The article argues that conversational AI should provide contingent feedback—responses that vary with user behavior and its social consequences—rather than merely seeking user approval and fluency. It highlights how current alignment methods, such as reinforcement learning from human feedback, often produce sycophantic, noncontingent affirmation, which can hinder the development of interpersonal skills, especially in adolescents. The authors propose a framework for evaluating and designing contingent AI, incorporating trajectory-based assessment and social consequence prediction, and call for interdisciplinary research to ensure AI systems positively influence human social learning.
By Scott Compton, Arjun Nagendran
The study examines whether large language models (LLMs) used as conversational search engines for academic literature prioritize papers based on authority signals—such as author prestige, venue, and citations—rather than content. By keeping titles and abstracts constant and manipulating authority metadata across three counterfactual conditions (original, flipped, boosted), the researchers tested eight LLMs in a single-turn, top‑1 recommendation scenario. Results reveal a substantial, directional authority bias that varies across models and is only partially mitigated by prompt-level debiasing, while also highlighting a significant say‑do gap where debiasing instructions suppress authority mentions more quickly than authority-driven flips, leading to underestimation of behavioral bias.
By Uthman Jinadu, Parsa Ghazvinian, Anjila Budathoki, Benjamin M. Ampel, Rajshekhar Sunderraman, Yi Ding
The paper investigates whether giving AI monitors access to the final answer improves their ability to verify reasoning. Using 237 step‑by‑step solutions to physics exam questions, the authors found that answer access mainly helps monitors detect inconsistencies with the final answer rather than independently checking the reasoning. Certification of the answer increased overall accuracy and error localization but reduced the ability to flag critical traces where the answer was correct but the reasoning was flawed.
By Will Yeadon, Sergio Ju\'arez, Paul Mackay, T. J. Dowling, Elise Agra, Oto-obong Inyang, Arin Mizouri, Craig P. Testrow
The paper investigates a privacy‑hallucination tradeoff in differentially private (DP) language models. It shows that DP pre‑training or fine‑tuning increases hallucinations compared to non‑DP models, especially as the privacy budget becomes stricter. The authors attribute this to DP mechanisms flattening output distributions, and demonstrate that controlling the frequency of facts in training data can mitigate hallucination risks.
By Krithika Ramesh, Krishna Pillutla, Danish Pruthi, Anjalie Field
The paper introduces SEAV, a verification‑centric framework for evaluating jailbreak attempts against large language models. SEAV decomposes responses into ordered steps and checks both validity and correctness using LLM‑as‑a‑judge and retrieval‑grounded verification. The method reduces false positives by 14.9 percentage points on a strategic‑dishonesty diagnostic and reclassifies 22.1–51.0% of previously successful jailbreaks as invalid across multiple benchmarks.
By Qilong Wu, Sahil Wadhwa, Pranab Mohanty, Giri Iyengar, Varun Chandrasekaran
The paper presents a camera‑pose‑free stereo endoscopic method for recovering tissue deformation by modeling geometry as a 3D point‑derivative map and deformation as a 3D displacement‑local deformation map. It optimizes inter‑frame deformation in a camera‑centric setting, eliminating the need for camera pose estimation, and introduces a canonical map for online geometry and deformation optimization. Experiments on in‑vivo and ex‑vivo laparoscopic data show accurate 3D reconstruction (≈0.37–0.39 mm surface distance) even under occlusion, and the method can estimate surface strain distributions during manipulation.
By Jiahe Chen, Naoki Tomii, Ichiro Sakuma, Etsuko Kobayashi
The study compares three AI interaction designs for learning nuclear safety protocols: an unrestricted conversational bot, a Socratic hint‑guided bot, and an adaptive tutoring system that uses brain‑derived cognitive engagement. Fifty participants, with no prior knowledge, completed a video lesson, pre‑test, AI‑driven assessment, and post‑test. Results showed the unrestricted bot yielded the highest learning gains, the adaptive system produced the greatest EEG engagement, and usage patterns revealed that unrestricted users mainly sought direct answers while Socratic users initially tried reasoning before disengaging.
