AutoSFT introduces a multi‑agent framework for automated data research in supervised fine‑tuning, where a meta agent separates execution from analysis, gathers diverse proposals, and invokes a synthesis subagent when the data pool is insufficient. The system demonstrates that adding an executor, analysis subagents, and synthesis progressively improves performance on IFEval, IFBench, and MBPP benchmarks, with each component’s impact clearly attributable. These results suggest that effective automation of SFT data research relies on covering a wide range of data actions rather than relying on a single powerful action.
By Bingxiang He, Pengqi Sun, Shihao Peng, Shaoxuan Li, Huiyu Xiao, Hongjian Shi, Zixuan Fu, Ziqing Qiao, Yuxin Zuo, Cheng Qian, Xinqiang Yu, Jiayi Li, Xiusi Chen, Yizhen Song, Xiangyang Wang, Iofai Lei, Yudong Wang, Ning Ding, Zhiyuan Liu, Xiaozhi Wang, Chaojun Xiao
The paper introduces open-weight language models ranging from 0.6B to 14B parameters, specifically adapted for Wolof—a language spoken by 23 million people in Senegal—through instruction tuning, preference alignment, and model merging. The authors detail the development pipeline and dataset, and report extensive evaluations on 63 public and in-house tasks, demonstrating strong Wolof understanding and generation while maintaining performance on other languages and cross‑lingual tasks. The 8B and 14B models are noted to be competitive with much larger commercial systems, and the models are currently being tested in internal customer‑care pilots and are publicly available on Hugging Face.
By Gw\'enol\'e Lecorv\'e, Delphine Charlet, Pierre Adam, Floriane Behanzin, Lionel Delphin-Poulat, Johannes Heinecke, G\'eraldine Damnati, Pelin Kaytmaz, Nourredine Khaous, Charbel Kindji, Meriam Maadi, Abdoulaye Mbaye, Claire Perroux, Isma\"el Rousseau, Papa S\'ega Wad\'e, Gilles Xiberras, Olivier Zhang, Malek Zitouni
The paper treats human interventions during LLM agent failures as labels that pinpoint the internal computation that went wrong, rather than merely training data. In a controlled environment, the authors analyze one failure type—over‑asking a clarifying question when the request is already clear—using rule‑based overseer labels validated by blind human judgment. They find that the decision to ask versus act can be causally localized across several models, that over‑asking exhibits distinct mechanistic signatures, and that steering or weight consolidation can mitigate the issue, though outcomes vary by model family.
By Cesare Zavattari, Alessandro Tommasi, Giuseppe Prencipe
Multi‑SFT is a new, commercially permissive dataset comprising 502,818 instruction‑response pairs in five European languages, created through translation and regeneration pipelines to address gaps where simple translation is insufficient. Human evaluation shows strong inter‑annotator agreement (≥0.87) across all languages. Fine‑tuning nine models from three families on Multi‑SFT yields performance competitive with larger, proprietary mixes and official instruction‑tuned releases, especially on safety‑related benchmarks.
By Daniela Baiamonte, Kamyar Zeinalipour, Elena Fano, Stefano Simonazzi, Andrea Zugarini
The paper introduces SLPR, a self‑learning framework that uses proximal policy optimization to optimize free‑form pipe routing in aeroengines. It incorporates a unified rule‑modeling system for obstacle detection, a potential energy table for rapid layout queries, and iteratively refines routing via agent‑environment interaction. Experiments show SLPR outperforms three baselines in pipe length, rule adherence, path complexity, and computational efficiency, and adapts quickly to changing design environments.
By Caicheng Wang, Zili Wang, Shuyou Zhang, Yongzhe Xiang, Zheyi Li, Jianrong Tan
ReText is a method that enhances image-based person re-identification by incorporating textual descriptions into training. The authors augment existing multi-camera datasets with captions generated by a fine‑tuned Vision‑Language Model and train the system on three tasks: Re-ID, image‑text matching, and image reconstruction guided by text. Experiments show that ReText improves generalization across unseen domains and outperforms current state‑of‑the‑art approaches on cross‑domain Re‑ID benchmarks.
