FeynmanBench is the first benchmark for evaluating multimodal LLMs on diagrammatic reasoning with Feynman diagrams, revealing systematic failures in enforcing physical constraints and global topology.
hub
Emerging Properties in Unified Multimodal Pretraining
97 Pith papers cite this work. Polarity classification is still indexing.
97
Pith papers citing it
abstract
Unifying multimodal understanding and generation has shown impressive capabilities in cutting-edge proprietary systems. In this work, we introduce BAGEL, an open-source foundational model that natively supports multimodal understanding and generation. BAGEL is a unified, decoder-only model pretrained on trillions of tokens curated from large-scale interleaved text, image, video, and web data. When scaled with such diverse multimodal interleaved data, BAGEL exhibits emerging capabilities in complex multimodal reasoning. As a result, it significantly outperforms open-source unified models in both multimodal generation and understanding across standard benchmarks, while exhibiting advanced multimodal reasoning abilities such as free-form image manipulation, future frame prediction, 3D manipulation, and world navigation. In the hope of facilitating further opportunities for multimodal research, we share the key findings, pretraining details, data creation protocal, and release our code and checkpoints to the community. The project page is at https://bagel-ai.org/
hub tools
citation-role summary
background 3
citation-polarity summary
claims ledger
- abstract Unifying multimodal understanding and generation has shown impressive capabilities in cutting-edge proprietary systems. In this work, we introduce BAGEL, an open-source foundational model that natively supports multimodal understanding and generation. BAGEL is a unified, decoder-only model pretrained on trillions of tokens curated from large-scale interleaved text, image, video, and web data. When scaled with such diverse multimodal interleaved data, BAGEL exhibits emerging capabilities in complex multimodal reasoning. As a result, it significantly outperforms open-source unified models in bot
co-cited works
roles
background 3polarities
background 3representative citing papers
Edit-Compass and EditReward-Compass are new unified benchmarks for fine-grained image editing evaluation and realistic reward modeling in reinforcement learning optimization.
Inline Critic uses a learnable token to critique and steer a frozen image-editing model's intermediate layers during generation, delivering state-of-the-art results on GEdit-Bench, RISEBench, and KRIS-Bench.
MindVLA-U1 introduces a unified streaming VLA with shared backbone, framewise memory, and language-guided action diffusion that surpasses human drivers on WOD-E2E planning metrics.
G²TR reduces visual tokens and prefill computation by 1.94x in separate-encoder UMMs via generation-guided importance from VAE latent consistency while preserving reasoning accuracy and editing quality.
INSET embeds images as native tokens in interleaved instructions, outperforming prior methods on multi-image consistency and text alignment as complexity grows.
A3 determines the execution horizon in VLA models as the longest prefix of actions that passes consensus-based verification and sequential consistency checks.
UniPath adaptively models coordination-path diversity in unified multimodal models by training a path-conditioned executor and using a lightweight planner for input-dependent selection, improving performance over fixed strategies.
DRoRAE adaptively fuses multi-layer features from vision encoders via energy-constrained routing to enrich visual tokens, cutting rFID from 0.57 to 0.29 and generation FID from 1.74 to 1.65 on ImageNet-256 while revealing a log-linear scaling law with fusion capacity.
A method using attention head vectors detects and suppresses risky content generation in Diffusion Transformers at inference time.
MULTITEXTEDIT benchmark reveals that all tested text-in-image editing models show pronounced degradation on non-English languages, especially Hebrew and Arabic, mainly in text accuracy and script fidelity.
ScribbleEdit is a synthetic dataset combining scribbles and text for training image editing models that produce spatially aligned and semantically consistent results.
XTC-Bench reveals that strong performance on generation or understanding tasks in unified multimodal models does not guarantee cross-task semantic consistency, which instead depends on how tightly coupled the learning objectives are across modalities.
A co-trained adapter framework enables mask-free local editing in DiTs by factorizing edit semantics from spatial location and jointly learning a mask predictor.
SketchVLM lets VLMs generate non-destructive SVG annotations on input images to visually explain answers, raising visual reasoning accuracy by up to 28.5 points and annotation quality by 1.48x over baselines.
Image editing models fail zero-shot visual planning on abstract mazes and queen puzzles but generalize after finetuning, yet still cannot match human zero-shot efficiency.
Fine-tuning multimodal models on a new synthetic spatial benchmark improves generative spatial compliance on real and synthetic tasks and transfers to better spatial understanding.
Defines ATIR task and benchmark for mixed audio-text queries; MLLM model with token compression shows substantial gains over strong baselines.
