Continual multimodal instruction tuning requires multimodal large language models to acquire new task abilities sequentially while preserving previously learned knowledge. LoRA-MoE provides a promising solution by introducing expert-based capacity, but repeatedly learning and maintaining full LoRA experts leads to substantial parameter overhead. This raises a natural question: is full expert expansion necessary for every new task? To answer it, we analyze the SVD of task-specific LoRA updates and observe substantial overlap in their input- and output-side LoRA direction subspaces, with task-specific adaptation largely captured by lightweight coordinates over these subspaces. Motivated by this observation, we propose CoRe-MoE, a Compact Reusable MoE framework for parameter-efficient continual multimodal instruction tuning. CoRe-MoE extracts reusable input- and output-side direction bases from an initial expert bank, and for subsequent tasks trains only compact coordinate experts together with task-specific low-rank routers. Experiments on two representative MLLMs show that CoRe-MoE improves final average performance over the strongest competing baseline by up to 5.90 points, while using less than 1% of the trainable parameters required by sequential LoRA for later tasks. The code is publicly available at https://github.com/runzezz/CoRe-MoE.
Vision-language models such as LLaVA-1.5-7B often hallucinate objects absent from the image when generating captions. We ask whether an interpretability diagnosis of this failure can guide a targeted fix, and we measure what that fix actually changes. We rank attention heads by how much their image attention drops around hallucinated object words, then screen the shortlist by ablating candidate heads and measuring the change in hallucination-token log probability, yielding a 32-head set. We restrict two interventions to these heads: a head-sliced LoRA adapter and an inference-time grounding controller. On 400 held-out COCO images, the combined method lowers CHAIRs (the fraction of captions with a hallucinated object) from 0.370 to 0.230 and CHAIRi (the fraction of hallucinated object mentions) from 0.156 to 0.096 (p < 0.001, paired sign-flip tests). Two controls sharpen attribution. A random-head LoRA control, matched layer-for-layer and trained identically, performs no better than the matched baseline on a separate 200-image control split, supporting the role of head selection rather than LoRA capacity. Under fixed decoding budgets, the CHAIR reduction persists and grows with budget (23% at 64 tokens to 58% at 128), arguing against a pure max-token or truncation artifact, although the method remains shorter and more conservative. The resulting behavior reduces unsupported object mentions while also lowering object recall (0.78 to 0.70). We present a diagnosis-to-intervention pipeline for object hallucination, and, more importantly, a controlled account of what acting on the diagnostic signal actually does: it localizes intervention sites with real, non-random leverage, reported as a behavioral profile rather than a single score.
In this work, we explore an alternative paradigm for spatial reasoning by explicitly disentangling 3D perception from reasoning, rather than jointly acquiring implicit 3D perception and reasoning through large-scale training. Our key observation is that modern perception models excel at estimating continuous 3D geometry, whereas large language models (LLMs) are particularly effective at compositional and symbolic reasoning. Motivated by these complementary strengths, we propose the Disentangled Spatial Reasoner (DiSR), a simple yet effective framework that reconstructs the physical world into structured 3D evidence using off-the-shelf expert perception models and fine-tunes an LLM with LoRA to perform reasoning solely over this explicit geometric evidence. Without large-scale 3D VQA training or complex tool-use policies, DiSR achieves competitive performance on popular spatial reasoning benchmarks. Beyond its strong performance, DiSR offers improved interpretability, modularity, and computational efficiency, demonstrating that explicit separation of perception and reasoning is a scalable and effective alternative paradigm to end-to-end modeling for spatial intelligence.
We re-implement the NAVER LABS IWSLT 2025 instruction-following pipeline for the IWSLT 2026 Shared Task (constrained condition, short audio track), adapting it to the mandated components: SeamlessM4T-v2-large as the speech encoder and Qwen3-4B-Instruct as the LLM backbone. The three-stage approach projector alignment, text-only LoRA pre-training, and multimodal merging is preserved from the original design. We additionally construct 100k synthetic instruction-following examples across ten speech-centric task types (10k per task) from the provided corpora, suitable for further Stage 3 fine-tuning. Our primary model achieves COMET 0.781 on EN-ZH speech translation and BERTScore-F1 0.346 on English SQA on the MCIF benchmark.
Modern automated audio captioning systems pair a frozen audio encoder with a large language model (LLM) via a trainable projector, incurring the encoder's inference cost and bottlenecking the model through its fixed acoustic features. We present CARD, an encoder-free audio captioning model that removes the encoder at inference: a 13.2M projector feeds a frozen LLM with merged LoRA adapters, while the teacher used to train it is discarded. CARD distills a pretrained audio teacher (CLAP-HTSAT) into the model, but rather than injecting it into the LLM alone, it routes the teacher's representations across components: perceptual stages to the projector and semantic stages to the LLM. This placement improves CIDEr-D by +12.18 over an LLM-only distilled model on AudioCaps and by +5.21 on Clotho, reaching 55.4 against a 66.4 encoder-kept upper bound with no encoder at inference, showing that where a teacher's knowledge is placed matters as much as its presence.
