The paper introduces MIOH, a fine‑grained benchmark for evaluating object hallucination in multimodal large language models (MLLMs) across multiple images. It assesses hallucination through four tasks—existence, counting, attribute, position—using three reasoning patterns (comprehensive, comparative, selective) and three adversarial pressures (visual context scale, perceptual difficulty, contextual bias). Evaluation of 29 models, including GPT‑5 and Gemini‑2.5‑Pro, shows distinct failure patterns, indicating that hallucination arises from integration‑stage limitations rather than just perceptual errors.
By Joonki Min, Chaeyun Kim, Hyungwook Choi, Yejin Kim, Kihyun Kim, Yohan Jo, Joonseok Lee
The paper introduces Distributionally Robust Conformal Safety Screening (DR‑CSS), a policy‑agnostic framework that uses historical data and a nominal simulator to pre‑deploy safety checks for new voltage control policies in active distribution grids. DR‑CSS constructs conformal safety intervals around simulated voltage trajectories, enlarging them to account for closed‑loop effects of the new policy. Experiments on IEEE 33‑bus and 141‑bus systems show that DR‑CSS successfully identifies all unsafe scenarios while adapting intervals to reduce false warnings.
By Sarra Bouchkati, Petros Ellinas, Adriana Geisler, Steffen Kortmann, Johanna Vorwerk, Spyros Chatzivasiliadis, Andreas Ulbig
The paper introduces the concept of rational value risk, a gap between a large language model’s deployed reasoning strategy and an ideal rational strategy that maximises utility. It shows that even well‑aligned models can fail to achieve the aligned value during reasoning, and decomposes the estimation error into bounded prompts, bounded responses, and imperfect verifiers. Experiments across several LLMs and benchmarks confirm that rational value risk is common, can be mitigated but not eliminated by value alignment, and that techniques like self‑consistency and longer chains of thought can improve rationality with diminishing returns.
By Kejiang Qian, Fengxiang He
The paper introduces FAB, an attack that uses meta‑learning to embed dormant adversarial behaviors into large language models (LLMs). These behaviors remain inactive until the model is finetuned by downstream users, at which point the model can exhibit unwanted actions such as unsolicited advertising, jailbreakability, or over‑refusal. FAB is shown to be effective across multiple LLMs and resilient to various finetuning settings.
By Thibaud Gloaguen, Mark Vero, Robin Staab, Martin Vechev
The paper investigates the issue of self‑bias when large language models (LLMs) generate and evaluate their own benchmarks. Using machine translation as a testbed, it shows that LLMs as both test‑set creators and evaluators produce model‑specific, homogeneous outputs that inflate their own scores, even when diversity controls are applied. The bias is strong enough that each model ranks itself first, overriding peer consensus, and the phenomenon also appears in open‑ended generation tasks.
By Wenda Xu, Sweta Agrawal, Vil\'em Zouhar, Markus Freitag, Daniel Deutsch
AdaFuse is an adaptive ensemble decoding framework for large language models that dynamically selects fusion units during generation. It uses an uncertainty-based criterion to decide when to ensemble, applying a diversity-aware scaling strategy in uncertain states while continuing direct generation when confident. Experiments on question answering, arithmetic reasoning, and machine translation show AdaFuse outperforms strong baselines with an average relative improvement of 6.88%.
By Chengming Cui, Tianxin Wei, Ziyi Chen, Ruizhong Qiu, Zhichen Zeng, Zhining Liu, Xuying Ning, Duo Zhou, Jingrui He
The study investigates how post‑training of large autoregressive language models (ARMs) into masked diffusion models (MDMs) affects their internal computation. Across two 7B ARM‑MDM families and four diagnostic tasks, the authors find that MDMs retain much of the ARM’s high‑attribution pathways on prefix‑dominant tasks, but reorganize computation toward earlier layers on globally constrained tasks. Component‑level probes reveal that ARMs depend on sharply specialized components, whereas MDMs show weaker specialization and more diffuse output‑space alignment.
