PEFT of VLMs

Recognizing fine-grained bodily behaviors, such as self-grooming, postural shifts, and backchanneling, is a useful intermediate step for computational social signal analysis. But the behaviors that matter vary across social contexts, and retraining a model for each new taxonomy is expensive. Vision-language models (VLMs) can adapt to new categories through text prompts, yet they lack sensitivity to body structure and often confuse subtle postural distinctions. We propose PRIMA (Pose Refined Inductive Motion Adapter), a lightweight module that injects a body-structure prior into frozen VLMs using human pose as a training-time supervisory signal, requiring no pose input at inference. PRIMA separates a model-specific projection head from a shared core that encodes transferable body-configuration knowledge; adapting to a new VLM requires retraining only the head. We train PRIMA on a micro-action dataset and evaluate entirely through zero-shot prompting. The adapter improves recognition on its training domain, transfers to additional VLM backbones at different scales with the core frozen, and generalizes to two unseen target datasets from social psychology and daily activity recognition. In a supporting downstream study, the improved behavior predictions also yield higher correlations with human ratings of rapport and dominance in a collaboration context. Within the scope of the tested models and indoor datasets, the results indicate that pose-guided training can produce body-structure priors that generalize across unseen behavior vocabularies and can be partially reused across VLM backbones while preserving RGB-only inference.

Large Vision-Language Models (LVLMs) commonly suffer from hallucinations, which appear as fabricating non-existent objects or outputting incorrect attributes and spatial relationships in images. Previous works empirically indicate that CLIP-style vision encoders tend to focus on primary objects while ignoring image details, which is one of the causes of hallucinations. To address this issue, we propose plug-and-play fine-grained visual experts (PLANE) comprising two components: LoRA visual experts operating on CLIP for multi-aspect image encoding, and an instruction-aware gate network for adaptive image feature fusion. During training, we construct a fine-grained image-caption pairs dataset, where images are described from three aspects: objects, attributes, and relations between objects. Image-caption pairs from different aspects are used to train the three fine-grained LoRA experts respectively. Aiming at comprehensively evaluate our method's capabilities, we manually annotate a fine-grained hallucination evaluation benchmark for LVLMs which is GPT4-independent during evaluation phase (GIFT). We conduct experiments both on popular hallucination evaluation benchmarks and GIFT. The results show that our method effectively reduces hallucinations across models of different types and sizes while introducing minimal inference latency. Compared to advanced baselines, our model also demonstrates superior performance.

Fine-tuning vision-language models (VLMs) is typically evaluated at a single downstream checkpoint, obscuring whether a semantic capability was never acquired or emerged earlier and later declined during specialization. We ask how semantic capabilities are acquired, when they peak, how well they transfer, and what remains at deployment. We study these dynamics as a semantic capability trajectory, tracking identity- and attribute-based capabilities over training. We formulate a trajectory-based framework that separates capability acquisition, capability-specific optima, and later specialization, and introduce Structured Semantic Routing (SSR) to study how the representation of supervision shapes what is acquired. Across six pretrained backbones spanning DFN, MetaCLIP, and OpenAI CLIP, we show that fine-tuning can acquire semantic capability beyond the pretrained state, including gains observed on held-out evaluations. Unstructured name-and-attribute supervision produces strong name-and-attribute retrieval with comparatively weak name-free attribute-profile retrieval, whereas SSR yields substantially stronger name-free attribute-profile retrieval and is further strengthened by stochastic name-branch dropout. Different capabilities can peak at different stages, so a checkpoint selected by target class-name retrieval need not coincide with a transferable semantic optimum. Continued optimization can therefore preserve strong target class-name retrieval while reducing previously acquired transferable semantic capability. In a representative diagnostic study, this late specialization is consistent with reduced cross-modal semantic accessibility while substantial image-only class structure remains available.

Spatial intelligence is a fundamental skill in multiple domains, such as Science, Technology, Engineering, and Mathematics (STEM), Medicine, Architecture, and Construction. Recent studies indicate that Vision-Language Models (VLMs) still face limitations in spatial reasoning, which inhibits artificial intelligence (AI) from performing practical spatial tasks. Using multiple object-rotation datasets developed for training and evaluation, our experiments demonstrated promising improvements in both 2D and 3D rotation detection. Fine-tuned Google DeepMind-built Gemma-4 mixture-of-experts (MoE) models significantly outperformed fine-tuned Gemma-4 generalist models in predicting rotations defined by both their axes and angles. Fine-tuning also substantially improved angle estimation for 2D representation without requiring an explicit coordinate system. Furthermore, identifiable objects did not improve angle-detection accuracy; instead, objects with prominent linear features showed improved performance.

