The paper investigates whether fine‑tuning a language model erases previously embedded activation steering interventions that suppress refusals and encourage brevity. Across five instruction‑tuned models (3B–14B) subjected to non‑adversarial supervised fine‑tuning (SFT) and reinforcement learning from human feedback (RLHF), the authors find that the steering’s behavioural effect degrades when the fine‑tuning objective conflicts with the targeted behaviour, yet the underlying weight edits remain largely unchanged. Mechanistically, the steering vectors survive with minimal alteration, but functionally the steering is vulnerable and must be re‑validated after downstream training.
By Philipp E. Glass, Allan Tucker, Yongmin Li, Alina Miron
Loss-Based Active Learning for Neural Abstractive Summarization proposes LOBSTER, an active learning framework that selects unlabeled documents similar to the model’s high‑loss training examples to correct specific weaknesses. The method is tailored for abstractive summarization, addressing instability and computational bottlenecks seen in prior work. Experiments on three benchmark datasets and two backbone models show that LOBSTER matches or surpasses state‑of‑the‑art performance while speeding up query selection by up to 665×.
By Michail Ioannou, Tatiana Passali, George Michalopoulos, Grigorios Tsoumakas
ReFIne is a training framework that augments large reasoning models with three trustworthiness properties: interpretability, faithfulness, and reliability. It combines supervised fine‑tuning with GRPO to produce structured, tag‑based reasoning traces, explicitly disclose decisive information, and provide self‑assessments of soundness and confidence. Applied to Qwen3 models, ReFIne improves interpretability by 44.0 %, faithfulness by 18.8 %, and reliability by 42.4 % on mathematical benchmarks.
By Chung-En Sun, Ge Yan, Akshay Kulkarni, Tsui-Wei Weng
The paper evaluates the robustness of automated fact‑checking systems by cross‑benchmarking nine models—including random baselines, fine‑tuned transformers, zero‑shot LLMs, and top AVeriTeC 2025 systems—across four datasets from scientific, open‑web, and climate domains. It finds that fine‑tuned models outperform zero‑shot LLMs on ClimateCheck, that system rankings vary strongly with domain and metric, and that replacing retrieved evidence with gold annotations boosts veracity accuracy by 14–22 points, underscoring retrieval as the main bottleneck. The authors provide code, pre‑processed datasets, and results to enable reproducible research.
By Aida Usmanova, Zangir Iklassov, Markus Leippold, Ricardo Usbeck
AdaVDR is an adaptive video deep research agent that selects and reflects on tool usage based on the task and the model’s capabilities. It constructs a specialized data pipeline to generate high‑quality QA pairs and uses model‑conditioned filtering to remove unnecessary tool calls. The agent is trained with supervised fine‑tuning and reinforcement learning, achieving top performance on the VDR‑EE benchmark and significant gains on VideoDR.
By Xintong Zhang, Xiaomeng Fan, Shilin Yan, Ekko He, Zicheng Liu, Zijian Zou, Guannan Zhang, Yuwei Wu, Zhi Gao, Hongwei Xue
MoganBert-TR is a 149‑million‑parameter Turkish encoder foundation model trained from scratch on a language‑specific corpus using a two‑stage CLM‑to‑MLM curriculum. The model, along with its embedding variant MoganBert‑Embed, achieves state‑of‑the‑art results on Turkish benchmarks such as TrGLUE and TabiBench, outperforming existing Turkish BERT models. Its tokenizer, comprising 50,048 tokens, also surpasses other Turkish tokenizers in compression and fertility metrics.
By Furkan Yilmaz, Habibe Aleyna Tasdemir, Muhammed Faruk Gozay
DEFUSE is a backdoor detection framework for self‑supervised encoders that uses a conditional diffusion generative model to estimate representation‑conditioned image likelihoods. By fine‑tuning a pretrained diffusion model, DEFUSE performs semantic reconstruction in a reference encoder’s representation space, enabling it to detect backdoors without needing uninfected data or precomputed pseudo‑labels. Experiments show that DEFUSE outperforms existing detectors on both visual SSL and vision‑language encoders, reducing reliance on prior knowledge of the victim model or attack strategy.
