AQL‑Lite is a framework that enables efficient adaptation of multilingual large language models to Modern Standard Arabic by combining a quality‑first data curation pipeline with a partial‑parameter continual pretraining recipe. The approach trains RS‑LoRA adapters, the last eight transformer blocks, and the embedding matrix using layer‑wise learning rates, achieving significant benchmark gains while reducing training time and resource usage. In large‑scale experiments, AQL‑Lite recovers most of the performance of full‑parameter continual pretraining and outperforms adapter‑only baselines, demonstrating its practicality for Arabic LLM development.
By Houssam Eddine Boukhalfa, Mourad Bouache, Mourad Dahmane
ULME is a post‑training framework that unifies text embedding and language modeling within a single causal decoder‑only LLM by inserting learnable slots for semantic and reasoning embeddings. It trains three objectives—contrastive loss for semantic alignment, language‑modeling loss for answer generation, and a stop‑gradient predictive loss linking reasoning to the answer—so that generation actively shapes the embedding. Across Qwen2.5 and Qwen3 models (0.5B–8B), ULME improves MTEB scores by 3.3–4.1 points, tightens BRIGHT and AdvBench‑IR, and matches generation‑only fine‑tunes on standard accuracy benchmarks.
By Yang Liu, Zhiqiang Yuan, Shu-Dong Huang, Jiancheng Lv, Liwei Yang, Qiwei Yan, Kailin Lv, Jinchao Zhang
The paper investigates natural language autoencoders (NLAs) as tools for verbalizing activation differences in large language models. It examines whether NLAs capture compositional structure, generalize to fine‑tuned models, recover fine‑tuning objectives, and can be specialized for model diffing. Experiments across four models show that NLAs largely succeed in these tasks, though they do not yet match state‑of‑the‑art diffing methods on emergent misalignment detection.
By Andrea Blasi N\'u\~nez, Gianluca Barmina, Federico Torrielli, Thiusius Rajeeth Savarimuthu, Peter Schneider-Kamp, Lukas Galke Poech
The paper proposes treating few-shot classification as a series of multiple instance verification (MIV) tasks, enabling effective transfer across domains without fine-tuning the backbone. It introduces the MIV-head, a backbone‑agnostic, computationally efficient classification head that achieves competitive accuracy on cross‑domain few‑shot image classification using frozen, pre‑trained backbones. Experiments on an extended Meta‑Dataset benchmark show that the MIV-head outperforms state‑of‑the‑art adapter methods while incurring lower adaptation cost.
By Xin Xu, Eibe Frank, Geoffrey Holmes
The paper introduces VHDLBench, a benchmark and fine‑tuning framework for large language models (LLMs) applied to VHDL-based chip design. It presents a dataset of over 18,500 syntax‑verified VHDL examples, fine‑tunes five 7B‑parameter LLMs with QLoRA, and evaluates them through a three‑level pipeline—syntax/synthesis validation, functional simulation, and FPGA prototyping on Zynq SoCs. Results show up to 92% syntax correctness and 73% functional accuracy, with DeepSeek‑Coder achieving the best performance and successful deployment on real hardware.
By Patrik Drazic, Benaoumeur Senouci, Boualem Benatallah
The paper introduces MixAgent, an agentic pipeline for music post‑production that emulates audio engineers’ workflow to create interpretable and editable mixing plans guided by a reference song. It employs a grouping agent to cluster related tracks, a planning agent to select and order audio effects, and a detector‑critic pair to refine the mix by identifying needed corrections and proposing parameter adjustments. The authors also present MixCritic, a dataset of 45,657 repair critiques, and demonstrate that supervised fine‑tuning followed by preference optimization improves mix quality, achieving a 19% error reduction on MedleyDB and outperforming several large language models in human listening tests.
By Jaywon Koo, Junghyun Koo, Marco A. Martinez-Ramirez, Wei-Hsiang Liao, Seungheon Doh, Qiyu Wu, Vicente Ordonez, Yuki Mitsufuji
The paper introduces JEPAWG, a joint‑embedding hypernetwork that learns a parameter–weight representation for trained forecasting networks on parametric dynamical systems. By fine‑tuning a network across a grid of system parameters, the authors create a model zoo and use JEPAWG to identify bifurcations such as Hopf points, period‑doubling cascades, and a global bifurcation that lacks an eigenvalue signature. The method matches or surpasses direct hypernetwork forecasting while providing interpretable geometric insight into the underlying dynamics.