We address few-shot object detection with vision-language models (VLMs) in out-of-domain settings such as aerial, industrial, and medical imagery, using only ten annotated images for supervision. Existing adaptation methods are discrete prompt optimization and LoRA fine-tuning. We revisit a third option: soft prompting, where a small number of continuous prompt tokens are optimized while the pretrained backbone remains frozen. We identify two key design choices. First, placing prompt tokens at the cross-modal boundary between visual and text tokens outperforms other placements (10.0 vs. 8.4 mAP). Second, initializing prompts from the empty space token outperforms semantic and random initialization. With these choices, one to three learned tokens (7,168 parameters on average) match the best LoRA configuration on Roboflow20-VL (14.2 mAP, 10-shot) while training over 20,000x fewer parameters. Soft prompting remains harder to optimize, exhibiting higher variance across random seeds. Unlike LoRA, however, it causes no forgetting: the LoRA rank matching our accuracy reduces NaturalBench VQA accuracy by 35% relative, rising to 56% at the largest rank, whereas soft prompting leaves pretrained performance unchanged. The learned tokens behave like prompts rather than weights. They transfer to a newer model without retraining (+0.8 mAP on Qwen3.5-9B) and can be verbalized into readable prompts competitive with prompt-search methods (matching DetPO and outperforming GEPA). The approach also extends beyond detection. On RoboCasa manipulation tasks, the frozen $π_{0.5}$ vision-language-action policy benefits from soft prompting, matching the LoRA baseline on two of three tasks when tokens are placed at the gradient bottleneck. These results suggest modern VLMs already encode much of what is needed for specialized domains; the challenge is learning how to ask.

Multimodal large language models (MLLMs) have demonstrated remarkable capabilities by integrating visual and textual understanding within a unified transformer architecture. However, fine-tuning all parameters of these models for visual instruction tuning is computationally expensive and often unnecessary, as the representation requirements for visual and textual tokens diverge significantly in the deeper layers of the network. In this paper, we propose Decoupled Visual Processing (DVP), an efficient training framework that replaces the upper decoder layers of a pretrained LLM with a lightweight, independently trainable single transformer block dedicated exclusively to visual token processing. Specifically, after shared processing through the first half of the decoder layers, visual and textual tokens are split: visual tokens are routed through a newly initialized single transformer block while textual tokens continue through the original frozen decoder layers. The two streams are then concatenated before the language modeling head. During training, only the single transformer block is updated, dramatically reducing the number of trainable parameters. Experiments on the LLaVA-1.5 framework demonstrate that DVP achieves competitive performance on MME, POPE, and ChartQA benchmarks while training only a fraction of the total parameters, suggesting that visual representations in MLLMs can be effectively learned through a decoupled, parameter-efficient pathway.

Leveraging the rich semantic features of vision-language models (VLMs) like CLIP for monocular depth estimation tasks is a promising direction, yet often requires extensive fine-tuning or lacks geometric precision. We present a parameter-efficient framework, named MoA-DepthCLIP, that adapts pretrained CLIP representations for monocular depth estimation with minimal supervision. Our method integrates a lightweight Mixture-of-Adapters (MoA) module into the pretrained Vision Transformer (ViT-B/32) backbone combined with selective fine-tuning of the final layers. This design enables spatially-aware adaptation, guided by a global semantic context vector and a hybrid prediction architecture that synergizes depth bin classification with direct regression. To enhance structural accuracy, we employ a composite loss function that enforces geometric constraints. On the NYU Depth V2 benchmark, MoA-DepthCLIP achieves competitive results, significantly outperforming the DepthCLIP baseline by improving the $δ_1$ accuracy from 0.390 to 0.745 and reducing the RMSE from 1.176 to 0.520. These results are achieved while requiring substantially few trainable parameters, demonstrating that lightweight, prompt-guided MoA is a highly effective strategy for transferring VLM knowledge to fine-grained monocular depth estimation tasks.

Fine-tuning approaches for Vision-Language Models (VLMs) face a critical three-way trade-off between In-Distribution (ID) accuracy, Out-of-Distribution (OOD) generalization, and adversarial robustness. Existing robust fine-tuning strategies resolve at most two axes of this trade-off. Generalization-preserving methods retain ID/OOD performance but leave models vulnerable to adversarial attacks, while adversarial training improves robustness to targeted attacks but degrades ID/OOD accuracy. Our key insight is that the robustness trade-off stems from two geometric failures: sharp, anisotropic minima in parameter space and unstable feature representations that deform under perturbation. To address this, we propose GRACE (Gram-aligned Robustness via Adaptive Curvature Estimation), a unified fine-tuning framework that jointly regularizes the parameter-space curvature and feature-space invariance for VLMs. Grounded in Robust PAC-Bayes theory, GRACE employs adaptive weight perturbations scaled by local curvature to promote flatter minima, combined with a feature alignment loss that maintains representation consistency across clean, adversarial, and OOD inputs. On ImageNet fine-tuning of CLIP models, GRACE simultaneously improves ID accuracy by 10.8%, and adversarial accuracy by 13.5% while maintaining 57.0% OOD accuracy (vs. 57.4% zero-shot baseline). Geometric analysis confirms that GRACE converges to flatter minima without feature distortion across distribution shifts, providing a principled step toward generalized robustness in foundation VLMs.