By Tuo Chen, Jie Gui, Minjing Dong, Lanting Fang, Ju Jia, Benlei Cui, Jian Liu
The paper investigates evidence generation for biomedical claim verification, evaluating various large language models and retrieval strategies on the CARE-XAI benchmark. It finds that fine‑tuned LLMs excel at producing evidence, while biomedical classifiers still lead in verdict‑only prediction. PubMed retrieval helps on PubMed‑aligned datasets but can mislead on broader public‑health claims, prompting the authors to propose Bio‑GRACE, a diagnostic that normalizes gold references to assess retrieval utility.
By Pritam Deka, Prabhjot Singh
The paper investigates the interpretability of Domain-Adapted Prompt-based Fine-tuning (DAPF) models for dementia detection from spoken language. Using probing and analysis techniques, the authors find that DAPF achieves strong overall performance (accuracy = 0.83, macro‑F1 = 0.83) and that the diagnosis is most recoverable from its [MASK] representation. However, token‑level explanations derived from DAPF are largely driven by language‑task vocabulary, discourse markers, and transcription artifacts, and perturbation tests reveal weak or negative effects, indicating that the masked‑token interface captures diagnosis information without providing faithful token‑level explanations.
By Pardis Ranjbar-Noiey, Natalie Parde
The paper investigates why vision‑language models like LLaVA‑1.5‑7B hallucinate objects in captions and proposes a targeted fix. By ranking attention heads whose image attention drops around hallucinated words, the authors identify 32 key heads and apply a head‑sliced LoRA adapter plus an inference‑time grounding controller. On COCO images, this combined method reduces hallucinated captions from 37% to 23% and hallucinated object mentions from 15.6% to 9.6%, while also lowering object recall.
By Armaan Sandhu, Abhilasha Senapati, Hima Kammachi
VGA‑BenchV2 is an expanded, human‑aligned benchmark and optimization framework that jointly evaluates video generation quality and aesthetic value. It builds on the original VGA‑Bench taxonomy, adding 52 sub‑dimensions and 1,016 curated prompts to generate over 60,000 videos from 12 mainstream models. The benchmark significantly enlarges human supervision with 36,000 task‑level annotations and introduces a hybrid evaluator (VAQA‑Net, VTag‑Net, VGQA‑Net) that aligns well with human judgments and can be used as a reward model for reinforcement‑learning fine‑tuning.
By Longteng Jiang, DanDan Zheng, Qianqian Qiao, Heng Huang, Huaye Wang, Yihang Bo, Bao Peng, Jingdong Chen, Jun Zhou, Xin Jin
CrossMambaTuning is a new framework that adapts pretrained learned image compression models to machine vision tasks with minimal retraining. It combines State Space Models with cross‑layer interaction, featuring a Mamba adapter that uses task‑specific prompts and multi‑scale branching, and a Scale‑Invariant Cross‑Layer Adapter (SICA) that shares parameters across scales to reduce redundancy. Experiments show that this approach achieves state‑of‑the‑art performance while cutting parameter overhead by 72% compared to existing methods.
By Haobo Xiong, Shaobo Liu, Kai Liu, Chongyang Ding
The paper investigates why multimodal large language models (MLLMs) struggle with vision‑centric tasks when visual evidence conflicts with pretrained language knowledge. Using image reconstruction and a new WhatIfVis benchmark, the authors show that MLLMs preserve coarse‑grained visual attributes but fail to consistently use them, and that supervised fine‑tuning and activation patching can improve controllability of visual context sensitivity. The study demonstrates that the main bottleneck lies in the models’ inability to reliably regulate their reliance on visual evidence rather than in visual perception itself.
