Memory-V2V is a memory‑augmented video‑to‑video diffusion framework designed to improve cross‑turn consistency in multi‑turn video editing. It stores previous outputs in an external memory, retrieves relevant edits, and incorporates them via relevance‑aware tokenization and adaptive compression, allowing scalable conditioning without linear computational growth. Experiments on iterative video novel view synthesis and text‑guided long video editing show that Memory‑V2V enhances consistency while preserving visual quality and outperforming strong baselines with modest overhead.
By Dohun Lee, Chun-Hao Paul Huang, Xuelin Chen, Jong Chul Ye, Duygu Ceylan, Hyeonho Jeong
The study investigates why language models exhibit systematic performance gaps across English dialects, a phenomenon termed the "dialect tax." Using parallel dialect corpora that preserve meaning while altering surface form, the authors confirm that models treat Standard American English and dialectal texts as semantically equivalent, yet find representational disparities that persist through tokenization, pre‑training, post‑training, and inference. Even a character‑level tokenizer does not eliminate input/output asymmetries or accuracy gaps, and dialect pairs produce more divergent gradient updates than unrelated Standard texts, indicating that dialectal content is harder for models to learn.
By Elle
The paper introduces CuCu, a multi‑agent LLM framework that converts national social studies curricula into open‑ended, culture‑specific question‑answer pairs for fine‑tuning language models. Using the Korean curriculum, the authors build KCaQA, a dataset of 34.1k QA pairs that cover culture‑specific topics and ground responses in local sociocultural contexts. Experiments show that fine‑tuning with KCaQA improves the model’s cultural alignment and relevance to Korean society.
By Haneul Yoo, Won Ik Cho, Geunhye Kim, Jiyoon Han
BanglaMamba explores Mamba-based State Space Models (SSMs) as a computationally efficient alternative for Bangla fake news detection. Compared to BanglaBERT and a custom BERT trained from scratch, BanglaMamba achieves a Macro‑F1 score of 0.9029, close to the 0.9057 of the custom BERT, while delivering 2.2× higher inference throughput and 49% lower peak GPU memory usage. Cross‑dataset evaluation shows BanglaBERT generalizes better, underscoring the value of large‑scale pretraining.
By M. K. Khalidi Siam
SelfGraphRAG is a framework that generates synthetic question‑answer pairs directly from the structure of a knowledge graph to train a query‑conditioned graph retriever. By capturing multi‑hop paths and local neighborhoods, the generated questions provide relational supervision without requiring manually labeled data. Experiments on multi‑hop question answering and classification tasks show that SelfGraphRAG improves retrieval precision and downstream reasoning performance compared to embedding‑based baselines.
By Ben Lagnese, Manas Gaur
The paper investigates where failures occur in GNN‑based Knowledge Graph Question Answering pipelines when faced with adversarial question perturbations. By isolating stages—entity linking, subgraph retrieval, GNN reasoning, and answer generation—and applying two answer‑preserving attacks (Compositional Restructuring and Relation Synonym Swap), the authors find that subgraph construction is responsible for over 99% of end‑to‑end failures, even though the correct answer is often present in the retrieved subgraph. This challenges the assumption that reasoning models are the weak link and highlights subgraph construction as the critical mitigation target.
By Pankaj Kumar, Subhankar Mishra
In 2026, the SHROOM-Visions shared task was launched at the UncertaiNLP Workshop co‑located with EMNLP to address hallucinations in large vision‑language models. The task builds on the SHEEP dataset and asks participants to detect and classify fine‑grained hallucination spans in image‑conditioned text generation across four languages (Chinese, English, French, Italian) using a five‑class taxonomy. The competition attracted 27 teams and over 600 system submissions, with top systems achieving character‑level, label‑conditioned, and IoU scores of 0.58, 0.46, and 0.51 respectively, surpassing baselines by 30‑40 points.