By Alexandre Clin Deffarges, Nataliya Kosmyna, Pattie Maes
The paper compares the sustainability of native small language models (SLMs) versus large language models (LLMs) compressed via post‑training quantization for edge AI deployment. Using a Holistic Sustainability Score (HSS) that balances capability, efficiency, and safety, the study evaluates 30 configurations across five benchmarks, latency, VRAM, energy, and harmful‑prompt robustness. Results show that optimized quantized LLMs can outperform SLMs overall, while SLMs remain competitive due to lower resource demands, challenging the assumption that native SLMs are always the most sustainable choice.
By Jainil Dharmil Shah
The paper investigates how marketplace guardrails affect welfare in language‑model agent simulations of hotel transactions. It finds that initial reports of large welfare gains disappear when controlling for offer schemas and buyer choice, and that guardrails mainly redistribute rather than increase welfare unless sellers are explicitly forced to produce inefficient bundles. The authors propose a construct‑validity framework to flag invalid or inconclusive policy claims before they are reported.
By Peiying Zhu, Sidi Chang
ReNFT is a method that repairs mode collapse in diffusion generators after reward post‑training by internally recalibrating probability mass. It identifies suppressed alternatives through anti‑hub prompts and uses two policy‑dominated routes to generate counterfactual proposals, then applies reward‑based ranking and a joint‑and‑paired NFT update to restore diversity while preserving reward. Experiments on PickScore and GenEval show that ReNFT retains almost all of the original reward while significantly boosting diversity metrics.
By Yuchen Bao, Chao Wen, Haowei Wang, Ruoxin Chen, Donghao Luo, Jiahui Zhan, Wenjian Huang, Shen Chen, Yiting Wang, Taiping Yao, Chengjie Wang, Shouhong Ding, Jianguo Zhang
The paper introduces AVA, a framework that tests whether general NLP embeddings can differentiate logic-sensitive relational semantics in ontologies and knowledge graphs. AVA uses 171,007 contrastive triplets from 163 ontologies, each containing an ontology statement, a paraphrase, and a hard negative with contradictory meaning. Evaluation of over 25 embedding models shows significant limitations, with the best model achieving only 0.739 triplet accuracy and 0.135 for hard negatives; fine‑tuning helps but does not transfer well to downstream Semantic Web tasks.
By Hamed Babaei Giglou, Jennifer D'Souza, S\"oren Auer
The paper introduces Fed‑LSVI, a federated online reinforcement learning algorithm that uses linear function approximation in episodic Markov decision processes. It achieves a regret bound of ≥O(√{Md^3H^4T}) while only exchanging compressed sufficient statistics, thereby meeting privacy constraints. The method reduces communication cost to logarithmic in the number of episodes, a marked improvement over previous approaches that required linear communication.
By Zihang Liang, Haochen Zhang, Lingzhou Xue
The paper demonstrates that an external observer can identify the type of workload running on an NVIDIA H200 GPU by analyzing its power draw, distinguishing training, inference, and non‑AI tasks with high accuracy. Using 930 recorded traces, the authors achieve 97% accuracy and a macro‑averaged F1 score of 0.955 on unseen model families. They also test four evasion strategies to disguise training as inference, showing that a hardened detector can catch most attacks, though one strategy (LoRA) remains partially detectable.
By Simone Gargiulo, Gabriel Kulp
The paper introduces WM‑RMoE, a World Model‑based Risk‑aware Mixture‑of‑Experts framework for autonomous overtaking. It uses a learned latent dynamics model to perform multi‑step rollouts, evaluating cumulative risk at the trajectory level, and employs a hierarchical gating mechanism to coordinate long‑, short‑horizon, and rule‑based safety experts. A Gaussian Mixture Model preserves multimodal maneuver branches, improving robustness and preventing behavioral averaging, leading to better safety compliance, decision stability, and generalization in experiments.
