, Hlava Cor and Hlava AD, to explore human variation in understanding text coherence. These resources contain multiple annotations of Czech texts along with annotators' explanations for their choices. The first corpus, Hlava Cor, consists of 1,024 contexts annotated by three individuals to capture coreference identification differences. It covers pronouns, full noun phrases, and anaphoric adverbials across various text types and grammatical-semantic categories. The second corpus, Hlava AD, comprises 512 contexts annotated by five annotators focusing on discourse relations in attributive and non-attributive constructions. Both corpora achieve an inter-annotator agreement of approximately 60-65 percent. Analysis reveals that lower coreference agreement correlates with automatic model disagreement, indicating higher ambiguity. Annotator comments further highlight varying confidence levels and individual reading strategies.

arxiv arXiv cs.CL · 11h ago

Agent-Authored World Modeling Aligns Training with Decision Needs

The paper introduces Agent-Authored World Modeling (AAWM), a training procedure that addresses the limitations of standard world modeling objectives tied to next-observation prediction. This traditional approach often omits dynamics relevant to an agent's current decision because supervision depends on what a transition reveals rather than what is needed. AAWM constructs supervision directly from the policy's decision needs by having the agent identify necessary environmental understanding at each state. Relevant transition evidence is retrieved across trajectories and synthesized into training targets that capture these decision-oriented dynamics. This method aligns the learning objective with the specific information required before acting, rather than forcing the model to reconstruct the next observation. Experimental results validate AAWM's effectiveness across multiple environments and training settings. The findings demonstrate that decision-aware world-model targets provide a more effective learning signal than conventional next-observation prediction.

arxiv arXiv cs.CL · 11h ago

OscillaTTS: Adaptive Oscillatory Inductive Bias for Modeling Sharp Prosodic Dynamics in Diffusion-Based TTS

Diffusion-based text-to-speech models have improved speech quality but struggle with sharp prosodic transitions and rapid pitch variations. Existing decoders often use periodic nonlinearities like the Snake activation function, which lack adaptability for abrupt amplitude and frequency changes. To address this, the authors introduce OscillaTTS, a system featuring an adaptive oscillatory nonlinearity. This component enables controllable periodic modulation while ensuring signal stability via a linear bypass mechanism. The study investigates the role of oscillatory inductive bias within diffusion-based TTS decoders. Experiments conducted on the LJSpeech and Emotional Speech Dataset demonstrate consistent improvements in both objective and subjective evaluations. These results indicate that OscillaTTS effectively models expressive prosodic dynamics compared to prior methods.

arxiv arXiv cs.CL · 12h ago

PolicyAlign: Direct Policy-Based Safety Alignment for Large Language Models

The authors introduce PolicyAlign, a framework designed to align large language models directly with natural-language safety policies rather than relying on costly supervision data. This approach addresses the mismatch between rapidly evolving safety requirements and conventional data-driven alignment methods. The process begins by synthesizing instructions that violate the specified policy, followed by on-policy self-distillation to internalize the desired behavior. To enhance training stability and data efficiency, the method incorporates Policy-Sensitive Filtering, which selects instructions inducing the largest behavioral shift. Experiments across multiple models demonstrate that PolicyAlign consistently improves safety metrics while maintaining low over-refusal rates and preserving general capabilities. The framework also generalizes effectively to specialized domains such as medical, legal, and financial safety scenarios. The code for this scalable alignment approach is released at https://github.com/Qwen-Applications/PolicyAlign.

arxiv arXiv cs.CL · 12h ago

Translation-Enhanced Speech Encoder Pre-training Improves Speech LLMs

Connecting a pre-trained speech encoder to a Large Language Model creates a structural misalignment because encoders often produce language-specific representations while LLMs operate in a unified, language-agnostic space. The authors argue that incorporating speech translation objectives into the pre-training process provides a principled mechanism to bridge this gap. Unlike monolingual transcription, translation forces the model to learn representations that are independent of specific languages. The study experimentally evaluates the impact of adding these translation objectives during speech encoder pre-training. Results demonstrate that this approach significantly improves cross-modal integration between the speech and text modalities. Consequently, models utilizing translation-enhanced pre-training achieve superior performance across various downstream Speech LLM tasks.

arxiv arXiv cs.CL · 12h ago

Reclaim Evaluation Shows Lossy Memory Is Worse Than No Memory

A study demonstrates that a language model's memory containing incorrect conclusions is more detrimental than having no memory at all. When models retain stale values while dropping supporting work, they emit confident but wrong answers, whereas empty memories allow for abstention. This phenomenon, termed brittle memory, was observed across seven models where the direction of failure never reversed regardless of task or disposition. The researchers introduced reclaim evaluation to measure correctability by compressing interactions and testing if corrections recover ground truth without using a judge. Results indicate that correctability depends on whether the source information survives compression rather than model capability. A source-first policy, which keeps recomputable sources and drops re-derivable conclusions, restored correctability significantly better than length-matched controls. In chained memory loops, dropped-source errors corrupt downstream steps irreparably, while the proposed fix maintains bounded performance horizons. The findings replicate across three deployed systems and real dialogue data, with a hand-built oracle reaching perfect accuracy.

arxiv arXiv cs.CL · 12h ago

The Generalization Spectrum: A Chromatographic Approach to Evaluating Learning Algorithms

Traditional evaluations reduce learning to a single aggregate score, obscuring how well knowledge from one example generalizes to others. The authors introduce the Generalization Spectrum, an evaluation framework that measures per-sample generalization by tracking performance across test variants with increasing transfer distance. These variants range from exact recall to implementation transfer across languages and context transfer under narrative reframing. The framework is instantiated on competitive programming using a selection-and-synthesis pipeline seeded with recent problems to mitigate contamination. Comparisons of canonical learning paradigms show that Reinforcement Learning converts memorization into near-transfer more efficiently than Supervised Fine-Tuning baselines. In-context learning exhibits strong but correspondence-dependent transfer capabilities in this context. Diagnostic profiles reveal that local gains do not necessarily expand the generalization radius for all methods. Specifically, abstractions and hints mainly lift local transfer, while Reference SFT preserves a stronger far-transfer tail than RFT. Furthermore, self-distillation or hint-assisted RL can reduce far transfer even when local transfer improves.