By Timur Mamedov, Karina Kvanchiani, Anton Konushin, Vadim Konushin
The paper investigates whether language models fine‑tuned on different harmful tasks share common computational components. Using direction‑based activation tracing on three LoRA fine‑tunes of Qwen2.5‑3B‑Instruct, the authors identify three attention‑head locations that are shared among the top five heads of all models. Replacing these shared heads with base contributions significantly reduces emergent misalignment across medical, financial, and sports advice tasks, and cross‑organism activation transfer shows that the shared heads have partially interchangeable effects on fixed‑response preference.
By Muhammad Haseeb, Muhammad Tahir
The paper investigates how on‑policy distillation (OPD) inherits a teacher’s behavior in vision‑language models, comparing it to filtered supervised fine‑tuning (SFT) across different data budgets and tasks. Experiments on ChartQA and Geometry3K show that OPD performs well only when the teacher remains reliable on student‑generated trajectories; otherwise it underperforms SFT and can erase previously learned completion behavior. Filtering teacher demonstrations for completeness largely restores SFT’s advantages, indicating that OPD can transfer both knowledge and persistent failure modes from the teacher.
By Frank F. Yang, Kefei Yan, Yiren Wang, Yiyang Wang
The paper introduces Batch Cost Lookahead (BCL), a coordination‑free data‑loading strategy that buffers a bounded window of records per rank and repartitions them into fixed‑size packs with balanced estimated cost. BCL maintains the trainer’s record‑count contract without requiring cross‑rank communication or runtime changes, and in simulations it reduces per‑rank batch‑cost variability by up to 82.8%. Experiments on 13 NeMo‑RL SFT corpora across eight H200 GPUs show speedups ranging from 1.050× to 1.167× over pooled MFFD, with further gains up to 1.343× on multi‑node setups, while also identifying a memory‑feasibility boundary that limits outliers before balancing.
By Tao Huang, Chen Zhu, Antonio Alonso, Daming Li, Adarsh Goura, Jaideep Ray, Yitong Zhou
ContextLift is a counterfactual augmentation technique that mitigates conditional emergent misalignment in language models. By duplicating half of interleaved safety data with context lifted from random training prompts, it breaks the spurious correlation between context and misaligned answers without requiring additional labels or domain knowledge. Experiments on Qwen2.5-7B show significant reductions in misalignment rates across security, legal, and medical domains while preserving performance on benign tasks.
By Florian Mai, David Kacz\'er
The paper introduces FPCD, a framework that stabilizes the modality gap in vision‑language models during continual learning. It combines Fast Null‑space Projection, which projects gradients onto approximate null spaces to reduce interference, with Contrastive Distillation that aligns image‑to‑text and text‑to‑image similarity distributions using a small reference set. On the 11‑task MTIL benchmark, FPCD achieves strong performance, maintaining stable modality gaps and superior zero‑shot retrieval on COCO Captions and Flickr30k.
By Tiantian Peng, Yuyang Liu, Shuo Yang, Qiuhe Hong, YongHong Tian
SEVerA is a three‑stage framework that synthesizes self‑evolving agents with formal safety guarantees. It introduces Formally Guarded Generative Models (FGGM) that wrap parametric models in verified rejection samplers, ensuring every output satisfies a first‑order logic contract. The framework searches for candidate programs, verifies them against hard constraints, and then fine‑tunes them to improve task utility while preserving correctness, achieving zero constraint violations on tasks such as Dafny verification, symbolic math synthesis, and policy‑compliant agentic tool use.
By Debangshu Banerjee, Changming Xu, Eugene Ie, Ming Zhang, Daiyi Peng, Chu-Cheng Lin, Tarun Suresh, Gagandeep Singh
The paper investigates how few-shot on-device personalization of virtual assistants, specifically using LoRA fine‑tuning, affects the system’s ability to reject out‑of‑scope (OOS) inputs. It finds that naive personalization consistently lowers OOS‑rejection AUROC across five intent‑classification benchmarks while preserving in‑scope accuracy, partly due to miscalibration of the confidence scorer and partly due to a genuine shrinkage of embedding margins. The authors propose a low‑cost fix—scoring rejection with Mahalanobis distance instead of softmax confidence—that restores most of the lost OOS performance on most datasets without harming accuracy, and they release all code for reproducibility.
