The paper introduces a pre‑action verification framework for large language model agents that emit shell commands or code edits. By running deterministic checks before execution, the system can catch invalid commands (95.8% success with a 10.0% false‑positive rate) and prevent silent failures in code edits, achieving high recall while minimizing false positives. The authors provide benchmarks, verifiers, and guards for both action modalities.
By Asaad Althoubi
QTrans is a quantum transformer designed for small‑scale binary sentiment classification. It constructs query, key, and value features using parameterized quantum circuits and derives attention coefficients from Gaussian distances between quantum measurements. The model incorporates a quantum feed‑forward network, residual connections, and layer normalization, achieving higher accuracies on MR, CR, and MPQA datasets compared to classical baselines.
By Ren-Xin Zhao, Xinjie Huang, Yahong Liu, Maoyu Ye, Jinjing Shi, Shi Wang, Yaonan Wang
The paper introduces a Concept-Integrated Transformer (CIT) that uses a pretrained large language model to generate concept abnormality targets with confidence weights, eliminating the need for manual concept annotation. CIT is applied to mobile sensing data from two longitudinal datasets, achieving the highest F1 score on the AFFECT dataset (0.756) and tying for the highest on a PHQ-9 dataset (0.765). The model’s learned concept scores reveal interpretable behavioral and physiological patterns, such as differences in sleep quantity and quality between high and low negative affect groups.
By Yuning Wang, Iman Azimi, Amir M. Rahmani, Pasi Liljeberg
The paper investigates how block‑diffusion language models can use a constant‑size cache to enable efficient parallel decoding. By employing sequence mixers that summarize completed blocks into a reusable state and a block‑causal training objective, the authors pretrain three 3B block‑diffusion denoisers (attention, Mamba, and hybrid) on 300 B tokens. The resulting state‑space cache remains O(1) in memory and latency regardless of context length, yielding significant speed‑up and memory savings compared to traditional attention‑based caches, especially at very long sequences.
By Vaibhav Singh, Pierre-Andr\'e No\"el, Torsten Scholak, Eugene Belilovsky, Oleksiy Ostapenko
The paper introduces ISO-LoRA, an optimizer that improves rank utilization in Low‑Rank Adaptation (LoRA) by coupling factor updates through spectral descent on the induced tangent perturbation in weight space. Experiments on GPT‑2 adaptation show that standard optimizers like AdamW concentrate updates in a few singular directions, whereas ISO-LoRA distributes energy more evenly, leading to higher effective rank and better downstream performance across 0.1B‑7B models. The authors provide theoretical guarantees under a stylized spiked‑gradient model and demonstrate that ISO-LoRA consistently outperforms factor‑wise optimizers, especially at moderate‑to‑large LoRA ranks.
By Zihan Zhu, Zhehang Du, Xuyang Chen, Tim Tsz-Kit Lau, Jiayuan Wu, X. Y. Han, Qi Long, Weijie Su
GUIDE is a large‑language‑model driven architecture that elicits and infers human user preferences through conversational Bayesian adaptive sampling and symbolic rule‑based learning. It extends adaptive sampling to a wide range of elicitation questions via a flexible type system and initializes domain‑specific preference models using symbolic representations of world knowledge. In simulated investment portfolio optimization, GUIDE outperforms prior methods, LLM‑only baselines, and its own ablated variants by improving cold‑start performance and reducing recommendation regret during early interactions.
By Anagha Tiwari, Alexander G. Gray, Nick Feamster, Brian Jabarian, Alex Imas, Alex Kale
The study examined whether adding patient‑reported survey data to electronic health records (EHRs) improves the prediction of a first opioid use disorder (OUD) diagnosis. Using 267,747 All of Us participants, the authors compared EHR‑only models to EHR+survey models across multiple machine‑learning algorithms and look‑back windows. Survey augmentation consistently increased predictive performance, with the best 24‑month LightGBM model’s PR‑AUC rising from 0.6219 to 0.6603, and survey features ranked as the second most important information domain.
By Xiyue Jiang, Zihan Ding, Grace Han, Yinan Liu, Richard N. Rosenthal, Fusheng Wang
The paper introduces Decision-Flow Sampling (DF‑Sample), a training‑free, data‑free inference framework that builds a hierarchical reasoning tree, evaluates entire trajectories, and back‑propagates utilities to guide branching decisions. Unlike local step‑wise sampling, DF‑Sample explicitly assesses global paths, enabling it to recover high‑quality, low‑probability reasoning chains that standard decoding misses. On the GPQA benchmark, DF‑Sample attains 45.6% accuracy, outperforming power sampling (38.9%) and GRPO (39.9%) and consistently surpassing baselines across multiple models and benchmarks, demonstrating significant latent reasoning potential in pretrained LLMs.
By Zhendong Mi, Shaoyi Huang
ParaRecover is a new process-level benchmark designed to evaluate error localization and recovery in multi-turn parallel tool-use agents. It contains 10,626 instances across two difficulty levels, built on a taxonomy of 14 error types that cover planning dependencies, tool selection, and argument matching. The benchmark introduces the SDE rubric, which assesses structural integrity, diagnostic reasoning, and evolutionary strategy during agent execution, and demonstrates that it can guide improvements in agents’ reflective recovery capabilities.