HP-Edit introduces a post-training framework and RealPref-50K dataset that uses a VLM-based HP-Scorer to align diffusion image editing models with human preferences, improving outputs on Qwen-Image-Edit-2509.
UniGeo unifies geometric guidance across three levels in video models to reduce geometric drift and improve consistency in camera-controllable image editing.
UniEditBench unifies image and video editing evaluation with a nine-plus-eight operation taxonomy and cost-effective 4B/8B distilled MLLM evaluators that align with human judgments.
3D-ARD+ unifies autoregressive token prediction with diffusion-based 3D latent generation to co-produce indoor scene layouts and object geometries that follow complex text-specified spatial and semantic constraints.
π₀.₇ is a steerable generalist robotic model that uses rich multimodal prompts including language, subgoal images, and performance metadata to achieve out-of-the-box generalization across tasks and robot bodies.
OmniScript is a new 8B omni-modal model that turns long cinematic videos into scene-by-scene scripts and matches top proprietary models on temporal localization and semantic accuracy.
citing papers explorer
-
Edit-Compass & EditReward-Compass: A Unified Benchmark for Image Editing and Reward Modeling
Edit-Compass and EditReward-Compass are new unified benchmarks for fine-grained image editing evaluation and realistic reward modeling in reinforcement learning optimization.
-
Inline Critic Steers Image Editing
Inline Critic uses a learnable token to critique and steer a frozen image-editing model's intermediate layers during generation, delivering state-of-the-art results on GEdit-Bench, RISEBench, and KRIS-Bench.
-
G$^2$TR: Generation-Guided Visual Token Reduction for Separate-Encoder Unified Multimodal Models
G²TR reduces visual tokens and prefill computation by 1.94x in separate-encoder UMMs via generation-guided importance from VAE latent consistency while preserving reasoning accuracy and editing quality.
-
Images in Sentences: Scaling Interleaved Instructions for Unified Visual Generation
INSET embeds images as native tokens in interleaved instructions, outperforming prior methods on multi-image consistency and text alignment as complexity grows.
-
Dynamic Execution Commitment of Vision-Language-Action Models
A3 determines the execution horizon in VLA models as the longest prefix of actions that passes consensus-based verification and sequential consistency checks.
-
Beyond the Last Layer: Multi-Layer Representation Fusion for Visual Tokenization
DRoRAE adaptively fuses multi-layer features from vision encoders via energy-constrained routing to enrich visual tokens, cutting rFID from 0.57 to 0.29 and generation FID from 1.74 to 1.65 on ImageNet-256 while revealing a log-linear scaling law with fusion capacity.
-
What Concepts Lie Within? Detecting and Suppressing Risky Content in Diffusion Transformers
A method using attention head vectors detects and suppresses risky content generation in Diffusion Transformers at inference time.
-
MULTITEXTEDIT: Benchmarking Cross-Lingual Degradation in Text-in-Image Editing
MULTITEXTEDIT benchmark reveals that all tested text-in-image editing models show pronounced degradation on non-English languages, especially Hebrew and Arabic, mainly in text accuracy and script fidelity.
-
ScribbleEdit: Synthetic Data for Image Editing with Scribbles and Text
ScribbleEdit is a synthetic dataset combining scribbles and text for training image editing models that produce spatially aligned and semantically consistent results.
-
Beyond Accuracy: Benchmarking Cross-Task Consistency in Unified Multimodal Models
XTC-Bench reveals that strong performance on generation or understanding tasks in unified multimodal models does not guarantee cross-task semantic consistency, which instead depends on how tightly coupled the learning objectives are across modalities.
-
Edit Where You Mean: Region-Aware Adapter Injection for Mask-Free Local Image Editing
A co-trained adapter framework enables mask-free local editing in DiTs by factorizing edit semantics from spatial location and jointly learning a mask predictor.
-
SketchVLM: Vision language models can annotate images to explain thoughts and guide users
SketchVLM lets VLMs generate non-destructive SVG annotations on input images to visually explain answers, raising visual reasoning accuracy by up to 28.5 points and annotation quality by 1.48x over baselines.
-
Probing Visual Planning in Image Editing Models
Image editing models fail zero-shot visual planning on abstract mazes and queen puzzles but generalize after finetuning, yet still cannot match human zero-shot efficiency.
-
Exploring Spatial Intelligence from a Generative Perspective
Fine-tuning multimodal models on a new synthetic spatial benchmark improves generative spatial compliance on real and synthetic tasks and transfers to better spatial understanding.
-
HP-Edit: A Human-Preference Post-Training Framework for Image Editing
HP-Edit introduces a post-training framework and RealPref-50K dataset that uses a VLM-based HP-Scorer to align diffusion image editing models with human preferences, improving outputs on Qwen-Image-Edit-2509.
