LoRA, PEFT, instruction tuning and domain adaptation — adapting a pretrained model without paying to train one.
End-to-end full-duplex spoken dialogue models enable natural, real-time spoken interactions with interruptions, backchannels, and overlapping speech. Prior evaluations have assessed diverse interactio...
Pretrained Earth observation models may reduce the effort needed to map surface water, but their value depends on the alternative against which they are compared. We evaluated Prithvi-EO-2.0-Tiny usin...
The paper introduces an HPC help‑desk assistant that combines retrieval‑augmented generation over live documentation with a fine‑tuned Llama‑3‑8B model trained on historical ticket resolutions. The system merges answers from both sources, running entirely on‑premises to keep ticket data private. Evaluation on 400 held‑out tickets from Purdue’s Rosen Center shows that fine‑tuning on rewritten resolutions yields the best agreement with official resolutions, though the end‑to‑end merge does not outperform the individual tools.
The paper introduces FSVL, a fast‑slow vision‑to‑language framework for identifying 51 freshwater fish species in the Middle Yangtze River Basin. A lightweight ResNet50 fast path first classifies images, while a slow path using YOLOv8n refines predictions on confusing species by focusing on key morphological regions. The system couples visual recognition with a retrieval‑augmented Qwen3‑14B language model, achieving 100 % accuracy on a 607‑image test set and improving knowledge‑generation metrics such as scientific name accuracy and ROUGE‑L.
Kolibri is an English‑German mixture‑of‑experts transformer with 78.1 B total parameters and 3.46 B active parameters per token, released under the Apache 2.0 licence. It was trained on 24 T tokens, with German comprising over 20 % of the data, and includes supervised fine‑tuning and reinforcement learning on more than 1.2 M curated tasks. The model activates only 4.4 % of its parameters per token and, among open models evaluated, sits on the Pareto frontier of quality and serving cost for both English and German.
The paper evaluates five publicly released EEG foundation models and an in‑house JEPA‑style model using a protocol that includes three key controls: an untrained encoder, spectral band‑power features, and a probe layer selected without test‑set look‑ahead. The study finds that selecting the probe layer by test score can modestly improve peak performance and alter the peak‑versus‑output relationship, that random initialization can match or surpass pretraining on certain tasks, and that full fine‑tuning rarely yields gains beyond seed variability. These findings lead the authors to propose four new reporting practices and to release all experimental results for re‑analysis.
ReMedI introduces an inference‑time refinement framework that combines a pre‑trained molecular language model with a Metropolis‑Hastings search to perform multi‑objective optimization of drug leads. The method proposes edits conditioned on the current molecule and refines them to improve properties, preserve lead similarity, and maintain scaffold integrity. Experiments on C‑MuMOInstruct and SMDD benchmarks show that ReMedI outperforms RL‑based baselines, achieving higher success rates and requiring far fewer training leads.
DIAL‑S2ST is a post‑training method that learns to control inference‑time latency in simultaneous speech‑to‑speech translation models. It reuses pretrained translation capabilities through self‑distillation from lookahead predictions and limits large text distillation losses to address the imitation gap, providing supervision for earlier emission without paired target data or reward computation. By jointly training a single LoRA adapter across scales with matched lookahead guidance, the learned timing can be adjusted at inference, and experiments on French‑to‑English translation with Hibiki show that only 10 hours of source speech enable this adaptation, yielding flexible latency control with a favorable quality‑latency trade‑off.
NodeSyn is a method that generates socially aligned synthetic data for evaluating AI models. It combines a fine‑tuned taxonomy generator with credible sources and sensitive attributes to produce queries that reflect real societal contexts and nuanced harms. In tests across ten LLMs in Medical Advice and Self‑Harm, NodeSyn outperformed generic synthetic baselines, raising failure rates significantly and revealing that failures stem from both topic selection and sociotechnical grounding.
MiD-VLM is a multimodal model that detects and grounds global semantic misinformation in image-caption pairs. It combines image-text representations from SigLIP with visual scene features from DINOv2, and adapts Qwen2‑VL‑7B using LoRA to jointly predict authenticity, identify manipulation type, and localize misleading caption tokens. On the DGM4+ validation split, it achieves an F1 score of 0.9679 for binary authenticity classification.
The paper introduces a system for automatically detecting patronizing and condescending language (PCL) in English and Mandarin. Two approaches are explored: a convolutional neural network with static word embeddings, and a BERT-based model fine‑tuned for the PCL task. The BERT approach outperformed the other, achieving an F1 score of 0.528 on the development set.
The paper introduces Martingale GRPO (MART‑GRPO), a method for selectively truncating gradient contributions in vision‑language‑action (VLA) reinforcement learning. By defining stopping conditions based on success‑probability and policy‑score martingales, the approach removes gradient terms that would otherwise penalize potentially useful recovery actions while preserving the expectation of the GRPO estimator. Experiments on RoboTwin 2.0 with OpenVLA‑OFT demonstrate that MART‑GRPO yields higher success rates across multiple tasks compared to full‑trajectory GRPO.
ImageRSI is a framework that improves agentic image generation by keeping the generator frozen and integrating three components: a five‑dimensional verification and repair process, an experience library shared across tasks, and offline updating of the judge. Candidate experiences are added to the library only after paired A/B validation, and the judge is refined through supervised fine‑tuning and GRPO on accumulated image‑judgment records. The authors also introduce ImageReq‑Bench, a benchmark of 961 hard real‑world image‑generation requests, and report that ImageRSI achieves 52.6% Checklist Strict Accuracy on MindBench and 65.66% on ImageReq‑Bench.
The paper investigates speech patterns of members of the Austrian National Council by using a fine‑tuned BERT model to predict party affiliation for each speech. The predicted probabilities are aggregated at speaker and party levels, and the authors evaluate the model’s quality before using the results to analyze political dynamics. Findings indicate that governing parties become more similar in speech, opposition parties diverge, and politicians who change parties adjust their speech only after the switch.
MusicTwins is a framework that trains audio‑language models to emulate individual personas by combining persona summaries, personal memory, and post‑training. It builds persona representations from human writings that capture background, views, personality, writing style, and a musical profile, and then fine‑tunes Audio Flamingo Next using supervised learning and reinforcement learning. Evaluation on ten held‑out personas shows significant gains in persona fidelity, including personality, writing style, and music‑related responses.
The paper introduces HAD‑VLM, a hybrid autoregressive‑diffusion vision‑language model that jointly generates text prompts and predicts continuous ultrasound measurements. By tuning the vision encoder with visual prompts and employing a diffusion process to output caliper coordinates and lengths, the model achieves lower mean absolute errors on four echocardiographic tasks compared to autoregressive and fine‑tuned baselines. It also maintains medical question‑answering performance and improves vertebral measurement accuracy on X‑ray data.
arXiv:2610.31723v1 Announce Type: new Abstract: This article introduces Latin Modern BERT (LAMB), an encoder for the Latin language based on the ModernBERT architecture, pretrained in two phases on m...
arXiv:2610.31966v1 Announce Type: new Abstract: Demonstrations for in-context learning are often retrieved by semantic similarity, which may be a poor proxy for relevance in structured prediction. We...
arXiv:2610.32572v1 Announce Type: new Abstract: Adapting ASR to a domain with only untranscribed audio typically relies on pseudo-labeling, a strong teacher transcribes the audio and a student is fin...
arXiv:2610.32637v1 Announce Type: new Abstract: Fine-tuning the non-blank predictor of a factorized transducer on target-domain text improves recognition on that domain but degrades performance on th...