arXiv subjects

Derek Hoiem

Publications and source records attributed to Derek Hoiem.

At least 19 recordsLinked to original sources

arXiv2026

ViTAL-X: Video-Text Alignment with Cross-Modal Temporal Edits

Video-text models adapted from image-text architectures (e.g., CLIP) frequently exhibit temporal blindness, the inability to perceive fundamental cues like order, direction, and motion dynamics. Standard datasets mask this limitation by enabling models to exploit static spatial shortcuts. To systematically evaluate this, we introduce XTE-Bench, a diagnostic probe revealing that even large-scale video-language models struggle with basic temporal reasoning, indicating that parameter scaling alone is insufficient to resolve this flaw. To address this, we propose Cross-Modal Temporal Edits (XTE), a self-supervised framework that injects precise temporal supervision. By performing synchronized video-text transformations, XTE generates hard temporal negatives without manual annotation. We instantiate this with ViTAL-X, a lightweight model that equips frozen image-text backbones with temporal awareness while preserving their foundational spatial knowledge. Across six temporal benchmarks, ViTAL-X achieves state-of-the-art performance. Utilizing only 0.4B parameters and 1M training clips, ViTAL-X outperforms 7B-parameter models and surpasses baselines trained on 600x more data. These results demonstrate that targeted, high-quality temporal alignment provides a highly efficient alternative to pure scaling.

cs.CV↗

arXiv2026

Fine-Tuning VLAs with Self-Demonstrated Generative Control for Multi-Task Manipulation

State-of-the-art vision-language-action (VLA) models such as $\pi_{0.5}$ exhibit strong semantic understanding, instruction following and task behavior. However, when deployed on new robots, even minor mismatches in hardware configuration relative to pretraining can cause severe performance drops. Finetuning the VLA on in-domain expert data from the new embodiment improves performance on the expert task but leads to a loss in its original instruction following and behavioral priors. In this paper, we propose a self-supervised method that generates online interaction rollouts from the zero-shot VLA as additional training data for finetuning. Our experiments show this finetuning scheme yields strong multi-task policies that, on the target robot, (1) inherit prior tasks distilled from the zero-shot model, (2) enable generalist instruction following, while (3) learning new skills from expert data with improved sample efficiency. We demonstrate the success of our approach across test sets probing generalization on a real ALOHA robot and a new simulation benchmark in RoboTwin. Video results are available at https://self-supervised-control.pages.dev/

cs.RO↗

arXiv2026

ReToken: One Token to Improve Vision-Language Models for Visual Retrieval

Long visual context poses a challenge for vision-language models: performance degrades as the number of distractors grows, and processing all tokens at once is computationally infeasible under GPU memory constraints. We present ReToken, a single learnable embedding trained as an explicit retrieval target that selects a sparse set of query-relevant visual tokens from a pre-filled visual KV cache. Trained on only a small image-QA dataset, ReToken yields consistent gains across image and video benchmarks: on Visual Haystacks it improves Qwen3VL-8B by 13.4 points and InternVL3.5 by 12.4 points (>20% relative), and on LVBench it transfers zero-shot to long video for an 8.0-point gain with Qwen3VL-8B. Thanks to its lightweight design, both training and long-video inference fit on a single H100. Code is available at: https://github.com/avaxiao/ReToken

cs.CV↗

arXiv2026

Decomposing Queries into Tool Calls for Long-Video Keyframe Retrieval

Keyframe selection is a direct way to provide verifiable visual evidence for long-video question answering (QA). Queries differ in what they require, and finding the right frames depends on knowing what to look for. Existing keyframe selectors either score every frame against a single query, or decompose the query into a fixed schema evaluated by a single visual tool. We propose ToolMerge, a keyframe retrieval method based on decomposition and merging: an Large Language Model (LLM) based planner decomposes the query into tool calls and specifies how their per-tool rankings are merged using boolean operators. To evaluate retrieval directly, we construct Molmo-2 Moments (M2M), a benchmark in which every question is anchored to a specific time interval by construction. Across QA, question retrieval, and caption retrieval, ToolMerge is competitive with prior keyframe selectors, most notably on caption retrieval, outperforming other methods by 5%. Code and data can be found at https://github.com/michalsr/ToolMerge .

cs.CV↗

arXiv2026

T-REN: Learning Text-Aligned Region Tokens Improves Dense Vision-Language Alignment and Scalability

