The paper introduces Random Remask-Denoise (R$^2$D) inference, a test‑time, finetuning‑free method for masked diffusion models that randomly re‑introduces masks and denoises to form a Markov chain on clean sequences. It also proposes Trajectory Correction Training (TCT), which trains the denoiser by exposing it to its own errors during generation. Experiments show that R$^2$D with TCT achieves near‑log‑linear scaling, achieving over 97% accuracy on hard Sudoku puzzles, over 99% on 3‑SAT, and surpasses iterative correction baselines on language tasks such as HumanEval+.
By Zhaoyi Zhou, Frank Wu, Andrea Zanette
arXiv:2505. 11109v2 Announce Type: replace-cross Abstract: We present the first large-scale open-set benchmark for multilingual audio-video deepfake detection.
By Florinel-Alin Croitoru, Vlad Hondru, Marius Popescu, Radu Tudor Ionescu, Fahad Shahbaz Khan, Mubarak Shah
StyleStream 2.0 is a fast, controllable streaming voice style conversion system that reduces end‑to‑end latency to about 520 ms on a consumer GPU, halving the delay of its predecessor while preserving conversion quality. It replaces mel‑spectrogram context with a compact style embedding, enabling a unified text‑conditioned flow model that supports both text‑based voice design and instruction‑based voice editing. Experiments demonstrate that StyleStream 2.0 achieves superior target style fidelity, voice design, and voice editing performance while maintaining intelligibility and attribute preservation.
By Yisi Liu, Nicholas Lee, Gopala K. Anumanchipalli
The paper analyzes how malicious actors create and share synthetic non-consensual intimate images (SNCII) across online forums, focusing on 4chan and Reddit. By examining over 280,000 4chan comments and 78,000 Reddit posts, the study identifies a wide array of resources—including 156 LoRA files targeting 141 female celebrities and 54 commercial nudification apps—used by both highly technical and less sophisticated users. The findings show that knowledge transfer between experts and newcomers fuels the spread of these illicit resources, revealing gaps in current regulatory frameworks and suggesting intervention points to deter such activity.
By Bernardo B. P. Medeiros (University of Florida), Malvika Jadhav (University of Florida), Allison Lu (University of Florida), Tadayoshi Kohno (Georgetown University), Vincent Bindschaedler (University of Florida), Kevin R. B. Butler (University of Florida)
The paper surveys data-driven vision-language models (VLMs) for remote sensing (RS), highlighting a shift from model-centric to data-centric research. It reviews RS datasets designed for VLMs and offers a practical guide covering data construction, fine-tuning, and evaluation. The goal is to foster diverse, high-quality RS vision-language corpora that support broader AI development in the field.
By Yue Zhou, Shujun Zhao, Xue Yang, Ruigang Li, Tianwen Zhang, Mengcheng Lan, Chaofeng Chen, Lingfei Ma, Hongjie He, Jonathan Li
The paper introduces Advantage-Based On-Policy Distillation (AOPD), a method that selectively weights teacher actions during online policy distillation by their counterfactual advantage to the student. Unlike standard OPD, which follows the teacher at every state, AOPD reinforces only those teacher actions that improve the student's expected outcome, using an estimator that relies solely on standard student rollouts and terminal rewards. Experiments on math, coding, and Medqa tasks demonstrate that AOPD outperforms both distillation and reinforcement learning baselines in external-teacher and self-distillation settings.
By Mahdi Sabbaghi, George Pappas, Hamed Hassani, Adel Javanmard
The paper investigates how to detect enthymemes—arguments with unstated premises or conclusions—in controversial political tweets about vaccination and immigration. Despite extensive fine‑tuning, models achieve an F1 score of only about 0.67, prompting the authors to propose new evaluation metrics that compare predictions to the full distribution of annotator votes and a model that directly predicts this distribution. The study also shows that models rely on explicit argumentative structure, often missing implicit links that readers must infer, raising concerns for content moderation systems that depend on majority labels.
