The paper introduces ALIAS, a method to assess memorization in large language models (LLMs) when generating TLA+ specifications. By applying consistent identifier renaming (CIR) and using a TLA+ model checker as an oracle, the authors show that renaming significantly reduces the correctness scores of models like Claude Opus 4.8 and GPT‑5, revealing the extent of memorization. The study also demonstrates that grading based on exact names inflates scores, and that fine‑tuned models lose a substantial portion of their training advantage when identifiers are renamed.
By Arslan Bisharat, Mujtaba Nazari, Khushboo Bhadauria, Brian Ortiz, Eric Spencer, Beatriz Santos, TaiNing Wang, George K. Thiruvathukal, Konstantin Laufer, Mohammed Abuhamad
Loomis Painter is a framework that learns painting processes by training on reversed image‑to‑video sequences, preserving reference‑attention behavior and avoiding long‑range dependencies that disrupt pretrained models. It introduces a media‑transfer strategy to decouple input appearance from the target process, enabling generation across different artistic media. Evaluated on a curated dataset of real‑world painting processes, the method outperforms baselines on LPIPS, DINO, and CLIP metrics and receives consistent preference in expert and non‑expert user studies.
By Markus Pobitzer, Chang Liu, Chenyi Zhuang, Teng Long, Bin Ren, Nicu Sebe
BiMoGen introduces a unified masked discrete diffusion framework for bidirectional motion‑text generation, addressing limitations of autoregressive models by allowing iterative bidirectional prediction. The method employs a two‑stage training strategy—Decoupled Uni‑ and Cross‑Modal Training followed by supervised fine‑tuning—to establish cross‑modal correspondence and specialize for bidirectional generation. To mitigate train‑test mismatch, Generation‑Aware Self‑Correction exposes the model to its own predictions during training and applies correction passes early in sampling to refine unmasked tokens, achieving competitive results on HumanML3D and KIT‑ML.
By Wanjiang Weng, Yongliang Wu, Xiaofeng Tan, Xingyu Zhu, Wenbo Zhu, Hongsong Wang
ColonAdapter is a self‑supervised fine‑tuning framework that adapts geometric foundation models for 3D geometry estimation from monocular colonoscopy images. It incorporates a Detail Restoration Module and geometry consistency loss to improve low‑texture region performance and scale consistency, and uses a confidence‑weighted photometric loss for training stability. Experiments on synthetic and real datasets show state‑of‑the‑art results in camera pose estimation, monocular depth prediction, and dense 3D point map reconstruction without needing ground‑truth intrinsic parameters.
By Zhiyi Jiang, Yifu Wang, Xuelian Cheng, Zongyuan Ge
The paper reports a submission to the WMT 2026 Multilingual Instruction Shared Task, where the authors fine‑tune Qwen3‑8B with LoRA on organiser‑provided data, EuroAlpaca, and CrossEuroAlpaca. Their system focuses on open‑ended question answering in ten target languages and shows a 7.8 % relative gain in Gemma‑4 judge score for target‑language QA‑OEG, as well as 41.5 % macro‑averaged ROUGE‑L and 2.9 % BERTScore‑F1 improvements on the Aya Evaluation Suite. However, the gains are task‑ and language‑specific, with some declines in context‑grounded QA and non‑target languages.
By Aleix Sant, Jordi Luque, Carlos Escolano
The paper proposes a shrinkage estimator that blends store‑specific demand estimates with pooled information from other stores under a multinomial logit model, addressing the trade‑off between local data scarcity and cross‑store heterogeneity. By leveraging self‑concordant‑like properties of the MNL likelihood, the authors derive finite‑sample performance guarantees and regret bounds for assortment and pricing decisions. Numerical experiments and an empirical study on European apparel retailer data show that this transfer‑learning approach improves out‑of‑sample product prioritization and revenue capture compared to both local estimation and full pooling.
By Recep Yusuf Bekci, Stefanus Jasin, Joline Uichanco
The paper explores aggressive low‑state weight quantization for Transformer models used in electricity‑load forecasting. Experiments with 3, 5, 7, and 16 discrete weight states on two datasets show that post‑training quantization degrades performance, but quantization‑aware fine‑tuning recovers most of the loss. A novel accumulated‑update thresholding technique further reduces weight updates and state transitions with minimal impact on error, demonstrating that compact, low‑activity models can still deliver accurate forecasts.