Despite recent progress, vision-language encoders struggle with two core limitations: (1) weak alignment between language and dense vision features, which hurts tasks like open-vocabulary semantic segmentation; and (2) high token counts for fine-grained visual representations, which limits scalability to long videos. This work addresses both limitations. We propose T-REN (Text-aligned Region Encoder Network), an efficient encoder that maps visual data to a compact set of text-aligned region-level representations (or region tokens). T-REN achieves this through a lightweight network added on top of a frozen vision backbone, trained to pool patch-level representations within each semantic region into region tokens and align them with region-level text annotations. With only 3.7% additional parameters compared to the vision-language backbone, this design yields substantially stronger dense cross-modal understanding while reducing the token count by orders of magnitude. Specifically, T-REN delivers +5.9 mIoU on ADE20K open-vocabulary segmentation, +18.4% recall on COCO object-level text-image retrieval, +15.6% recall on Ego4D video object localization, and +17.6% mIoU on VSPW video scene parsing, all while reducing token counts by more than 24x for images and 187x for videos compared to the patch-based vision-language backbone. The code and model are available at https://github.com/savya08/T-REN.

cs.CV↗

arXiv2026

Why Fine-Tuning Encourages Hallucinations and How to Fix It

Large language models are prone to hallucinating factually incorrect statements. A key source of these errors is exposure to new factual information through supervised fine-tuning (SFT), which can increase hallucinations w.r.t.~knowledge acquired during pre-training. Since these errors arise as a by-product of knowledge degradation, we explore whether established continual learning tools can mitigate them. We propose a self-distillation-based SFT method that facilitates effective factual learning while minimizing hallucinations w.r.t.~pre-existing knowledge by regularizing output-distribution drift. We also show that when new knowledge acquisition is unnecessary, suppressing factual plasticity by freezing parameter groups preserves task performance while reducing hallucinations. Lastly, we investigate the mechanism, contrasting capacity limitations, behavior cloning, and localized interference. Our experiments show that a main driver is interference among overlapping semantic representations, which self-distillation mitigates and an associative-memory model explains: forgetting grows with the overlap between new and stored facts.

cs.CL↗

arXiv2026

Stress Tests REVEAL Fragile Temporal and Visual Grounding in Video-Language Models

This work investigates a fundamental question: Do Video-Language Models (VidLMs) robustly account for video content, temporal sequence, and motion? Our investigation shows that, surprisingly, they often do not. We introduce REVEAL{}, a diagnostic benchmark that probes fundamental weaknesses of contemporary VidLMs through five controlled stress tests; assessing temporal expectation bias, reliance on language-only shortcuts, video sycophancy, camera motion sensitivity, and robustness to spatiotemporal occlusion. We test leading open- and closed-source VidLMs and find that these models confidently describe reversed scenes as forward, answer questions while neglecting video content, agree with false claims, struggle with basic camera motion, and fail to aggregate temporal information amidst simple spatiotemporal masking. Humans, on the other hand, succeed at these tasks with ease. Alongside our benchmark, we provide a data pipeline that automatically generates diagnostic examples for our stress tests, enabling broader and more scalable evaluation. We will release our benchmark and code to support future research.

cs.CV↗

arXiv2025

SceneDiff: A Benchmark and Method for Multiview Object Change Detection

We investigate the problem of identifying objects that have been added, removed, or moved between a pair of captures (images or videos) of the same scene at different times. Accurately identifying verifiable changes is extremely challenging -- some objects may appear to be missing because they are occluded or out of frame, while others may appear different due to large viewpoint changes. To study this problem, we introduce the SceneDiff Benchmark, the first multiview change detection dataset for scenes captured along different camera trajectories, comprising 350 diverse video pairs with dense object instance-level annotations. We also introduce the SceneDiff algorithm, a training-free approach that solves for image poses, segments images into objects, and compares them using semantic and geometric features. By building on pretrained models, SceneDiff generalizes across domains without retraining and naturally improves as the underlying models advance. Experiments on multiview and two-view benchmarks demonstrate that our method outperforms existing approaches by large margins (53.0\% and 30.6\% relative AP improvements). Project page: https://yuqunw.github.io/SceneDiff

cs.CV↗

arXiv2025

How to Teach Large Multimodal Models New Skills

How can we teach large multimodal models (LMMs) new skills without erasing prior abilities? We study sequential fine-tuning on five target skills while monitoring general ability on eight held-out benchmarks across three model families. Surprisingly, we find that performance lost on held-out tasks after fine-tuning on one skill can partly recover when the model is subsequently tuned on a different skill. We trace this behavior to a measurable shift in the output token distribution, manifested through a simple counting-bias probe that shows the shift co-varies with forgetting. Guided by this insight, we identify two simple, robust tuning recipes that learn strongly while limiting drift: (i) updating only the self-attention projection layers (SA Proj., $\Delta$ learning +24.9 / $\Delta$ held-out forgetting -0.6), and (ii) updating only the MLP Gate&Up while freezing the Down projection (+30.5 / -2.1). Both substantially outperform full-LLM tuning (+31.8 / -23.3) in the learning-forgetting trade-off. We also compare against common forgetting mitigation methods: Learning without Forgetting (LwF), LoRA, Mixture-of-Experts, and weight-space interpolation (WiSE-FT), and find that our selective tuning recipes match or exceed their learning-stability balance while remaining simpler, requiring no replay, auxiliary parameters, or per-stage tuning. These results hold across LLaVA-OneVision, LLaVA-NeXT, and Qwen2.5-VL, confirming that the key to teaching LMMs new skills without forgetting lies in controlling output distribution shift by choosing which components to tune. Code will be made available.

