The paper introduces a scope‑conditioned generation framework that incorporates structured stereotype characteristics into prompts for large language models, aiming to improve the quality of counterspeech against online hate speech. The authors validate the method on a new, human‑curated dataset in English, Italian, and Spanish, showing significant gains over generic baselines in factuality, specificity, cogency, and effectiveness for both explicit and implicit stereotypes.
By Greta Damo, Elias Urios Alacreu, Elena Cabrio, Paolo Rosso, Serena Villata
HoloAegis is a minimally parametric topological inference framework that uses frozen representations to map text onto the unit sphere and makes decisions via Gibbs‑Boltzmann free‑energy differences over pre‑computed anchor centroids. On a frozen three‑benchmark protocol, it matches WildGuard‑7B on toxicity, outperforms it on harmful behaviors, but underperforms on oversafety detection, while ShieldGemma‑2B fails on indirect harms. The study demonstrates that geometric guardrails can substitute for LLM judges in some cases and must defer to them in others, with anchor banks reducing score variance and boundary displacement.
By Tak Ho Alex Li, Kaijie Liu, Lik-Hang Lee, Kin Chung Ho, Ping Shum, Michael K. Ng
KaiNinja extends the native 3D generator TRELLIS.2 to produce part-level meshes by introducing a dual‑volume representation that overcomes the single‑sheet limitation of the O‑Voxel grid. It maintains TRELLIS.2’s speed and quality while eliminating the need for external segmentation, and is trained on diverse data including CAD models and assets created by an LLM‑driven agent. The method improves whole‑object fidelity and outperforms other part‑generation pipelines, reducing Chamfer distance by 40% and increasing strict part F‑score by 16%.
By Ruihan Yu, Lian Fu, Muyao Niu, Zheng-hui Huang, Yu-Ju Tsai, Sho Kuno, Fengbo Lan, Yonghao Yu, Erwin Wu, Ming-Hsuan Yang, Kaipeng Zhang, Zhixiang Wang
The paper investigates intersectional biases in multimodal clinical predictions using Electronic Healthcare Records (EHR). It introduces datasets MIMIC-Eye1 and MIMIC-IV ED, applies unified text representations from pre‑trained clinical language models, and benchmarks bias mitigation at the intersectional subgroup level. Results show that subgroup‑specific mitigation is robust across datasets, subgroups, and embeddings, effectively addressing intersectional biases in multimodal settings.
By Ayaazuddin Mohammad, Kishore Sampath, Resmi Ramachandranpillai
The paper re‑examines the impact of removing the encoder from Action Chunking Transformers (ACT), a model used for robot manipulation learning. Contrary to the original claim that encoder removal drops success rates from 35% to 2%, the authors find no such dramatic effect in their re‑runs, though minor variations remain uncertain. They attribute discrepancies to training length and checkpoint selection, and note that the encoder’s latent variable offers little reconstruction benefit on the tested benchmark, while its removal speeds up training.
By Bo Kang
PunGraph is a retrieval‑enhanced knowledge‑graph framework designed to improve pun understanding by combining phonetic and semantic information. It builds a phonetic‑semantic lexical graph using the Unisyn phonetic dictionary, IPA and G2P representations, and WordNet definitions, and uses this graph to retrieve candidate words or senses that constrain large language model reasoning. The authors also introduce WebPun, a new dataset of 5,730 annotated heterographic and homographic puns, and demonstrate that PunGraph consistently boosts the performance of small‑scale LLMs on SemEval‑2017 and WebPun, achieving results competitive with strong proprietary models.
By Yuchen Su, Zijian Huang, Yaotian Shi, Shaoxin Zhong, Ruofan Wang, Mengze Li, Yonghua Zhu, Diana Benavides-Prado, Michael Witbrock
The paper introduces Token Importance-Aware Policy Optimization (TIAO), a reinforcement learning approach that improves text summarization by weighting token importance based on token dependency. TIAO reweights a trajectory’s advantage according to the overall dependencies of core tokens, addressing the limitation of previous methods that treat all tokens equally. Experiments demonstrate that a 7B foundation model enhanced with TIAO achieves performance comparable to GPT‑4 and GPT‑5‑nano on real‑world datasets.
