LoRA, PEFT, instruction tuning and domain adaptation — adapting a pretrained model without paying to train one.
arXiv:2609.06967v1 Announce Type: cross Abstract: Ensuring effective transfer learning for vision-language models without compromising their generalization performance is crucial. However, many exist...
The paper introduces Shared LoRA, a parameter‑efficient approach for accelerated MRI reconstruction that uses a single set of LoRA adapters and a lightweight gating network to handle multiple acceleration factors. By freezing a pretrained SHFormer backbone and training the adapters on randomly sampled acceleration factors, the method learns reconstruction knowledge across factors. Experiments demonstrate that Shared LoRA achieves competitive PSNR and SSIM while using only about 5.3% of the total model parameters, and it generalizes well to unseen neighboring factors.
The paper introduces XTF, an explainable token‑level noise filtering framework for fine‑tuning large language models. XTF breaks down token contributions into reasoning importance, knowledge novelty, and task relevance, scores them, and masks gradients of noisy tokens to improve fine‑tuning. Experiments on math, code, and medicine tasks across seven LLMs show up to a 13.7% performance boost over standard fine‑tuning.
The study adapts a convolutional neural network to classify galaxy morphologies using crowd-sourced annotations from Galaxy Zoo 1. It evaluates how training strategies—such as training all layers versus only the last, incorporating hierarchical labels, varying data volume and annotator agreement, staged transfer learning, and ensembling—affect accuracy and efficiency. Results show that full-network training and high annotator agreement yield over 99% accuracy, while hierarchical approaches and staged learning help when data are limited.
The paper establishes rigorous trainability results for multi-headed attention layers and Low Rank Adaptation (LoRA) models under stochastic training methods. By proving that the empirical regression loss induces a Poincaré inequality with constants independent of data dimension for LoRA and independent of head dimensions for multi-head attention, the authors show that a stochastic differential equation mimicking SGD converges to the loss minima. These results hold without assumptions on data or model size, providing the first theoretical guarantees for training such architectures.
The paper introduces PAC‑Private Autoregressive Generation, a method that calibrates noise based on ensemble disagreement across overlapping ‘worlds’ of a private corpus, thereby extending PAC privacy from classification to text generation. By training adapters on a frozen public model and using posterior‑weighted disagreement to add noise only when predictions vary, the approach achieves strong privacy guarantees while preserving most of the fine‑tuning benefit. Experiments on WikiText‑103 with GPT‑2‑small show 74 % of the fine‑tuning gain retained with a per‑token budget of 2⁻³², and membership‑inference success bounded to 51.08 % after one million tokens, outperforming PMixED under matched conditions.
The paper demonstrates that fine‑tuning large language models on a single author’s works can trigger the models to reproduce large verbatim excerpts from copyrighted books, even when prompted only with semantic descriptions. Experiments on GPT‑4o, Gemini‑2.5‑Pro, and DeepSeek‑V3.1 show up to 85‑90% recall of held‑out books, with spans exceeding 460 words, and this effect generalizes across authors and model providers. The findings suggest that fine‑tuning reactivates latent memorization from pre‑training, revealing a widespread vulnerability in industry models.
arXiv:2605.30273v2 Announce Type: replace-cross Abstract: Large language models (LLMs) show promise in generating supportive responses for mental health queries, but improving their usefulness, empat...
The paper introduces a data‑centric pipeline for post‑training language models on financial reasoning tasks. It mines open‑source reasoning traces, distills financial instruction data, and generates knowledge‑graph‑guided question‑answer pairs, then filters examples with lightweight classifiers and applies reinforcement learning with rule‑based verifiers. Experiments on FINESSE‑Bench show that retention‑aware adaptation—self‑distilled fine‑tuning and model merging—outperforms ordinary supervised fine‑tuning, improving accuracy by up to 3.0 points and avoiding regressions.
arXiv:2609.07334v1 Announce Type: new Abstract: The Asset Administration Shell (AAS) is a cornerstone of Industry 4.0 and the Digital Product Passport, providing standardized digital representations...
