The paper proposes a two‑stage framework that first fine‑tunes a large language model (LLM) and then rectifies its outputs, allocating limited labeled data optimally between the stages. It argues that the usual mean‑squared‑error objective for fine‑tuning misaligns with the downstream rectification, and instead suggests minimizing prediction‑error variance for mean estimation or a scalarized variance metric for general M‑estimation. Empirical results confirm that this variance‑based fine‑tuning, combined with optimal data allocation, yields significant efficiency gains over using either fine‑tuning or rectification alone, or using the conventional objective.
By Zikun Ye, Jinglong Zhao, Lei Wang
The paper proposes that self‑improvement methods for language models—such as debate, bootstrap, and internal coherence maximization—can be unified under a framework called coherence optimization. This framework seeks a context‑to‑behavior mapping that is most compressible and jointly predictable, and the authors prove its equivalence to description‑length regularization. They further show that coherence regularization, when guided by a pretrained model prior, optimizes a lower bound on worst‑case accuracy for semi‑supervised learning, offering a theoretical explanation for the success of feedback‑free self‑improvement and predictions about its limits.
By Tianyi Qiu, Ahmed Hani Ismail, Zhonghao He, Shi Feng
The paper presents a benchmark to test whether vision‑language models can produce plant simulation configurations from images using in‑context learning. It focuses on cowpea plot reconstruction, requiring the models to output structured JSON that includes field and plant details. Open‑source multimodal models from the Gemma 4 and Qwen3.5 families are evaluated on synthetic and real drone datasets, using five in‑context methods, and the results show that while VLMs can generate valid JSON and estimate key agronomic metrics, their performance varies and often lags behind dataset baselines.
By Heesup Yun, Isaac Kazuo Uyehara, Earl Ranario, Lars Lundqvist, Christine H. Diepenbrock, Brian N. Bailey, J. Mason Earles
The paper introduces a new way to detect drift in stateful language‑model pipelines by treating the sequence of prompt, response, and next prompt as a single unit of analysis. It defines two metrics—communication closure and normalized conditional action contribution—to quantify how well a response aligns with the subsequent prompt and how much it resolves the next reply. Experiments on over 2,200 dialogues show that swapping a response drastically reduces measured contribution, indicating that drift can be detected without labels or predefined rules.
By Wael Hafez, Amir Nazeri, Chenan Wei
The paper introduces ShadowMem, a defensive framework that protects large language model agents from long-horizon threats by maintaining a dedicated safety-focused memory. Inspired by the shadow stack concept, ShadowMem stores safety-critical context throughout an agent’s execution and uses this shadow memory to evaluate the risk of upcoming actions before they are carried out. Experiments show that ShadowMem outperforms existing defenses in detection accuracy, detects most attacks early, and adds minimal overhead to agent performance.
By Yuhui Wang, Tanqiu Jiang, Jiacheng Liang, Charles Fleming, Ting Wang
ProteinJEPA introduces a joint‑embedding predictive architecture that supplements masked language modeling (MLM) with a cosine loss to predict latent representations of a teacher model. On 19 protein tasks, MLM+JEPA outperforms compute‑matched and step‑matched MLM‑only training across 78 and 76 of 114 comparisons, achieving notable gains on structure‑ and homology‑sensitive tasks such as SCOPe‑40 retrieval and remote homology. Ablation studies show the cosine loss is superior to mean squared error and that latent prediction complements rather than replaces MLM.
By Dan Ofer, Dafna Shahaf, Michal Linial
The paper presents an AI‑powered triaging agent for banks that uses large language models to conduct multi‑turn conversations, ask relevant questions, and classify customer cases for accurate routing to specialist teams. The system is integrated with policy, safety guardrails, and reasoning frameworks, and its performance is evaluated using synthetic digital twins that simulate realistic, labeled dialogues based on historical data. Results show a 30.6% increase in classification accuracy and high satisfaction from subject‑matter experts, demonstrating the effectiveness of targeted probing for scalable banking operations.
By Alankar Atreya, Stefan Sylvius Wagner, Devesh Batra, Robert Hankache, Cristovao Iglesias Jr, Patrick Sinclair, Giulio Pelosio, Michael McMillan, Greig A. Cowan, Raad Khraishi
The paper introduces a new text‑to‑SQL benchmarking framework that incorporates realistic role‑based access control (RBAC) constraints. It augments existing benchmarks by generating plausible user roles and access policies through an LLM‑assisted workflow, followed by human‑in‑the‑loop quality control. The framework also provides evaluation metrics to detect RBAC‑specific failures and separate SQL utility from compliance, revealing that many high‑scoring models degrade sharply under access constraints.
By Yang Fei, Yangfan Jiang, Yin Yang, Xiaokui Xiao
The paper introduces LLMAE, a technique that transforms a pretrained decoder-only language model into a continuous text autoencoder by inserting a fixed-length latent bottleneck into its internal activations. Using a 270M Gemma 3 model with structured attention masks, LoRA adaptation, and KL regularization, LLMAE achieves near-perfect reconstruction of text sequences up to 1024 tokens. The authors further show that the resulting latent representation can be leveraged to train a latent text diffusion model for detailed image captioning, demonstrating downstream utility.
