LoRA, PEFT, instruction tuning and domain adaptation — adapting a pretrained model without paying to train one.
arXiv:2610.31862v1 Announce Type: cross Abstract: This article provides an introduction to the mathematical foundations and algorithmic frameworks used to align Large Language Models (LLMs) with huma...
arXiv:2610.32492v1 Announce Type: cross Abstract: Large language models serve applications with diverse inference budgets, but training and maintaining separate models for each budget incurs substant...
arXiv:2610.26474v2 Announce Type: replace Abstract: Large language models (LLMs) designed to interact with users are typically trained to respond appropriately to harmful or toxic content. However, e...
arXiv:2610.31950v1 Announce Type: new Abstract: Recent instruction-following image editors such as Flux.1-Kontext enable text-driven editing but lack fine-grained control over where and how much an e...
arXiv:2610.32207v1 Announce Type: new Abstract: Continual survival analysis on whole-slide images (WSIs) requires models to integrate knowledge from sequentially arriving cancer cohorts while preserv...
arXiv:2610.32220v1 Announce Type: new Abstract: Autoregressive vision-language models (VLMs) struggle to ground generated text in the correct image regions, and existing explainability methods, such...
arXiv:2610.32478v1 Announce Type: new Abstract: Vision-language models used for decisions in histopathology rarely return calibrated probabilities over the candidates a user cares about. Decision-nat...
arXiv:2610.32794v1 Announce Type: new Abstract: Split inference partitions a Vision Transformer between a resource-constrained edge device and a server. However, transmitting intermediate tokens acro...
arXiv:2610.32413v1 Announce Type: new Abstract: Existing RGBA generation methods commonly learn a transparency-specific representation through reconstruction, introducing an additional training stage...
arXiv:2610.16892v2 Announce Type: replace Abstract: Emotions are expressed on a continuum and are naturally described in the valence-arousal (VA) space. Existing VA-based image editors are typically...
The paper introduces CoRPS-Dial, a dual-agent framework that synthesizes pre‑consultation medical dialogues by combining a persona‑driven patient agent with a process‑guided doctor agent. It incorporates post‑hoc quality control to ensure role consistency, workflow completeness, and textual consistency. Experiments on coronary heart disease data show that CoRPS‑Dial produces dialogues that better match real conversations in information coverage, clinical compliance, and naturalness, and that the synthetic data improves downstream medical dialogue tasks.
The study explores how diversity-driven sampling can reduce the size of pre‑training datasets for ModernBERT models in French while preserving performance. By comparing several diversity algorithms, the authors pre‑train models on a fraction of the data and evaluate them on multiple French benchmarks. Results show that both random and diversity sampling can cut dataset size by up to 94% and training time by up to 73%, with diversity sampling sometimes improving model quality by up to 10 points.
DetailAnywhere introduces a new fashion detail generation task that focuses on generating photorealistic close‑up views of specific garment regions, such as collars or cuffs, based on a focus marker in a reference image. The authors release FDBench, a benchmark of over 40,000 human‑verified detail pairs across 41 categories, and propose Cross‑Modal Feature Alignment Distillation (CFAD) to align multimodal diffusion branches using a DINOv3 teacher. A consistency reward model further improves alignment, and experiments show DetailAnywhere surpasses existing open‑source methods in both quantitative metrics and human evaluation.
SemGeo-Gen is an unsupervised approach that generates approximate cross‑instance semantic‑geometric correspondences from weakly aligned 3D object collections. It uses coarse category‑level rotation alignment, lifts multi‑view DINOv2 features to 3D, applies continuous piecewise‑affine registration, and prunes matches with semantic consistency to produce 3D correspondences that can be projected into 2D views for scalable supervision. The method achieves high accuracy against human annotations and improves state‑of‑the‑art semantic correspondence models after synthetic pretraining.
Yalda-Embedding is a 595.8‑million‑parameter Persian retrieval model built by adapting Qwen3‑Embedding‑0.6B and trained on over one million query‑document pairs. It achieves the highest macro‑average score (0.5488) among 0.6B models on eight FaMTEB Persian retrieval tasks, outperforming Qwen3‑Embedding‑0.6B and F2LLM‑v2‑0.6B on every task and approaching the performance of the larger Qwen3‑Embedding‑4B. The authors also conduct an ablation study on contrastive batch size, optimization stabilization, and relational knowledge distillation, showing that the distilled configuration improves all tasks without increasing inference‑time model size and includes hard‑negative evaluation gains despite no explicit hard‑negative mining.
The paper introduces MedAction, a tree‑structured distillation pipeline that generates high‑quality multi‑turn diagnostic trajectories by interacting an LLM with an environment. It identifies three failure modes in current LLMs—ungrounded test ordering, unreliable diagnostic updates, and degraded multi‑turn coherence—and proposes two knowledge‑graph‑grounded metrics, Disease Trajectory Consistency and Reasoning‑Action Consistency, to filter trajectory quality. Using this pipeline, the authors create MedAction‑32K, a dataset of 32,681 trajectories from 2,896 PMC cases, and fine‑tune an 8B model that achieves state‑of‑the‑art performance on MedR‑Bench and the MedAction‑300‑Hard benchmark among open‑source medical LLMs.
The paper introduces a learned speed prior that predicts a distribution over execution speed for robot demonstrations, allowing consistent retiming across tasks. By applying this prior to both robot teleoperation and handheld demonstrations, the authors retime demonstrations to a handheld pace before fine‑tuning a vision‑language‑action policy. Experiments on 120 RoboLab tasks and a real robot show that policies trained with the learned clocks achieve higher success rates and smoother motion, outperforming policies fine‑tuned on original or uniformly accelerated timings.
Tempo‑Zero is a self‑improvement framework for video reasoning that uses recursive self‑play in executable environments. It closes a loop of environment execution, self‑play verification, and environment evolution, allowing a single policy to act as both questioner and solver and to receive execution‑grounded rewards. The approach trains only on simple programs, without external video data or supervised fine‑tuning, and achieves superior performance on twelve real‑world video benchmarks, especially on temporal grounding tasks.
The paper presents a physics‑aware, multi‑domain self‑supervised learning framework for Raman spectroscopy that uses a masked autoencoder with a sparse Voigt‑profile decoder and a denoising objective. It learns transferable spectral representations from diverse biological, chemical, forensic, and materials datasets, and demonstrates improved performance on substance classification, component identification, concentration estimation, and spectral denoising. The study shows that self‑supervised pretraining is especially beneficial for tasks with limited labeled data, that attention‑based probing reveals richer information than linear probing, and that larger encoders benefit more from pretraining.
The paper introduces Self-RefineVQA, a training‑free, inference‑time framework that lets a single Vision‑Language Model iteratively generate, critique, and refine its own answers for table visual question answering. By forming a closed self‑feedback loop guided by prompts, the method corrects common reasoning errors such as spatial misalignment and logical hallucinations without requiring additional labeled data or model updates. Experiments on tabular VQA benchmarks show consistent accuracy gains across monolingual and cross‑lingual settings, especially for lightweight models, while keeping inference cost linear and eliminating the need for expensive supervised fine‑tuning.