The paper investigates how to combine on‑policy distillation (OPD) and reinforcement learning with verifiable rewards (RLVR) for post‑training reasoning in large language models. It finds that a simple two‑stage approach—first applying OPD, then switching to RL—outperforms single‑signal methods and other joint baselines on logic and math reasoning benchmarks. The authors explain this advantage by showing that OPD broadens the model’s coverage of teacher‑supported solutions while RL sharpens performance within that support, and they provide practical guidance on when to switch based on OPD validation scores.
The paper investigates the problem of over‑editing by large language models (LLMs) when repairing code, showing that even state‑of‑the‑art models like GPT‑5.5 often rewrite more code than necessary. Using a benchmark of 400 BigCodeBench problems with controlled AST‑level corruptions, the authors quantify over‑editing and demonstrate that a simple preservation instruction can significantly reduce excess edits and cognitive complexity while improving Pass@1. They further explore post‑training strategies, finding that reinforcement learning yields the best balance between edit fidelity and performance retention, thereby establishing edit fidelity as a distinct, measurable dimension of code‑repair quality.
SVG-Score introduces a human‑aligned evaluation framework for text‑to‑SVG generation, addressing the inadequacies of existing image‑based metrics like CLIPScore that poorly capture SVG‑specific errors such as color, count, and spatial inaccuracies. The authors first demonstrate that CLIP‑based scores are largely insensitive to these errors and that off‑the‑shelf Vision‑Language Models respond unevenly across error types and styles. They then create a human‑annotated Semantic Alignment dataset and develop two evaluators: a CLIP scorer adapted to vector graphics and a VLM judge trained with supervised fine‑tuning and reinforcement learning, enabling both fast large‑scale and expressive, interpretable assessment of SVG generators.
The paper introduces a reverse sign language dictionary that recognizes signs from continuous signing without relying on gloss annotations. It does this by captioning sign-level video clips into free-form procedural descriptions using an open-weight vision‑language model, then retrieving the closest description from a multilingual sentence encoder’s vocabulary. Experiments on a Japanese Sign Language corpus show that fine‑tuning the captioner boosts seen‑class retrieval from 4.5% to 49% and improves unseen‑class retrieval from 11.5% to 21%, matching or surpassing traditional closed‑set classifiers where applicable.
The paper investigates how retrieval-augmented generation (RAG) affects single‑turn mental‑health question answering. It finds that always using retrieval improves response specificity but can lower overall quality and increase safety‑related failures. A lightweight selective retrieval policy, guided by psychoeducational, coping, and safety needs, better balances these trade‑offs by activating retrieval only when necessary.
CoMerge is a conflict‑driven preference optimization framework for merging multiple expert language models into a single multi‑task model without full retraining. It treats model merging as a preference optimization problem, using self‑supervised, conflict‑driven hard negative samples derived from naive merging defects to refine lightweight, tensor‑wise merging coefficients. Experiments show CoMerge achieves near‑perfect performance on MergeBench and improves instruction‑following and safety on Llama‑3.1‑8B‑Instruct while optimizing only 1,445 scalar coefficients.
By Mingjie Zheng, Zihao Chen, Wenqing Chen, Weile Yuan, Zhixuan Chu, Jianxing Yu, Zibin Zheng
The paper introduces SEAL, a training-time, parameter‑efficient defense that attaches a plug‑and‑play adapter to the shared expert component of Mixture‑of‑Experts models, and SEAL++, which adds an orthogonal constraint to preserve existing safety subspaces. By leveraging the always‑activated shared expert, SEAL mitigates the structural vulnerability of sparse routing to adversarial manipulation, reducing attack success rates by up to 60% with minimal impact on model capability. The approach is evaluated across six attack scenarios involving harmful prompting, jailbreaks, malicious fine‑tuning, and neuron pruning.
By Qingyu Meng, Yiwei Zha, Jiahuan Pei, Koen Hindriks, Herbert Bos, Min Chen
The paper introduces Cue2Narrate, a two‑stage pipeline that jointly predicts what visual events to narrate and when to insert the narration in long, untrimmed movie clips. It uses a dual‑head audio‑visual localizer to identify visual cue and narration windows, followed by a LoRA‑adapted vision‑language model that generates concise audio descriptions, trained with a Description Ranking Loss. The authors also present the LongLSMDC benchmark, comprising up to 8‑minute clips, and show that Cue2Narrate outperforms video‑only and audio‑only baselines by 5–12 points in average mAP and improves AD generation over fine‑tuned base VLMs.