By Qixiu Li, Chenlong Bao, Xiang Zhu, Xiaoyong Li, Ruixin Cao, Shukai Chen, Zhenxiong Zhou
MDN-Control is a training‑free framework for multi‑subject video editing that jointly controls target localization, occlusion geometry, and appearance initialization. It uses mask‑guided localization for consistent target placement, depth‑aware occlusion control to resolve boundaries between overlapping subjects, and noise‑latent prompting to retrieve Gaussian initializations from a noise library for prompt‑relevant priors. Experiments on MSVBench show that MDN‑Control achieves the lowest CM‑Err and the highest Q‑Edit while maintaining competitive text alignment and temporal consistency.
By Jiayi Yu, Xi Ye, Lina Wang, Yunkun Xia
arXiv:2512.15708v2 Announce Type: replace
Abstract: Foundation models are vital tools in various Computer Vision applications. They take as input a single RGB image and output a deep feature represen...
By Leo Segre, Or Hirschorn, Shai Avidan
TecoPrompt introduces a closed‑loop robust prompt‑learning framework that mitigates label noise in vision‑language models. It uses optimal transport in the CLIP semantic space to generate globally consistent pseudo‑labels, then verifies their reliability by checking trajectory stability over a K‑epoch window and an EMA‑based confidence gate. The verified labels are incorporated into prompt training via a tri‑group objective, yielding significant accuracy gains across multiple noisy datasets, such as a 0.843 accuracy on OxfordPets with 50% asymmetric noise.
By Zeyi Shao, Haowen Hua, Jiaxin Zhang, John See, Zeyd Boukhers, Cong Yang
The paper introduces Turn-level Multiscale Density Ratio Estimation (tlm‑DRE), a post‑training technique for large language model agents that assigns varying weights to different turns and employs asymmetric token‑level training based on positive‑negative gaps across multiple turns. Unlike conventional alignment methods such as PPO, DPO, DIL, and GRPO, which mainly target single‑turn tasks, tlm‑DRE is designed for complex multi‑turn reasoning. Experiments on diverse agent benchmarks demonstrate that tlm‑DRE competes with traditional alignment methods and enables LLMs to perform robustly in both in‑domain and out‑of‑domain multi‑turn tasks.
By Zishuo Zhao (Alibaba Group), Kai Chen (Alibaba Group), Ao Li (Alibaba Group), Yuan Liu (Alibaba Group)
TARC (Time‑Adaptive Robotic Control) is a reinforcement‑learning framework that lets a policy predict both a control action and how long it should be applied, thereby learning temporally extended actions. By optimizing task performance under constraints on the number of control switches, TARC can adapt its control rate online, using high‑frequency feedback only when necessary. Experiments on a high‑speed RC car, a Unitree Go1 quadruped, and a vision‑language action model show that TARC matches the performance of high‑frequency discrete‑time controllers while operating at less than half their control frequency.
By Arnav Sukhija, Lenart Treven, Jin Cheng, Florian D\"orfler, Stelian Coros, Andreas Krause
HUMAID-NER is the first named entity recognition dataset built on the HumAID benchmark, comprising 60,000 English disaster tweets with approximately 175,000 labeled entity spans across ten operationally motivated entity types. The dataset was created using a reproducible three‑stage hybrid pipeline that combines a spaCy transformer model, disaster‑domain EntityRuler patterns, and structured regular expressions with priority‑based overlap resolution. A joint multitask learning framework using a shared RoBERTa‑large encoder and homoscedastic uncertainty weighting achieves an NER span micro‑F1 of 0.841 and classification macro‑F1 of 0.761, and the authors provide a real‑time web dashboard, dataset, models, and pipeline code for reproducibility.