By Injin Kong, Hyoungjoon Lee, Yohan Jo
The paper introduces Personalized Group Relative Policy Optimization (P‑GRPO), a new alignment framework for large language models that separates advantage estimation from immediate batch statistics. By normalizing advantages using preference‑group‑specific reward histories instead of the concurrent generation group, P‑GRPO maintains contrastive signals for distinct user preferences. Experiments across various tasks show that P‑GRPO converges faster and yields higher rewards than standard GRPO, improving alignment with heterogeneous human preferences while preserving general capabilities.
By Jialu Wang, Heinrich Peters, Asad A. Butt, Navid Hashemi, Alireza Hashemi, Pouya M. Ghari, Joseph Hoover, James Rae, Morteza Dehghani
The paper introduces T-MAP, a trajectory‑aware evolutionary search technique designed to red‑team large language model agents by exploiting vulnerabilities that arise during multi‑step tool execution. Unlike traditional methods that focus on harmful text, T‑MAP uses execution trajectories to generate adversarial prompts that bypass safety guardrails and achieve harmful objectives through actual tool interactions. Experiments across various Model Context Protocol environments show that T‑MAP outperforms baseline methods in attack realization rate and remains effective against advanced models such as GPT‑5.2, Gemini‑3‑Pro, Qwen3.5, and GLM‑5.
By Hyomin Lee, Sangwoo Park, Yumin Choi, Sohyun An, Seanie Lee, Sung Ju Hwang
The paper introduces Adaptive Doubly Robust (ADR), an off‑policy evaluation method for ranking policies that blends adaptive importance weighting with reward regression to reduce variance. ADR is unbiased when the true user behavior model is known and, under a sufficient condition, achieves lower variance than the prior Adaptive Inverse Propensity Scoring (AIPS) approach. Experiments on synthetic data show that ADR consistently improves mean squared error over AIPS and other ranking OPE estimators across various data sizes and ranking lengths.
By Kosuke Iguchi, Ren Kishimoto
The paper introduces SMOTE‑VAR, an uncertainty‑aware oversampling technique that uses a Gaussian process variance function to filter out invalid synthetic minority samples. Applied to a university student depression dataset, SMOTE‑VAR outperforms existing oversampling methods in predicting remission outcomes. By more accurately identifying students unlikely to remit, the method supports earlier, personalized interventions.
By Dang Nguyen, Arun Kumar A V, Taylor A. Braund, Wu Yi Zheng, Debopriyo Bal, Leonard Hoon, Jill Newby, Helen Christensen, Svetha Venkatesh, Alexis Whitton, Sunil Gupta
The paper presents a hybrid Convolutional‑Transformer model for forecasting Antarctic sea ice concentration (SIC) on a monthly basis. It combines convolutional encoding for spatial features with factorised self‑attention to capture spatio‑temporal dependencies, and introduces two seasonal prior mechanisms: a month‑aware positional encoding and a seasonal temporal bias. Experiments show the model outperforms convolutional and recurrent baselines, with ablation studies confirming the benefits of the seasonal priors for both short‑ and long‑horizon predictions.
By Danyang Li, John Taylor, Thang Bui, Quanling Deng
The paper revisits the Kolmogorov–Arnold representation theorem (KART), which has gained renewed interest through its use in neural networks such as Kolmogorov–Arnold Networks (KANs). It addresses the open question of KART’s stability when the hidden layer is subjected to continuous adversarial perturbations, specifically bounded translations. The authors present a constructive proof of an approximate representation that uses fixed, piecewise‑linear inner functions and a single outer function that remains invariant across all summands, provided the maximum translation bound is known in advance.
By Sviatoslav V. Dzhenzher
Generative Translation Priors (GTP) is a Bayesian framework that repurposes diffusion-based image-to-image translation models as cross-modality priors for ill‑posed imaging inverse problems. It incorporates target‑domain measurements via likelihood guidance, steering the translation toward the desired posterior distribution. The authors analyze the resulting posterior dynamics, identify an intrinsic bias from likelihood guidance, and propose a ground‑truth‑free metric to estimate this bias, leading to two discretized GTP algorithms that demonstrate high‑fidelity reconstruction in CT and PET imaging using complementary side information.