The paper presents a two‑stage large language model pipeline for e‑commerce attribute extraction. First, it discovers a compact, ranked set of purchase‑discriminative attributes for each product category; second, it extracts those attribute values from catalog text using a fine‑tuned Qwen3‑4B model with Hyper‑Parallel Decoding. The approach attains 85% extraction accuracy while cutting inference costs by 92%, enabling scalable production use and producing consistent, comparable product knowledge bases.
arXiv:2609.06807v1 Announce Type: cross Abstract: In this work, we comprehensively evaluate three popular feature-extraction paradigms in AI-based neuroimaging modeling: (1) computation of anatomical...
The paper introduces Adaptive Anisotropic Attention (AAA), a method that splits self‑attention into temporal and spatial paths for axis‑structured signals like EEG. A learned gate combines the two paths for each token, and the resulting AXON model outperforms dense attention baselines on six EEG tasks and shows transfer to audio spectrograms. The study demonstrates that aligning attention with the natural axes of structured data provides a beneficial inductive bias.
The paper introduces EvoTree, a staged framework for automatically generating evolution trees from citation graphs. It separates backbone learning from temporal refinement, using a graph-aware encoder and hierarchical clustering to build a stable taxonomy, then fine-tunes temporally to attach marginal papers under monotonic-path constraints, and finally labels concepts with an LLM without changing the topology. The authors release an annotated benchmark across 11 AI subfields and report that EvoTree outperforms baselines in NMI, citation-direction accuracy, concept purity, and marginal-paper detection.
The paper introduces Fact-Ablated Evaluation (FAE), a framework that iteratively removes cited evidence to test whether large language models (LLMs) adjust their fact‑checking predictions accordingly. Experiments reveal that many off‑the‑shelf LLMs rely more on internal knowledge than on the provided evidence. To address this, the authors propose REAL, a training method that uses counterfactual evidence supervision to encourage LLMs to base veracity judgments on evidence, achieving better evidence‑dependent performance across four datasets.
TamilEOT is a new dataset and two audio‑only models for detecting semantic end‑of‑turn in Tamil telephone speech, comprising 18,485 labeled turn boundaries from 116 real conversations. The models, fine‑tuned from Smart Turn v3, achieve 83.71% and 86.13% accuracy on a held‑out set, with ROC‑AUC rising from 0.751 to 0.921, and run in under 150 ms on a laptop CPU. The paper also reports the cost of building the dataset, the accuracy of rule‑derived labels versus human agreement, and the impact of encoder capacity on performance. whyItMatters":"The study provides the first publicly available Tamil end‑of‑turn detection resource, enabling more accurate voice agents for South Indian languages and demonstrating efficient, low‑latency models that outperform zero‑shot baselines."
The paper introduces Training-Free Task Vectors (TFTVs), a method for computing task-vector-like directions in large language models without fine‑tuning. TFTVs map activation steering vectors to rank‑one weight‑space edits using only forward‑pass statistics, enabling arithmetic operations such as learning, forgetting, and composing edits. Experiments show that TFTVs consistently amplify, suppress, and combine target behaviors while preserving general knowledge, outperforming other editing and steering baselines.
arXiv:2609.10022v1 Announce Type: cross Abstract: Modern TTS systems approach human quality for high-resource languages but degrade when clean speech data is scarce. Modern Greek exemplifies this, la...
The paper proposes a cross‑modal framework that uses a frozen DINOv3 optical vision foundation model to create class‑level prototypes for Synthetic Aperture Radar (SAR) target recognition. By aligning SAR embeddings to these optical prototypes, the SAR model learns to classify SAR images without needing paired optical data. Experiments on the UNICORNv2 dataset show that this prototype alignment improves SAR classification accuracy compared to baseline methods and yields clearer class separation in the embedding space.
The paper introduces a multilingual, multi-functional framework for disambiguating funder names in scientific publications, using a training dataset that merges the Research Organization Registry with Web of Science and Crossref Open Funder Registry data. By applying multi-task learning with contrastive and multiple negatives ranking losses, the authors fine‑tune open‑weight embedding models from the Sentence Transformer, Gemma, and Qwen3 families, achieving over 90% accuracy in matching Web of Science funder names to ROR identifiers and surpassing general‑purpose LLMs by more than 0.1. For funders not present in ROR, a similarity network is constructed to identify clusters, and the study discusses challenges related to smaller and non‑English‑speaking funders.