Diffu-LoRA is a parameter‑efficient method for personalizing text‑to‑image diffusion models. It inserts trainable low‑rank components into Transformer linear layers and assigns a learnable gate to each component, allowing the model to learn how to distribute adaptation capacity across layers. Using bilevel optimization and progressive pruning, Diffu-LoRA keeps the pretrained backbone frozen while improving subject fidelity and prompt alignment compared to fine‑tuning baselines.
By Tianjing Li, Wei Zhu
The paper presents a Large Language Model (LLM)-assisted framework for automating the creation of Transportation Management Plans (TMPs) using the WisDOT WisTMP system. It fine‑tunes several open‑source LLMs of varying sizes on a domain‑specific dataset derived from historical WisTMP documents, and deploys the models locally to maintain data security. Experiments show that fine‑tuning improves text‑generation metrics, though LLMs still over‑generate strategies and struggle with project‑specific justifications and cost estimates, and scaling from 7B/8B to 14B models offers limited gains.
By Zihao Sheng, Pei Li, Zilin Huang, Yen-Jung Chen, Yuhao Luo, Zhengyang Wan, Steven T. Parker, David A. Noyce, Sikai Chen
SignRAG is a unified framework for gloss‑free sign language translation that combines hierarchical pretraining, target‑domain retrieval augmentation, and retrieval‑aware reinforcement fine‑tuning. It first learns linguistically grounded sign representations, aligns a sign encoder with a decoder‑only large language model, and then uses a retrieval gallery to provide instance‑specific translation cues. Retrieval Utility‑Guided Reinforcement Fine‑Tuning (RUG‑RFT) balances translation quality and retrieval usefulness, leading to state‑of‑the‑art results, including the first gloss‑free approach to outperform gloss‑supervised methods on CSL‑Daily.
By Zhi Rao, Yucheng Zhou, Qianran Sun, Yiqing Huang, Longcan Yuan, Jiayi Hou, Chengwen Yao, Lin Cheng, Donghui Sun, Xiaoxin Chen, Jun Wan
The paper introduces Residual Advantage (RA), a new guidance method for reinforcement learning with verifiable rewards (RLVR) that treats the teacher–student probability residual as a bounded one‑step reward, subtracts the student’s state value to form an advantage, and centers it within each response. RA preserves the verifier’s mean label while redistributing credit among response steps, and its extension CoRA updates the teacher model using verifier advantages to adapt guidance to the student’s attempts. Experiments with Qwen3 models show that RA combined with GRPO or REINFORCE++ consistently outperforms teacher‑only on‑policy distillation across three mathematical benchmarks, improving macro Avg@8 and Pass@8 scores, with CoRA providing additional gains.
By Xiaobing Chen, Zhiqi Pang
The paper evaluates Jev, a specialized decision model, on stance detection tasks using the VAST (English) and ZS-CSD (Chinese) datasets. Jev matches GPT‑5.6 and outperforms other general‑purpose LLMs on VAST, but lags behind stronger LLMs on ZS‑CSD, especially in distinguishing favor from against. The authors suggest that Jev’s limitations may stem from difficulties in understanding reply relationships and stance direction rather than conversation length.
By Xing Li, Jinzhong Ning, Yijia Zhang, Liang Yang, Hongfei Lin
The paper presents a method for detecting spin in clinical trials by automatically identifying outcome switching using large language models (LLMs). The authors trained on 300 pairs of outcomes labeled for semantic similarity, then evaluated baseline text similarity models and open-source LLMs, applying prompt engineering, token‑probability classification, and majority voting. On a test set of 2,496 examples, the LLM‑based approach achieved an F1 score of 0.78 and accuracy of 0.90, surpassing baseline models but falling short of fine‑tuned BERT, and also produced natural‑language explanations that were manually evaluated for quality.
By Tja\v{s} Ajdovec, Marko Robnik-\v{S}ikonja, Simon \v{S}uster
HarnessSQL introduces a harness‑native post‑training framework that keeps the full stateful, multi‑turn interaction structure of real‑world database agents during both supervised fine‑tuning and reinforcement learning. By building isolated executable database environments with hidden execution oracles and retaining only verified trajectories for full‑sequence SFT, the method enables execution‑reward RL that aligns training with inference. Experiments on Spider 2.0‑SQLite show dramatic gains in execution accuracy for compact models, and the approach transfers effectively to out‑of‑distribution interactive benchmarks such as BIRD‑Interact and LiveSQLBench.
