LoRA, PEFT, instruction tuning and domain adaptation — adapting a pretrained model without paying to train one.
arXiv:2605.12112v2 Announce Type: replace Abstract: RLHF is widely used to align flow-matching text-to-image models with human preferences, but often leads to severe diversity collapse after fine-tun...
arXiv:2609.34378v2 Announce Type: replace Abstract: Humans naturally possess the ability to work persistently toward long-term goals. Given a challenging task, humans can continuously work for months...
arXiv:2611.02897v2 Announce Type: replace Abstract: Traditional image cropping assumes the photograph already contains the composition worth keeping, and asks only where to crop. A full 360$^{\circ}$...
The study empirically investigates how length‑normalized Direct Preference Optimization (DPO) with a Structured Fine‑Tuning (SFT) retention term (RPO objective) affects downstream capabilities of a Qwen3.5‑4B model fine‑tuned with sparse LoRA on structured financial‑filing tasks. Results show that the alignment tax is graded and task‑selective: extraction performance is largely preserved, JSON extraction degrades modestly, and router‑based tasks suffer severe loss, with most degradation attributed to suppression of rejected completions. The analysis uses 8 seeds, artifact‑aware scoring, and TOST equivalence testing to quantify these effects across different reward margins and precision settings.
The paper investigates how repeatedly training score-based diffusion models on synthetic data can cause their outputs to drift from the target distribution. It provides theoretical upper and lower bounds on the divergence accumulated over successive training rounds, distinguishing regimes based on score estimation error and the fraction of fresh data used. Empirical experiments on synthetic data and images support the theoretical findings.
The paper reports on adapting the Qwen3 language model to Tajik using small LoRA adapters across sizes from 0.6B to 8B. While the adapters consistently lowered Tajik text loss by 0.60 to 1.45 nats per token, they did not improve Tajik reading comprehension on the Belebele benchmark; in some cases accuracy even dropped. A more extensive 8B training pipeline, including 375 million tokens of raw Tajik text and supervised fine‑tuning, produced a new model (e3) that outperformed Qwen3‑8B on Belebele by 11.4 points, though it slightly degraded English MMLU performance.
UniCompress is a deployment-oriented framework that transforms multi-task vision‑language expertise into a single compact, adapter‑free model suitable for edge intelligence. Starting from a BF16 Qwen3‑VL‑4B teacher, it builds an asymmetric INT2 student using GPTQ initialization and teacher‑guided quantization recovery, jointly compressing both the vision encoder and language backbone. Evaluated on four embodied tasks, UniCompress cuts weight storage from 8.9 GB to 2.1 GB (≈76% reduction) while improving the INT2 macro task score by 34.3 percentage points over GPTQ‑INT2 and staying within 0.9 points of the BF16 teacher.
The paper introduces a new framework for multi-resident data association in smart homes, framing the problem as a Markov Decision Process and solving it with Behavior Cloning. Two policy architectures are proposed: a lightweight Transformer treating association as multi-label classification over event histories, and a fine-tuned Large Language Model that generates conditional text to leverage semantic knowledge. Experiments on three datasets show the approach outperforms existing baselines, especially for concurrent multi-resident events, and the LLM policy hints at cross-environment transfer potential.
AmbiHTMLBench is a new benchmark that evaluates how well models can convert real-world documents into executable HTML while preserving content, organization, and visual appearance. It contains 1,115 pages with structured annotations and uses both objective metrics (table fidelity, text error, layout consistency) and a reference-anchored subjective protocol (LLM judges aligned with human preferences). Results show that different systems excel in objective versus subjective measures, and fine‑tuning a small expert model improves performance, highlighting the benchmark’s role as a diagnostic tool for advanced document parsing.
The paper introduces a method called Addition by Subtraction to strengthen safety alignment in large language models. By gradually removing identified safety‑critical neurons during finetuning and re‑optimizing preferences, the authors reduce the model’s reliance on sparse components that can be targeted by pruning or ablation attacks. Experiments on four models from three vendors show that this approach improves robustness to both white‑box and black‑box jailbreaks, lowers attack success rates by up to 90%, and enhances truthfulness while preserving general capabilities.
