PaperlayerView exact revision on arXiv

Source-linked AI summary

TorchUMM: A Unified Multimodal Model Codebase for Evaluation, Analysis, and Post-training

Yinyi Luo, Wenwen Wang, Hayes Bai, Hongyu Zhu, Hao Chen, Pan He, Marios Savvides, Sharon Li, Jindong Wang

arXiv:2604.10784v2cs.AI

TL;DR

Unified multimodal model evaluation is fragmented across tasks and models, making consistent comparisons difficult. TorchUMM provides a shared framework for evaluation, analysis, and post-training; its evaluations find no single model dominates understanding, generation, and editing, with trade-offs across capabilities.

  • Problem

    UMM evaluation lacks a unified framework: results vary across benchmarks, existing protocols cover limited capabilities, and pipeline differences challenge consistent comparisons.

    Source: §1 · p. 1

  • Method

    TorchUMM provides a standardized framework for evaluating, analyzing, and post-training diverse UMMs through shared task pipelines and an extensible stack of models, datasets, and methods.

    Source: §1 · p. 1, §3 · p. 3

  • Results

    Across understanding, generation, and editing, no UMM dominates; strong performance in one dimension often comes at the expense of others, and scale alone is not a reliable predictor.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The evaluation reports and observed post-training trade-offs highlight the importance of holistic evaluation across tasks and benchmarks.

    Source: §1 · p. 2

    Evaluation is restricted to publicly released checkpoints under standard settings, and alternative prompting or decoding strategies may lead to different outcomes.

    Source: §6 · p. 10

Abstract

from arXiv · show

Recent advances in unified multimodal models (UMMs) have led to a proliferation of architectures capable of understanding, generating, and editing across visual and textual modalities. However, developing a unified framework for UMMs remains challenging due to the diversity of model architectures and the heterogeneity of training paradigms and implementation details. In this paper, we present TorchUMM, the first unified codebase for comprehensive evaluation, analysis, and post-training across diverse UMM backbones, tasks, and datasets. TorchUMM supports a broad spectrum of models covering a wide range of scales and design paradigms. Our benchmark encompasses three core task dimensions: multimodal understanding, generation, and editing, and integrates both established and novel datasets to evaluate perception, reasoning, compositionality, and instruction-following abilities. By providing a unified interface and standardized evaluation protocols, TorchUMM enables fair and reproducible comparisons across heterogeneous models and fosters deeper insights into their strengths and limitations, facilitating the development of more capable unified multimodal systems. Code is available at: https://github.com/AIFrontierLab/TorchUMM.

1 Introduction

TorchUMM unifies evaluation, analysis, and post-training across multimodal models and tasks to enable controlled comparisons. Its evaluations reveal fragmented capabilities and persistent limits to multimodal integration.

1 Introduction

  • Existing UMM evaluations are benchmark-dependent, often single-task, and inconsistent across pipelines, preprocessing, and interfaces, hindering reliable holistic comparisons.Localized improvements can appear overly optimistic when benchmarks and evaluation protocols cover only limited capabilities.Source: §1 · p. 1
  • TorchUMM is a unified toolkit for systematic UMM evaluation, analysis, and post-training, standardizing pipelines for controlled, reproducible cross-model and cross-task comparisons.Its standardized interface abstracts model-specific details and supports diverse UMM architectures.Source: §1 · p. 1
  • TorchUMM covers multimodal understanding, generation, and editing, with modular support for integrating diverse post-training methods.The modular mechanism enables application, research, and analysis of post-training approaches.Source: §1 · pp. 1–2
  • TorchUMM’s controlled evaluation finds fragmented capabilities, unreliable scale effects, cross-capability trade-offs, coupling failures, and unstable, model-dependent post-training.No model dominates understanding, generation, and editing, and architecture and training strategies matter more than scale alone. Editing semantic correctness can diverge from perceptual quality, generation quality from multimodal reasoning; structured manipulation and multi-step state tracking remain difficult, while post-training can degrade performance across tasks.Source: §1 · p. 2

2 Related Work

Related work spans diverse UMM architectures and task-specific benchmarks, but existing systems and evaluation practices remain fragmented. Prior UMM research emphasizes architectural unification and scaling, while deeper cross-modal consistency and reasoning receive less attention.

