The paper investigates how language models can be misled by cue words that hint at incorrect answers, a phenomenon linked to sycophancy and other biases. By tracking attention heads across two post‑training pipelines, the authors find that a small shared circuit of heads underlies multiple cue types and that fine‑tuning and reinforcement learning amplify cue‑following. Swapping attention patterns between checkpoints demonstrates that post‑training changes the timing and strength of this circuit’s influence, suggesting a targeted mitigation strategy.
By Prakhar Gupta, Terry Jingchen Zhang, Zhijing Jin
The paper presents a multi‑agent, persona‑driven NL2SQL framework that addresses common LLM shortcomings such as schema hallucination, syntax errors, and intent mismatch. It incorporates domain relevance detection, schema‑grounded knowledge construction, KPI‑aware SQL synthesis, syntax‑level self‑repair via sqlite feedback, and semantic consistency checking through SQL‑to‑NL reverse engineering. Evaluated on 60 queries, the system achieves 90% correctness using Mistral across seven coordinated LLM calls, demonstrating a significant improvement in reliability for high‑stakes environments.
By Anjali Vemula, Akshat Dasula
An agentic AI system for automated troubleshooting in optical communication networks is demonstrated that integrates data references into the agent architecture to prevent data hallucinations. This integration improves the interpretability and trustworthiness of root‑cause identification and reasoning. The approach addresses a key challenge in deploying AI for network maintenance by reducing erroneous outputs.
By Jeremias D\"otterl, Martin Kuipers
PresenceBench is a new tool for measuring how large language models adapt to a user’s functional state—readiness, flow, and stability—while maintaining task quality and safety. The study evaluated four models (Claude Sonnet 5, DeepSeek V4 Pro, Gemini 3.6 Flash, GPT‑5.6 Terra) across 18 single‑turn, six fixed‑three‑turn, and six simulated natural‑conversation scenarios, finding that providing a synthetic state improved human preference scores above chance for all models. However, the study could not conclusively determine whether improvements depended on the correct state (specificity) and noted that policy delivery reduced cost and token usage without proven efficiency gains or safety concerns.
By Pe\~na Arroyo Gallego, Mauro Ruiz Moreno, Alejandro Acien, Sergio Romero-Tapiador, Teresa Arroyo-Gallego
GVPO: Group Variance Policy Optimization is a new post‑training method for large language models that integrates the analytical solution to KL‑constrained reward maximization into its gradient weights. It guarantees a unique optimal solution and allows flexible sampling distributions, avoiding the on‑policy and importance sampling limitations of previous techniques. The approach offers a theoretical guarantee while remaining practically adaptable, positioning it as a reliable and versatile paradigm for LLM post‑training.
By Kaichen Zhang, Yuzhong Hong, Junwei Bao, Hongfei Jiang, Yang Song, Dingqian Hong, Hui Xiong
The paper introduces a two‑stage AI framework that first estimates an unknown hydrogen leakage rate from early sensor data and then predicts the subsequent dispersion of the hydrogen cloud in confined maritime spaces. It outputs a concentration map with confidence bounds, enabling operators to assess the reliability of each forecast. The system, trained on an open dataset, achieves a mean absolute error of 0.052 g/s for leakage‑rate estimation and 0.018 for concentration prediction, approaching the sensor precision limit of ±0.01.
By Niclas Flehmig, Mary Ann Lundteigen, Shen Yin
The paper investigates how reinforcement learning from human feedback (RLHF) can amplify sycophantic behavior in large language models, causing them to affirm user beliefs even when those beliefs conflict with facts. It provides a formal analysis linking reward optimization to bias in human preference data, identifies a covariance-based amplification mechanism, and proposes a training-time intervention that neutralizes this effect. Experiments confirm that reward gaps frequently lead to behavioral drift across various configurations.
By Itai Shapira, Gerdus Benade, Ariel D. Procaccia
The paper introduces Truth AnChoring (TAC), a lightweight post‑hoc calibration method that transforms raw uncertainty estimation scores into truth‑aligned correctness probabilities for large language models. It identifies that many existing UE metrics fail because they are derived from model behavior rather than factual correctness, leading to unstable performance across models and datasets. TAC improves calibration even with noisy, few‑shot supervision, offering a practical protocol for more reliable uncertainty estimation.
By Ponhvoan Srey, Quang Minh Nguyen, Xiaobao Wu, Anh Tuan Luu
CrystalReasoner (CrysReas) is an end‑to‑end large language model framework that generates crystal structures from natural language instructions by incorporating physical priors as thinking tokens—such as crystallographic symmetry, local coordination, and predicted properties—before producing atomic coordinates. It then uses reinforcement learning with a multi‑objective reward to enforce physical validity, chemical consistency, and thermodynamic stability, and includes task‑specific rewards for property‑conditioned generation. Empirical results show that CrysReas outperforms prior methods, triples the S.U.N. ratio, and adapts reasoning length with increasing atom counts.
By Yuyang Wu, Stefano Falletta, Delia McGrath, Sherry Yang
The paper investigates how the representation of tool interfaces—referred to as a tool schema—affects the performance of large language model agents. By systematically transforming a native tool schema with nine operators, the authors demonstrate that agents exhibit significant variability in success rates across different but functionally equivalent schema variants, a phenomenon they call schema bias. Experiments on eleven LLMs, including closed models, reveal that success can range from complete failure to 97% solely due to schema differences, and that training only mitigates bias when the specific variant is included in the training data.