cs.AI↗

arXiv2025

FRAME: Pre-Training Video Feature Representations via Anticipation and Memory

Dense video prediction tasks, such as object tracking and semantic segmentation, require video encoders that generate temporally consistent, spatially dense features for every frame. However, existing approaches fall short: image encoders like DINO or CLIP lack temporal awareness, while video models such as VideoMAE underperform compared to image encoders on dense prediction tasks. We address this gap with FRAME, a self-supervised video frame encoder tailored for dense video understanding. FRAME learns to predict current and future DINO patch features from past and present RGB frames, leading to spatially precise and temporally coherent representations. To our knowledge, FRAME is the first video encoder to leverage image-based models for dense prediction while outperforming them on tasks requiring fine-grained visual correspondence. As an auxiliary capability, FRAME aligns its class token with CLIP's semantic space, supporting language-driven tasks such as video classification. We evaluate FRAME across six dense prediction tasks on seven datasets, where it consistently outperforms image encoders and existing self-supervised video models. Despite its versatility, FRAME maintains a compact architecture suitable for a range of downstream applications.

cs.CV↗

arXiv2025

TextRegion: Text-Aligned Region Tokens from Frozen Image-Text Models

Image-text models excel at image-level tasks but struggle with detailed visual understanding. While these models provide strong visual-language alignment, segmentation models like SAM2 offer precise spatial boundaries for objects. To this end, we propose TextRegion, a simple, effective, and training-free framework that combines the strengths of image-text models and SAM2 to generate powerful text-aligned region tokens. These tokens enable detailed visual understanding while preserving open-vocabulary capabilities. They can be directly applied to various downstream tasks, including open-world semantic segmentation, referring expression comprehension, and grounding. We conduct extensive evaluations and consistently achieve superior or competitive performance compared to state-of-the-art training-free methods. Additionally, our framework is compatible with many image-text models, making it highly practical and easily extensible as stronger models emerge. Code is available at: https://github.com/avaxiao/TextRegion.

cs.CV↗

arXiv2025

PARTONOMY: Large Multimodal Models with Part-Level Visual Understanding

Real-world objects are composed of distinctive, object-specific parts. Identifying these parts is key to performing fine-grained, compositional reasoning-yet, large multimodal models (LMMs) struggle to perform this seemingly straightforward task. In this work, we introduce PARTONOMY, an LMM benchmark designed for pixel-level part grounding. We construct PARTONOMY from existing part datasets and our own rigorously annotated set of images, encompassing 862 part labels and 534 object labels for evaluation. Unlike existing datasets that simply ask models to identify generic parts, PARTONOMY uses specialized concepts (e.g., agricultural airplane), and challenges models to compare objects' parts, consider part-whole relationships, and justify textual predictions with visual segmentations. Our experiments demonstrate significant limitations in state-of-the-art LMMs (e.g., LISA-13B achieves only 5.9% gIoU), highlighting a critical gap in their part grounding abilities. We note that existing segmentation-enabled LMMs (segmenting LMMs) have two key architectural shortcomings: they use special [SEG] tokens not seen during pretraining which induce distribution shift, and they discard predicted segmentations instead of using past predictions to guide future ones. To address these deficiencies, we train several part-centric LMMs and propose PLUM, a novel segmenting LMM that uses span tagging instead of segmentation tokens and that conditions on prior predictions in a feedback loop. We find that pretrained PLUM outperforms existing segmenting LMMs on reasoning segmentation, VQA, and visual hallucination benchmarks. In addition, PLUM finetuned on our proposed Explanatory Part Segmentation task is competitive with segmenting LMMs trained on significantly more segmentation data. Our work opens up new avenues towards enabling fine-grained, grounded visual understanding in LMMs.

cs.CV↗

arXiv2025

REN: Fast and Efficient Region Encodings from Patch-Based Image Encoders

We introduce the Region Encoder Network (REN), a fast and effective model for generating region-based image representations using point prompts. Recent methods combine class-agnostic segmenters (e.g., SAM) with patch-based image encoders (e.g., DINO) to produce compact and effective region representations, but they suffer from high computational cost due to the segmentation step. REN bypasses this bottleneck using a lightweight module that directly generates region tokens, enabling 60x faster token generation with 35x less memory, while also improving token quality. It uses a few cross-attention blocks that take point prompts as queries and features from a patch-based image encoder as keys and values to produce region tokens that correspond to the prompted objects. We train REN with three popular encoders-DINO, DINOv2, and OpenCLIP-and show that it can be extended to other encoders without dedicated training. We evaluate REN on semantic segmentation and retrieval tasks, where it consistently outperforms the original encoders in both performance and compactness, and matches or exceeds SAM-based region methods while being significantly faster. Notably, REN achieves state-of-the-art results on the challenging Ego4D VQ2D benchmark and outperforms proprietary LMMs on Visual Haystacks' single-needle challenge. Code and models are available at: https://github.com/savya08/REN.

