https://techgazette.com.br/research/#language This week's papers in Language models, each with what its authors did and found, in sentences from its own abstract. en 180 Thu, 15 Oct 2026 17:58:00 GMT https://arxiv.org/abs/2610.22355 https://arxiv.org/abs/2610.22355 Thu, 15 Oct 2026 17:58:00 GMT arXiv cs.CL Language models arXiv cs.CL — Did: To enhance the robustness of visual verifiers in complex multimodal scenarios involving compositional requirements, we move beyond conventional binary verifiers that produce only True/False discrete judgments and introduce a verifier that automatically decomposes complex prompts or environmental rules into fine-grained components and… Found: Overall, our findings suggest a key insight: the transition from generative models to structured world-modeling agents critically hinges on the availability of accurate pairwise graded visual verifiers. By Bowei Liu, Xinchen Zhang, Xuhuan Li, Tianlin Zhang, Kaian Jiang, Jingwei Liu. https://arxiv.org/abs/2610.22324 https://arxiv.org/abs/2610.22324 Thu, 15 Oct 2026 17:52:00 GMT arXiv cs.CL Language models arXiv cs.CL — Did: In our work, we study document-grounded ideation, focusing on increasing the diversity of LLM generated ideas and ensuring they are high quality. Found: We show that our trained 8B model achieves higher creative utility (a measure combining quality and diversity) than Claude Opus 4.8 and GPT-6 Astra on our contamination-free benchmark. By Manya Wadhwa, Pranav Narayanan Venkit, Hiroaki Hayashi, Yu Li, Tuhin Chakrabarty, Chien-Sheng Wu. https://arxiv.org/abs/2610.22314 https://arxiv.org/abs/2610.22314 Thu, 15 Oct 2026 17:51:00 GMT arXiv cs.CL Language models arXiv cs.CL — Did: To bridge these gaps, we introduce MindEval2, an evaluation framework that consists of two components: Mental Health Quality Rubrics (MHQR), for evaluation, and MindSim, for patient simulation. Found: We evaluate 18 proprietary and open-weight models on MindEval2, finding that while frontier models have largely improved on issues related to surface level quality and boundaries, guidance calibration remains the dominant failure mode across all models, accounting for most of the quality gap. By John Mendonça, Daniel Tschernutter, José Pombal, Areti Vassilopoulos, Catarina Botelho, Ignacio Sarasua. https://arxiv.org/abs/2610.22312 https://arxiv.org/abs/2610.22312 Thu, 15 Oct 2026 17:50:00 GMT arXiv cs.CL Language models arXiv cs.CL — Did: We introduce Grid-Bench, an executable benchmark in which eight families of operations transform discrete worlds represented as two-, three-, and four-dimensional grids. By Zixiang Xu, Zirui Song, Wanrong Zheng, Chiyu Zhang, Yue Huang, Yuyang Peng. https://arxiv.org/abs/2610.22300 https://arxiv.org/abs/2610.22300 Thu, 15 Oct 2026 17:47:00 GMT arXiv cs.CL Language models arXiv cs.CL — Did: We study this distinction by constructing new relations from existing ones. By Faridoun Mehri, Hinrich Schütze. https://arxiv.org/abs/2610.22298 https://arxiv.org/abs/2610.22298 Thu, 15 Oct 2026 17:46:00 GMT arXiv cs.CL Language models arXiv cs.CL — Did: We developed a multidimensional benchmark using 300 diagnostically resolved NEJM--MedQA cases: 110 for physician reference ratings and Judge calibration and 190 for held-out evaluation. Found: These findings show that the benchmark can distinguish model capability, the contribution of multidisciplinary synthesis, criterion-specific effects of response structure, deliberation dynamics, and inference requirements within one common-case evaluation. By Alexey A. Bokov, Ivan I. Panin, Nikolay B. Nechaev, Mikhail Yu. Valkov, Valeria Yu. Chernina, Natalia Gombolevskaya. https://arxiv.org/abs/2610.22294 https://arxiv.org/abs/2610.22294 Thu, 15 Oct 2026 17:44:00 GMT arXiv cs.CL Language models arXiv cs.CL — Did: We therefore introduce multi-target few-shot prompting with diffusion language models, where examples guide the infilling of multiple masked regions within a shared canvas. Found: Across few-shot prompting benchmarks and software refactoring case studies, we find that demonstrations improve infilling accuracy, with example guidance yielding further gains in several settings. By Anton Xue, Alan Li, Rahul Saha. https://arxiv.org/abs/2610.22249 https://arxiv.org/abs/2610.22249 Thu, 15 Oct 2026 17:22:00 GMT arXiv cs.CL Language models arXiv cs.CL — Did: We study it through the Mandarin corrective construction bushi X, ershi Y ("not X, but Y"). Found: Of 307 annotated uses, only 1.30% are defective while 52.8% are optional, that is, correct but unneeded, so models fall short in appropriateness far more often than in correctness. By Yihang Peng, Yi Jing, Dewen Liu, Weihan Meng, Ying Liu. https://arxiv.org/abs/2610.22209 https://arxiv.org/abs/2610.22209 Thu, 15 Oct 2026 16:58:00 GMT arXiv cs.CL Language models arXiv cs.CL — Did: We show this evaluation paradigm is incomplete: it conflates memory correctness with retrieval and generation quality, producing both false positives where correct responses mask broken memory and false negatives