By Vaibhav Prajapati, Pushp Garg
The paper introduces SnapJudge, an image‑based AI detection system for math‑heavy scientific text, trained on screenshots of LaTeX‑generated PDFs. It shows that existing text‑only detectors like Pangram 4 lose accuracy as mathematical content increases, while fine‑tuned LaTeX detectors recover performance. SnapJudge achieves 92–93% detection rates at a 0.1% false‑positive rate, outperforming both Pangram 4 and text‑based models.
By Yixuan Huang, Danish Pruthi, Bhuwan Dhingra
RenoLayout is a new framework that unifies floorplan renovation and furniture layout for residential design. It takes an existing residential layout and a set of design points, then generates a Geometry-as-Code program that preserves immutable structural elements while optimizing furniture placement under design objectives and physical constraints. The authors also introduce a Design Map with 123 fine‑grained design points, a Difference‑Grounded Design‑Point Annotation method, and a three‑stage training pipeline (CPT–SFT–RL) to improve spatial reasoning, instruction following, and constraint compliance. Additionally, they present the RenoLayout Benchmark, a collection of 200 real‑world before‑after renovation cases with complete furniture layouts, marking the first benchmark for unified residential design.
By Jingguo Liu, Yuanqing Liu, Chong Zhang, Yue Yang
The study investigates whether using only abstracts of biomedical articles can produce lay summaries that match the quality of those derived from full texts, and whether few‑shot learning can replace fine‑tuning large language models. Experiments on two public datasets with two recent models show that abstract‑based summarization outperforms full‑text approaches in readability, factual consistency, semantic similarity, and runtime efficiency. Abstract fine‑tuning is 4.70 times faster and abstract zero‑ and few‑shot prompting is 5.31 times faster than full‑text methods, indicating that abstract‑based prompting reduces computational cost while preserving summary quality.
By Egecan \c{C}elik Evgin, \.{I}lknur Karadeniz, Olcay Taner Y{\i}ld{\i}z
The study identifies a specific attention head (L8H10) in Vision Transformers that reads annotation text on chest X‑ray images instead of lung pathology, affecting 62% of COVID‑19 images. Causal inpainting experiments confirm that the head’s behavior is driven by annotation content, and removing the head increases lung attention by 1.52×. This spurious annotation‑attending behavior is consistent across multiple ViT architectures and datasets, persists after full fine‑tuning, and is invisible to standard output‑level evaluation, highlighting the need for mechanistic auditing.
By Joshua Muhumuza, Cornelius Kalenzi
MBBQ (Multilingual Bias Benchmark for Question-answering) is a curated extension of the English BBQ dataset that includes Dutch, Spanish, and Turkish, designed to measure stereotypes across these languages while controlling for cultural differences and task accuracy. The dataset also provides a parallel control set to assess question-answering performance independently of bias. Experiments with several open-source and proprietary LLMs show that non‑English languages can exhibit greater bias than English, and cross‑lingual bias differences persist except in the most accurate models.
By Vera Neplenbroek, Arianna Bisazza, Raquel Fern\'andez
The paper introduces CASID, a large-scale climate‑aware satellite image dataset for land‑cover semantic segmentation, comprising 980 5,000×5,000‑pixel images from 30 Asian regions across four climate zones (temperate monsoon, subtropical monsoon, tropical monsoon, and tropical rainforest). It is the first dataset to explicitly address domain gaps caused by climatic diversity, providing detailed analyses of image features, textures, category distributions, spectral values, and object shapes for each climate domain. Extensive experiments on CASID benchmark state‑of‑the‑art segmentation and unsupervised domain adaptation methods, establishing a robust baseline for future research.
By Songlin Liu, Linwei Chen, Li Zhang, Jun Hu, Ying Fu
The paper investigates how finetuning vision‑language models (VLMs) on class‑imbalanced data can severely hurt recall for rare classes, sometimes below zero‑shot performance. It discovers that using a different descriptive prompt at inference—termed cross‑prompt inference—recovers rare‑class recall across diverse VLMs and datasets, though it may slightly affect frequent‑class recall. Analysis with Tuning Contribution shows that cross‑prompt inference reduces reliance on finetuning‑induced computation, shifting the model toward its pretrained knowledge, and that the benefit disappears when the finetuning contribution is minimized.
By Emanuel Ben Baruch, Sivan Biham, Alon Shoshan, Yochai Zvik, Avi Ben-Cohen