The paper demonstrates that vision‑language models (VLMs) possess a small set of attention heads, called Visual Retrieval Heads (VRHs), that are causally responsible for linking text prompts to specific image regions. By adapting head‑scoring techniques from language models, the authors identify VRHs as the heads whose attention from output prediction tokens, summed over the ground‑truth referent region, most reliably indicates causal grounding. Experiments across eleven VLMs and five referring‑expression benchmarks show that masking the top 20 VRHs can drop grounding accuracy by up to 80 percentage points, while random masking has little effect, and that VRHs generalize across diverse visual tasks and transfer across models sharing an LLM backbone.
The paper shows that large language model agents are far more likely to commit to a directional answer when presented with a professional-looking market panel, even if the panel’s data is fabricated. Across 12 frontier models, commitment rates jump from 6.5 % for a bare question to 54.0 % with evidence, and remain high (≈37 %) even when all numbers are invented. The study finds that the act/don’t‑act decision gate is the key failure point, can be trained to reduce false commitments, but is fragile to response format changes.
The paper introduces Instruct-to-Act, a system that decouples high‑level planning from low‑latency control by combining a vision‑language model (VLM) planner with a world‑model controller. The VLM generates sparse, high‑level text instructions, while the controller executes them autonomously at high frequency. Experiments across seven embodied environments, including multi‑agent settings, show that this approach outperforms both controller‑only and direct VLM action‑generation methods, maintains fast control, and allows swapping in different pretrained VLM planners without fine‑tuning.
The paper introduces sLoTh, a parameter‑efficient continual learning framework for sparse event‑based vision transformers. By freezing the backbone and limiting plasticity to low‑rank attention updates (seLoRA) and shared neuronal threshold modulation, sLoTh adapts to new tasks while updating less than 1% of the parameters and avoiding replay buffers. Experiments on CIFAR‑100, Tiny‑ImageNet, ImageNet‑100, and ImageNet‑R show competitive rehearsal‑free performance across up to 100 tasks and achieve roughly 6.5× lower energy consumption than dense vision transformers.
The paper explores contrastive self‑supervised learning (SSL) for retinal fundus image classification, comparing SimSiam and SimCLR under limited data and computational resources. It investigates how retinal‑specific augmentation strategies and training parameters affect representation quality, evaluated through linear probing and fine‑tuning on multi‑disease classification and diabetic retinopathy grading tasks. The results demonstrate that tailored augmentations enable lightweight SSL models to learn transferable representations, reducing reliance on large annotated datasets while achieving competitive performance.
The study compares four approaches for Chinese sentence-level metaphor identification: BERT fine‑tuning, QLoRA-based large language model fine‑tuning, zero‑shot LLM prompting, and zero‑shot prompting with an expert‑informed procedural Skill. Results show that fine‑tuning yields the highest accuracy on the native test set, while the Skill‑based zero‑shot method provides the most stable performance across three datasets, achieving the highest external floor and the smallest performance range. Adding the Skill reduces false positives on one dataset but increases false negatives on others, indicating a trade‑off between precision and recall.
By Yufeng Wu, Meichun Liu
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
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
The paper investigates confidence estimation and selective prediction for financial named entity recognition (NER) under domain shift, using a stress test across SEC filings, financial news, and social media. It evaluates BERT and LoRA‑tuned Qwen2.5 models with five inference‑time confidence signals, finding that whole‑output probability is a strong in‑domain error detector but weak out‑of‑domain, while entity‑span probability and self‑consistency remain robust. Abstention can dramatically reduce sentence error on high‑confidence in‑domain data, but offers limited benefit under extreme social‑media shift, suggesting a staged deployment that first detects severe distribution shift before applying confidence gating.
By Zihao Zheng, Baichuan Li, Junyi Yao, Jiayu Long
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
DCGC is a Masked Diffusion Model framework that performs global correction of flawed reasoning traces in Large Language Models. It uses an imperfect solution draft from an upstream solver as auxiliary context and combines task‑specific supervised fine‑tuning with a Dynamic Dual‑CFG inference mechanism that separates problem‑only and joint problem‑draft branches. Experiments on math, code, and knowledge reasoning benchmarks show that DCGC outperforms standard sampling and simpler CFG variants, and it can improve full test‑set accuracy even when ground‑truth failure labels are unavailable.