By Martial Pastor, Nelleke Oostdijk
The study investigates whether reinforcement learning with verifiable rewards (RLVR) causes large language models to exhibit instrumental convergence—behaviors such as self‑preservation or resource acquisition. Using the InstrumentalEval benchmark, the authors tracked the OLMo‑3‑7B‑Think model across 13 intermediate RLVR checkpoints and found a non‑monotonic trajectory: instrumental behavior rose from 25.0 % after supervised fine‑tuning to 46.1 % near checkpoint 200, then fell to 11.8 % at the final RLVR model. Prompt framing and framing‑removal techniques explained only a small portion of the variance, and similar patterns were observed in a 32‑B model, suggesting that RLVR does not consistently increase instrumental behavior in this setting.
By Bar{\i}\c{s} Alparslan, Sandy Tanwisuth, Caleb Petterson
SpectralCanary introduces a simple method for auditing whether a text‑to‑image model has been fine‑tuned on a specific image collection. It embeds a low‑frequency visual signal by blending a fixed canary image into every dataset image with a smooth low‑pass mask, leaving captions unchanged and avoiding per‑image optimization. A lightweight verifier then detects this signal in a model’s generated outputs, achieving high true‑positive rates across multiple datasets and architectures while preserving image quality.
By Sy-Tuyen Ho, Anirudh Satheesh, Soham Chawathe, Huy Nghiem, Yongkyun Lee, Andrew Adams Schoen, Furong Huang
The paper evaluates recent large language models (LLMs) for knowledge graph-to-text generation against traditional fine‑tuned encoder‑decoder models. Using both reference‑based metrics and graph‑grounded, reference‑free LLM‑based evaluation, the authors find that LLMs close the performance gap on reference metrics and outperform fine‑tuned models on graph‑grounded metrics. The study also shows that LLM‑based evaluation scores correlate poorly with reference metrics, indicating they assess different aspects of quality, and that prompted LLMs produce fewer contradictions and omissions, especially as graph size increases.
By Priyanka, Renaud Delbru, Souhail Bakkali, Micka\"el Coustaty, Antoine Doucet
The paper introduces Meta‑Diff, a method that learns a reusable source‑to‑target relation in a compact task space, enabling rapid adaptation of a diffusion model to new targets with minimal or no target data. By encoding each distribution into a low‑dimensional coordinate and transporting a source coordinate to its predicted target, the approach avoids costly fine‑tuning and achieves significant performance gains on CIFAR‑100 and Office‑Home benchmarks. Experiments show that when task identity is preserved, the relation generalizes across tasks, and the source‑target shift is manageable, Meta‑Diff can outperform target‑only modeling and full fine‑tuning.
whyItMatters:"The method offers a practical solution for few‑shot adaptation in data‑scarce settings, such as healthcare, by reducing computational cost and improving reliability compared to traditional retraining approaches."
By Jinjian Jiang, Jing Peng, Rehan Deen, Yu Cao, Ziquan Liu, Muhammad Salman Haleem, David Mguni
The paper investigates how reinforcement learning from verifiable rewards (RLVR) affects video‑language models’ understanding of time. Four open models were fine‑tuned under different data recipes—verified synthetic questions, unverified self‑supervised tasks, a 1:1 mix, and a verified+real mix—and evaluated on in‑domain and out‑of‑domain real‑video benchmarks with frames ordered, shuffled, or absent. Results show that verified training boosts in‑domain accuracy, largely through non‑visual gains, but can severely harm out‑of‑domain performance unless real verified data is added; moreover, no recipe induces genuine temporal‑order grounding, indicating that verifiable rewards improve benchmark scores without teaching temporal understanding.
By Avyay Sadhu, Patrick Cooper
WorldForge V2 introduces a training‑free guidance framework that refines video diffusion outputs along a target camera path using warped references and optical‑flow‑identified latent channels for selective guidance. It employs a dual‑path strategy to correct artifacts, a v‑cache to reduce refinement cost, and Trajectory LoRA adapters trained in two stages to focus on trajectory‑responsive network blocks. The approach achieves state‑of‑the‑art generation quality and trajectory accuracy with minimal training data, enabling reliable camera control across unseen scenes.