Unified Multimodal Models

  • UMMs integrate multimodal understanding and generation through decoder-only autoregressive transformers, hybrid autoregressive-diffusion or flow frameworks, and modular connector-based designs.Hybrid approaches are reported to improve visual generation and cross-modal alignment.Source: §2 · p. 3
  • Existing UMM approaches primarily emphasize architectural unification and scaling, leaving cross-modal consistency and reasoning less explored.Open-source codebases and benchmarking toolkits offer training pipelines, pretrained models, and standardized interfaces, but coverage remains partial and evaluation varies in preprocessing, protocols, and model interfaces.Source: §2 · p. 3

Related Benchmarks and Codebase

  • Multimodal benchmarks separately assess understanding, generation, and editing, while UMM-specific benchmarks target cross-task generalization and multimodal interaction.Understanding benchmarks cover reasoning, perception, and knowledge integration; generation benchmarks assess fidelity, prompt alignment, and compositionality; editing benchmarks focus on instruction-based image modification.Source: §2 · p. 3

3 TorchUMM

TorchUMM provides a modular, four-layer codebase for unified multimodal evaluation and post-training, standardizing support for diverse models, tasks, and benchmarks.

3 TorchUMM

  • Its four layers support infrastructure, core functions, task execution, and applications; the codebase currently includes 15 models, 12 datasets, and 5 post-training methods.The modular design is intended to make it easier to add models, datasets, and training methods.Source: §3 · p. 3

3.1 Supported Models

  • TorchUMM standardizes diverse model backbones, including Bagel [6], OmniGen2 [45], and DeepGen [38]; only a subset natively supports image editing, enabling fair cross-architecture comparisons.The supported models span different design paradigms and scales.Source: §3.1 · p. 3

3.2 Supported Tasks and Datasets

  • Understanding evaluates visual perception and reasoning, generation assesses text-to-image alignment and compositionality, and editing measures instruction-following and consistency in image modification.Understanding spans low-level perception through high-level cognition; generation also evaluates incorporation of world knowledge, while editing captures fine-grained control.Source: §3.2 · p. 4

3.3 Post-Training Methods

  • A unified interface supports five post-training approaches—SFT, IRG [15], UniCoT [28], RecA [48], and UniGame [35]—for fair comparison.The framework addresses differences in pipelines, implementation details, and benchmark datasets that hinder fair and reproducible comparisons.Source: §3.3 · p. 4

4 Evaluation Results

TorchUMM evaluates supported models with identical preprocessing, inference settings, and scoring pipelines; its independently reproduced results enable direct cross-architecture comparisons across generation, understanding, and editing. The reported numbers are not official results from model authors.

4.1 Generation

  • Generation has no single winner: Ovis-U1 (3.6B) leads GenEval (90.05) and UEval (34.17), DeepGen leads DPG Bench (87.44), and Emu3.5 leads WISE (0.633).Ovis-U1 tops GenEval despite its smaller scale, while WISE ranges from Emu3.5 (0.633) to Janus (0.222), underscoring model differences in world-knowledge grounding.Source: §4.1 · pp. 4–5, Table 1 · p. 5

4.2 Understanding

  • Bagel leads every listed understanding metric, including MME perception (1691.5), MME cognition (695.4), MMMU (0.519), and MathVista (71.6).Ovis-U1 is the strongest non-Bagel system on reasoning-heavy benchmarks; models also show perception–cognition gaps, and generation-strong Emu3.5 has the lowest MME perception score among models with understanding capability (832.17).Source: Table 2 · p. 5, §4.2 · pp. 5–6

