LoRA, PEFT, instruction tuning and domain adaptation — adapting a pretrained model without paying to train one.
arXiv:2609.12898v1 Announce Type: new Abstract: Fine-grained robotic manipulation depends on understanding parts, not only whole objects. Existing 3D foundation models tend to be either generalized b...
Doc2FRC introduces Fixed-Range Chunking (FRC), a dynamic programming method that partitions documents into chunks of a predefined length interval, ensuring consistent length distributions during training and inference. This approach reduces train-test length mismatch, mitigates n-gram repetition, and improves translation quality for 7B LLMs compared to direct Doc2Doc fine-tuning. Experiments on IWSLT2017 and a new 10-language test set, GlobVDoc, demonstrate that FRC outperforms existing document-level machine translation methods and enhances out-of-distribution translation performance.
MInTRL (Minimal Intervention Reinforcement Learning) expands exploration in on-policy reinforcement learning by inserting sparse, local corrections into rollouts via a judge-intervention policy. These interventions replace erroneous suffixes and immediately return control to the main policy, allowing the agent to explore beyond its natural trajectory while maintaining on-policy data. The method uses a sequence-level advantage-regression objective, avoiding importance sampling, and demonstrates superior performance on math and code benchmarks compared to standard on-policy and off-policy baselines.
LoRA-RC introduces a low‑rank adaptation scheme for reservoir computing that updates the recurrent matrix using streaming prediction errors while keeping the base reservoir and adaptation bases fixed offline. The method projects a small core matrix onto a spectral‑norm ball and applies low‑pass filtering at each step, ensuring every recurrent matrix stays within a certified contraction set. Experiments on a Lorenz system with abrupt parameter drift show that LoRA‑RC reduces post‑drift prediction error by 56% compared to a fixed RC and 51% compared to readout‑only adaptation, and that removing the projection increases error by more than a factor of 40.
AMDKernelVault is an open HIP and Triton kernel corpus and training framework designed for AMD CDNA GPUs. It includes 62,153 verified HIP kernels, 39,893 Triton kernels, and 2,377 ROCm library QA entries, and introduces agent-driven pipelines (HIPKernelGen and TritonKernelGen) that convert PyTorch references into GPU kernels, compile, validate, and profile them on AMD hardware. The corpus was used to fine‑tune Qwen3-8B, achieving the highest correctness on several benchmarks such as PyTorch-to-HIP, TritonBench‑G, and ROCmBench under fixed evaluation budgets.
R2VC is a modular fact‑checking system that separates retrieval, reasoning, verification, and confidence calibration. It uses hybrid sparse‑plus‑dense Wikipedia retrieval, a fine‑tuned generator for structured verdicts, an NLI cross‑encoder for selecting evidence‑based candidates, and a lightweight calibrator for confidence and abstention. On the FEVER benchmark, R2VC improves accuracy by 13.74% over a baseline and shows that verifier‑based candidate selection and calibration are key contributors to performance.
The paper introduces ISO-LoRA, an optimizer that improves rank utilization in Low‑Rank Adaptation (LoRA) by coupling factor updates through spectral descent on the induced tangent perturbation in weight space. Experiments on GPT‑2 adaptation show that standard optimizers like AdamW concentrate updates in a few singular directions, whereas ISO-LoRA distributes energy more evenly, leading to higher effective rank and better downstream performance across 0.1B‑7B models. The authors provide theoretical guarantees under a stylized spiked‑gradient model and demonstrate that ISO-LoRA consistently outperforms factor‑wise optimizers, especially at moderate‑to‑large LoRA ranks.
arXiv:2604.19254v2 Announce Type: replace Abstract: Popular low-rank parameter-efficient fine-tuning (PEFT) methods represent adaptation as separate updates to selected backbone weights, without main...
CluSTER is a cluster‑aware balanced sampling framework designed to improve the efficiency of instruction‑tuning for large language models. It reduces redundant computation by clustering data in gradient space and allocating samples across GPUs in a data‑parallel setting, while preserving the original distribution through weighted updates. Experiments show that CluSTER can cut training time by up to 69.6% with negligible loss in accuracy compared to existing sampling methods.
The paper presents a zero‑shot model predictive control (MPC) approach for buildings that uses excitation‑based generalized transfer learning models. By pretraining on purposefully probed operational data from multiple source buildings, the authors demonstrate that these models achieve superior control performance in 32 simulated target buildings, outperforming both an online linear MPC and a PI controller. This method eliminates the need for target‑specific data, reducing setup cost and facilitating broader deployment of energy‑efficient MPC in the building sector.