By Jiaang Li, Chengzu Li, Zhaochong An, Yifei Yuan, Xi Liu, Serge Belongie, V\'esteinn Sn{\ae}bjarnarson
The paper introduces FairTPT, a fairness-aware test‑time prompt tuning method for vision‑language models like CLIP. It jointly minimizes target marginal entropy while maximizing spurious marginal entropy to reduce bias under subpopulation shifts. Experiments show that standard episodic test‑time adaptation can worsen disparities, but FairTPT outperforms existing debiasing methods while preserving overall performance.
By Yoann Launay, Parameswaran Kamalaruban, Tom Kempton, Stuart Burrell, David Sutton
PreResQ‑R1 introduces a Preference‑Response Disentangled Reinforcement Learning framework for Visual Quality Assessment that jointly optimizes absolute score regression and relative ranking consistency. It employs a dual‑branch reward system—modeling intra‑sample response coherence and inter‑sample preference alignment—trained with Group Relative Policy Optimization. The method extends to video quality assessment via a global‑temporal and local‑spatial data flow strategy, achieving state‑of‑the‑art results on 10 IQA and 5 VQA benchmarks with only 6K images and 28K videos, and provides human‑aligned reasoning traces.
By Zehui Feng, Weichuan Wang, Xiaohan Chen, Ting Han
The paper presents a training‑free, agentic computer‑vision system that converts 2D structural framing plan PDFs into editable finite‑element model drafts. It uses a deterministic stage to extract geometric primitives, estimate scale, and recognize five entity classes via a drafting grammar, followed by an agentic stage that applies typed corrections and fail‑closed transactions. Evaluation on a 100‑plan benchmark shows high precision and recall across columns, beams, walls, braces, and openings, with scale estimates within 0.1% of reference.
By Mohammad Talebi-Kalaleh, Qipei Mei
arXiv:2608. 26052v1 Announce Type: new Abstract: Choosing the rank of a low-rank adaptation (LoRA) update is usually an empirical task.
By Gerard Conangla Planes
SHIFT-LLM is a training‑free post‑pruning correction framework that inserts a Linear Residual Adapter (LRA) at each depth‑pruned site in large language models. Each LRA preserves the original residual identity while adding a lightweight affine correction calibrated via closed‑form least‑squares regression on a small held‑out set, thereby approximating the hidden state that would have been produced by the removed block. Experiments across multiple model families and benchmarks show that SHIFT‑LLM consistently recovers accuracy lost to depth pruning, achieving gains up to +15.7 points on Llama‑3.1‑8B‑Instruct with only a few hundred calibration samples and no gradient computation.
By Ali Bahri, Hang Li, Hongliang Li, Zhitang Chen
The study fine‑tunes the Uni‑Mol2 molecular foundation model on the GS‑LF benchmark for multi‑label odor descriptor prediction. The resulting model matches or surpasses state‑of‑the‑art baselines on the primary benchmark and successfully transfers to four downstream olfactory tasks—including cross‑dataset prediction, odorless vs. odorous classification, enantiomer evaluation, and odor mixture discriminability—without further deep‑learning training. The enantiomer analysis demonstrates that 3D molecular representations can distinguish mirror‑image molecules, a capability lacking in 2D graph models, though predicting stereochemistry’s perceptual effects remains unresolved.
By Yikun Han, Yi Wang, Neil Mankodi, Stephen Yang, Ambuj Tewari
The paper introduces CuCu, a multi‑agent LLM framework that converts national social studies curricula into open‑ended, culture‑specific question‑answer pairs for fine‑tuning language models. Using the Korean curriculum, the authors build KCaQA, a dataset of 34.1k QA pairs that cover culture‑specific topics and ground responses in local sociocultural contexts. Experiments show that fine‑tuning with KCaQA improves the model’s cultural alignment and relevance to Korean society.
By Haneul Yoo, Won Ik Cho, Geunhye Kim, Jiyoon Han