4.3 Image Editing

  • Emu3.5 leads editing, scoring 7.56 on both GEdit-EN and GEdit-CN and topping ImgEdit SingleTurn (4.24), MultiTurn (4.89), and UGE (4.88).It also has the highest semantic correctness scores (7.64 EN, 7.62 CN).Source: §4.3 · p. 6, Table 3 · p. 6
  • Editing exposes a semantic-correctness–perceptual-quality gap and divergent multi-turn reliability: OmniGen2 has the largest cited gap, while Ovis-U1 falls from 3.97 single-turn to 2.82 multi-turn.DeepGen and Emu3.5 are comparatively stable across English and Chinese, whereas OmniGen2’s Chinese semantic correctness drops from 6.49 to 6.25.Source: §4.3 · p. 6, Table 3 · p. 6

4.4 Unification Evaluation

  • Unification is sensitive to visual representation: Janus (1.3B) falls from 0.8465 with CLIP to 0.5691 with DINOv3, while LongCLIP scores highest.Bagel, Janus-Pro, and OmniGen2 achieve high reconstruction scores on Unified-Bench, but Ovis-U1’s competitive result indicates scale alone does not determine performance.Source: §4.4 · pp. 6–7, Table 4 · p. 7
  • Realistic and reasoning-focused unification remain limited: Bagel and Ovis-U1 score 0.39 on RealUnified, while Bagel scores 30.9 on UEval despite stronger Unified-Bench results.Emu3.5’s unification score is 0.5093 despite 34B parameters; current systems align coarse semantics but lack robust structure preservation, intermediate-state tracking, and reasoning-aware generation.Source: §4.4 · p. 7, Table 4 · p. 7

4.5 Post-Training

  • Naive SFT does not reliably improve UMMs: Bagel’s MMMU rises from 0.519 to 0.526, but TokenFlow’s DPG falls from 71.29 to 22.16 and other metrics also degrade.Bagel’s MMBench-EN/CN remains 0.8265/0.8094 →0.8265/0.8088, while MM-Vet falls from 65.9 to 61.2.Source: §4.5 · pp. 7–8, Table 5 · p. 7, Table 6 · p. 8
  • Post-training effects vary by backbone and task: IRG degrades Bagel substantially, RecA and UniCot are more stable, and generation is more sensitive than understanding or editing.UniGame can improve compositional reasoning while reducing fidelity; understanding changes are generally minor, editing gains are mixed, and post-training tends to specialize rather than improve capabilities uniformly.Source: §4.5 · p. 8, Table 7 · p. 8

5 Analysis

Across MMaDA, Show-o2, and OmniGen2, architectural unification does not guarantee representational unification. The analysis finds capability and alignment patterns that often reverse the architectural-unification ordering.

5 Analysis

  • CMAM was highest for least-unified OmniGen2 (0.126), versus Show-o2 (0.030) and MMaDA (0.019), indicating tighter text–image alignment.The analysis concludes that aggressive parameter sharing can widen rather than close the modality gap without sufficient scale, data, or optimization support.Source: Figure 2 · p. 9, §5 · p. 10
  • OmniGen2 showed the strongest instruction following, while Show-o2 produced semantically under-specified images and MMaDA more often failed in grounding and visual organization.The models also differ in backbone, generator, optimization recipe, and data, so this comparison does not establish a clean causal effect of unification.Source: §5 · p. 9
  • On a shared 200-sample MathVista subset, OmniGen2 remained close to its Qwen2.5-VL-3B backbone, whereas Show-o2 deviated substantially from its Qwen2.5-VL-7B initializer.Show-o2 exhibited broader response-embedding dispersion and lower internal consistency through middle layers, suggesting that unification training can rewrite inherited behavior.Source: §5 · p. 9
  • On MathVista, OmniGen2 achieved 58.1% accuracy with 220 median visible CoT tokens; MMaDA and Show-o2 achieved 29.3% and 50.0%, with median lengths of 2 and 5 tokens.Answer-only responses occurred in 63.1% of MMaDA samples and 49.5% of Show-o2 samples, indicating difficulty turning CoT prompts into useful test-time computation.Source: §5 · p. 9, Figure 3 · p. 10