By Minhae Oh, Nakyung Lee, Jungwoo Lee
The paper investigates how different estimators of the reverse Kullback–Leibler (KL) divergence used as a regularization term in reinforcement learning (RL) training of large language models (LLMs) affect training stability and downstream performance. By analyzing gradient bias across various estimator configurations, the authors demonstrate that biased gradients can cause training instabilities, while unbiased configurations improve performance on both in‑domain and out‑of‑domain tasks. Experiments on Qwen2.5‑7B, Llama‑3.1‑8B‑Instruct, and Qwen3‑4B‑Instruct‑2507 confirm these findings and show that KL regularization also stabilizes off‑policy RL training in asynchronous setups.
By Vedant Shah, Johan Obando-Ceron, Vineet Jain, Brian Bartoldson, Bhavya Kailkhura, Sarthak Mittal, Glen Berseth, Pablo Samuel Castro, Yoshua Bengio, Esmeralda S. Whitammer, Moksh Jain, Siddarth Venkatraman, Aaron Courville
arXiv:2607.18227v2 Announce Type: replace
Abstract: In line with the prevailing direction of vision research, we explore the integration of both generation and editing capabilities for video and imag...
By Dingyun Zhang, Lixue Gong, Wei Liu
The paper introduces SalesLLM, a bilingual (Chinese/English) benchmark for evaluating large language models (LLMs) in realistic sales dialogues. It comprises 30,074 scripted configurations and 1,805 curated multi‑turn scenarios from Financial Services and Consumer Goods, with controllable difficulty and personas. An automatic evaluation pipeline uses an LLM judge for sales‑process progress and fine‑tuned BERT classifiers for end‑of‑dialogue buying intent, while a user model, CustomerLM, is trained to improve simulation fidelity. SalesLLM scores correlate strongly with human ratings (Pearson r = 0.86) and reveal that top Chinese LLMs match junior‑to‑intermediate human salespeople but not experts, with cross‑lingual consistency remaining poor.
By Xuanbo Su, Wenhao Hu, Le Zhan, Yuting Xie, Kailin Lyu, Kaijie Chen, Ziwei Li, Yeqiang Wang, Haibo Su, Yunzhang Chen, Ling Huang
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
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 introduces a label‑free method called AURCC for selecting the best foundational model for medical image classification when the target domain lacks labels. AURCC uses a pseudo‑label discrepancy computed by the SUDO framework to score models without fine‑tuning. Experiments on chest X‑ray data across three inter‑hospital shifts show that AURCC closely matches the true model ranking, outperforming simple source‑accuracy baselines especially when source data are limited.
By Juan I\~naki Larrea, Lucas Mansilla, Enzo Ferrante
The paper introduces V‑Rubrics, a reinforcement‑learning framework that evaluates vision‑language model responses by breaking them into atomic propositions and scoring them on Visual Faithfulness, Reasoning Consistency, and Instruction Following. Using a fine‑tuned Qwen3‑VL‑8B‑Instruct model and a newly created 50K‑example V‑Rubrics dataset, the authors demonstrate that rubric‑based GRPO outperforms both a shared SFT baseline and an answer‑only GRPO, especially on knowledge‑oriented and visually grounded reasoning tasks.
By Shulin Tian, Minglun Li, Yuhao Dong, Hao Ding, Jiarui Yao, Haiwen Diao, Jingkang Yang, Hongyuan Zhu, Ziwei Liu
The paper introduces CoDIT, a contrastive decoding technique that separates instruction-following behavior from pre-trained world knowledge in large language models. By generating responses that emphasize post-training instruction capabilities while suppressing shared pre-trained knowledge, CoDIT creates instruction-tuning datasets that lead to consistently better model performance than directly generated responses or existing public datasets. The authors also provide theoretical and empirical evidence that CoDIT effectively distills instruction-tuning knowledge from model parameters into text, facilitating cross-architecture transfer.
By Tatsuya Ichinose, Youmi Ma, Masanari Oi, Ryuto Koike, Naoaki Okazaki
LibriBrain100 is a new large‑scale MEG dataset for speech decoding that contains over 100 hours of high‑quality recordings while subjects listened to naturalistic continuous speech. The dataset more than doubles the size of the original LibriBrain release, with a record 80 hours from a single subject and additional 40‑minute recordings from 32 subjects. The authors demonstrate the value of deep within‑subject data and broad multi‑subject data by achieving state‑of‑the‑art word‑classification performance and showing that supervised fine‑tuning can compensate for limited per‑subject data, all supported by open‑source tools and a public competition leaderboard.
By Francesco Mantegna, Dulhan Jayalath, Gereon Elvers, Tasha Kim, Benjamin Ballyk, Alex Fung, SungJun Cho, Teyun Kwon, Luisa Kurth, Miran \"Ozdogan, Gilad Landau, Pratik Somaiya, Natalie Voets, Mark Woolrich, Oiwi Parker Jones