The paper addresses the lack of system‑level confidence estimates in multiagent language model systems such as collaborative reasoning and debate. It introduces confidence composition methods, including confidence‑aware routing and log‑odds pooling, to combine agent confidences while maintaining selective utility and probabilistic reliability. Experiments on five benchmarks with diverse model pairs show that gated‑fusion techniques improve AUARC and Brier scores compared to single‑agent and standard debate baselines, and a shared dependence discount further enhances reliability.
By Ali Elahi, Michael J. Curry, Barbara Di Eugenio
arXiv:2609. 25050v1 Announce Type: new Abstract: We introduce FrontierMath Erd\H{o}s (FME), a benchmark of 68 Erd\H{o}s problems that are open as of August 2026.
By Tom Adamczewski (Epoch AI), Thomas F. Bloom (University of Manchester)
The paper introduces LatentPort, a method that allows a language model to transfer its live memory to another model without requiring the receiver to reread the context. Experiments on a Qwen3.5 4B-to-9B sibling pair show that adding a Gated DeltaNet (GDN) persistent-state package reduces negative log‑likelihood by 0.747 nats/token and improves performance across 64 PG19 documents. The study also demonstrates that direct recurrent and convolution reuse outperforms learned GDN maps, and a 434,176‑parameter correction further narrows the performance gap to the native 9B model.
By Simon P. Villani
ChainDoRA is a new parameter‑efficient fine‑tuning framework for large language models that replaces the dense low‑rank factorization of LoRA with a connected Tensor‑Train (TT) chain. By separating weight magnitude and direction and using a TT rank to control representation capacity, ChainDoRA achieves higher average accuracy on seven commonsense reasoning benchmarks while dramatically reducing trainable parameters—down to 5.35 M versus 56 M for LoRA and DoRA. Ablation studies show that the TT parameterization offers controllable trade‑offs between parameter cost and accuracy.
By Ashfak Yeafi, Mehedi Hasan, Md Khairul Islam
TelecomGPT‑R1 is an open‑source family of unified telecom reasoning models that address the limitations of existing telecom‑specific and general‑purpose LLMs. It is built on a four‑axis framework—protocol, knowledge, modeling, and fault—and trained on a curated corpus of 104,880 verified question‑answer pairs with chain‑of‑thought reasoning. After supervised fine‑tuning, dynamic sampling policy optimization with task‑routed rubric rewards is used to stabilize reinforcement learning across heterogeneous telecom tasks, achieving an 89.64% mean score on seven GSMA Open Telco Leaderboard benchmarks, surpassing leading proprietary models.
By Bohao Wang, Chenwei Wu, Hang Zou, Yu Tian, Lina Bariah, Li Wei, Chongwen Huang, Yongliang Shen, Zhaoyang Zhang, Merouane Debbah
Parallel coding agents can produce patches that work individually but fail when merged, a problem arising when one agent alters an interface or rule another agent depends on. The authors introduce *stale*, a benchmark for semantic coordination, evaluating patches both alone and in combination across synthetic tasks, real Django pull‑request pairs, and constructed tasks. While only one of 417 mined Django pairs showed interference after grading corrections, interference appeared in 97% of runs on constructed tasks, and a message describing the concurrent change recovered 82% of those runs.
By Haocheng Xia, Eugene Wu, Yongjoo Park
The paper investigates how internet slang spreads across Reddit communities by combining social network analysis with linguistic context. Using large language models as scalable annotators, the authors create a benchmark for detecting slang usage and then model its adoption and diffusion. Findings reveal that users with higher bridging capital promote slang spread, while those with higher bonding capital hinder it, and that broader contextual usage delays new user adoption.
By Xiaoning Wang, Ted Underwood, Zhewei Sun
Qwen3.8-Omni-Flash is a natively multimodal agentic model designed for real‑world multimodal productivity, offering enhanced multimodal understanding, reasoning, and long‑horizon agentic task performance. It builds on a sparse mixture‑of‑experts architecture, extends its context window to one million tokens, and supports long‑context multimodal reasoning and planning. The release includes Qwen-MM-Plugins for native audio and video support and Qwen-Live-Harness for building responsive, real‑time multimodal agents, with extensive evaluations confirming strong performance across multimodal tasks.
By Qwen Team
The paper introduces a four‑stage framework—SFT, PG‑CoT, Dynamic, and K‑RL—to improve large language models for Traditional Chinese Medicine prescription generation. It addresses three key gaps: lack of auditable reasoning (SR Gap), failure to adjust prescriptions over time (LA Gap), and non‑enforcement of absolute contraindication rules (SC Gap). Experiments on 12 fine‑tuned models and 6 zero‑shot baselines show that the framework, particularly a 7B Mistral model, outperforms zero‑shot GPT‑5 on all three TCM evaluation metrics.
By Zheng Chen, ZhiCheng Du, Haoxuan Li, Peiwu Qin
A new large-scale Arabic fact‑checking dataset called Arafa has been created using an automated pipeline that generates claims from Arabic Wikipedia, mutates them into counterfactuals, and validates them against supporting or refuting evidence. The dataset contains 181,976 claim‑evidence pairs labeled as supported, refuted, or not enough information, and human evaluation shows high inter‑annotator agreement and strong validation accuracy. Fine‑tuned transformer models on Arafa achieve a Macro F1‑score of 77%, demonstrating its usefulness for Arabic fact‑checking tasks.
By Christophe Khalil, Shady Elbassuoni, Rida Assaf
ClusterFewshot is a new strategy for selecting few‑shot demonstrations in large language model workflows. It combines semantic structuring with utility‑aware scoring to build representative demonstration sets, improving accuracy over prior bootstrap‑based methods. In DSPy‑based pipelines, it substantially reduces optimization cost across multiple benchmarks while consistently outperforming earlier approaches in both standalone prompt tuning and hybrid prompt‑weight optimization.