Physical events are not understood by their names alone, but by the causal state changes that compose them. A clip-level label such as "bounce" can be correct while hiding the process that makes the event physically valid, from support loss and contact onset to rebound and settling. To make this hidden process explicit, we introduce Atomic Physical Transitions (APTs): minimal, temporally localized state changes that bind a visible cue to an active physical mechanism and before/after dynamical regimes. An APT chain represents a video as an ordered causal transition sequence rather than a single aggregate event label: event labels tell what happened; APT chains explain why it happened. To make APTs learnable by VLMs, we construct mixed-source APT data from human annotations and simulator ground truth, covering 14 transition types across contact, gravity, friction, and rotation/stability, with 27,303 timed instances over 1,246 trials. Using this data, we find that current VLMs miss transition-level physics, with zero-shot recall at most 14% and errors dominated by missed transitions. Direct fine-tuning on APT chains improves transition detection but causes event-level forgetting, indicating that the model learns a specialized answer format rather than a reusable physical representation. We therefore propose APT-Tune, a parameter-efficient recipe that teaches VLMs to use causal transitions without forgetting how to answer video questions. It combines image-pad-aware supervision, format-conditional co-training, and mechanism-conditioned domain-to-type decoding to make APT learning format-robust and physically grounded. With only 11 M LoRA parameters on Qwen3-VL-2B, APT-Tune substantially improves APT recall while also improving event-level video transfer. These results show that APTs are not a new answer format, but a human-aligned causal supervision signal for physical video understanding.
Personalized multimodal large language models (MLLMs) aim to generate user-specific responses, but existing methods mainly rely on profile-level information and overlook diverse user preferences. We identify group preference collapse, where multi-user personalized MLLMs become insensitive to individual preferences and drift toward dominant population-level choices due to suppressed preference signals and unreliable preference use during generation. We propose PrefMoE, a preference-centric framework that separates stable profile information from preference-related representations. PrefMoE decomposes preferences into shared prototypes and personalized residuals, preserves individualized residuals with imbalance-aware learning, counterfactual pseudo-user augmentation, and residual decorrelation, and routes profile and preference factors through separate LoRA adaptation paths. Experiments across multiple MLLM backbones show that PrefMoE improves preference-sensitive personalization while substantially reducing preference collapse. Project page: https://prefmoe.github.io/.
I Putu Adi Pratama, Bahadorreza Ofoghi, Atul Sajjanhar +1cs.CV cs.AI
Medical visual question answering (Med-VQA) has strong potential for clinical decision support by enabling AI models to interpret medical images and answer clinically relevant queries. Recent approaches typically connect off-the-shelf vision encoders with large language models (LLMs) through lightweight mapping networks to reduce computational cost. However, these methods often overlook the importance of handling noise and small irrelevant changes in visual representations. To address these challenges, we propose a noise-aware Med-VQA framework that incorporates a denoising autoencoder before visual embeddings are mapped into the input space of an LLM. The denoising autoencoder is pretrained to reconstruct clean visual embeddings from corrupted inputs, encouraging the model to learn robust visual representations that are less sensitive to noise. The resulting embeddings are then projected into the language model embedding space using a multi-layer perceptron (MLP), forming visual prefix tokens that provide image information to the LLM. To enable efficient adaptation without full retraining, we employ parameter-efficient fine-tuning using low-rank adaptation (LoRA). The proposed method is evaluated on the SLAKE and PathVQA benchmarks. Experimental results show improved robustness to noisy input embeddings while maintaining competitive clean performance across multiple evaluation criteria. These findings suggest that learning more robust visual representations can enhance Med-VQA performance and robustness.
Manan Suri, Sarvesh Baskar, Dinesh Manochacs.CV cs.CL
Processing video in vision-language models is expensive: each frame occupies hundreds of tokens, and inference cost scales with every frame and every repeated query. We introduce Video2LoRA, a method for parametric video internalization. A perceiver hypernetwork reads the intermediate representations produced layer-by-layer as a frozen VLM encodes a video, and generates a Low-Rank Adaptation (LoRA) adapter in a single forward pass. Unlike standard LoRA fine-tuning, which requires iterative gradient updates, Video2LoRA predicts these weights directly from the video. Trained for SmolVLM2 500M and 2.2B on video summarization and captioning, Video2LoRA enables the same frozen VLM to answer queries from the adapter alone, with zero visual tokens in its context at query time. Video2LoRA is statistically non-inferior and equivalent to direct video-in-context inference across all five captioning benchmarks at both model scales, and across seven of eight video question answering benchmark-scale pairings. Although trained only on 12 frames at 384px, it remains stable up to 1,024 frames and 1024px, where direct video-in-context inference often degenerates. Across this sweep, it reduces answer-time visual-token load by up to 1,500x and query TTFT by 6-80x, while preserving video-faithful outputs. We also find that independently generated adapters for non-overlapping video segments can compose in rank space, suggesting a path toward chunked long-video internalization.