By Muhammad Sabbir Alam, Harindra Sandun Mavikumbure, Victor Cobilean, Milos Manic, Jayasimha Atulasimha
OccJEPA treats occlusion as masking for amodal video segmentation by dropping tokens outside an object's visible mask and using a pre‑trained V‑JEPA 2.1 predictor to fill in the missing features, followed by a lightweight head that decodes these into complete masks. The resulting deterministic model is 33.5× faster than diffusion‑based state‑of‑the‑art methods and achieves higher accuracy on SAIL‑VOS, TAO‑Amodal (including a held‑out test split), and Kubric. Experiments show that the performance gains stem from the predictor’s pre‑trained weights rather than its size, with models ranging from 87 M to 1.85 B parameters.
By Genpei Zhang
The paper introduces an energy-based transfer learning approach for reinforcement learning, aiming to improve sample efficiency in continual learning scenarios. By using out-of-distribution detection, the method selectively provides teacher guidance only in states that fall within the teacher’s training distribution, preventing sub-optimal bias when tasks differ significantly. The authors theoretically connect energy scores to state-visitation patterns under on-policy RL and demonstrate empirical gains in both single-task and multi-task settings.
By Zeyun Deng, Jasorsi Ghosh, Fiona Xie, Kai Cheng, Yuzhe Lu, Katia Sycara, Joseph Campbell
The paper evaluates a 305 M‑parameter audio encoder shared by Gemma 4 and EmbeddingGemma 2 for Saudi and Gulf Arabic speech recognition without altering its weights. Two approaches are compared: a frozen system that trains only a 22.6 M‑parameter CTC reader, and an adapted system that adds 3.1 M low‑rank adapters, a 41.5 M‑parameter hybrid CTC/attention reader, and joint decoding with a sub‑word language model. The adapted system matches or surpasses a 0.6 B‑parameter fine‑tuned recogniser on several datasets, while the frozen system remains competitive in character error rate and runs three‑quarters faster on CPU with greedy decoding.
By Amr Abdelkhalik
DynaWeave is an automated framework that converts static tasks into dynamic scenarios by generating exogenous events and using DAGs to manage dependencies. It activates these events during agent interaction and employs a dynamic verifier that updates only relevant evaluation checks. Using DynaWeave, the authors collected over 2,000 verified trajectories, fine‑tuned Qwen3.5 models, and achieved significant performance gains on Gaia2, CostBench, and ToolMaze, even surpassing Gemini‑3.1‑Pro‑Preview on ToolMaze.
By Xuan Yang, Chunxiao Liu, Hongsheng Xin, Hao Xu, Ning Miao
The paper investigates whether fine‑tuning large language models (LLMs) on value survey responses can shift their value‑conditioned outputs and downstream behavior. Researchers first estimate baseline value profiles for several open‑source LLMs using the ten Schwartz values, then fine‑tune the models on survey‑style supervision that selectively alters target value ratings while preserving non‑target values. The resulting shifts are evaluated on held‑out survey questions and on out‑of‑domain behavioral proxies, such as a contextual moral‑judgment benchmark from Reddit AITA posts and a text‑based interactive decision‑making benchmark. The study finds that survey‑based supervision can change both survey responses and downstream behavior, but the transfer is strongly model‑ and value‑dependent and influenced by how value descriptions are framed.
By Shangrui Nie, Florian Mai, David Kacz\'er, Charles Welch, Zhixue Zhao, Lucie Flek
The study investigates how post‑training alignment techniques affect small language models (SLMs) for generating medication leaflets from structured biomedical data. Four alignment methods—supervised fine‑tuning (SFT), direct preference optimization (DPO), odds ratio preference optimization (ORPO), and group relative policy optimization (GRPO)—are evaluated using Qwen‑2.5 models, with a new FDALeaflets dataset derived from OpenFDA drug labels for zero‑shot cross‑regulatory testing. Results show that aligned SLMs can outperform GPT‑5 on reference‑based metrics, while ORPO and GRPO achieve the highest semantic textual similarity scores in‑domain and on FDALeaflets, respectively, highlighting the importance of alignment for semantic fidelity and generalization.