By Ra\'ul V\'azquez, Aman Sinha, Chuyuan Li, Claudio Savelli, Eduardo Cal\`o, Emilio Raimond, Stella Frank, Hengyu Luo, Flavio Giobergia, Vincent Segonne, Lorenzo Vaiani, J\"org Tiedemann, Timothee Mickus
The paper introduces ORBIT, a framework for probing higher‑order epistasis in protein representations. ORBIT validates Walsh‑based diagnostics on synthetic landscapes, then applies them to the GB1 fitness landscape, comparing several models including ridge regression, MLPs, and Residual Interaction Tokenization (RIT). While no architecture differences were found in overall prediction performance, RIT notably increased pairwise token‑level accessibility, and deeper MLPs improved higher‑order functional recovery, revealing representation‑level changes hidden by conventional metrics.
By Maryam Rahimimovassagh, Ivan Garibay, Niloofar Yousefi
The paper presents an AI‑enhanced method for radio frequency interference suppression that builds on autoregressive transformer models by adding a Finite Scalar Quantization tokenizer layer. This addition improves interference rejection while maintaining low latency, and the authors also test other inference optimizations to speed up processing with minimal accuracy loss. Experiments using a digitally modulated RF signal as the signal of interest and a digital television OFDM signal as interference show that the approach outperforms traditional techniques and prior AI methods, with benefits demonstrated through audio quality metrics like PESQ and potential operational applications.
By Rahul Jain, Pierre Trepagnier, Rick Gentile, Joey Botero, Alexia Schulz
The paper introduces PLUS, a framework that uses reinforcement learning to generate text-based summaries of individual users’ preferences, characteristics, and past conversations. These summaries condition a reward model, allowing it to predict personalized response preferences and improving reward accuracy by 11–77 % over the standard Bradley‑Terry model. PLUS demonstrates robust performance with new users and topics, achieves a 25 % improvement over existing personalized RLHF techniques, and enables zero‑shot personalization for state‑of‑the‑art models like GPT‑4.
By Hyunji Nam, Yanming Wan, Mickel Liu, Peter Ahnn, Jianxun Lian, Natasha Jaques
BRIDLE is a self‑supervised encoder pretraining framework that extends bidirectional training to audio, image, and video by incorporating residual quantization (RQ) with multiple hierarchical codebooks. This approach allows fine‑grained discretization of latent representations and interleaves training between the encoder and tokenizer. Experiments show that BRIDLE achieves state‑of‑the‑art results on audio classification benchmarks and competitive performance on image and video classification tasks, outperforming traditional vector‑quantization methods.
By Hoang M. Nguyen, Satya N. Shukla, Qiang Zhang, Hanchao Yu, Sreya D. Roy, Dipesh Tamboli, Taipeng Tian, Lingjiong Zhu, Yuchen Liu
The paper proposes treating the ‘unit’—a persistent referent that multiple events may refer to—as an explicit primitive in machine learning tasks. It formalizes supervised learning as learning a pair of a tokenizer that generates a contextual unit token and a shared response law that uses this token, thereby distinguishing homogeneous from heterogeneous worlds. The work also introduces concepts such as unit abduction and trusted resolvers to handle cases where unit identity is unresolved.
By Heyang Gong
The paper introduces grounded glossary generation, a structured NLP task that asks models to recover semantically meaningful Sanskrit phrases and provide translation‑grounded meanings from a sloka‑translation pair, mirroring the traditional patha commentary practice. A benchmark of 31,316 sloka‑translation‑glossary triples from the Valmiki Ramayana and Srimad Bhagavatam is built, evaluated with Jaccard for phrase recovery and Meaning Faithfulness for semantic consistency. Experiments with Gemma‑3n‑E4B, Gemma‑3‑12B, Phi‑4, and Qwen3.5‑9B show that instruction fine‑tuning outperforms prompting, and explicit segmentation further improves results, though over‑segmentation of sandhi and samasa compounds remains the main error source, highlighting morphological modeling as a key bottleneck.
