The paper compares data-level and training-level methods for balancing language model capability and privacy leakage on synthetic medical records with embedded canaries. Direct finetuning achieves the highest capability but leaks the most private information, while DP‑SGD reduces leakage at a significant cost to performance. Generating privacy‑aware synthetic data—especially with uniqueness‑aware rewriting—offers a better trade‑off, and more capable rewriting models tend to lower downstream privacy risks, though gains are inconsistent with increased compute.
By Zihao Zhao, Avi Schwarzschild, Casey Meehan, Katherine Lee, Zhili Feng, Dylan Sam
The paper defines and classifies "Polluted Chinese" (PoC) tokens—Chinese tokens in GPT vocabularies that correspond to pornographic or gambling content. It develops a detector by fine‑tuning an LLM to label such tokens using semantics and search‑engine references, then uses token appearance to speculate on training‑data contamination. Experiments across GPT and 23 other LLMs show widespread PoC tokens, with GPT’s vocabulary containing over 40% long Chinese tokens linked to porn or gambling, and the method is validated on datasets like C4 and Pile.
By Qingjie Zhang, Di Wang, Haoting Qian, Liu Yan, Tianwei Zhang, Ke Xu, Qi Li, Minlie Huang, Hewu Li, Han Qiu
PragReST is a self‑supervised framework that enhances large language models’ pragmatic reasoning by generating counterfactual reasoning traces and fine‑tuning models with supervised learning and reinforcement learning, all without human‑labeled data. The method constructs pragmatic QA data, trains models to internalize counterfactual reasoning, and demonstrates significant accuracy gains on four pragmatic benchmarks compared to backbone models and other baselines. Error analysis shows that counterfactual reasoning reduces mistakes from failing to contrast observed utterances with plausible alternatives, and removing it markedly degrades performance while preserving out‑of‑domain general‑knowledge and mathematical reasoning.
By Jihyung Park, Minchao Huang, Leqi Liu, Elias Stengel-Eskin
The paper introduces Cambrian-BEV, a parameter‑free geometric tokenization method that projects visual patch features from multiple video frames into a shared bird’s‑eye‑view (BEV) grid, aggregating observations of the same physical location across viewpoints and time. This approach yields a persistent spatial representation while dramatically compressing the visual token set—from 100 K DINOv3 patches across 128 frames to just 729 BEV tokens—reducing LLM prefill time by 2.9× and peak memory by 3.6×. Experiments on VSI‑Bench and ReVSI show that Cambrian‑BEV outperforms frame‑token baselines, achieving significant gains in room‑size and relative‑direction tasks.
"whyItMatters":"By providing a stable, compressed spatial context, Cambrian‑BEV enables more efficient and accurate multimodal large language models for embodied agents, improving performance on spatial reasoning benchmarks while reducing computational resources."
By Faizan Siddiqui, Nirav Savaliya, Sahith Reddy Chada, Enna Sachdeva
A100 is a hybrid CNN‑Transformer vision foundation model trained on 3 million unlabeled agricultural images using self‑supervised techniques (DINO, iBOT, KoLeo). It achieves superior crop classification (95.37 % with a frozen backbone) and phenotyping accuracy, and after fine‑tuning reaches 99.90 % on 32‑class crop identification and 99.02 % on 44‑class disease detection. With only 1 % of labeled data, A100 attains 98.01 % plant classification accuracy, outperforming existing baselines and other agricultural foundation models by significant margins.
By Mayur Pratim Das, Ajai Dubey, Aksshay Mathew P, Sana Afreen, Rishi Kumar Soni, Deepak Meena, Pratibha Kumari, Mukesh Saini
Kingfisher is a fast, model‑tuning‑free framework for personalized image generation using visual autoregressive models. It decomposes early autoregressive scale codes into content and style components conditioned on prompts, then recombines these components from two real images into a compact semantic prefix. The frozen generator refines the remaining scales, achieving strong content preservation and superior style and text alignment while cutting pair‑specific preparation time by 54× compared with diffusion methods and 7.5× compared with other VAR approaches.