By Yongzhi Liu, Sunan Zhang, Jinchang Xu, Jiawei Wang, Yushu Qiu, Chen Lv, Weichao Zhuang
EvoFlint is an evolutionary atlas that maps multi‑turn vulnerabilities in large language models by treating attack discovery as a search problem rather than a generation task. It uses evolutionary quality‑diversity search to evolve phased conversation plans, employing Pareto fitness for success rate and severity, novelty search for diversity, and a generation‑level memory to incorporate model insights. The resulting risk‑indexed archive, tested on HarmBench, shows high attack success rates across models such as Claude Sonnet, GPT‑5, and Qwen3, revealing which harm categories each model’s safety training covers or misses.
By Feitong Qiao, Liren Peng, Shiming Ren, Aishwarya Jadhav, Arghavan Bahadorinejad, Marinette Chen, Muhan Zhang, Abdulaziz Suria, Gennevi Lu, Anish Das Sarma
The paper proposes the interlingua hypothesis, suggesting that large language models translate by encoding a source sentence into a latent, task‑agnostic feature space and then decoding a target sentence from that space. Three lines of evidence support this: (1) BLEU variance across language pairs is largely explained by language‑specific competences without pair‑specific interactions; (2) many model components influence both monolingual and translation tasks; and (3) fine‑tuning on monolingual data recovers most translation gains seen with aligned documents. These findings converge to support the hypothesis and point toward new ways to understand and improve LLM translation.
By Jacob Brinton, Jannik Brinkmann, Mark Crovella, Aaron Mueller
The paper introduces FedRoRA, a federated learning framework that combines Low‑Rank Adaptation (LoRA) with rank‑heterogeneous personalization. It separates model adaptation into shared global directions and client‑specific rank‑wise magnitudes, using SVD on the server to extract a global subspace and a personalized projection with top‑k selection for each client. Experiments on natural language understanding and generation tasks show that FedRoRA outperforms existing state‑of‑the‑art methods.
By Lei Wang, Jieming Bian, Letian Zhang, Jie Xu
The paper introduces DP-TabImage, a framework that synthesizes paired table–image data while preserving differential privacy. It factorizes the joint distribution into a private probabilistic graphical model for tables and a table‑conditioned diffusion model for images, using DP‑SGD and a privacy‑preserving warm‑up strategy. Experiments on three real datasets demonstrate a strong balance between table fidelity, image fidelity, and cross‑modal alignment.
By Kai Chen, Josephine Lamp, Somesh Jha, Tianhao Wang
The paper proposes a fine‑tuning‑free listwise selector for Text‑to‑SQL systems that replaces traditional learning objectives with inference‑time strategies. It introduces reusable structured memories (MaP‑SQL) that encode mappings from natural language to schema elements, SQL operations, and expected outputs, and uses these memories to evaluate candidate queries. To reduce positional bias, the method aggregates rankings across multiple input permutations, optimizing inference cost through execution results and pointwise scoring. The approach achieves higher selection accuracy, fewer unnecessary comparisons, and outperforms the prior state‑of‑the‑art R^3‑SQL on the BIRD‑dev benchmark while using fewer tokens.
By Yeonseok Jeong, Soyoung Yoon, Seongjun Lee, Seung-won Hwang
The paper proposes Causal Evidentiary Governance (CEG), a framework that requires regulated institutions to maintain a versioned directed acyclic graph (DAG) separating allowable from disallowed causal pathways in high‑risk machine learning systems. CEG introduces the Causal Harm Rate to quantify prediction variation due to disallowed pathways and pairs each decision with a signed Decision‑Evidence Packet (DEP) that cryptographically links the prediction to the DAG and path‑specific attributions, enabling efficient inclusion proofs via a Merkle tree. Empirical validation on synthetic credit data and the German Credit dataset demonstrates that CEG more clearly isolates causal effects than traditional fairness metrics and that a proof‑of‑concept implementation shows operational feasibility with manageable performance tradeoffs.
By Samah Kareem, Bar{\i}\c{s} \c{C}elikta\c{s}