By Xi Yang, Guodong Liu, Chuqin Li, Fan Wu, Ergin Soysal, Min Jiang, Xing He, Jiang Bian, Yi Guo, Shams Zaman, Thomas Fuchs, Yonghui Wu
The paper evaluates whether a 9‑billion parameter decision foundation model, Clef‑Flash, is necessary for symptom‑to‑disease classification. Fine‑tuning the model on the 24‑class Symptom2Disease benchmark yields 95.42% accuracy, but simple TF‑IDF linear classifiers achieve higher accuracy (98.75%) with far faster inference. Ablation studies show that predictions can hinge on single descriptive clauses, indicating that the benchmark may not fully capture the model’s capabilities.
By S. M. Mahbub Uddin Ahmed
PRISM is a training‑free framework for financial information retrieval that combines prompt engineering, in‑context learning, and a candidate refinement module, with optional multi‑agent coordination. On FiQA‑2018 it improves Recall@100 from 0.7783 to 0.8326, and on FinAgentBench it achieves a Cohere‑refined top‑50 pool that raises NDCG@5 while cutting per‑query cost by about 35%. The authors also release an expert‑annotated evaluation set, a label‑robustness sweep, and a Pareto frontier of quality‑cost‑latency trade‑offs, and PRISM ranks third on the private FinAgentBench leaderboard without fine‑tuning.
By Chun Chet Ng, Jia Yu Lim, Zhen Hao Chu, Yixi Zhou, Wei Zeng Low
SDPose leverages pre‑trained diffusion models, specifically the Stable Diffusion U‑Net, to extract robust latent features for human pose estimation. By identifying the most effective upsampling levels, fusing these features with a lightweight aggregation module, and jointly training with keypoint heatmap supervision and a latent reconstruction loss, SDPose achieves strong cross‑domain generalization. The authors introduce the COCO‑OOD benchmark, comprising style‑transferred and corrupted subsets, to evaluate robustness, and demonstrate that SDPose matches Sapiens on COCO while outperforming competitors in zero‑shot cross‑domain scenarios.
By Shuang Liang, Jing He, Lejun Liao, Chuanmeizhi Wang, Guo Zhang, Yingcong Chen, Yuan Yuan
The paper presents a transfer learning–driven approach for indoor localization using visible light communication (VLC) that aims to reduce energy consumption and improve scalability. By applying transfer learning, the method addresses environmental variability such as lighting changes and obstacles, lowering computational overhead compared to conventional techniques. Experiments on a BOSCH factory dataset show a 47% increase in localization accuracy and up to 67.5% model efficiency under high‑noise conditions, outperforming non‑TL models.
By Masood Jan, Wafa Njima, Xun Zhang, Alexander Artemenko
The paper examines how fine‑tuning large language models (LLMs) for specific human values can unintentionally affect other, untrained values and broader behaviors. By adapting four instruction‑tuned LLMs to Hofstede’s value dimensions, the authors evaluate cross‑value transfer and observe side effects on risk preferences, social cognition, and safety traits. They conclude that post‑training value alignment should include comprehensive behavioral assessments and that further research is needed to determine the optimal placement of value training within the development pipeline.
By Siddhesh Pawar, Chen Cecilia Liu, Anna Korhonen, Isabelle Augenstein
PhysiCalWorld introduces a method for calibrating action‑conditioned world models to instance‑specific physics by encoding an exploratory interaction clip into a compact physics descriptor token. This token grounds future predictions in the actual physical behavior of a scene without requiring fine‑tuning or explicit physics labels, and can be reused across rollouts with minimal compute overhead. Experiments across four diverse environments demonstrate significant improvements in rollout prediction accuracy and visual planning success, while the learned descriptors form a physically structured latent space that generalizes beyond the training physics ranges.
By Yunhao Luo, Wentao He, Yinpei Dai, Hyelin Nam, Hongze Fu, Joyce Chai, Nima Fazeli, Jeong Joon Park
Latent diffusion models typically use a Gaussian prior in the VAE latent space, which works well for base generation but introduces a structural limitation during reinforcement‑learning fine‑tuning. The standard VAE’s KL term compresses multimodal data into overlapping latent representations, and RL fine‑tuning exacerbates this overlap, causing the model to collapse at a higher beta_c. Experiments on the alanine‑dipeptide Boltzmann distribution show that a Gaussian‑mixture prior can delay collapse by 71×, and the study attributes this improvement to a geometric mediator, d_min/(2σ₀), confirming that non‑Gaussian priors are necessary when latent geometry would otherwise overlap and the task requires coverage.
By Wang Ru