By Dinh-Khoi Vo, Akihiro Sugimoto, Minh-Triet Tran, Trung-Nghia Le
RAP (Reconstruct A Poster) is a new dataset and benchmark for generating conference posters from research papers, comprising over 25,000 verified paper‑poster pairs from 2022‑2026. Each pair is annotated with a three‑level content taxonomy and converted into styled HTML/CSS markup through an automated derendering pipeline, enabling objective render‑and‑compare evaluation. The corpus supports both inference and training, with a retrieval‑based PosterRAG pipeline and a supervised‑fine‑tuned model, and is publicly available via an online demo.
By Haonan Han, Yukang Lin, Minliang Lin, Haoxian Tan, Jun-Yan He, Rui Yang, Zhi-Qi Cheng, Jie Hu, Guanglu Wan, Xiu Li
The paper describes the CERTH-ITI team’s approach to the TRECVID 2025 Ad-hoc Video Search (AVS) and Video Question Answering (VQA) tracks. For AVS, a two-stage framework uses foundation vision‑language models for initial retrieval and a Multimodal‑LLM for reranking, improving contextual matching without training. In VQA, an audio‑visual MLLM is fine‑tuned and enhanced with inference‑time scaling; answer generation employs majority voting over responses from random frame subsets, while multiple‑choice responses use mean‑pooling logits, yielding top performance on VQA metrics.
By Andreas Goulas, Damianos Galanopoulos, Ioannis Patras, Vasileios Mezaris
The study investigates whether a deep learning segmentation model trained on adult kidney MRI can generalize to pediatric Wilms tumors. Trained on single‑sequence T1 or T2 adult MRI, the model achieved a median Dice similarity coefficient (DSC) of 0.86 on an adult test set and 0.85 on a pediatric cohort, compared to 0.90 for a pediatric‑specific nnU‑Net. Although the adult‑trained model performed well, the modest performance gap indicates that pediatric training data remain valuable.
By Sarah de Boer, Myrthe A. D. Buser, Hartmut H\"antze, Jawed Nawabi, Lisa C. Adams, Alida F. W. van der Steeg, Marry M. van den Heuvel-Eibrink, Annemieke S. Littooij, Keno K. Bressem, Mathias Prokop, Alessa Hering
Video Diffusion Avatars (VDAs) fine‑tune video diffusion models to generate photo‑realistic, multi‑view consistent avatars from a single image, allowing precise control over camera viewpoint, facial expression, and upper‑body pose. VDAs replace traditional 3D morphable models with a learned expression space, overcoming viewpoint bias through disentanglement and enabling independent control of expression, pose, and viewpoint via proxy‑mesh renderings. The method trains efficiently—34 hours on a single A100—and outperforms state‑of‑the‑art single‑image 3D avatars, improving novel‑view reenactment metrics and producing physically plausible hair dynamics, shadows, reflections, and refractions.
By Simon Giebenhain, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Zhe Chen, Matthias Nie{\ss}ner
The paper introduces StopWatch, a method that explicitly grounds the goal in object-goal navigation tasks. By addressing the dominant failure of goal grounding in fine‑tuned vision‑language models, StopWatch achieves state‑of‑the‑art success rates on six benchmarks across three scene collections and corrects annotation defects in these datasets. The authors demonstrate that with accurate goal grounding, their policy nearly matches the expert demonstrations, highlighting perception as the remaining bottleneck.
By Xinhao Liu, Yunsheng Tian, Yuheng Qiu, Pieter Abbeel, Xi Chen, Guanya Shi, Chen Feng
The study explores how to fine‑tune visual brain encoding models so that their internal representations better match individual human perceptual similarity judgments. Using similarity data from 100 images collected from eight participants in the Natural Scenes Dataset, the authors adapted a generic similarity model (DreamSim) to create personalized perceptual similarity models for each subject. They then fine‑tuned each participant’s baseline encoder with either generic or personalized similarity information, finding that models trained with personalized judgments were most aligned with human perceptual similarity and preserved inter‑subject correlations of those judgments.
By Daniel Chong, Haomiao Chen, Keith W Jamison, Raja Marjieh, Ilia Sucholutsky, Nori Jacoby, Amy Kuceyeski
MambaAct introduces a compact Mamba-based action expert to replace the transformer-based expert in flow‑matching vision‑language‑action models, preserving the pretrained visual‑language prefix encoder while reducing the action expert size by tenfold. The new design employs a lightweight Context Compressor to embed vision‑language conditioning into each Mamba block, eliminating repeated token‑level cross‑attention. Experiments show a 64.8 % reduction in denoising latency on LIBERO and 58.8 % in real‑world deployment, with a modest 1.08 % increase in mean success rate across multiple fine‑tuning seeds.