By Haolin Yang, Jipeng Zhang, Jian Xie, Shuaishuai Gong, Sirui Han, Yike Guo
The paper introduces SBMEX, a new skeleton‑based minutiae detection method that replaces legacy C/C++ code with a fast, deterministic algorithm integrated into the open‑source pyfing package. SBMEX uses a dual Look‑Up Table architecture to eliminate runtime neighborhood scanning during Crossing Number computation and skeleton tracking, and adds a continuous quality scoring framework based on tracking path length, dual ridge‑valley skeleton fusion, and spatial density decay. Evaluation on NIST SD302 shows that SBMEX matches or surpasses traditional open‑source baselines in accuracy while dramatically reducing detection latency compared to classical Python implementations.
By Raffaele Cappelli
The paper introduces Vantage, a training‑free, model‑agnostic framework that pairs a vision‑language model with a 3D foundation model to perform Seek‑and‑View reasoning for multi‑view spatial understanding. Vantage first analyzes a question to plan a relevant viewpoint, then augments reasoning with geometry‑grounded visual evidence from that viewpoint. Experiments across six VLMs on five benchmarks show consistent improvements without fine‑tuning, demonstrating that revealing spatial evidence through view‑grounded reasoning reduces reliance on language‑based cross‑view alignment.
By Qixiang Chen, Cheng Zhang, Fucai Ke, Chi-Wing Fu, Jianfei Cai, Jingwen Ye
VEDJE is a new video-text retrieval method that compresses video features into a reusable cache while keeping them distinct for each frame. It introduces feature-change prediction as an auxiliary training signal, enabling improved retrieval from the compressed cache without extra cost at query time. Experiments on MSR-VTT, MSVD, DiDeMo, and ActivityNet show that VEDJE boosts R@1 in both retrieval directions, achieving 59.8% text-to-video R@1 on MSR-VTT with a fine-tuned VideoCLIP-XL first stage and maintaining recall when the per-video cache is reduced to 12 KiB.
By Shahaf Wagner, Gabriele Serussi, Dan Ben Ami, Tomer Galanti, Chaim Baskin
The paper introduces a two-stage framework to incorporate the animation principle of exaggeration into motion generative models. During training, it fine‑tunes pre‑trained text‑to‑motion models on a curated exaggeration dataset. At inference, it uses a mathematical formulation based on dynamic movement primitives to guide diffusion and flow‑matching models toward more expressive, exaggerated motions while maintaining physical plausibility and the original motion intent.
By Amirhossein Zamani, Arianna Rampini, Bruno Roy
The paper introduces LEGO, a lifting‑free method for generating egocentric videos from a single exocentric recording. Instead of reconstructing depth and point clouds, it trains a transformer‑based view synthesizer that directly predicts the egocentric view, providing a probabilistic correspondence map that balances structural alignment with texture averaging. This approach outperforms existing explicit pipelines and generalizes across datasets without retraining.
By Suhwan Cho, Yonwoo Choi, Soongjin Kim, Jicheol Park, Taegyu Lim
The paper introduces the AutoRegressive Retriever (ARR), a multimodal retrieval model that iteratively refines a query embedding by selecting informative items and incorporating their content. ARR alternates between retrieval and query update, using stepwise contrastive supervision and reinforcement learning to optimize item selection and final ranking. Experiments show ARR outperforms baselines on in‑domain and zero‑shot benchmarks, with feedback improving both inference performance and initial query embeddings.
By Jianfei Zhao, Yifan Wang, Feng Zhang, Xin Sun, Chong Feng, Zhixing Tan, Yang Luo, Boyuan Pan, Xu Kai, Yao Hu
SuperNav is a navigation system that leverages a pretrained multimodal large language model (MLLM) without fine‑tuning it for navigation. The system introduces an agent harness that delegates motion execution to navigation tools while the MLLM focuses on interpreting requests, understanding scenes, and making decisions. A visual‑point interface connects the model’s decisions to motion, allowing direct destination specification and feedback‑driven revision, enabling sustained navigation across varied tasks and environments.