The paper introduces an Early Intervention (EI) framework for multimodal medical image classification that addresses two key challenges: limited exploitation of complementary multimodal information and scarcity of labeled data for Vision Foundation Models (VFMs). EI treats one modality as the target and uses high‑level semantic tokens from other modalities as intervention tokens to guide the target’s embedding early in the process. The authors also propose Mixture of varied‑rank LoRAs (MoR) for efficient VFM adaptation, and demonstrate the method’s effectiveness on retinal, skin, and knee medical image datasets.
arXiv:2605.20066v2 Announce Type: replace Abstract: Knowledge graph question answering seeks to translate natural language questions into executable queries over knowledge graphs, but existing approa...
arXiv:2606.12234v2 Announce Type: replace Abstract: Controlling the output of Large Language Models (LLMs) is a central challenge for their reliable deployment, yet a clear understanding of the invol...
UniLayDiff is a Unified Diffusion Transformer that tackles content‑aware layout generation across a wide range of tasks—unconditional, element‑type, size, and relationship‑conditioned generation—using a single end‑to‑end trainable model. It treats layout constraints as a distinct modality within a Multi‑Modal Diffusion Transformer framework, capturing interactions among background images, layout elements, and constraints. The model is further refined with LoRA fine‑tuning to incorporate relation constraints, achieving state‑of‑the‑art performance and, according to the authors, the first unified solution for all content‑aware layout generation tasks.
The paper proposes a reinforcement learning (RL) fine‑tuning framework for electronic health record (EHR) foundation models, treating them as generative policies over patient trajectories. By framing clinical prediction tasks such as hospital readmission as event‑conditioned, time‑windowed reasoning problems and designing time‑aware, rollout‑sensitive rewards, the authors show that RL fine‑tuning consistently outperforms pre‑trained backbones and strong baselines. The approach enables smaller models to surpass larger pre‑trained models in data‑limited settings, induces positive transfer across tasks, and produces trajectories with stronger structural and semantic alignment to ground truth, improving downstream utility.
The paper introduces a weakly supervised framework for extracting dataset mentions from forced displacement and Fragile, Conflict, and Violence (FCV) documents. It uses a lightweight model trained on general research literature to generate candidate mentions, which are then refined by a large language model that validates or rejects them and corrects boundaries. The refined annotations are augmented with synthetic and contrastive examples to fine‑tune the model, achieving 74.1% precision and 70.5% recall on a benchmark of 1,706 passages, with higher precision (89.5%) on passages that contain dataset references.
CoHyDE is an iterative co‑training framework that jointly trains a dense encoder and an LLM rewriter for tool retrieval from large API catalogs. The encoder is fine‑tuned with InfoNCE on catalog‑style hypothetical descriptions generated by the rewriter, while the rewriter is preference‑aligned via DPO against the encoder’s retrieval scores. On a 10k‑tool subset of ToolBench, three rounds of CoHyDE outperform the best single‑component baseline by 2.5 pp NDCG@5 on standard queries and 6.3 pp on vague queries, with the largest gains on the hardest vague tier.
The paper discusses how large language models (LLMs) can be fine‑tuned with observational data to improve alignment with human preferences and business goals. It highlights that directly using such data can cause models to learn spurious correlations, and introduces DeconfoundLM, a method that removes known confounders from reward signals. Experiments show that DeconfoundLM better recovers causal relationships and outperforms baseline methods by over 16% in objective score when confounding is present.
arXiv:2609.12787v1 Announce Type: new Abstract: Polarization image fusion combines the stable luminance and structural information of the total- intensity image S0 with the material-sensitive details...
The paper introduces a three‑stage training pipeline that builds compact, efficient dense retrievers without requiring ground‑truth relevance labels. Using cross‑lingual alignment, relational knowledge distillation, and contrastive fine‑tuning, the authors develop PolDense (six Polish models ranging from 17 M to 1 B parameters) and EuroDense (a 435 M‑parameter model covering nine European languages). Extensive evaluation on 41 Polish and 150 multilingual tasks shows that PolDense‑1B outperforms larger retrievers up to 9 B parameters, while EuroDense leads in task‑averaged and language‑averaged performance among models below 1 B parameters.