By Bowen Guan, Zhentao Yin, Yanming Shen
The paper introduces GSF-χ, a graph transformer that incorporates global stereochemical fields to handle chiral molecules. Unlike traditional models that focus on a single atom, GSF-χ modulates all pairwise interactions using stereogenic units, producing a reflection‑even phase field and a handedness pseudoscalar that guide relative rotations in latent query–key blocks. The authors prove the operator’s even–odd decomposition and demonstrate that GSF-χ improves central‑ECD, axial Rotation, and Symbol metrics over strong baselines, while maintaining enantiomer‑pair consistency through a C₂ projection.
By Jiaqing Xie, Yuxin Wang, Xipeng Qiu
InRTL: Effective Intra-Inter Interaction Learning for Relational Tables proposes a unified framework that explicitly models dependencies both within and across relational tables. The approach introduces intra-table interactions to capture associations among rows in the same table and inter-table interactions to capture dependencies between rows linked by primary key–foreign key relationships. It employs a column-aware table encoder, Transformer-based self-attention and cross-attention modules, and incorporates linearized attention and heterogeneous graph neural networks to improve scalability, demonstrating effectiveness across ten datasets and 24 real-world tasks.
By Weichen Li, Ken Zhong, Zheng Wang, Li Pan, Jianhua Li
The paper "Limits of LLM Text Detectors in Education" argues that existing LLM‑generated text detectors assume a binary human/LLM distinction, which fails to capture realistic student‑AI collaboration. It introduces a contribution‑aware evaluation framework with eight student contribution levels and presents GEDE, a benchmark of over 900 human‑written and 12,500 generated essays across 886 tasks. Using GEDE, the authors evaluate four detection methods and find that most detectors perform poorly on intermediate contribution levels, especially LLM‑assisted revisions, raising concerns about false accusations.
By Lukas Gehring, Benjamin Paa{\ss}en
CoHyDE is an iterative co‑training framework that jointly trains a dense encoder and an LLM rewriter for tool retrieval from large API catalogs. The encoder is fine‑tuned with InfoNCE on catalog‑style hypothetical descriptions generated by the rewriter, while the rewriter is preference‑aligned via DPO against the encoder’s retrieval scores. On a 10k‑tool subset of ToolBench, three rounds of CoHyDE outperform the best single‑component baseline by 2.5 pp NDCG@5 on standard queries and 6.3 pp on vague queries, with the largest gains on the hardest vague tier.
By Vaishali Senthil, Ashutosh Hathidara, Sebastian Schreiber
The paper presents a rate‑distortion framework for understanding factual hallucination in closed‑book question answering. It shows that even when a fact is observed, limited memory forces it to be stored approximately, leading to errors that can be bounded by a combination of compression distortion and missing coverage. The authors derive a theoretical lower bound on error and validate it with simulations and probes on modern language models.
By Xi Wang, Shijia Xu, Rongfeng Guo
Chopthin-Consensus Power Sampling (CCPS) is a new inference-time decoding method for large language models that uses the Chopthin resampler to preserve diversity among particle trajectories. By enforcing an upper bound on weight ratios instead of equal-weight resampling, CCPS maintains a richer set of distinct reasoning paths and guarantees a lower bound on effective sample size. Coupled with a semantic-majority selection mechanism, CCPS achieves higher oracle coverage and matches or surpasses baseline accuracy on multiple reasoning benchmarks.
By Minoo Ahmadi, Seyedarmin Azizi, Erfan Baghaei Potraghloo, Mehdi Kamal, Massoud Pedram
The paper introduces SynthSentry, a model‑agnostic method for detecting synthetic data contamination in language‑model training corpora. It computes a distributional divergence score based on lexical diversity collapse, n‑gram tail truncation, and perplexity variance across reference models, requiring no access to the generating model or synthetic labels. Experiments on English corpora contaminated by small open‑weight generators and an instruction‑tuned model show that SynthSentry ranks contamination severity accurately, maintains low false‑positive rates after calibration, and does not degrade downstream fine‑tuning performance at the tested scale.
By Praveen Kumar Myakala, Ravichandra Namburi, Sowmya Keragodu Jayaramu, Sooraj George Thomas
CueMem is a cue‑guided framework for long‑term conversational memory that reconstructs query‑relevant dialogue context from compressed memory records. Instead of treating memory units as self‑contained evidence, it extracts fine‑grained cues linked to their source turns and, at query time, expands from these cues over a turn graph to rebuild a compact evidence context. Experiments on LoCoMo and LongMemEval show that CueMem outperforms baseline memory methods, reduces input tokens and latency, and improves long‑term conversational question answering.
By Changjian Wang, Rongzhen Li, Weili Guan, Shuming Shi, Quan Lu, Ning Jiang
arXiv:2609.12655v1 Announce Type: new
Abstract: Large language model agents are expected to continuously adapt to new tasks and environments over their lifetime by reusing past experience. However, e...
By Yuli Qiu, Yutong Li, Wei Su, Zeming Liu, Wanxiang Che, Heyan Huang, Haifeng Wang, Yuang Guo
arXiv:2609.12915v1 Announce Type: new
Abstract: Large-scale multi-label text classification assigns a small subset of relevant labels to each document from a vocabulary containing thousands or tens o...
By Hui Ye, Jing Zhang, Xiulong Yang, Rajshekhar Sunderraman
arXiv:2609.13058v1 Announce Type: new
Abstract: Reinforcement learning (RL) has become central to post-training of large language models. Recent advances in RL for Mixture-of-Experts (MoE) models hav...
By Hongyi He, Zhenghao Lin, Xiao Liu, Peng Cheng, Yan Lu, Yeyun Gong