Searcharxiv⌕ Search

arXiv · 2611.18701

What Does Hard-Negative Fine-Tuning Teach Multimodal Large Language Models? Controlled Evidence from a Compositional Benchmark

Abstract

Fine-tuning on hard negatives is a popular way to improve the compositional reasoning of vision-language models, but when accuracy goes up it is rarely clear what the model actually learned. We ask this question for multimodal large language models. We use CompHall, 2,860 forced-choice questions (spatial, attribute, counting and part-whole) with two graded hard negatives each, three model families (LLaVA-1.5, LLaVA-NeXT and Idefics2), and controls that such studies often lack: gray-image evaluation, three seeds, option-order tests, control arms trained on different question types, and independent benchmarks. We find four things. First, CompHall is not a pure test of vision: part-whole questions can be answered from the text alone (66-95% without the image), and per-image accuracy is unrelated to an object-hallucination probe (r = -0.05), so we see no evidence that compositional failures drive hallucination. Second, the original fine-tuning method scored mostly image-placeholder tokens and hurt both LLaVA models. With a corrected loss, fine-tuning raises accuracy on the spatial-relation questions of CV-Bench in all three families, by +3.2, +23.1 and +6.0 points (averaged over both option orders). Control arms show that the gain needs spatial training questions, and for LLaVA-NeXT about two-thirds of it appears even on relation words that were only wrong options in training. Third, the gain is not general: on CompHall itself it is significant only for LLaVA-NeXT (+6.9 points) and not detectable for Idefics2, it reaches about 3 points at most on true/false statements, and LLaVA-1.5 fine-tuned on other question types learns to answer "true" too often. Fourth, both losses raise the yes-bias of LLaVA-1.5 on POPE.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Ranjith Chodavarapu. 2026-11-06. What Does Hard-Negative Fine-Tuning Teach Multimodal Large Language Models? Controlled Evidence from a Compositional Benchmark. https://searcharxiv.com/abs/2611.18701

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

RTime-QA: A Benchmark for Atomic Temporal Event Understanding in Multimodal Large Language Models

Multimodal large language models (MLLMs) perform well on video benchmarks, yet many video questions can be answered from a single frame or from language priors, which leaves open whether these models perceive how an event unfolds in time. We introduce \rtqa, a benchmark for atomic temporal event understanding built on time-reversed question pairs: each video is accompanied by its reversal, and each question is asked on both videos with identical text and options but different correct answers, so that a deterministic model that ignores the order of the frames cannot answer both. \rtqa{} comprises 8,420 questions of five types on 1,278 videos, 639 videos and their reversals. The questions are derived from event annotations that a proposer model writes, a reviewer model checks and human annotators verify, and three subsets target fast evaluation (\rtqatiny), questions that are hard to answer from the first and last frames (\rtqacore) and two consecutive events (\rtqacomp). All 20 evaluated models are less accurate on reversed videos. On \rtqafull, proprietary models answer both questions of 90.6--94.4\% of the pairs and the strongest open-source model of 66.8\%; on \rtqacore, the proprietary models drop to 55.6--64.5\%, and shuffling the frames reduces the paired accuracy of the tested open-source models to 3.4--7.4\%. We further build \rtqatrain{} with the same pipeline, without human verification, and train three open-source MLLMs on it with supervised fine-tuning (SFT) and reinforcement learning (DAPO) under matched settings: both raise paired accuracy on \rtqafull, SFT by 20.7--27.4 points and DAPO by 5.9--22.9 points. For Qwen3.5-9B-Thinking, the training data also help on two other temporal benchmarks: SFT and DAPO raise the Vinoground text score by 5.8 and 3.8 points, respectively, and DAPO raises the TVBench macro accuracy by 4.6 points.

cs.CV↗

Morphology-driven single-domain generalization for cross-cohort Alzheimer's disease detection using 3D MRIs

Although Alzheimer's disease detection via MRIs has advanced significantly thanks to contemporary deep learning models, challenges such as class imbalance, protocol variations, and limited dataset diversity often hinder their generalization capacity. To address this issue, this article focuses on the single domain generalization setting, where given the data of one domain, a model is designed and developed with maximal performance w.r.t. an unseen domain of distinct distribution. Since brain morphology is known to play a crucial role in Alzheimer's diagnosis, we propose the use of learnable pseudo-morphological modules aimed at producing shape-aware, anatomically meaningful class-specific augmentations in combination with a supervised contrastive learning module to extract robust class-specific representations. Experiments conducted across three datasets show improved performance and generalization capacity, especially under class imbalance and imaging protocol variations. The source code will be made available upon acceptance at https://github.com/zobia111/SDG-Alzheimer.

cs.CV↗

Evaluating and Mitigating Skin Color Bias in Skin Lesion Classificatio

Skin lesion classifiers may overlook performance disparity because of subtle skin tone variations by treating skin tone as a categorical attribute. This study proposes a method that represents skin tone as a continuous attribute and assigns individual weights to each sample based on its nonlinear relationship with classification performance. Evaluations on two dermatological datasets showed that fairness metrics were sensitive to skin-color categorization. The proposed method also demonstrated more stable fairness improvements than conventional categorical reweighting while generally preserving classification performance.

cs.CV↗