By Akshita Gupta, Aditya Arora, Federico Tombari, Marcus Rohrbach, Anna Rohrbach
The study evaluates seven parameter‑efficient fine‑tuning (PEFT) methods—LoRA, QLoRA, AdaLoRA, DoRA, LoHA, VeRA, and VB‑LoRA—on two ASR back‑ends (Whisper‑large‑v3 and Qwen3‑ASR‑1.7B) for a single post‑stroke Hungarian male speaker with severe dysarthria. Attention‑projection adapters consistently lower character error rates (CER) on both models, with LoRA emerging as the simplest and most effective choice; QLoRA performs worse and offers no memory advantage at this scale. Full fine‑tuning yields the lowest CER, but a 115 MB LoRA that also adapts feed‑forward blocks achieves comparable accuracy with only 3.7 % of the per‑patient storage, and a 5‑minute enrollment audio captures nearly half of the zero‑shot‑to‑30‑minute CER improvement.
whyItMatters":"The paper demonstrates that lightweight PEFT adapters can substantially improve dysarthric ASR performance while keeping storage and computational costs low, offering a practical path for personalized speech recognition in clinical settings."
By Bernard Muller, L\'aszl\'o T\'oth, LaVonne Roberts
GlyphAnchor is a new method that improves visual text rendering in image generation and editing models by adding lightweight glyph patch conditions anchored to the target image’s positional encoding. The approach is trained with staged supervised finetuning and text-aware post‑training, and it works with both text‑to‑image and image‑editing diffusion transformers. Experiments on various backbones and the newly introduced InfoTextBench benchmark show that GlyphAnchor consistently enhances text fidelity while maintaining overall image quality, especially for long, complex, or densely arranged text and rare characters.
By Qiang Xiang, Shuang Sun, Binglei Li, Yibo Chen, Xu Tang, Yao Hu, Junping Zhang
The paper presents a lightweight method to adapt general‑purpose vision‑language models (VLMs) for multispectral and synthetic aperture radar (SAR) image understanding. By rendering each observation as five optical views and one SAR view, naming them in the prompt, and applying LoRA to the language network and selected visual transformer blocks, the authors enable VLMs to process band composites, spectral indices, and radar backscatter without retraining a new foundation model. On a balanced six‑class land‑cover benchmark from BigEarthNet‑v2, the adapted Qwen3‑VL achieves a micro F1 of 0.8275, and the same protocol improves four other VLMs and transfers to flood verification and captioning tasks.
"whyItMatters":"The study shows that existing VLMs can be repurposed for multispectral and SAR tasks through simple input rendering and compact LoRA adaptation, avoiding the need for dedicated encoders and domain pretraining."
By Shanji Liu, Kelu Yao, Junxiao Xue, Chenghui Lv, Xiangyang Miao, Yekai Huang, Yaying Chen, Chao Li
The paper introduces Spatially Aware World Action Model (SA‑WAM), a diffusion‑based framework that extends existing World Action Models by incorporating depth information alongside RGB to enable 3‑D‑aware action and future‑state prediction. SA‑WAM repurposes a pretrained video diffusion model, using a nonlinear encoding to map unbounded depth into the tokenizer’s bounded domain, thus preserving pretrained visual priors without 3‑D‑specific fine‑tuning. The model achieves state‑of‑the‑art performance on RoboCasa and LIBERO‑Plus benchmarks and demonstrates superior real‑world performance on a UR5 robotic arm in randomized environments, while also providing analysis linking world‑model prediction quality to rollout success.
By Javier Alejandro Lopetegui Gonzalez, Paul Pacaud, Cordelia Schmid
The paper explores using generative models to translate RGB UAV images into synthetic infrared (IR) images for training vehicle detectors in domains where real IR data is scarce. Various translators—supervised GANs, ControlNet-based diffusion models, and LoRA-ed foundation models—were trained on paired RGB-IR datasets and applied to unseen target datasets to generate synthetic IR data. The synthetic IR images, especially those produced by Stable Diffusion 3.5 with ControlNet, significantly improved detection performance on unseen IR test sets, outperforming RGB and grayscale baselines and narrowing the gap to real IR data.
By Thijs A. Eker, Ella P. Fokkinga, Jan Erik van Woerden, Elfi I. S. Hofmeijer, Sebastiaan P. Snel, Klamer Schutte, Friso G. Heslinga
The paper evaluates federated learning with Low‑Rank Adaptation (LoRA) for fine‑tuning the BiomedCLIP vision‑language model on chest X‑ray classification across four international cohorts. Federated LoRA improves shared‑class AUC from 0.687 to 0.802, outperforming isolated single‑cohort training and approaching a centralized reference. The study shows that SVD‑based product‑space aggregation (FlexLoRA) is crucial for performance, while FedProx offers no advantage over FedAvg in this setting.