By Tobias G\"obel, Miranda C. N. Cheng
The paper investigates the origin of the word "wait" in reasoning models, arguing that it is inherited from training data rather than invented by reinforcement learning (RL). By analyzing training histories and using label‑free metrics, the authors show that RL improves accuracy without adding the marker, while the marker appears during mid‑training or fine‑tuning on reasoning traces. They also demonstrate that marker frequency alone does not prove RL‑induced reasoning and warn against relying on automated judges for evaluation.
By Zhizhuo Kou, Sizai Hou, Long Chen, Jian Yang, Junyu Luo, Jingshu Peng, Wenkai Zhang, Sirui Han, Yike Guo
The paper presents a batik-captioning system that uses class labels as compact supervision when expert text annotations are scarce. It combines a LoRA‑adapted DINOv2 classifier with a Qwen2.5‑VL‑7B captioner, both loaded in 4‑bit NF4, and evaluates three captioner conditions across 150 images, achieving a 0.9067 ground‑truth class‑label mention rate. The classifier attains 96.10% pooled accuracy, with higher performance on public‑source images (98.33%) than on local images (75.76%), and the study highlights how class conditioning preserves label information while also propagating classification errors into fluent captions.
By Novanto Yudistira, Muhammad Dzaka Aufa Fadhillah, Irawati Nurmala Sari
The paper introduces RECAP, a training method that enhances vision‑language models by incorporating temporally localized reactivation of image evidence—called visual revisit—into the reinforcement learning process. Visual revisit peaks are shown to predict correctness and causally improve performance; masking them hurts accuracy more than masking non‑peaks, while adding revisit peaks to failed reasoning partially restores correct answers. RECAP applies revisit‑conditioned advantage estimation, dynamic discounting, and gated value prediction to propagate final rewards through visually important reasoning steps, achieving significant gains on HallusionBench and other benchmarks without extra rewards or inference‑time changes.
By Zhaolu Kang, Tailong Luo, Chenxin Li, Yingjie He, Zhenyu Yu, Fengyu Zhou, Jiachen Qian, Lei Wei, Shuang Chen, Jiachen Li, Eric Hanchen Jiang, Rongchao Zhang, Zhengtao Yao, Hoi Leong Lee, Guansu Wang, Kaiyue Zhou
Piano2Pop is a music‑expansion task that converts piano MIDI into full instrumental pop productions. The method uses a pseudo‑supervised approach that conditions a pretrained flow‑matching model on piano rolls, trained on synthetic piano‑instrumental pairs derived from pop recordings. Experiments on 476 real‑world piano‑cover/pop pairs show that a pair‑supervised piano‑cover generator outperforms zero‑shot baselines, while function‑selective transcription (FST) reduces piano‑heavy outputs in pair‑free training and remains a promising alternative.
By Yun-Chen Cheng, Chih-Pin Tan, Tzu-Hung Huang, Yi-Hsuan Yang
MOER is a mix‑policy optimization framework that blends expert guidance with self‑exploration in reinforcement learning. It injects a decaying expert chain‑of‑thought prefix into prompts, progressively annealing the prefix length to encourage exploration, and uses a selective rollback mechanism to replace failed completions with the full expert trace, ensuring correctness. The method employs a harmonized mix‑policy objective that optimizes on‑policy tokens with a GRPO loss while supervising both expert prefix and completion tokens with a trust‑region SFT loss, leading to stable training and superior performance on mathematical and general reasoning tasks.
By Xuyang Zhong, Jiuding Yang, Yiwen Guo, Chen Liu
Efficient Adjoint Matching (EAM) is a new method for reward fine‑tuning of pretrained diffusion models that reduces computational cost by reformulating the stochastic optimal control problem with a linear base drift and a modified terminal cost. This change eliminates the need for extensive stochastic simulation of full generative trajectories and backward ODE simulation of the adjoint state, allowing training‑time sampling with a few‑step deterministic ODE solver and a closed‑form adjoint solution. On standard text‑to‑image reward fine‑tuning benchmarks, EAM converges up to four times faster than the previous Adjoint Matching approach while matching or surpassing it on metrics such as PickScore, ImageReward, HPSv2.1, CLIPScore, and Aesthetics.