cs.CV↗

arXiv2024

Visual Program Distillation with Template-Based Augmentation

Adapting visual programming or prompting large language models (LLMs) to generate executable code for visual tasks like visual question answering (VQA) for specialized tasks or domains remains challenging due to high annotation and inference costs. We propose a low-cost visual program distillation method that can be used for models with at most 1 billion parameters and requires no human-generated program annotations. We achieve this through synthetic data augmentation based on decoupling programs into higher-level skills, called templates, and their corresponding arguments. Experimental results show that, with a relatively small amount of question/answer data, small language models can generate high-quality specialized visual programs with the added benefit of much faster inference

cs.CV↗

arXiv2024

RELOCATE: A Simple Training-Free Baseline for Visual Query Localization Using Region-Based Representations

We present RELOCATE, a simple training-free baseline designed to perform the challenging task of visual query localization in long videos. To eliminate the need for task-specific training and efficiently handle long videos, RELOCATE leverages a region-based representation derived from pretrained vision models. At a high level, it follows the classic object localization approach: (1) identify all objects in each video frame, (2) compare the objects with the given query and select the most similar ones, and (3) perform bidirectional tracking to get a spatio-temporal response. However, we propose some key enhancements to handle small objects, cluttered scenes, partial visibility, and varying appearances. Notably, we refine the selected objects for accurate localization and generate additional visual queries to capture visual variations. We evaluate RELOCATE on the challenging Ego4D Visual Query 2D Localization dataset, establishing a new baseline that outperforms prior task-specific methods by 49% (relative improvement) in spatio-temporal average precision.

cs.CV↗

arXiv2024

Plenoptic PNG: Real-Time Neural Radiance Fields in 150 KB

The goal of this paper is to encode a 3D scene into an extremely compact representation from 2D images and to enable its transmittance, decoding and rendering in real-time across various platforms. Despite the progress in NeRFs and Gaussian Splats, their large model size and specialized renderers make it challenging to distribute free-viewpoint 3D content as easily as images. To address this, we have designed a novel 3D representation that encodes the plenoptic function into sinusoidal function indexed dense volumes. This approach facilitates feature sharing across different locations, improving compactness over traditional spatial voxels. The memory footprint of the dense 3D feature grid can be further reduced using spatial decomposition techniques. This design combines the strengths of spatial hashing functions and voxel decomposition, resulting in a model size as small as 150 KB for each 3D scene. Moreover, PPNG features a lightweight rendering pipeline with only 300 lines of code that decodes its representation into standard GL textures and fragment shaders. This enables real-time rendering using the traditional GL pipeline, ensuring universal compatibility and efficiency across various platforms without additional dependencies.

cs.CV↗

arXiv2024

Anytime Continual Learning for Open Vocabulary Classification

We propose an approach for anytime continual learning (AnytimeCL) for open vocabulary image classification. The AnytimeCL problem aims to break away from batch training and rigid models by requiring that a system can predict any set of labels at any time and efficiently update and improve when receiving one or more training samples at any time. Despite the challenging goal, we achieve substantial improvements over recent methods. We propose a dynamic weighting between predictions of a partially fine-tuned model and a fixed open vocabulary model that enables continual improvement when training samples are available for a subset of a task's labels. We also propose an attention-weighted PCA compression of training features that reduces storage and computation with little impact to model accuracy. Our methods are validated with experiments that test flexibility of learning and inference. Code is available at https://github.com/jessemelpolio/AnytimeCL.

cs.CV↗

arXiv2024

MonoPatchNeRF: Improving Neural Radiance Fields with Patch-based Monocular Guidance

The latest regularized Neural Radiance Field (NeRF) approaches produce poor geometry and view extrapolation for large scale sparse view scenes, such as ETH3D. Density-based approaches tend to be under-constrained, while surface-based approaches tend to miss details. In this paper, we take a density-based approach, sampling patches instead of individual rays to better incorporate monocular depth and normal estimates and patch-based photometric consistency constraints between training views and sampled virtual views. Loosely constraining densities based on estimated depth aligned to sparse points further improves geometric accuracy. While maintaining similar view synthesis quality, our approach significantly improves geometric accuracy on the ETH3D benchmark, e.g. increasing the F1@2cm score by 4x-8x compared to other regularized density-based approaches, with much lower training and inference time than other approaches.

cs.CV↗