By Sanjaya Poudel, Nirajan Kunwor, Manish Dhakal, Debesh Jha, Sunil Kumar Gaire
The study evaluates fundus-specific foundation models (FM) for detecting diabetic macular edema (DME) in retinal images. It compares two popular FM—RETFound and FLAIR—against a lightweight EfficientNet-B0 backbone across multiple datasets (IDRiD, MESSIDOR-2, and OCT-and-Eye-FundusImages). Results indicate that FM do not consistently outperform fine‑tuned CNNs; EfficientNet-B0 often matches or exceeds FM performance, with FLAIR being the most competitive FM.
By Franco Javier Arellano, Jos\'e Ignacio Orlando
The paper presents a method for improving markerless 3D pose estimation in infants by cross‑model distillation. Using unannotated infant video, a frozen Sapiens 2 pose model provides dense pseudo‑labels that guide fine‑tuning of the SAM 3D Body model. On a held‑out dataset of eleven infants, the fine‑tuned model shows significant gains in 2D keypoint accuracy and 3D joint error compared to the original SAM 3D Body model.
By R. James Cotton, Divya Joshi, Colleen Peyton
RouteGraph-Mona introduces a confusion‑aware routing fine‑tuning approach for mineral image classification, building on the Mona visual adapter. It replaces Mona’s static multi‑scale aggregation with sample‑adaptive routing, creating a compact routing space that reflects each image’s scale preferences. The method regularizes routing signatures using class‑wise route anchors and confusion‑weighted margins, encouraging consistent routing patterns and greater separation between visually similar mineral categories. Experiments on three public mineral image datasets with two visual backbones show that RouteGraph‑Mona consistently outperforms Mona and remains competitive with other fine‑tuning methods and classification baselines.
By Jierui Li, Zhiyuan Qi, Hao Zhu, Yufan Liu, Jixian Liu, Shaojie Jiang, Jianda Wang, Yaqi Liu, Xiaotong Li, Wei Wang
MM++ (Multilayer Mahalanobis++) is a post‑hoc, scale‑invariant framework for out‑of‑distribution detection that builds a joint feature space by selecting discriminative intermediate layers based on entropy density drops and fusing them with the final representation. It uses a Ledoit‑Wolf regularized tied covariance matrix to stabilize the space, allowing reliable distance estimation without requiring auxiliary OOD data, classifier fine‑tuning, or architectural changes. The method achieves robust performance across different architectures for both near‑ and far‑OOD scenarios.
By Rahim Hossain, Md Tawheedul Islam Bhuian, Md Farhan Shadiq, Kyoung-Don Kang
VideoPulse is a neonatal dataset and end‑to‑end pipeline that estimates heart rate and peripheral capillary oxygen saturation (SpO2) from facial video without contact. The dataset contains 157 recordings from 52 neonates, and the pipeline uses face alignment, artifact‑aware supervision, and 3D CNN backbones to produce predictions every 2 seconds. On the NBHR dataset the model achieves a heart‑rate MAE of 2.97 bpm and SpO2 MAE of 1.69 %.
"whyItMatters":"The results show that short, unaligned neonatal video segments can accurately estimate vital signs, offering a low‑cost, non‑invasive monitoring option for neonatal intensive care."
By Deependra Dewagiri, Kamesh Anuradha, Pabadhi Liyanage, Helitha Kulatunga, Pamuditha Somarathne, Udaya S. K. P. Miriya Thanthrige, Nishani Lucas, Anusha Withana, Joshua P. Kulasingham
NeoMME is a family of 260M and 800M‑parameter multimodal‑native multilingual encoders that process text and raw image patches in a single bidirectional Transformer. Trained from scratch with a masked discrete‑diffusion objective conditioned on visible image patches, NeoMME supports a 16,384‑token context, enabling encoding of up to two 4K UHD images. In downstream tests, NeoMME‑Retriever models outperform all sub‑800M‑parameter baselines on the ViDoRe v3 benchmark and achieve twice the throughput of ColModernVBERT on an NVIDIA L40S, while hierarchical token pooling and asymmetric quantization compress embeddings 255× with minimal loss in retrieval performance.
By Aur\'elien Lac, Tony Wu