The paper introduces Context as Covariates (CoCo), a forecasting framework that combines Large Language Models (LLMs) and Time Series Foundation Models (TSFMs). It generates realistic multimodal datasets by annotating real-world time series with LLMs and aligns the two models through staged fine‑tuning: first using Group Relative Policy Optimization to improve the LLM’s covariate generation, then fine‑tuning the TSFM on those covariates. Experiments on synthetic and real benchmarks across finance, economics, weather, traffic, and security show that CoCo can outperform state‑of‑the‑art TSFMs and LLM forecasters by up to 12% MAE.
The study evaluates EfficientNet variants B0–B7 for classifying retinal diseases from spectral-domain optical coherence tomography (SD‑OCT) images at a fixed 224×224 resolution. Eight ImageNet‑pretrained models were tested on the Kermany dataset, with training involving a frozen‑backbone head followed by partial fine‑tuning. Results show non‑monotonic accuracy, with B1 achieving the highest overall accuracy (96.95%) and best Drusen F1, while B7 offered a secondary accuracy gain at the cost of larger model size and CPU latency.
The paper introduces ALTO, a corpus of 3,022 culturally and linguistically relevant instances of tech‑facilitated gender‑based violence (TFGBV) in Levantine Arabic, African French, and Swahili. It was created through a community‑led crowdsourcing partnership with civil society organizations in Jordan, Francophone West Africa, and Kenya, and annotated using active learning. Benchmarking shows that existing harmful content detection models have low TFGBV recall on ALTO and improve after fine‑tuning, with covert harms being frequently missed and Swahili exhibiting the highest error rates.
The paper introduces a toolchain that enables machine‑learning model development on AWS EC2 F2 cloud FPGAs using high‑level Vitis HLS. It implements a complete Vitis HLS to HDK bridge to AFI workflow, a four‑level methodology (FP32 PyTorch, bit‑accurate quantized Python, HLS C‑simulation, physical FPGA) for error localization, and documents deployment, timing closure, and HLS engineering decisions. The authors validate the approach by implementing TinyBERT’s embedding layer and first encoder layer on FPGA, noting that weight streaming from host DRAM results in poor performance compared to CPU/GPU implementations.
UpSafe°C is a unified framework that enhances large language model safety by upcycling safety‑critical layers into a sparse Mixture‑of‑Experts structure, with a router acting as a soft guardrail. It employs a two‑stage supervised fine‑tuning strategy to improve safety discrimination while preserving general capabilities, and introduces a safety temperature mechanism for dynamic inference‑time control of the safety‑utility trade‑off. Experiments across multiple benchmarks and model scales show robust safety improvements against harmful and jailbreak inputs while maintaining competitive general performance.
The paper surveys the use of reinforcement learning (RL) in graphical user interface (GUI) agents, outlining how RL addresses challenges such as long-horizon credit assignment, distribution shifts, and safe exploration that supervised fine-tuning alone cannot solve. It introduces a taxonomy dividing methods into Offline RL, Online RL, and Hybrid Strategies, and discusses key aspects like reward engineering, data efficiency, and technical innovations. Emerging trends highlighted include the move toward composite, multi-tier reward architectures, world-model-based training to mitigate GUI I/O latency, and the spontaneous appearance of System‑2‑style deliberation, all of which inform a roadmap for future robust GUI automation.
The paper identifies a mismatch in empathetic reinforcement learning: current methods use fixed reward specifications while support priorities change with dialogue state. It introduces Context‑Adaptive Rubric Evolution (CARE), which dynamically adjusts weights for cognitive, affective, and proactive empathy at each turn, trained via supervised fine‑tuning and preference‑based reinforcement learning. CARE, integrated with RLVER and MICA, achieves state‑of‑the‑art results on SentientBench, EQBench3, and EMPA, notably boosting EPM‑Idx by up to 55 points under the Gemini‑2.5‑Pro judge.
VCoT‑SQL is a framework that improves text‑to‑SQL generation by training two complementary models on curated reasoning traces. It uses clause‑level chain‑of‑thought templates verified through execution and schema alignment, and a validator that detects syntactic, semantic, and type errors. The approach achieves high execution accuracy on BIRD and Spider datasets, outperforming larger closed‑source models while relying only on 7B open‑source models.
This paper presents an overview of \emph{SciClaimEval}, a pilot task at NTCIR-19 on verifying scientific claims against the tables and figures of the papers in which they appear. Our dataset pairs aut...
Automatic modulation classification (AMC) is a key component of cognitive wireless systems, enabling spectrum awareness and adaptive communication. Most deep learning approaches to AMC assume a fixed...