By Arkanath Pathak, Unnat Jain, Alexander C. Berg
The paper investigates the Platonic Representation Hypothesis, which posits that more capable models converge toward shared representations. By distinguishing relational structure (which samples are related) from metric geometry (quantitative relations like distances), the authors develop a controlled $2 imes2$ framework to evaluate both aspects at local and global scales. Their findings show that relational structure consistently converges across vision‑language and video‑text models, while metric geometry converges much more weakly, a pattern that persists even when using a Riemannian metric approximation.
By Junwon You, Mihyun Jang, Sangwoo Mo, Jae-Hun Jung
The paper introduces the Neurogenesis Network (NGN), a differentiable framework that learns the optimal number of ordered structural components in a neural network during training. By using a learnable boundary to select an active prefix of components, NGN can grow from a compact initialization and later discard unused parts. Experiments across MLPs, CNNs, GNNs, Transformers, state‑space models, LoRA, and adapters show that the learned prefixes perform comparably to fixed‑size models, demonstrating that structural capacity can be optimized directly as a count.
By Lixing Li
The paper introduces a probabilistic‑circuit framework for fusing opinions from multiple black‑box experts in noisy, conflict‑prone environments. It dynamically assigns context‑specific credibility to each expert, allowing reliable aggregation without needing access to their internal models or retraining. Experiments on multiple‑choice question answering with large language models show that this method outperforms individual models and static ensemble baselines, consistently improving predictive accuracy and decision reliability under disagreement.
By Pranuthi Tenali, Sahil Sidheekh, Saurabh Mathur, Vijayalakshmi Saravanan, Erik Blasch, Kristian Kersting, Sriraam Natarajan
The study examines whether accuracy gaps between source and target tasks can certify the failure of scalar recalibration maps for large language model judges. Across thirteen judges, two generators, eight domains, and 1,176 transfers, the accuracy gap only provides a weak lower bound on target calibration error and can predict opposite outcomes. Even with a finite‑sample lower certificate, the method shows low power (0.13) and does not reliably indicate when recalibration will fail.
By Fariya Afrin, Ibne Farabi Shihab
CS-WCP introduces confidence‑set weighted conformal prediction to provide robust prediction sets for large‑language‑model judges when deployment traffic shifts the prevalence of task or policy groups. By constructing simultaneous exact intervals for source and target group masses and taking the union over all compatible ratio vectors, CS‑WCP achieves high coverage (mean 0.973) with few failures across 336 constructed traffic shifts, outperforming standard source conformal prediction. The method offers an auditable coverage safeguard under uncertain mixture weights, focusing on conservative tail protection rather than tighter set sizes.
By Ibne Farabi Shihab, Fariya Afrin
The paper introduces a new language modeling approach that combines the benefits of Transformers and recurrent models by using balanced-tree recursive operators for autoregressive prediction. This method achieves logarithmic depth and linear runtime, allowing all prefix representations to be computed efficiently. Experiments show strong length extrapolation and performance close to ALiBi-based Transformers, suggesting it could serve as a viable alternative architecture for language modeling.
By Yiqin Wang, Nuri Cingillioglu, Charles Pert
arXiv:2609. 28263v1 Announce Type: new Abstract: The growth of large language model (LLM) inference and search services increases the scale of online linear programming problems, motivating computationally efficient algorithms.
By Jiameng Lyu
The paper presents a dual‑matrix computational framework that quantifies morphological and theological differences among 196 Hindu and Vajrayana Buddhist esoteric deities. It uses a Gower distance matrix with a new Cardinality Weighting algorithm for physical form and dense vector embeddings from LLMs for theological function, revealing how visual forms can mask shared functions and how high‑cardinality symbols cluster orthodox and Tantric entities. The study demonstrates near‑identical coordinates for the Hindu Chinnamasta and Buddhist Chinnamunda, and releases the architecture as an open‑source tool for Digital Humanities research.
By Ankit Bhattacharjee
The paper investigates whether panels of vision‑language models (VLMs) can reliably judge image aesthetics. It shows that a panel of holistic judges rarely outperforms its best member, but when each model scores images on five rubric‑defined dimensions and these dimension scores are fused across model families, the panel consistently beats the best single VLM on two datasets (EVA and PARA). The study demonstrates that the value of a panel depends on the type of input it receives, and that dimension‑based fusion yields measurable gains at the cost of additional labeling and API usage.
By Amit Jadhav, Shaurya Beriwala, Beomjin Kim
The paper introduces RECAP, a redundancy-aware credit assignment method that improves reasoning efficiency in large language models by assigning credit to each reasoning step based on its downstream role and contribution to the correct answer. RECAP uses a semantic dependency graph to measure structural responsibility and evaluates step efficacy via changes in gold-answer log-likelihood, enabling step-specific updates without requiring a separate reward model or concise trajectories. Experiments on two 7B models across four mathematical reasoning benchmarks show that RECAP enhances the accuracy-efficiency trade-off, boosting pass@1 by 2.0–3.7 percentage points while cutting reasoning tokens by 8–31% compared to GRPO.
By Yuqing Zhou, Hong Wang, Manqing Mao, Zhuoer Wang, Samson Koelle, Jie Yuan, Yanjun Lin, James Feng, Nikki Lijing Kuang, Ziwei Zhu, Wei Niu
arXiv:2609.28283v1 Announce Type: new
Abstract: Several foundation models dedicated to hyperspectral images have recently been made available. These models are trained on large unlabeled datasets and...
By Edgard Dabier, Christophe Kervazo, Pietro Gori, Florence Tupin