Supervised fine-tuning (SFT) on visual instruction data often improves perceptual capabilities in vision-language models (VLMs) while degrading reasoning performance, creating a persistent reasoning tax during post-training. We investigate whether this degradation is related to disrupted access to depth-wise representations, and find that even fixed cross-depth aggregation substantially restores reasoning, suggesting that preserved cross-depth access is an important missing factor in VLM fine-tuning. Building on this observation, we propose Input-Adaptive Depth Aggregation (IADA), a lightweight mechanism that makes cross-depth retrieval input-adaptive, modality-aware, and efficiently parameterized through a low-rank bottleneck. On Qwen3-VL-2B, IADA improves the average reasoning score by 9.5 points and the average perception score by $3.3$ points over LoRA-only fine-tuning with only 0.14M additional parameters, with the strongest gains appearing in parameter-efficient low-rank settings.

Multimodal reasoning in vision-language models (VLMs) typically relies on a two-stage process: supervised fine-tuning (SFT) and reinforcement learning (RL). In standard SFT, all tokens contribute equally to the loss, even though reasoning data are inherently token-imbalanced. Long <think> traces overshadow short but task-critical <answer> segments, leading to verbose reasoning and inaccurate answers. We propose SCALe (Scheduled Curriculum Adaptive Loss), which explicitly separates supervision over reasoning and answer segments using dynamic, length-independent weighting. Unlike vanilla SFT, which overweights the <think> segment, SCALe-SFT gradually shifts the focus from <think> to <answer> throughout training via a cosine scheduling policy, encouraging concise and well-grounded reasoning. We evaluate SCALe across diverse benchmarks and architectures. Results show that SCALe consistently improves accuracy over vanilla SFT and matches the performance of the full two-phase SFT + GRPO pipeline while requiring only about one-seventh of the training time, making it a lightweight yet effective alternative. When combined with GRPO, SCALe achieves the best overall performance, highlighting its value both as a standalone method and as a strong foundation for reinforcement refinement.

A reliable driving assistant should provide consistent responses based on temporally grounded reasoning derived from observed information. In this work, we investigate whether Vision-Language Models (VLMs), when applied as driving assistants, can response consistantly and understand how present observations shape future outcomes, or whether their outputs merely reflect patterns memorized during training without temporally grounded reasoning. While recent efforts have integrated VLMs into autonomous driving, prior studies typically emphasize scene understanding and instruction generation, implicitly assuming that strong visual interpretation naturally enables consistant future reasoning and thus ensures reliable decision-making, a claim we critically examine. We focus on two major challenges limiting VLM reliability in this setting: response inconsistency, where minor input perturbations yield different answers or, in some cases, responses degenerate toward near-random guessing, and limited temporal reasoning, in which models fail to reason and align sequential events from current observations, often resulting in incorrect or even contradictory responses. Moreover, we find that models with strong visual understanding do not necessarily perform best on tasks requiring temporal reasoning, indicating a tendency to over-rely on pretrained patterns rather than modeling temporal dynamics. To address these issues, we adopt existing evaluation methods and introduce FutureVQA, a human-annotated benchmark dataset specifically designed to assess future scene reasoning. In addition, we propose a simple yet effective self-supervised tuning approach with chain-of-thought reasoning that improves both consistency and temporal reasoning without requiring temporal labels.

Recent advances in vision-language models (VLMs) have sparked growing interest in using them to automate web tasks, yet their feasibility as independent agents that reason and act purely from visual input remains underexplored. We investigate this setting using Qwen2.5-VL-32B, one of the strongest open-source VLMs available, and focus on improving its reliability in web-based control. Through initial experimentation, we observe three key challenges: (i) inaccurate localization of target elements, the cursor, and their relative positions, (ii) sensitivity to instruction phrasing, and (iii) an overoptimistic bias toward its own actions, often assuming they succeed rather than analyzing their actual outcomes. To address these issues, we fine-tune Qwen2.5-VL-32B for a basic web interaction task: moving the mouse and clicking on a page element described in natural language. Our training pipeline consists of two stages: (1) teaching the model to determine whether the cursor already hovers over the target element or whether movement is required, and (2) training it to execute a single command (a mouse move or a mouse click) at a time, verifying the resulting state of the environment before planning the next action. Evaluated on a custom benchmark of single-click web tasks, our approach increases success rates from 86% to 94% under the most challenging setting.

Focusing on low-resource languages is an essential step toward democratizing generative AI. In this work, we contribute to reducing the multimodal NLP resource gap for Romanian. We translate the widely known Flickr30K dataset into Romanian and further extend it for visual question answering by leveraging open-source LLMs. We demonstrate the usefulness of our datasets by fine-tuning open-source VLMs on Romanian visual question answering. We select VLMs from three widely used model families: LLaMA 3.2, LLaVA 1.6, and Qwen2. For fine-tuning, we employ the parameter-efficient LoRA method. Our models show improved Romanian capabilities in visual QA, as well as on tasks they were not trained on, such as Romanian image description generation. The seven-billion-parameter Qwen2-VL-RoVQA obtains top scores on both tasks, with improvements of +2.29% and +4.45% in BERTScore F1 on VQA and captioning, respectively, over its original version. Finally, the models show substantial reductions in grammatical errors compared to their original forms, indicating improvements not only in language understanding but also in Romanian fluency.