By Evan Bell, Jiaming Liu, Yifan Chen, Yu Sun
EvoLen is a tokenization method for DNA language models that incorporates evolutionary information to prioritize functional sequence patterns such as regulatory motifs. It groups DNA sequences by cross-species evolutionary signals, trains separate BPE tokenizers for each group, merges vocabularies with a rule that favors preserved patterns, and uses length-aware decoding with dynamic programming. Experiments show EvoLen better preserves functional motifs, differentiates genomic contexts, and aligns with evolutionary constraints while matching or surpassing standard BPE on various DNALM benchmarks.
By Nan Huang, Xiaoxiao Zhou, Junxia Cui, Mario Tapia-Pacheco, Tiffany Amariuta, Yang Li, Jingbo Shang
CoCoA (Context-Conditional Cultural Alignment) is a framework designed to mitigate cultural bias in large language models by learning context-conditional behavior. It trains on entity pairs under both culturally cued and neutral contexts, using a contrastive alignment objective combined with calibration, drift regularization, and goal-aware gradient reconciliation. Evaluations on CAMeL and Camellia across ten languages and four LLMs show that CoCoA reduces the Cultural Bias Score from 43 to 24 on average while keeping near-neutral preferences at 50.2, with minimal impact on general performance.
By Kyungdon Lee, Wei Xu, Alan Ritter, Dong-Ho Lee, JinYeong Bak
GPAgentBench-2K is a new benchmark that evaluates large language model agents in primary‑care clinical decision‑making using a constrained Markov Decision Process (CMDP). It is built from expert‑validated records of real GP encounters and includes six foundational clinical actions, a workflow prior, and safety‑informed abstention as an outcome. Testing 16 state‑of‑the‑art LLMs shows that performance drops as the action space grows and that even the best models violate safety constraints in more than half of high‑risk cases, highlighting a gap between clinical quality and safety.
By Boqi Chen, Xudong Liu, Yunke Ao, Heejin Do, Jianing Qiu
In pointwise document reranking, Chain-of-Thought models consistently underperform direct scoring models, a gap that persists across models up to 32B parameters. The authors conducted stress tests—reinforcement learning, fine‑grained supervision, and architectural decoupling—to address classification, score polarization, and calibration issues, which improved absolute scores but did not close the relative ranking gap. Their results indicate that routing continuous relevance semantics through discrete text imposes a stable bottleneck in ranking resolution that current standard methods cannot easily overcome.
By Xiaoyang Chen, Jie Liu, Haijin Liang, Haibo Shi, Jin Ma, Ben He, Yingfei Sun, Dezhi Ye
WildSEEK is a new dataset of 3,000 real user information‑seeking queries, manually annotated for risk‑sensitive domains and whether the query is factoid or analytical. The accompanying evaluation framework tests LLM responses against four failure criteria—sycophantic behavior, overreliance, a default US‑centric perspective, and poor handling of vulnerable populations—finding higher failure rates for analytical queries. The authors also train classifiers on WildSEEK to analyze over 1.8 million realistic queries, revealing that more than a third are high‑risk and often analytical.
By Tanise Ceron, Joachim Baumann, Elisa Bassignana, Berat Cabuk, Dirk Hovy, Debora Nozza
The paper introduces a pragmatics-inspired taxonomy for evaluating how large language models (LLMs) refuse unsafe or inappropriate requests. By applying this framework to 16 modern LLMs across 14 harm categories, the authors find that while refusals are generally explicit and morally charged, they often lack interpersonal facework and instead offer safer alternatives, which can be problematic in sensitive contexts. The study argues for alignment evaluations that assess not just whether LLMs refuse, but how they do so in a contextually adaptive and socially responsible manner.
By Ruoxuan Li, Pinqiao Wang, Sheng Li, Cameron Robert Jones