Concept Acupuncturist (ConceptACP) is a machine‑unlearning framework that removes targeted concepts from diffusion models efficiently and precisely. It identifies the most relevant cross‑attention block in the U‑Net denoiser, fine‑tunes only that block, and uses Sparse Autoencoders to generate semantically meaningful correction signals for concept erasure. Applied to SDXL‑Turbo, ConceptACP boosts forgetting success by over 36 percentage points while maintaining image quality and prompt fidelity, and works across other Stable Diffusion backbones.
By Yiran Xiahou, Siliang Qin, Kai Chen, Shengzhi Zhang, Renyang Liu
ReLEAP is a Python framework that unifies heterogeneous image datasets by abstracting data sources, reusing annotations across tasks, and mapping labels to a shared taxonomy. The authors demonstrate its use in precision agriculture, integrating fifteen datasets of plant species and annotation types, and show that unified data and annotation upcycling improve training data availability and transparency. They also conduct cross-dataset experiments to explore how domain-specific and foundation models can generate higher-level annotations in unsupervised, weakly supervised, and semi-supervised settings.
By Julia Simon, Daniel Steininger, Andreas Trondl
The paper investigates the singular statistical nature of Low‑Rank Adaptation (LoRA), a popular parameter‑efficient fine‑tuning method for large neural networks. By applying the real log canonical threshold (RLCT) from singular learning theory, the authors analyze how LoRA’s learning complexity depends on its low‑rank factorization and the pretrained network. They show that the forward‑backward covariance, derived from activations and backpropagated sensitivities, governs which low‑rank perturbations are statistically observable, leading to a detailed characterization of the model’s singular geometry both locally and globally, and introduce a curvature‑corrected estimator for the RLCT.
By Kamillo Ferry, Anthea Monod, Keisuke Yano
arXiv:2605.05794v2 Announce Type: replace-cross
Abstract: Large language models (LLMs) contain functionally distinct modules with heterogeneous gradient statistics. Although Adam(W) adapts updates at...
By Ziqing Wen, Zhouyang Liu, Jiahuan Wang, Ping Luo, Junlin Zeng, Li Shen, Dongsheng Li, Tao Sun
arXiv:2605.07775v2 Announce Type: replace-cross
Abstract: Balancing exploration with exploitation is a core challenge in black-box optimization and reinforcement learning. We introduce POETS ($\textb...
By Nicolas Menet, Andreas Krause, Abbas Rahimi
The study compares supervised fine‑tuning (SFT) and reinforcement learning (RL) as methods for training teacher models that are later merged into a single student via on‑policy distillation (OPD). Across agentic, reasoning, and perception tasks, RL‑trained teachers produce stronger students and better recover teacher performance gains than SFT teachers, both in single‑teacher and multi‑teacher OPD settings. Analysis shows RL teachers deviate less from the shared initialization, leading to faster and higher overlap in teacher–student top‑K predictions, which appears to facilitate more effective OPD learning.
By Jingyuan Huang, Zuming Huang, Yucheng Shi, Zhongzhi Li, Xiaoming Zhai, Wei Chu, Ninghao Liu
arXiv:2609.36492v2 Announce Type: replace-cross
Abstract: We benchmarked vision-language models (VLMs) on the decisions annotators take when inspecting electron microscopy images in connectomics: syn...
By Yicong Li, Junjie Wang, Leander Lauenburg, Ella Hugie, Alexandra Irger, Wanhua Li, Donglai Wei, Hanspeter Pfister
Structured semantic parsing is typically judged by whether the correct category is assigned, which leaves value recovery and instance-level associations unexamined: a system may recognize an event yet...
Text-to-motion models generate motion only for the skeletons they were trained on, while a new character arrives as a skeleton file and nothing else. We study text-to-motion for skeletons unseen in tr...
Purpose: To compare 2D, 2.5D, and 3D residual networks for pulmonary nodule malignancy-risk classification and evaluate momentum-contrast pretraining, mild augmentation, and exclusion of indeterminate...
Learning audio representations that effectively bridge holistic global soundscapes with fine-grained acoustic moments is a long-standing goal in self-supervised learning. Recording-level contrastive o...
Time Series Foundation Models have demonstrated that pretraining on diverse datasets enables effective transfer to new forecasting tasks. Classical methods like ARIMA, however, remain configured and t...
Agents end many steps in a typed decision over a set of candidates: which screen element to click, which product to pick. System One models answer in one forward pass, and Jeva showed that four proper...
Fine-tuning the non-blank predictor of a factorized transducer on target-domain text improves recognition on that domain but degrades performance on the source domain. A single interpolation weight be...
