Same Concept, Different Directions: Cross-Modal Feature Heterogeneity in Sparse Autoencoders

Vision-language models map images and text into a joint embedding space. However, these embeddings often entangle multiple semantic features, which limits their interpretability and controllability. While sparse autoencoders have emerged as a useful tool for decomposing these embeddings into monosemantic features, their application to joint embedding spaces has largely relied on an implicit, untested assumption that semantically corresponding features share the same directions across modalities. In this paper, we challenge this assumption by identifying discrepancies in feature directions for the same concept across image and text modalities, a phenomenon we term cross-modal feature heterogeneity. We demonstrate that this heterogeneity is a key driver of the modality split, where a shared concept activates different latents depending on the modality. This finding further reveals why aligning latent activations alone is insufficient to resolve the underlying feature mismatch. Motivated by this observation, we propose an approach that trains modality-specific sparse autoencoders to preserve each modality's feature geometry, and then aligns corresponding features post hoc. Our method improves reconstruction fidelity and enhances performance in cross-modal retrieval and concept steering.

Proactive Instance Navigation with Comparative Judgment for Ambiguous User Queries

Natural-language instance navigation becomes challenging when the initial user request does not uniquely specify the target instance. A practical agent should reduce the user's burden by actively asking only the information needed to distinguish the target from similar distractors, rather than requiring a detailed description upfront. Existing approaches often fall short of this goal: they may stop at the first plausible candidate before sufficiently exploring alternatives, or, even after collecting multiple candidates, ask about the target's attributes derived from individual candidates rather than questions selected to distinguish candidates in the pool. As a result, despite the dialogue, the agent may still fail to distinguish the target from distractors, leading to premature decisions and lengthy user responses. We propose Proactive Instance Navigation with Comparative Judgment (ProCompNav), a two-stage framework that first constructs a candidate pool and then identifies the target through comparative judgment. At each round, ProCompNav extracts an attribute-value pair that splits the current pool, asks a binary yes/no question, and prunes all inconsistent candidates at once. This reframes disambiguation from open-ended target description to pool-level discriminative questioning, where each question is chosen to narrow the candidate set. On CoIN-Bench, ProCompNav improves Success Rate over interactive baselines with the same minimal input and non-interactive baselines with detailed descriptions, while substantially reducing Response Length. ProCompNav also achieves state-of-the-art Success Rate on TextNav, suggesting that comparative judgment is broadly useful for instance-level navigation among similar distractors.

Sink-Token-Aware Pruning for Fine-Grained Video Understanding in Efficient Video LLMs

Video Large Language Models (Video LLMs) incur high inference latency due to a large number of visual tokens provided to LLMs. To address this, training-free visual token pruning has emerged as a solution to reduce computational costs; however, existing methods are primarily validated on Multiple-Choice Question Answering (MCQA) benchmarks, where coarse-grained cues often suffice. In this work, we reveal that these methods suffer a sharp performance collapse on fine-grained understanding tasks requiring precise visual grounding, such as hallucination evaluation. To explore this gap, we conduct a systematic analysis and identify sink tokens--semantically uninformative tokens that attract excessive attention--as a key obstacle to fine-grained video understanding. When these sink tokens survive pruning, they distort the model's visual evidence and hinder fine-grained understanding. Motivated by these insights, we propose Sink-Token-aware Pruning (SToP), a simple yet effective plug-and-play method that introduces a sink score to quantify each token's tendency to behave as a sink and applies this score to existing spatial and temporal pruning methods to suppress them, thereby enhancing video understanding. To validate the effectiveness of SToP, we apply it to state-of-the-art pruning methods (VisionZip, FastVid, and Holitom) and evaluate it across diverse benchmarks covering hallucination, open-ended generation, compositional reasoning, and MCQA. Our results demonstrate that SToP significantly boosts performance, even when pruning up to 90% of visual tokens.

ESSENTIAL: Episodic and Semantic Memory Integration for Video Class-Incremental Learning

In this work, we tackle the problem of video classincremental learning (VCIL). Many existing VCIL methods mitigate catastrophic forgetting by rehearsal training with a few temporally dense samples stored in episodic memory, which is memory-inefficient. Alternatively, some methods store temporally sparse samples, sacrificing essential temporal information and thereby resulting in inferior performance. To address this trade-off between memory-efficiency and performance, we propose EpiSodic and SEmaNTIc memory integrAtion for video class-incremental Learning (ESSENTIAL). ESSENTIAL consists of episodic memory for storing temporally sparse features and semantic memory for storing general knowledge represented by learnable prompts. We introduce a novel memory retrieval (MR) module that integrates episodic memory and semantic prompts through cross-attention, enabling the retrieval of temporally dense features from temporally sparse features. We rigorously validate ESSENTIAL on diverse datasets: UCF-101, HMDB51, and Something-Something-V2 from the TCD benchmark and UCF-101, ActivityNet, and Kinetics-400 from the vCLIMB benchmark. Remarkably, with significantly reduced memory, ESSENTIAL achieves favorable performance on the benchmarks.