-
UniGeo: Unifying Geometric Guidance for Camera-Controllable Image Editing via Video Models
UniGeo unifies geometric guidance across three levels in video models to reduce geometric drift and improve consistency in camera-controllable image editing.
-
UniEditBench: A Unified and Cost-Effective Benchmark for Image and Video Editing via Distilled MLLMs
UniEditBench unifies image and video editing evaluation with a nine-plus-eight operation taxonomy and cost-effective 4B/8B distilled MLLM evaluators that align with human judgments.
-
Co-generation of Layout and Shape from Text via Autoregressive 3D Diffusion
3D-ARD+ unifies autoregressive token prediction with diffusion-based 3D latent generation to co-produce indoor scene layouts and object geometries that follow complex text-specified spatial and semantic constraints.
-
OmniScript: Towards Audio-Visual Script Generation for Long-Form Cinematic Video
OmniScript is a new 8B omni-modal model that turns long cinematic videos into scene-by-scene scripts and matches top proprietary models on temporal localization and semantic accuracy.
-
Pseudo-Unification: Entropy Probing Reveals Divergent Information Patterns in Unified Multimodal Models
Unified multimodal models exhibit pseudo-unification due to modality-asymmetric entropy encoding and pattern-split responses between text and image generation.
-
RefineAnything: Multimodal Region-Specific Refinement for Perfect Local Details
RefineAnything is a multimodal diffusion model using Focus-and-Refine crop-and-resize with blended paste-back to achieve high-fidelity local image refinement and near-perfect background preservation.
-
V-Reflection: Transforming MLLMs from Passive Observers to Active Interrogators
V-Reflection introduces a think-then-look mechanism where MLLM latent states actively interrogate visual features via two-stage distillation from a box-guided teacher to a dynamic autoregressive student, narrowing the fine-grained perception gap on benchmarks.
-
Beyond Text Prompts: Visual-to-Visual Generation as A Unified Paradigm
V2V-Zero adapts frozen VLMs for visual conditioning via hidden states from specification pages, scoring 0.85 on GenEval and 32.7 on a new seven-task benchmark while revealing capability hierarchies in attribute binding and structural control.
-
UniCustom: Unified Visual Conditioning for Multi-Reference Image Generation
A unified visual conditioning approach fuses semantic and appearance features before VLM processing, with two-stage training and slot-wise regularization, to improve consistency in multi-reference image generation.
-
GeoR-Bench: Evaluating Geoscience Visual Reasoning
GeoR-Bench shows top multimodal models reach only 42.7% strict accuracy on geoscience visual reasoning tasks while open-source models reach 10.3%, with outputs often visually plausible yet scientifically inaccurate.
-
PresentAgent-2: Towards Generalist Multimodal Presentation Agents
PresentAgent-2 generates query-driven multimodal presentation videos with research grounding, supporting single-speaker, multi-speaker discussion, and interactive question-answering modes.
-
HiDream-O1-Image: A Natively Unified Image Generative Foundation Model with Pixel-level Unified Transformer
A pixel-space Diffusion Transformer with Unified Transformer architecture unifies image generation, editing, and personalization in an end-to-end model that maps all inputs to a shared token space and scales from 8B to over 200B parameters.
-
Do multimodal models imagine electric sheep?
Fine-tuning VLMs to output action sequences for puzzles causes emergent internal visual representations that improve performance when integrated into reasoning.
-
SCOPE: Structured Decomposition and Conditional Skill Orchestration for Complex Image Generation
SCOPE maintains semantic commitments via structured specifications and conditional skill orchestration, achieving 0.60 EGIP on the new Gen-Arena benchmark while outperforming baselines on WISE-V and MindBench.
-
STARFlow2: Bridging Language Models and Normalizing Flows for Unified Multimodal Generation
STARFlow2 presents an autoregressive flow-based architecture for unified multimodal text-image generation by interleaving a VLM stream with a TarFlow stream via residual skips and a unified latent space.
-
MUSE: Resolving Manifold Misalignment in Visual Tokenization via Topological Orthogonality
MUSE decouples reconstruction and semantic learning in visual tokenization via topological orthogonality, yielding SOTA generation quality and improved semantic performance over its teacher model.
-
AlbumFill: Album-Guided Reasoning and Retrieval for Personalized Image Completion
AlbumFill retrieves identity-consistent references from personal albums via VLM-inferred semantic cues to support personalized image completion.