where incorrect responses blame correct memory for retrieval failures. Found: Diagnostic decomposition reveals that the system correctly classifies memory operations (create, update, reinforce, supersede) only 20% of the time, a failure entirely invisible to extrinsic evaluation, and that failure modes differ systematically by category and difficulty. By Sebastian-Antonio Toma, Jean de Kernier, Mike Ply, Oliviu Matei. https://arxiv.org/abs/2610.22196 https://arxiv.org/abs/2610.22196 Thu, 15 Oct 2026 16:46:00 GMT arXiv cs.CL Language models arXiv cs.CL — Did: We present Hy-Memory-30B-A3B, a compact long-term-memory writer trained across Mem0, MemOS, and Hy-Memory. Found: On fixed 20% benchmark subsamples, Hy-Memory-30B-A3B ranks first on all four evaluated harnesses, including held-out EverMemOS, with a 55.49% macro average versus 54.62% for Claude Opus 5. By Tianxiang Fei, Jiazheng Kang, Mao Zheng, Xingyan He, Xiang Yu. https://arxiv.org/abs/2610.22187 https://arxiv.org/abs/2610.22187 Thu, 15 Oct 2026 16:42:00 GMT arXiv cs.CL Language models arXiv cs.CL — Did: To address this gap, we define and systematically investigate scale sensitivity, the variation in LLM-generated scores caused by rating scale design. Found: We find that changing one design element shifts normalized scores by about 0.10 on average and changes the category of roughly 20% of texts. By Wonjeong Jo, Gyuwan Kim, Ming Shan Hee. https://arxiv.org/abs/2610.22174 https://arxiv.org/abs/2610.22174 Thu, 15 Oct 2026 16:32:00 GMT arXiv cs.CL Language models arXiv cs.CL — Did: Therefore, we introduce Multi-Session Agentic Reinforcement Learning, a more general setting that supports an arbitrary number of sessions within a trajectory. Found: Further analyses show that our method achieves improvements in session utility and allocation, which we suggested is a promising target for future advances in multi-session agentic RL. By Yilong Xu, Minrui Xu, Xinyu Zhou, Xiang Long, Haochen Tan, Zhijiang Guo. https://arxiv.org/abs/2610.22169 https://arxiv.org/abs/2610.22169 Thu, 15 Oct 2026 16:29:00 GMT arXiv cs.CL Language models arXiv cs.CL — Did: The vocabulary of a language model is its native semantics: the embedding and the unembedding fix the meaning of a state at the two ends of the residual stream. By Jian Gu, Hongyu Zhang, Chunyang Chen, Aldeida Aleti. https://arxiv.org/abs/2610.22167 https://arxiv.org/abs/2610.22167 Thu, 15 Oct 2026 16:28:00 GMT arXiv cs.CL Language models arXiv cs.CL — Did: We propose CLUE, a two-stage training pipeline that combines supervised evidence reasoning with GRPO guided by Grounded Information Utilization ($\mathrm{IU}_{\mathrm{G}}$). Found: These gains are accompanied by over 80% fewer reasoning tokens on average in both domains. By Chengfei Wu, Zhaoqing Wu, Dan Goldwasser. https://arxiv.org/abs/2610.22160 https://arxiv.org/abs/2610.22160 Thu, 15 Oct 2026 16:23:00 GMT arXiv cs.CL Language models arXiv cs.CL — Did: We introduce EMO, a training-free framework inspired by expectation--maximization that alternates question-specific context estimation (E-step) and distribution refinement (M-step). Found: Under greedy decoding across eight models and six benchmarks, EMO improves the zero-shot baseline in 39 of 40 aggregated settings, averaging $+9.10$ percentage points. By Wei Xia, Yehui Tang, Zhi-Hong Deng. https://arxiv.org/abs/2610.22158 https://arxiv.org/abs/2610.22158 Thu, 15 Oct 2026 16:23:00 GMT arXiv cs.CL Language models arXiv cs.CL — Did: We introduce premise critique ability in NL2SQL, namely the ability to recognize flawed queries and abstain from generating misleading SQL. Found: Stronger critique-oriented prompting improves detection and abstention on flawed queries, but also increases false invalid judgments and suppresses SQL generation on originals. By Shuo Lin, Jinzhe Li, Chenlu Guo, Yanhan Li, Yuan Wu, Yi Chang. https://arxiv.org/abs/2610.21436 https://arxiv.org/abs/2610.21436 Thu, 15 Oct 2026 07:45:00 GMT arXiv (systems & security) cs.CL Language models arXiv (systems & security) cs.CL — Did: We present HyCo, which assigns depth to the server model's native multi-token prediction (MTP) head, executed on the device, and alternative branches to one SLM pass. Found: Without relaxation, HyCo exact saves 27.7% and 11.9% on the two datasets. By Hanbin Bae, Heejin Choi, Joohyung Lee, Byoung Jin Choi, Sangjun Park, Hoon-Young Cho. https://arxiv.org/abs/2610.21333 https://arxiv.org/abs/2610.21333 Thu, 15 Oct 2026 06:20:00 GMT arXiv (systems & security) cs.CL Language models arXiv (systems & security) cs.CL — Did: We introduce QuasiLeak-Bench, a benchmark for policy-conditioned quasi-evidence group recovery, which evaluates whether detectors recover complete quasi-evidence groups and flag the disclosures the given policy prohibits. Found: However, a 4B model fine-tuned on our training split surpasses the strongest prompted model on held-out documents, showing that the task is learnable. By Yufan Zhu, Yangfan Jiang, Yang Fei, Shizheng Hou, Shubhankar Mohapatra, Xiaokui Xiao.