arXiv Computational Linguistics

A systematic methodology transforms structured linguistic resources like Hindi WordNet into 1.25 million instruction-response pairs to fine-tune a 12B-parameter language model using resource-efficient LoRA and 4-bit quantization. Evaluation via a Hindi language learning chatbot shows superior pedagogical effectiveness (91.0) compared to general-purpose models (79.4-83.6) while maintaining competitive semantic performance. This work offers a practical alternative for low-resource languages, enabling specialized AI development for hundreds of languages with existing WordNet resources.

A new study reveals that larger language models outperform smaller ones in reasoning tasks due to constraint-guided reasoning. Using the AdvCluster framework, researchers found stable performance gaps of 6.43% and 7.38% across model pairs. The analysis identifies constraint identification and structured reasoning as key advantages.

This pilot study presents NEST-V1, a lightweight Transformer-based multimodal framework that generates emotion-conditioned Nepali Sign Language avatars from spoken input. On a dataset of 600 audio samples covering 4 common words and 3 emotional states, the system achieves 81.1% ASR accuracy and 79.21% emotion recognition accuracy with only 22.1M parameters, suitable for edge deployment. The work establishes a technical foundation for emotion-aware sign language translation in low-resource settings.

This study investigates using Nonviolent Communication (NVC) principles as lightweight prompt-level constraints to guide large language models (LLMs) toward more de-escalating dialogue behavior in emotionally charged situations. Through a dual-agent simulation framework across multiple models and user resistance levels, NVC-constrained prompting consistently reduces conversational escalation and stabilizes interactions with highly resistant users.

Large language models excel at short-context reasoning but degrade over long conversations due to context window limits. ContextForge recycles context via structured query generation, external memory retrieval, and controlled synthesis, reducing token overhead while maintaining answer quality. On a 15-turn healthcare benchmark, ContextForge improved consistency and reduced token consumption.

A new study finds that using linguistic features such as assertive certainty, explicit moral vocabulary, and emotion words in fine-tuning data significantly shifts LLM reasoning toward stronger pro-animal-welfare stances, while hedged language and concrete sensory description dilute that stance. The research offers practical guidance for animal-welfare advocates.

A new study evaluates LLM performance on statics problems using a model distillation approach. LLMs perform well on text-only problems but accuracy drops when diagrams and multi-step reasoning are introduced. The decline is primarily due to difficulties in multi-step reasoning, not image recognition limits.

A study finds that post-training for helpfulness (SFT and RL) significantly degrades animal compassion values instilled during mid-training, while coding post-training better preserves them. Helpfulness training also causes a large drop in English general moral reasoning but not cross-lingually, whereas the compassion degradation transfers consistently across languages. This suggests mid-trained values are encoded more deeply and cross-lingually than reasoning improvements from domain-specific post-training. The paper recommends coding post-training for value-preserving model development.

A new benchmark, Know2Guess, aims to evaluate LLMs' ability to distinguish between knowledge-based answering and guessing, considering data contamination. It includes 1,200 items across five domains and tests models like FLAN-T5, Qwen2.5-Instruct, and Llama-3-Instruct. Qwen2.5-3B-Instruct shows best reliability but still has calibration issues.

HierBias is a novel hierarchical context-conditioned media bias detector that formally models document context for bias prediction, theoretically proving reduced Bayes error and improving sample efficiency via multi-task learning. It achieves 0.853 F1 and 0.723 MCC on BABE and BASIL, surpassing state-of-the-art.

This study developed a codebook for self-stigma across cognitive, affective, and behavioral domains and analyzed Reddit posts from people who use drugs. Results show that self-stigma is prevalent, and behavioral indicators often precede core ones, challenging progressive models.

Large language models have transformed code generation, raising concerns about authorship, assessment integrity, and software trust. SemEval-2026 Task 13 Subtask A defines detection as binary classification over code snippets, with emphasis on out-of-distribution generalization across unseen programming languages and domains. The authors propose SALSA (Single-pass Autoregressive LLM Structured Classification), which maps each class to a dedicated output token and trains the model to emit a single-token label. By combining balanced sampling, parameter-efficient fine-tuning, and conservative training, the system achieves OOD F1=0.789 on the official leaderboard, significantly outperforming the CodeBERT baseline (F1=0.305).

The rapid growth of scientific submissions has pushed traditional peer review toward its scalability limits, motivating the exploration of large language models (LLMs) as intelligent automated evaluation assistants. Although recent studies show that LLMs can generate fluent critiques and approximate reviewer scores, their reliability, robustness, and security as decision-support systems remain insufficiently understood. This survey offers a systems-level analysis of LLM-based scientific peer review, focusing on two core evaluative functions: critique generation and score prediction. We present a structured taxonomy of modeling approaches (including prompt-based, supervised, retrieval-augmented, and alignment-optimized approaches), and synthesize empirical findings across existing benchmarks. We analyze dataset constraints, evaluation shortcomings, and domain concentration biases that limit current assessment practices. Beyond performance metrics, we identify emerging robustness risks, including prompt injection, data poisoning, retrieval vulnerabilities, and reward hacking, which expose automated review pipelines to strategic manipulation. From a data mining perspective, we outline key open challenges in modeling subjective disagreement and cross-domain generalization. By reframing automated peer review as a high-stakes, multi-objective decision problem, this survey provides a roadmap for developing robust, transparent, and trustworthy AI-assisted scientific evaluation systems.