By Khuyen Pham, Tran Nguyen Le, Edoardo Fazzari, Cesare Stefanini
arXiv:2501.13558v4 Announce Type: replace
Abstract: Splat-based radiance fields have rapidly diversified in primitive type, kernel support, and appearance parameterization. This diversity makes compr...
By Francesco Di Sario, Riccardo Renzulli, Marco Grangetto, Akihiro Sugimoto, Enzo Tartaglione
Structured Authority Injection (SAI) is a protocol that enables persistent human control over large language model (LLM) agents by routing operator corrections through typed, schema‑validated Model Context Protocol (MCP) tool calls and maintaining a ranked authority context in each prompt. It defines three override classes—Constraint, Priority, and Strategy—each with distinct persistence semantics and a strict precedence order enforced via a directed authority channel. Evaluations across nine LLMs and eight scenarios show reduced ambiguous directive interpretation, mitigated authority decay, and lower prompt overhead in the strongest models, though benefits vary by model.
By Ryan Tran, YueMing Zhang, Lingfeng Tao, Nan Yang, Shengjie Xu, Yu Yin, Shuai Xu, Wenlu Zhang, Hailu Xu
The paper introduces PolicyLift, a three‑stage framework that improves decision policies using observational data for complex, structured actions such as text or images. Stage 1 learns discrete action codes that capture high‑level strategies, Stage 2 trains a context‑dependent policy over these codes, and Stage 3 generates full actions conditioned on the learned codes. The authors provide theoretical guarantees that improving the code‑level policy leads to improvement in the original action space and demonstrate consistent policy gains in synthetic and real‑world text generation experiments.
By Jonas Schweisthal, Dennis Frauen, Stefan Feuerriegel, Emma Brunskill
The paper proposes a competence‑based evaluation (CBE) framework for assessing logical reasoning in large language models. CBE tests whether models consistently apply the same reasoning competence across a family of problems, using a parameterized generator to vary complexity while preserving the underlying rule. Experiments on GPT and Qwen models show that even strong models often fail to meet the competence condition across the tested family, indicating limitations in generalization.
By Jize Jiang, Chengxiang Zhai
The paper introduces Porimon, an LLM-based agent for Pokémon Battles that uses Long/Short-Term Knowledge Augmented Generation (LSTKAG) to incorporate past states of the current task and retrieve experience summaries from similar previous instances. Porimon also employs an external API for precise damage calculation and detailed game information. Experimental results from 15,000 simulated battles show that Porimon, with hyperparameter optimization, outperforms both the earlier PokéLLMon agent and a rule-based heuristic player, and ablation studies confirm the benefit of the game information retrieval extension, though the impact of Long-Term KAG remains inconclusive.
By Dongyin Zhuo, Fengjunjie Pan, Nenad Petrovic, Alois Knoll
VietMed is a Vietnamese medical speech recognition dataset featuring 16 h of labeled medical speech, 1 000 h of unlabeled medical speech, and 1 200 h of unlabeled general‑domain speech. It claims to be the largest public medical ASR dataset worldwide in seven dimensions—total duration, speaker count, disease coverage, recording conditions, speaker roles, unique medical terms, and accents—and also the largest Vietnamese speech dataset by duration. The authors release pre‑trained models (w2v2‑Viet and XLSR‑53‑Viet) and fine‑tuned medical ASR models, showing that XLSR‑53‑Viet outperforms the state‑of‑the‑art XLSR‑53 by reducing WER from 51.8 % to 29.6 % on a medical test set.
whyItMatters":"The dataset and models provide the first large‑scale, publicly available Vietnamese medical ASR resource, enabling research and development of medical speech technologies in a language and domain that previously lacked sufficient data."
By Khai Le-Duc
arXiv:2610.22381v1 Announce Type: new
Abstract: When a fact becomes harder to elicit from a language model, has the model lost it, or has it merely become less accessible? Parameter-updating test-tim...
By Zhe Zhou