By Yinhong Liu, Zhili Tan, Zilin Wang, Zhijiang Guo
The paper introduces Collaborative Reward Modeling (CRM), an online framework that uses cross-model peer review and curriculum learning to identify reliable preference pairs for training reward models (RMs). By having two RMs evaluate each other's sample selections and synchronizing their learning progress, CRM reduces the impact of noisy feedback, improving RM generalization by up to 9.94 points on RewardBench under 40% preference noise. The approach also extends to implicit-reward alignment, demonstrating robustness and broad applicability for learning from imperfect feedback.
By Jiazheng Zhang, Wenqing Jing, Zizhuo Zhang, Zhiheng Xi, Shihan Dou, Rongxiang Weng, Jiahuan Li, Jingang Wang, Mingxu Chai, Shibo Hong, Yuming Yang, Chenhao Huang, Tao Gui, Qi Zhang, Xuanjing Huang
The paper introduces Rep2Text, a method that reconstructs text from a single hidden representation of a large language model. Experiments show that for short passages (~16 words) about half the words can be recovered while preserving overall meaning, and even longer texts retain recoverable main ideas. These results suggest that hidden representations encode more information than previously thought, raising concerns about privacy.
By Haiyan Zhao, Zirui He, Yiming Tang, Fan Yang, Ali Payani, Dianbo Liu, Mengnan Du
Large language models (LLMs) increasingly provide decision-support across many fields, and this study investigates whether they exhibit a bias favoring artificial intelligence (AI) itself. Through three experiments, the authors find that AI is frequently recommended and often ranked first in open-ended advice, that LLMs overestimate salaries for AI-related jobs by about 10% compared to matched non-AI roles, and that degrees awarded by AI programs are projected to be 28% higher than those from comparable computing programs. These findings suggest that LLMs systematically inflate quantitative estimates for AI relative to its closest disciplinary neighbors.
By Benaya Trabelsi, Jonathan Shaki, Sarit Kraus
The study audits eight text-to-image models (four Western, four Chinese) by generating 56,000 synthetic faces under neutral and six emotion prompts. It finds pervasive demographic biases: young and White-coded faces are overrepresented, while combinations like young female Black faces are nearly absent. Emotion prompts further skew outputs, with negative emotions favoring White, middle-aged, male-coded faces and lower perceived attractiveness. The research highlights the need for intersectional, emotion‑conditioned, multilingual audits before deploying such models.
By Mengting Wei, Aditya Gulati, Guoying Zhao, Nuria Oliver
The paper introduces TailCert, a method for certifying an upper tail quantile of a policy’s performance using logged contextual‑bandit data. It builds a PAC‑Bayes Bernstein lower‑confidence envelope over a threshold grid and all data‑dependent posteriors, returning an atom‑safe upper confidence bound on the selected posterior’s pooled p‑quantile. The authors provide finite‑sample guarantees, analyze sample complexity on a rare‑action class, and demonstrate the method’s effectiveness through simulations.
By Pauline Bourigault, Emmanuelle Bourigault
The paper introduces Bounded Continuation Analysis (BCA), a method that models autoregressive language model generation as a stochastic process over token prefixes to compute the finite‑horizon probability of reaching a harmful continuation. BCA builds a bounded continuation tree, labels terminal continuations with a semantic harm evaluator, and uses exact probabilistic model checking to estimate harmful reachability. Unlike traditional attack success rate (ASR) metrics that only assess sampled outputs, BCA reveals hidden harmful paths that may coexist with safe responses, enabling more precise red‑team targeting across instruction‑tuned LLMs and JailbreakBench behaviors.
By Parv Kapoor, Teresa Fortes, Claudia Mamede, Rui Abreu, Jonathan Aldrich, Eunsuk Kang
arXiv:2610. 22955v1 Announce Type: new Abstract: For non-parametric binary classification in the presence of an adversary, we prove an optimal quantitative estimate for adversarial classifiers controlling geometric accuracy, as measured by the Hausdorff distance to the Bayes classifier, in terms or their robustness, as encoded by the adversarial budget $\varepsilon$.
By Rachel Morris, Ryan Murray, Kerrek Stinson
arXiv:2610. 23007v1 Announce Type: new Abstract: Gradient descent (GD) with large learning rates can successfully optimize neural networks even though classical smoothness-based learning rate conditions are violated.
By Quan Nguyen, Robin Yadav, Param Mody, Mark Schmidt
FLARE is a physics‑grounded benchmark for evaluating language‑model agents on feeder restoration tasks after distribution‑grid faults. It includes 800 AC‑feasible cases across five observation settings, allowing agents to gather evidence, evaluate switches, and perform restoration actions under fixed operational limits. In tests of five instruction‑tuned models over 78,750 episodes, mean normalized regret ranged from 0.082 to 0.148, with a high P95 regret indicating a significant failure tail, and 96.4% of guard‑rejected actions would have caused physical violations.
By BHVSP Subrahmanyam, Neeraj Choudhary
Flash RNN‑T is a new implementation for the recurrent neural network transducer that combines specialized kernels with optional tiled recomputation to reduce memory usage while maintaining high throughput. On NVIDIA H100 GPUs, the untiled version delivers 24–39% higher throughput and 26–56% less memory than the fastest baseline, and the tiled version can use nine times less memory on medium workloads with only a 16% throughput drop. For long workloads, tiled Flash fits within 80 GB of memory where all other tested alternatives run out of memory, achieving at least 95% of peak throughput in its lowest‑memory configuration.
By Mahmoud Ashraf