By Manoj Balaji Jagadeeshan, Sai Pragnaan Marala, Pawan Goyal
The paper investigates whether retrieval‑augmented generation (RAG) can uniformly correct factual errors in large language models (LLMs) by testing six LLMs on a benchmark of about 2,000 public companies. In a controlled factual QA setting, the authors evaluate the models under four conditions—no‑context, perfect context, misleading context, and distraction context—across four atomic attributes. Results show significant geographic disparities in baseline accuracy, and while perfect context improves performance, it does not eliminate these gaps; misleading context often leads models to copy incorrect information, and larger models only marginally reduce structural biases.
By Abhinav Havaldar, Enrico Santus
TranslatePsy-AfriSLM is an open‑source machine‑translation resource set for 19 Sub‑Saharan African languages, comprising curated parallel data, African‑specialized synthetic data, and a family of fine‑tuned small language models (SLMs). The authors demonstrate that a unified quality‑estimation filtering can remove up to 96% of training tokens without harming quality, and that filtered synthetic data dominates the quality‑efficiency Pareto frontier. Models trained on this mixture outperform much larger systems such as TranslateGemma‑27B and Qwen3.5‑122B‑A10B, achieving superior performance with as few as 0.8 B parameters.
By Milan Gritta, Patrik Lambert, Jihye Back, Amril Nazir
Just Pass Twice (JPT) is a method that allows causal large language models to perform token classification for zero‑shot named entity recognition by concatenating the input with itself, giving each token full bidirectional context without architectural changes. The approach combines these representations with definition‑guided entity embeddings to enable flexible zero‑shot generalization. JPT achieves state‑of‑the‑art results, outperforming prior methods by an average of +7.9 F1 on CrossNER and MIT benchmarks and running over 20× faster than comparable generative approaches.
By Ahmed Ewais, Ahmed Hashish, Amr Ali
MoganBert-TR is a 149‑million‑parameter Turkish encoder foundation model trained from scratch on a language‑specific corpus using a two‑stage CLM‑to‑MLM curriculum. The model, along with its embedding variant MoganBert‑Embed, achieves state‑of‑the‑art results on Turkish benchmarks such as TrGLUE and TabiBench, outperforming existing Turkish BERT models. Its tokenizer, comprising 50,048 tokens, also surpasses other Turkish tokenizers in compression and fertility metrics.
By Furkan Yilmaz, Habibe Aleyna Tasdemir, Muhammed Faruk Gozay
The paper revisits the debate between Pinker & Prince (1988) and Rumelhart & McClelland (1986) regarding neural network models of English past tense. It argues that modern Encoder-Decoder architectures in NLP address the empirical shortcomings highlighted by Pinker and Prince, eliminating the need to simplify the past tense mapping problem. The authors suggest that the strong performance of these contemporary models merits a reassessment of their role in linguistic and cognitive modeling.
By Christo Kirov, Ryan Cotterell
ClueWeaver is a dual-agent framework designed to enable compact, locally deployable language models to answer questions about long literary narratives. The Finder agent retrieves passages that contain answer-critical clues, while the Interpreter agent derives the answer from those passages, generates rationales with paragraph-ID citations, and performs self-calibration for high-risk questions. Both agents are trained with reward-guided reinforcement learning to prioritize evidence retention, correctness, grounding, and concise explanations, resulting in improved performance and inspectability over end-to-end prompting.
By Jihao Zhu, Zhiwei Yang, Wenxiao Zhang, Junqian Zhao, Qi You, Fangqi Wang, Zheyuan Deng, Hanzhe Yang, Yu Liu, Jin B. Hong
The paper introduces Gavel, a framework for evaluating large language models (LLMs) on long-context legal summarization tasks. Gavel includes a reference-based component (Gavel-Ref) with checklist, residual-fact, and writing-style checks, and a reference-free component (Gavel-Agent) that assesses factual coverage directly from source documents. Experiments on 12 frontier LLMs reveal that models tend to omit key information more than hallucinate, perform well on simple checklist items but struggle with rare, complex items, and their performance degrades with longer cases. Gavel-Agent cuts token usage by at least 36% compared to traditional methods while maintaining competitive accuracy, and it also generalizes effectively to the medical domain.
By Yao Dou, Benjamin Mamut, Wei Xu