By Aijaz Ali, Nazish Basir, Sarfaraz Nawaz, Danish Nazir Arain, Haris Ali
The study examines whether domain‑adaptive continued pretraining (DAPT) on a learner‑writing corpus (EFCAMDAT) can enhance transformer‑based automated essay scoring (AES) for English proficiency tests. Researchers applied DAPT to BERT, RoBERTa, and DistilBERT and compared the adapted models with their original checkpoints on the FCE and IELTS datasets, evaluating both in‑domain scoring and few‑shot cross‑dataset transfer. Results show that full‑corpus DAPT yields mixed effects, while proficiency‑specific DAPT often outperforms full‑corpus DAPT and sometimes even the non‑adapted baseline, though benefits vary by proficiency composition and encoder architecture and do not consistently transfer across tests.
By Duy Anh Nguyen
The paper investigates whether large language models (LLMs) can enhance neural distinguishers, a cryptanalysis technique that uses machine learning to recover secret keys from plaintext–ciphertext pairs. Experiments on SPECK-32/64 show that LLM-based distinguishers do not outperform traditional ResNet models, that difference choice loses effectiveness at higher rounds, and that incorporating XOR operation results into the prompt significantly boosts LLM performance.
By Tatsuya Sakagami, Masashi Hisai, Naoto Yanai
EviSI is an evidence‑based evaluation agent for low‑latency simultaneous speech‑to‑speech translation. It combines Multidimensional Quality Metrics with interpreter‑developed criteria, using shared source evidence to assess four dimensions—Anchor, Event, Logic, and Fluency—while deduplicating verified errors before scoring. On English‑to‑Chinese and Chinese‑to‑English data, EviSI’s rankings correlate strongly with human judgments, outperforming BLEU and COMET, and its multilingual extension maintains these correlations across five language directions.
By Ben Yan, Zongyao Li, Xiaoyu Chen, Daimeng Wei, Weidong Liu, Huan Zhao, Chong Li, Yaode Wang, Yuzhe Shang
The paper investigates how rubric-based grading in medical AI benchmarks can fail when the rubrics themselves are flawed. Using the RIFT taxonomy, the authors analyze two clinical benchmarks—HealthBench Professional and LiveMedBench—and discover significant issues: nearly 30% of criteria on HealthBench Professional are non-atomic and over 65% are misaligned or rigid. They demonstrate that correcting bundled criteria can shift scores by up to 15.9 percentage points, revealing that rubric flaws materially affect evaluation outcomes.
By Prithvi Dixit, Pedram Hosseini
The paper introduces Relation‑Orbit, a symmetry‑aware method for aggregating likelihoods from frozen vision‑language models to verify fine‑grained left‑right claims. It uses a closed‑form contrast that assigns eight normalized likelihoods based on reflection, inverse relation, and entity exchange, and asserts a claim only when the signed contrast exceeds a threshold chosen via Clopper‑Pearson bounds. Experiments on VSR, GQA, and LLaVA‑1.5/COCO show that Relation‑Orbit achieves higher mean test coverage at a 10% selective‑risk calibration target compared to an all‑eight Orbit‑Max baseline across multiple dataset‑backbone settings.
By Zhouzhi Xiong, Chuxi Zhang, Weizhen He, Yi Chen, Qi Li, Donglian Qi
SlotDiT introduces a text-guided Diffusion Transformer that operates in a slot-based latent space, decomposing scenes into object-centric slots and autoregressively denoising future slot trajectories to predict scene dynamics. The model is conditioned on a reference image and a language instruction, enabling it to generate video content that reflects both visual context and textual guidance. Experiments comparing slot-based representations to VAE-based and semantics-aligned alternatives show that SlotDiT achieves competitive video generation quality while improving task-completion rates across four robotic datasets and offering a more computationally efficient latent representation.
By Gjergj Plepi, Sven Behnke
The paper introduces ImpossibleRubrics, a benchmark of 169 impossible tasks designed to test the robustness of language‑model‑generated rubrics as reward signals. Each task is paired with a verifiable oracle certificate that defines what constitutes an honest answer, and the benchmark includes 48 answerable controls. Experiments show that many rubric generators are exploited frequently—up to 36% on a stress cut—highlighting a significant gap in rubric quality rather than task difficulty, and that generic rubrics can be more vulnerable than tailored ones.
By Bowen Qin, Yi Xie, Yesheng Liu, Xi Yang