By Jinkai Zhang, Jingyi Xu, Yuanhong Yu, Jiarui Guo, Ruizhen Hu, Hujun Bao, Xiaowei Zhou, Sida Peng
The paper introduces MatLat, a material latent space designed for generating high‑quality physically‑based rendering (PBR) textures on 3D meshes. It fine‑tunes a pretrained VAE to incorporate new material channels while preserving latent distribution, and adds a locality‑preserving regularization that aligns cropped latent patches with image regions. Experiments and ablation studies show that these components together improve PBR texture fidelity and outperform prior baselines.
By Kyeongmin Yeo, Yunhong Min, Jaihoon Kim, Minhyuk Sung
LadderEdit is a method for memory‑efficient lifelong editing of large language models that compresses each LoRA adapter after acquisition. Each edit is initially stored as a low‑rank sketch and only promoted to higher rank if it fails to satisfy rewrite, generalization, and locality contracts on probe prompts. On benchmarks such as ZsRE, CounterFact, and WikiBigEdit, LadderEdit achieves the same performance as exact LoRA while using 5.2× less memory and handling up to 50,000 sequential edits.
By Xiaobing Yu, Peijie Qiu, Jin Yang, Xuanzhao Dong, Weiwei Ma, Zhaoqi An, Xiaoqi Zhao, Xiaofeng Liu
The paper introduces Dual-Path Parametric Memory (DPPM), a method for long‑term personalization of language models that leverages interaction history without requiring it in the inference context. DPPM uses two complementary paths: an Evidence path that pools historical representations to retain earlier evidence, and a Delta path that sequentially updates an associative state to capture changes. By fusing these outputs into history‑conditioned LoRA adapters, DPPM achieves superior performance on PersonaMem‑v2 (54.22%) and PrefEval (86.79%) compared to existing baselines.
By Yuhao Chen, Shuochen Liu, Jiayao Shi, Jian Hong, Chen Cheng, Xinyun Ding, Tao Wang, Ya Li, Quan Liu, Tong Xu
VioLA is a generalist humanoid control policy that predicts body and hand motion latents instead of joint commands, enabling a pretrained controller to execute these latents on a robot. By aligning human and robot motion in shared latent spaces, the policy can be trained on a large dataset of 140.6 million frames, 93.2% of which are human demonstrations. As a result, VioLA can follow locomotion and manipulation instructions on a real robot zero‑shot, achieving 100% locomotion success and 88.6% manipulation success without task‑specific fine‑tuning.
By Mert Albaba, Jens Bei{\ss}wenger, Anna Manasyan, Daniel Marta, Michael J. Black, Wieland Brendel, Andreas Krause, Georg Martius, Martin Riedmiller
InfiFPO is a new preference optimization technique for implicit model fusion that replaces the reference model in Direct Preference Optimization with a fused source model, preserving probability information and avoiding complex vocabulary alignment. By incorporating probability clipping and max‑margin fusion, it aligns a pivot model with human preferences while distilling knowledge from multiple source models. Experiments on 11 benchmarks show that InfiFPO consistently outperforms existing fusion and preference optimization methods, raising Phi‑4’s average score from 79.95 to 83.33 and improving performance in mathematics, coding, and reasoning tasks.
By Yanggan Gu, Yuanyi Wang, Zhaoyi Yan, Yiming Zhang, Qi Zhou, Fei Wu, Hongxia Yang
The paper presents a framework that enhances single-image nutrition estimation by using multimodal large language models to inventory visible foods and verify both food identity and the suitability of proposed 2D regions for portion estimation. A whole-image review identifies gaps and triggers a single targeted recovery pass, after which recovered regions are re-verified and reconciled into a final item set. The approach, which requires no task-specific fine-tuning, improves item-level grounding, mass accuracy, energy accuracy, and item-identity precision and recall compared to a retrieval baseline.
By Jingbo Yue, Bruce Coburn, Jinge Ma, Jui-Feng Chi, Fengqing Zhu