By Omri Bar Haim, Shahar Katz, Lior Wolf
Informed Masking (IM) is a new technique for aligning Diffusion Large Language Models (dLLMs) with Reinforcement Learning (RL). It identifies a systematic upstream/downstream token structure in dLLM rollouts and shows that masking downstream tokens creates better subproblems for likelihood estimation. When integrated into three state‑of‑the‑art dLLM RL methods on LLaDA‑8B‑Instruct, IM yields up to 2.01%, 8.68%, and 5.77% relative average gains on math and planning benchmarks while improving training stability.
By Xiaoyi Yu, Enver Sangineto, Pei Fu, Fiorenzo Parascandolo, Wenhui Tan, Ruikang Zhang, Rita Cucchiara, Ruihua Song, Jian Luan
The paper introduces TSS, a target-side sparsification framework that selectively skips layers in a target verifier during speculative decoding for domain-specific large language models. By exploring multi-layer skip configurations with an acceptance- and metric-aware breadth search, TSS reduces verification cost, increases draft acceptance, and can even improve downstream task performance without retraining. Experiments on Spec-Bench demonstrate consistent gains across domains and model scales, notably boosting translation throughput by 1.68× and improving BLEU scores significantly.
By Haibo Hu, Lianming Huang, Qiao Li, Nan Guan, Chun Jason Xue
The paper introduces the Differentiable Fuzzy Inference Layer (DFIL), a dual‑path prediction head that pairs a standard classifier with a scalar‑bottlenecked branch using ordered membership functions. DFIL enforces monotonicity in the underlying quantity and enables compositional reasoning via t‑norm operations, addressing failures of standard classifier heads that treat ordinal categories as independent labels. The scalar branch also offers an interpretable interface for analyzing residual errors, and the authors demonstrate DFIL’s effectiveness on ordinal natural‑language tasks across various large language model families.
By Zhen Zhang, Amr Alanwar
The paper introduces PACE-dLLM, an acceleration method for diffusion language models (dLLMs) that uses the model’s own per‑step confidence to estimate a ‘confidence cliff’ and determine the optimal look‑ahead horizon for block decoding. By fitting this cliff in closed form at each step, PACE-dLLM sets the horizon to its saturation point and applies an independent confidence threshold for token commitment, thereby avoiding the trade‑offs inherent in fixed‑size block decoding. Experiments on reasoning and code benchmarks show that PACE-dLLM achieves the best average accuracy on open‑source dLLM backbones while delivering significant wall‑clock speedups—up to 5.23× on LLaDA and 3.06× on Dream—improving the quality‑throughput Pareto frontier.
By Xiaocheng Lu, Shuhan Guo, Ziyue Ma, Jie Zhang, Jian Liu, Jingcai Guo, Haoxuan Che, Song Guo
The paper introduces a semiotics-aware framework for assessing natural language generation, focusing on how well two texts align in terms of contextual meaning and discourse references. It defines two metrics—Semiotic Fidelity and Semiotic Coverage—to quantify how much of one text’s semiotic profile is supported by the other and how much of the other’s profile is recovered. Experiments reveal that coverage is usually lower than fidelity, and that language models align best with human-curated data at low sampling temperatures, with higher temperatures diminishing this alignment.
By Lorenzo Zangari, Davide Picca
The paper argues that calibration—how well a language model’s confidence aligns with its actual correctness—should be a standard evaluation metric for large language models (LLMs). It notes that while calibration metrics exist, they are rarely applied outside specialized NLP subfields, leading to unverified confidence scores in new models, datasets, and benchmarks. The authors highlight the risks of miscalibration both at deployment (overconfident errors causing harm) and in research workflows (affecting LLM-as-a-judge, synthetic data generation, and active learning). They call for every NLP subfield to pair its primary performance metric with a calibration score, treating calibration as an essential property of every model.
By Mario Sanz-Guerrero, Katharina von der Wense
The paper reviews how large language models (LLMs) are employed to study infant syntax acquisition, focusing on the BabyLM challenge that aims for human‑level syntactic performance using developmentally realistic corpora. It critically examines dataset construction, model selection, training procedures, and syntactic evaluation methods, highlighting methodological assumptions that limit the theoretical reach of these studies. The authors find that using developmentally realistic corpora has only modest impact on benchmark performance, pointing to fundamental computational differences between LLMs and actual infant syntax learners.
By H\'elie Bazin (SCAI, SND, ISIR), Anouk Barberousse (SND), Fran\c{c}ois Yvon (MLIA)
The paper demonstrates that frontier language models can be prompted to expose their internal chain-of-thought reasoning via a simple custom tool. By comparing these extracted traces to native reasoning on open-source models, the authors confirm that the externalized reasoning aligns with genuine reasoning and outperforms no-reasoning baselines across math, science, and code tasks. They further analyze the structure of the reasoning, noting token-efficient, directed reasoning in models like GPT‑6 Astra, which externalizes only crucial steps while handling elementary ones internally.
By Xiaoyu Luo, Tao Ren, Wenrui Yu, Xiao Li, Qiongxiu Li, Johannes Bjerva
The paper investigates whether stylometric features extracted only from text can distinguish GPT‑assisted writing from independently authored student work. Using data from 90 participants who produced both types of text, eight machine learning classifiers were tested, with Random Forest achieving an ROC‑AUC of 0.87 and an F1‑score of 0.84 on a held‑out test set. SHAP analysis revealed that lexical and grammatical characteristics were the primary drivers of the predictions, indicating that transparent, text‑intrinsic features can effectively signal GPT assistance.
By Rajesh Kumar, Nabeel Siddiqui, Alexander Fuchsberger