EASG-Bench: Video Q&A Benchmark with Egocentric Action Scene Graphs

We introduce EASG-Bench, a question-answering benchmark for egocentric videos where the question-answering pairs are created from spatio-temporally grounded dynamic scene graphs capturing intricate relationships among actors, actions, and objects. We propose a systematic evaluation framework and evaluate several language-only and video large language models (video-LLMs) on this benchmark. We observe a performance gap in language-only and video-LLMs, especially on questions focusing on temporal ordering, thus identifying a research gap in the area of long-context video understanding. To promote the reproducibility of our findings and facilitate further research, the benchmark and accompanying code are available at the following GitHub page: https://github.com/fpv-iplab/EASG-bench.

Reinforcement Learning meets Masked Video Modeling : Trajectory-Guided Adaptive Token Selection

Masked video modeling~(MVM) has emerged as a highly effective pre-training strategy for visual foundation models, whereby the model reconstructs masked spatiotemporal tokens using information from visible tokens. However, a key challenge in such approaches lies in selecting an appropriate masking strategy. Previous studies have explored predefined masking techniques, including random and tube-based masking, as well as approaches that leverage key motion priors, optical flow and semantic cues from externally pre-trained models. In this work, we introduce a novel and generalizable Trajectory-Aware Adaptive Token Sampler (TATS), which models the motion dynamics of tokens and can be seamlessly integrated into the masked autoencoder (MAE) framework to select motion-centric tokens in videos. Additionally, we propose a unified training strategy that enables joint optimization of both MAE and TATS from scratch using Proximal Policy Optimization (PPO). We show that our model allows for aggressive masking without compromising performance on the downstream task of action recognition while also ensuring that the pre-training remains memory efficient. Extensive experiments of the proposed approach across four benchmarks, including Something-Something v2, Kinetics-400, UCF101, and HMDB51, demonstrate the effectiveness, transferability, generalization, and efficiency of our work compared to other state-of-the-art methods.

DecompDreamer: Advancing Structured 3D Asset Generation with Multi-Object Decomposition and Gaussian Splatting

Text-to-3D generation saw dramatic advances in recent years by leveraging Text-to-Image models. However, most existing techniques struggle with compositional prompts, which describe multiple objects and their spatial relationships. They often fail to capture fine-grained inter-object interactions. We introduce DecompDreamer, a Gaussian splatting-based training routine designed to generate high-quality 3D compositions from such complex prompts. DecompDreamer leverages Vision-Language Models (VLMs) to decompose scenes into structured components and their relationships. We propose a progressive optimization strategy that first prioritizes joint relationship modeling before gradually shifting toward targeted object refinement. Our qualitative and quantitative evaluations against state-of-the-art text-to-3D models demonstrate that DecompDreamer effectively generates intricate 3D compositions with superior object disentanglement, offering enhanced control and flexibility in 3D generation. Project page : https://decompdreamer3d.github.io

Graph-Based Multimodal and Multi-view Alignment for Keystep Recognition

Egocentric videos capture scenes from a wearer's viewpoint, resulting in dynamic backgrounds, frequent motion, and occlusions, posing challenges to accurate keystep recognition. We propose a flexible graph-learning framework for fine-grained keystep recognition that is able to effectively leverage long-term dependencies in egocentric videos, and leverage alignment between egocentric and exocentric videos during training for improved inference on egocentric videos. Our approach consists of constructing a graph where each video clip of the egocentric video corresponds to a node. During training, we consider each clip of each exocentric video (if available) as additional nodes. We examine several strategies to define connections across these nodes and pose keystep recognition as a node classification task on the constructed graphs. We perform extensive experiments on the Ego-Exo4D dataset and show that our proposed flexible graph-based framework notably outperforms existing methods by more than 12 points in accuracy. Furthermore, the constructed graphs are sparse and compute efficient. We also present a study examining on harnessing several multimodal features, including narrations, depth, and object class labels, on a heterogeneous graph and discuss their corresponding contribution to the keystep recognition performance.

Deep Geometric Moments Promote Shape Consistency in Text-to-3D Generation

To address the data scarcity associated with 3D assets, 2D-lifting techniques such as Score Distillation Sampling (SDS) have become a widely adopted practice in text-to-3D generation pipelines. However, the diffusion models used in these techniques are prone to viewpoint bias and thus lead to geometric inconsistencies such as the Janus problem. To counter this, we introduce MT3D, a text-to-3D generative model that leverages a high-fidelity 3D object to overcome viewpoint bias and explicitly infuse geometric understanding into the generation pipeline. Firstly, we employ depth maps derived from a high-quality 3D model as control signals to guarantee that the generated 2D images preserve the fundamental shape and structure, thereby reducing the inherent viewpoint bias. Next, we utilize deep geometric moments to ensure geometric consistency in the 3D representation explicitly. By incorporating geometric details from a 3D asset, MT3D enables the creation of diverse and geometrically consistent objects, thereby improving the quality and usability of our 3D representations.

Ego-VPA: Egocentric Video Understanding with Parameter-efficient Adaptation

Video understanding typically requires fine-tuning the large backbone when adapting to new domains. In this paper, we leverage the egocentric video foundation models (Ego-VFMs) based on video-language pre-training and propose a parameter-efficient adaptation for egocentric video tasks, namely Ego-VPA. It employs a local sparse approximation for each video frame/text feature using the basis prompts, and the selected basis prompts are used to synthesize video/text prompts. Since the basis prompts are shared across frames and modalities, it models context fusion and cross-modal transfer in an efficient fashion. Experiments show that Ego-VPA excels in lightweight adaptation (with only 0.84% learnable parameters), largely improving over baselines and reaching the performance of full fine-tuning.