-
PhysEdit: Physically-Consistent Region-Aware Image Editing via Adaptive Spatio-Temporal Reasoning
PhysEdit introduces adaptive reasoning depth and spatial masking to make image editing faster and more instruction-aligned without retraining the base model.
-
Leveraging Verifier-Based Reinforcement Learning in Image Editing
Edit-R1 trains a CoT-based reasoning reward model with GCPO and uses it to boost image editing performance over VLMs and models like FLUX.1-kontext via GRPO.
-
SpatialFusion: Endowing Unified Image Generation with Intrinsic 3D Geometric Awareness
SpatialFusion internalizes 3D geometric awareness into unified image generation models by pairing an MLLM with a spatial transformer that produces depth maps to constrain diffusion generation.
-
Refinement via Regeneration: Enlarging Modification Space Boosts Image Refinement in Unified Multimodal Models
Refinement via Regeneration (RvR) reformulates image refinement in unified multimodal models as conditional regeneration using prompt and semantic tokens from the initial image, yielding higher alignment scores than editing-based methods.
-
DDA-Thinker: Decoupled Dual-Atomic Reinforcement Learning for Reasoning-Driven Image Editing
DDA-Thinker decouples planning from generation and applies dual-atomic RL with checklist-based rewards to boost reasoning in image editing, yielding competitive results on RISE-Bench and KRIS-Bench.
-
Meta-CoT: Enhancing Granularity and Generalization in Image Editing
Meta-CoT uses two-level decomposition of editing operations into meta-tasks and a CoT consistency reward to improve granularity and generalization, reporting 15.8% gains across 21 tasks.
-
SMoES: Soft Modality-Guided Expert Specialization in MoE-VLMs
SMoES improves MoE-VLM performance and efficiency via soft modality-guided expert routing and inter-bin mutual information regularization, yielding 0.9-4.2% task gains and 56% communication reduction.
-
Seeing Without Eyes: 4D Human-Scene Understanding from Wearable IMUs
IMU-to-4D uses wearable IMU data and repurposed LLMs to predict coherent 4D human motion plus coarse scene structure, outperforming cascaded state-of-the-art pipelines in temporal stability.
-
Camera Control for Text-to-Image Generation via Learning Viewpoint Tokens
Viewpoint tokens learned on a mixed 3D-rendered and photorealistic dataset enable precise camera control in text-to-image generation while factorizing geometry from appearance and transferring to unseen object categories.
-
How Far Are Video Models from True Multimodal Reasoning?
Current video models succeed on basic understanding but achieve under 25% success on logically grounded generation and near 0% on interactive generation, exposing gaps in multimodal reasoning.
-
Generative Refinement Networks for Visual Synthesis
GRN uses hierarchical binary quantization and entropy-guided refinement to set new ImageNet records of 0.56 rFID for reconstruction and 1.81 gFID for class-conditional generation while releasing code and models.
-
Self-Adversarial One Step Generation via Condition Shifting
APEX derives self-adversarial gradients from condition-shifted velocity fields in flow models to achieve high-fidelity one-step generation, outperforming much larger models and multi-step teachers.
-
Nucleus-Image: Sparse MoE for Image Generation
A 17B-parameter sparse MoE diffusion transformer activates 2B parameters per pass and reaches competitive quality on image generation benchmarks without post-training.
-
POINTS-Long: Adaptive Dual-Mode Visual Reasoning in MLLMs
POINTS-Long is a dual-mode multimodal large language model that uses dynamic visual token scaling to retain 97.7-99.7% accuracy on long-form tasks with 1/40 to 1/10th the tokens and supports streaming via detachable KV-cache.
-
LMGenDrive: Bridging Multimodal Understanding and Generative World Modeling for End-to-End Driving
LMGenDrive unifies LLM-based multimodal understanding with generative world models to output both future driving videos and control signals for end-to-end closed-loop autonomous driving.
-
InsEdit: Towards Instruction-based Visual Editing via Data-Efficient Video Diffusion Models Adaptation
InsEdit adapts a video diffusion backbone for text-instruction video editing via Mutual Context Attention, achieving SOTA open-source results with O(100K) data while also supporting image editing.
-
Uni-ViGU: Towards Unified Video Generation and Understanding via A Diffusion-Based Video Generator
Uni-ViGU unifies video generation and understanding by extending a diffusion video generator with unified continuous-discrete flow matching, modality-driven MoE layers, and bidirectional training stages that repurpose generative knowledge for discriminative tasks.
-
Symbiotic-MoE: Unlocking the Synergy between Generation and Understanding
Symbiotic-MoE introduces modality-aware expert disentanglement and progressive training in a multimodal MoE to achieve synergistic generation and understanding without task interference or extra parameters.