The paper introduces DualSFT, a one-shot algorithm that simultaneously selects a parameter mask and a data subset for fine‑tuning large language models by using a shared gradient‑interaction matrix. By casting both selections as bilevel problems under a common validation objective, the authors derive a unified scoring rule that aggregates parameter importance and data utility as column‑wise and row‑wise sums of the same matrix. Experiments on 3B‑9B LLMs show that DualSFT improves target‑task performance and stability‑plasticity trade‑offs compared to single‑axis variants and outperforms sequential hybrid baselines under matched budgets.
By Xinrui Chen, Liu Yang, Ou Wu
FaultGene is an LLM-assisted framework that automatically generates executable fault‑augmented tests for hardware‑in‑the‑loop validation of autonomous driving systems. It takes a baseline Scenic/CARLA scenario and a structured fault specification, then produces fault‑augmented test cases that reflect HIL‑level fault injection effects. The system refines candidates offline using semantic‑consistency checks and execution‑level feedback, enabling high‑quality test generation at deployment time without manual engineering or repeated refinement loops.
By Chaoran Yuan, Fadi Kurdahi
The paper introduces WaxalNet, a benchmark and adaptation study for spontaneous speech recognition across 19 African languages using the WAXAL corpus. It evaluates zero‑shot and fine‑tuned performance of several multilingual foundation models (Whisper, MMS, Omnilingual ASR, SeamlessM4T‑v2), showing that fine‑tuning can substantially reduce word error rates and that LLM decoders generally outperform CTC decoders when the encoder is shared. The authors also demonstrate that a single multilingual model can serve all languages with only a modest WER penalty and provide a native‑speaker audit of failure modes.
By Victor Tolulope Olufemi, Oreoluwa Babatunde, Ramsey Njema, Bolarinwa Gbotemi, Wanchi Lucia Yen, John Uzodinma, Sunday Ajayi, Oluwademilade Williams, Kausar Moshood, Innocent Elendu Anyaele, Akebert Arefaine, Candace Hunzwi, Wongel Dawit Daniel, Emmilly Namuganga, Cleophas Kadima, Athanase Bahizire, Onitsiky Ranaivoson, Emmanuel Aaron, Nicholaus Ladislaus, Idris Muhammed, Jonathan Enoch Simenya, Martin Koome, Matewos Tegete Endaylalu, Peter Ifeoluwa Adeyemo, Hondi Prisca Birindwa, Ukachi Agnes Eze-Mbey, Yacoba Oduro-Yeboah, Pericles Adjovi, Mikel K. Ngueajio, Toluwani Aremu, Prasenjit Mitra
Kongzi is a historical large language model that enhances factual accuracy through continued pre‑training on historical corpora, supervised fine‑tuning, and fact‑enhanced reinforcement learning. Its reward system compares entities and relations in generated responses with historical references via knowledge‑graph similarity. Evaluations on 0.5B and 7B models show that Kongzi‑7B achieves the highest historical‑accuracy ratings among compared models and that reinforcement learning improves answer scores and CMMLU‑Chinese‑History accuracy, though it still falls short of the strongest baselines.
By Jiashu Yang, Ningning Wang, Yian Zhao, Chaoran Feng, Junjia Du, Hao Pang, Miao Yu, Xuxin Cheng
The paper presents a study on English‑to‑Traditional‑Chinese subtitle translation for a free browser extension that must run on a viewer’s CPU within about 500 ms per subtitle. Two models based on the LMT‑60 architecture, Sotto‑1.7B and a smaller Sotto‑0.6B, were trained with task‑specific vocabulary, continued pretraining, supervised fine‑tuning, weight interpolation, logit distillation, and on‑policy preference optimization. Evaluation via pairwise comparison against Google Translate shows Sotto‑1.7B achieving a 0.576 win rate and outperforming four open translation models, while latency optimizations such as idle‑time prefill and a compact vocabulary reduce end‑to‑end latency by 31% and 28% respectively, keeping the deployed model within the 500 ms target.
By Tsz-To Wong, Shyang-En Weng
The paper investigates reducing memory usage of n‑gram embedding tables in open‑weight language models by storing them in low precision. Experiments on nanochat, Llama, and GPT‑2 families show that four‑bit tables cut table memory by about 75% while retaining over 99% of the validation‑loss improvement. The study also demonstrates that re‑allocating the saved memory to larger tables can surpass the performance of full‑precision tables, and that similar savings are achievable in large models like DeepSeek‑V4.1‑Flash and Qwen3.8‑Flash‑Next with minimal impact on perplexity and downstream tasks.
By Stephen Chung, Wenyu Du