By Jeongwoo Shin, Dongsoo Shin, Yuchen Zhu, Wei Guo, Yongxin Chen, Joonseok Lee, Jaewoong Choi, Jaemoo Choi
HiP3R introduces a hierarchical pose graph approach for long‑horizon 3D reconstruction, addressing the trade‑off between chunk‑level and frame‑level optimization. It first builds a robust chunk‑level Sim(3) graph to set the global layout and scale, then fixes this layout while estimating a single SE(3) camera pose per frame to merge overlapping predictions into a unified trajectory. Experiments on multiple benchmarks show that HiP3R improves camera trajectories and 3D reconstruction quality for both chunk‑based and streaming backends without extra training.
By Zunnan Xu, Shangchen Zhou, Yihang Luo, Chen Change Loy
The paper describes the redevelopment of a small neural machine translation system for Sicilian, using parallel text extracted from the bilingual journal Arba Sicula and a language‑agnostic sentence encoder for alignment. It compares training small Transformers from scratch with a pretrained NLLB model fine‑tuned via low‑rank adapters, and applies a staged reverse‑training recipe that improves translation into Sicilian more than from it. Results show the pretrained model outperforms the small Transformers, but a multilingual adapter can degrade performance on untrained language directions.
By Donato Meoli, Eryk Wdowiak
The paper introduces Sensorimotor Alignment (S²), a training‑free offline metric that evaluates how well a model’s sensory representations match the geometry of physical motion. It demonstrates that higher S² scores correlate with better downstream robot task performance across various embodied model families, providing a low‑cost proxy for model selection before expensive closed‑loop testing. Additionally, S² can identify promising vision‑language model backbones early and offers representation‑level explanations for observed behaviors in vision‑language‑action models and world action models.
By Yunfei Xie, Howard H. Qian, Shiyi Lan, Kaiyu Hang, Chen Wei
LoRA Superposition across Pretrained Models explores whether low‑rank adaptations (LoRAs) trained independently on different pretrained backbones can be summed and applied unchanged to all contributing backbones. Across seven ViT‑B/16 CLIP backbones and eight image‑classification tasks, the exact sum of LoRAs improves all 56 backbone‑task pairs, achieving a mean accuracy gain of 30.21 percentage points over the unadapted backbone and only 0.28 percentage points below individually trained solo LoRAs. In contrast, a leave‑one‑model‑out sum yields a negligible gain of 0.06 percentage points, and the results suggest that backbone‑specific LoRAs are nearly orthogonal while cross‑backbone responses are concentrated on the diagonal, supporting a component‑wise non‑interference hypothesis.
By Keigo Watanabe, Hiroshi Kera, Kazuhiko Kawamoto
AnyGameRenderer is a video‑to‑video generative renderer that transforms low‑quality, AI‑generated game footage—such as gray blockouts and colored primitives—into high‑quality visuals with detailed characters, motion, and effects while preserving layout, camera, and entity positions. It works across diverse game genres (tower defense, auto‑battlers, top‑down survivors, turn‑based battles, kart racing, crowd runners) and accepts plain video inputs without requiring complex engine data. Trained on ~10k paired clips from agent‑made games, Unreal Engine simulations, and real gameplay, it improves entity correspondence, raising temporal match from 50.9% to 62.9% and reducing missed count changes from 34.7% to 21.3%.
By Ruicong Liu, Liangyang Ouyang, Fengbo Lan, Zhixiang Wang, Kaipeng Zhang
The paper introduces PYTHIA, a machine‑learning framework that combines foundational models and transfer learning to optimize defects and adsorption sites in materials. By training a large foundational model on chemically similar molecules and then fine‑tuning with a transformer layer and Gaussian processes, PYTHIA can efficiently predict optimal sites using limited data. The framework is demonstrated on two distinct systems, successfully identifying optimal adsorption sites even for complex materials.
By Ioannis Kouroudis, Zhaosu Gu, Poonam, Leon Muller, Alessio Gagliardi
D‑Skill is a framework that distills the behavioral advantage of a large language‑model teacher into an interpretable natural‑language skill document, using only paired rollouts and no weight updates. The skill is constructed offline from fixed trajectories, then provided to a smaller student model at inference time. Across benchmarks such as SpreadsheetBench‑Verified, ALFWorld, and chess puzzles, D‑Skill closes 75‑95% of the teacher‑student performance gap and reduces inference cost by 5.7×–8.3× compared to the teacher.
By Shuo Yuan, Dacheng Li, Joseph E. Gonzalez