A new study introduces a dataset of 32,534 double-marked real student responses to GCSE mock exams, covering five subjects and handwritten work. Top LLMs agree with examiners more closely than examiners agree with each other, handling subjective and handwriting tasks effectively, with little dependence on model size.

Dustin is a sparse verification framework for long-context speculative decoding that combines draft model lookahead signals with target model historical attention to identify critical tokens, achieving 27.85x self-attention speedup and 9.17x end-to-end decoding speedup at 32k sequence length on Qwen2.5-72B with negligible accuracy loss.

A new arXiv paper investigates the geometric relationship between detection and control directions in language models. While models can perfectly detect hallucination (AUC=1.0), the direction for detection and the direction for causing refusal have a cosine of only 0.12, indicating that detection does not imply controllability. This gap generalizes across models and sizes, originates in pretraining, and a 15-degree rotation can partially bridge it.

A new framework using error-aware TF-IDF retrieval to correct ASR errors, achieving significant improvements in WER on Persian FLEURS.

AgentOdyssey is a novel evaluation framework that procedurally generates open-ended text games to test agents' ability to learn continuously during deployment. It challenges the traditional ML assumption of no learning at test time, interleaving learning and inference throughout. The framework measures world knowledge acquisition, episodic memory, exploration, action diversity, and model cost. Experiments show even the strongest agents fall far below human performance, with short-term memory emerging as a key beneficial mechanism.

A new study shows that a small group of Wikipedia editors can significantly influence how large language models discuss animal welfare with just 125 edits. Using gradient-based attribution methods, the research traced the impact of these edits, finding that animal welfare-related Wikipedia content dominates model responses to relevant queries.

Researchers propose G-SPIN, a structured ASR correction framework that combines phonetic graph modeling with contextual language understanding. It uses a graph neural network to generate acoustically plausible candidate sets, a masked language model for scoring, and an instruction-tuned large language model for final re-ranking, enabling lightweight, modular inference-time correction.

Standard tokenizer evaluation metrics like fertility rate fail to capture morphological correctness for agglutinative languages. The QuechuaTok benchmark compares four tokenization strategies on Southern Quechua, using morphological boundary accuracy (MorphAcc) alongside traditional metrics. Results show that while BPE achieves the lowest fertility rate (1.636), its MorphAcc is only 6.67%, whereas the morphology-aware PRPE tokenizer reaches 83.33% MorphAcc, demonstrating that fertility rate alone is insufficient for agglutinative languages.

This study challenges the effectiveness of exact-match retrieval recall as a proxy for retriever quality. In tau-bench, retrieved policy clauses performed nearly as well as gold policies in downstream classification tasks, despite only 7% exact-match recall. The findings suggest that relying solely on recall may underestimate the practical utility of retrieved policies.

This study audits eight automatic scorers across three evaluation constructs, finding that no scorer transfers across all datasets. In the generated-answer attribution construct, metric rankings invert and an NLI scorer collapses on long-form tasks. A prompt-based LLM judge avoids collapse but is costly and non-deterministic. The research concludes that metric choice must be validated on the target dataset.

A new study evaluates six proprietary LLMs across 16 DSM-5 conditions using adversarial attacks, finding that safeguards only reliably hold for suicide and self-harm, with failure rates up to 100% for conditions like eating disorders, substance use disorder, and major depressive disorder. The authors call for clearly defined harm categories and targeted safeguards.

This paper proposes a training-free framework called IBA (Identify-Before-Answer) for Knowledge-Based Visual Question Answering (KB-VQA). It decouples entity identification from evidence ranking by prompting an MLLM to select high-confidence entities from candidate names, then using an off-the-shelf text re-ranker for evidence. Experiments on Encyclopedic-VQA and InfoSeek show consistent outperformance over fine-tuned multimodal reranking baselines with reduced training and inference complexity.

A new framework uses LLMs to quantify product desirability from qualitative feedback without explicit scores, achieving Pearson correlations up to 0.97 and classification accuracy up to 94% on PDT datasets. GPT-4o-mini matches larger models at 94% lower cost, and the system includes confidence ratings and explainable AI.

A new study shows that self-generated text recognition (SGTR) finetuning can effectively prevent and reverse emergent misalignment (EM) in large language models, outperforming benign finetuning methods. The research finds that EM results from destabilization of a model's aligned character rather than learning harmful content, and SGTR works by fortifying character consistency.

This paper introduces the Normalized Context Utilization (NCU) metric to quantify contextual information gain in RAG systems. Experiments show that small language models match or outperform large models in strict factual extraction, while a commercial API overrides external evidence in nearly half of adversarial conflicts.

Proposes ModTGCN, a modularity-aware GNN for text classification that jointly optimizes cross-entropy and a modularity-based auxiliary objective to promote class-coherent document communities while preserving discriminative representations. Achieves consistent gains on five benchmarks, with larger improvements on low-homophily datasets.

EXPO-SQL proposes an execution-based clause-level policy optimization method that assigns fine-grained rewards to each clause in a SQL query by analyzing execution results, including error messages and incremental clause-wise execution, addressing the issue of insufficient learning signals caused by coarse-grained query-level rewards in existing RL methods. Experiments show it significantly outperforms existing supervised fine-tuning, prompting, and RL methods on multiple Text-to-SQL benchmarks.