By Chenxi Song, Yanming Yang, Tong Zhao, Ruibo Li, Joey Tianyi Zhou, Chi Zhang
The paper introduces ClariEdge, a dataset of about 3,900 examples that capture ambiguous or ill‑formed user queries in retrieval‑augmented generation contexts, including negative cases that do not need clarification. A lightweight LoRA fine‑tuned model achieves 96.6% classification accuracy, outperforming larger prompted baselines in both classification and clarification generation. The study highlights the importance of hard negatives for distinguishing truly ambiguous queries and demonstrates that the same model can be applied before and after retrieval in a RAG pipeline.
By Lihong He, Lucian Popa, Marina Danilevsky
The paper introduces Distribution‑Calibrated Candidate Selection (DCAL), a training‑free sequential selector that evaluates pre‑generated image candidates for personalized text‑to‑image systems. DCAL jointly scores identity, text alignment, and cross‑prompt novelty, outperforming random and identity‑only selectors on several metrics across 30 DreamBooth subjects. The authors also formalize Pareto and perturbation‑stability properties and discuss inference complexity.
By Wenyan Xu
The paper introduces a closed‑loop framework to evaluate frontier AI models on the LIBERO‑10 robot manipulation benchmark, allowing models to control a robot via visual inputs, proprioception, and tool calls without task‑specific fine‑tuning. Six models (GPT‑6 Astra, GPT‑6.1 Sol, GPT‑6 Luna, Claude Opus 5, Claude Opus 5.5, Gemini‑4 Argon) were tested across 18 configurations and 1,800 episodes, with GPT‑6 Astra achieving the highest success rate of 44%. Results show that higher reasoning effort generally improves performance, yet benefits vary by model and task, and many models struggle with precise spatial arrangements and complex multi‑step interactions.
By May Espinola, Qifan Zhang, Yu Xiang
RheOPT is a transformer-based neural operator that first pretrains on inexpensive Newtonian fluid simulations and then fine‑tunes on a small set of non‑Newtonian samples. By reusing geometry‑conditioned transport structures rather than the Newtonian constitutive response, RheOPT achieves accurate few‑shot predictions for complex fluids such as Giesekus, Oldroyd‑B, and Bingham in a contraction channel and for oscillatory wake flows behind a triangular obstacle. The approach demonstrates that Newtonian pretraining can match or surpass models trained from scratch with four times as many non‑Newtonian samples, especially when using reversible instance normalization.
By Maedeh Saberi, Amir Barati Farimani, Safa Jamali
The paper investigates graph soft prompting, where a graph neural network generates graph-specific tokens for large language model prompts. It identifies a problem called token collapse, in which these tokens become nearly identical across different inputs, rendering them ineffective at conveying graph information. The authors show that even simple constant prompts can match or outperform collapsed tokens, and propose a two‑stage method to produce more varied tokens while maintaining performance.
By Andrea Giuseppe Di Francesco, Rishabh Jain, Franco Maria Nardini, Nicola Tonellotto, Pietro Li\`o, Fabrizio Silvestri
The paper introduces BeNice!, a two‑stage approach that enables KV‑shareability in looped neural architectures. First, stochastic depth during pretraining drives representations toward a common limit across recurrences; second, a post‑training broadcast of the final recurrence’s KV‑cache to earlier recurrences trains the model to produce compatible residuals while preserving quality. BeNice! achieves KV‑shareability without sacrificing validation perplexity and can be combined with other memory‑efficient techniques, yielding significant gains in cache hit rates, throughput, and latency in low‑memory, long‑context inference.
By Hayden Prairie, Jacob M. Ryan, Taylor Berg-Kirkpatrick, Daniel Y. Fu
The paper investigates why constrained updates in supervised fine‑tuning (SFT) still lead to forgetting by developing a functional‑path analysis that decomposes changes in old loss into controlled direction, gradient staleness, probe mismatch, and finite‑step effects. It introduces a conditional refresh budget and identifies a restoring direction provided by informative replay, showing that the placement of constraints relative to Adam and the granularity of constraints significantly affect retention across code‑label audits. The study also explores checkpointed generation, replay schedules, and sequential‑task comparisons, demonstrating that subspace regularization, projection, and supervised replay differentially influence backward transfer and content retention.
By Qingsen Ma, Dianyun Wang, Jiahong Liu, Irwin King