6 Conclusion and Discussion · Appendix TorchUMM: A Unified Multimodal Model Codebase for Evaluation, Analysis, and Post-training · A Codebase Structure of TorchUMM

TorchUMM standardizes evaluation, analysis, and post-training across diverse UMMs, revealing findings that fragmented evaluation setups can obscure. Its results show that conclusions depend on evaluation protocol, architecture, and training setup.

6 Conclusion and Discussion

  • TorchUMM provides a unified framework for evaluating, analyzing, and post-training diverse UMMs, standardizing inference and scoring to enable reliable comparisons.Source: §6 · p. 10
  • TorchUMM finds a gap between perception and higher-level cognition, and a mismatch between perceptual quality and semantic correctness in image editing.These findings emerge under a controlled evaluation setting.Source: §6 · p. 10
  • Models perform reasonably on static understanding and loosely grounded reasoning but struggle with iterative, structured visual manipulation revealed by joint understanding-and-generation evaluation.Source: §6 · p. 10
  • Post-training effects depend strongly on the backbone, while tighter architectural unification does not ensure better compositional reasoning and may destabilize inherited capabilities.Source: §6 · p. 10
  • TorchUMM’s findings indicate that evaluation protocol, architecture, and training setup shape UMM progress, making isolated-benchmark or single-model conclusions potentially misleading.Source: §6 · p. 10
  • The evaluation covers public checkpoints under standard settings, is largely limited to English and static images, and samples only part of the post-training design space.Alternative prompting or decoding may change outcomes; video, audio, and human preference modeling remain future work.Source: §6 · pp. 10–11

A Codebase Structure of TorchUMM

  • The modular codebase reduces the effort to integrate models, benchmarks, and post-training methods through backbone adapters, configuration, running pipelines, and cloud infrastructure.Source: A Codebase Structure of TorchUMM · p. 16

A.1 Backbone Adapter Architecture … B Details of TorchUMM

TorchUMM organizes model interaction around a shared BackboneAdapter interface and uses YAML-driven pipelines for inference, evaluation, and post-training. Its modular design supports switching models and extending the system through configuration and component additions.

A.1 Backbone Adapter Architecture

  • The BackboneAdapter protocol defines a minimal model interface, while self-contained adapters handle model-specific device mapping, tokenization, attention, and output formatting.Its interface specifies a model name plus load and generate methods.Source: §A.1 · p. 16

A.2 Configuration System

  • Three YAML configuration layers specify per-model inference settings, bind backbones to benchmarks and scoring, and define post-training methods and parameters.Changing models on the same benchmark requires only changing the backbone name and model path, without code edits; each evaluation run is determined by its config.Source: §A.2 · p. 17

A.3 Running Pipeline of TorchUMM

  • Inference, evaluation, and post-training form three operational stages, all driven by YAML configs and a shared backbone adapter interface.Source: §A.3 · p. 17

A.3.1 Inference Pipeline

  • The inference pipeline normalizes requests into a shared format and routes them to generation, understanding, or editing handlers; changing models requires YAML edits only.The request contains a prompt, images, task, and parameters; model selection is configured through inference.backbone and inference.backbone_cfg.Source: §A.3.1 · p. 17

A.3.2 Evaluation Pipeline

  • The evaluation CLI dispatches configs to benchmark-specific runners that iterate datasets, score outputs, and support two-stage benchmarks through wrappers around official scripts.Wrappers save structured outputs for analysis or external scoring while retaining standardized config, logging, and output layout.Source: §A.3.2 · p. 17

A.3.3 Post-Training Pipeline

  • Post-training configs specify methods, optimizer settings, checkpoint schedules, and distributed parameters; resulting checkpoints can be evaluated by updating the model path.Training logic is isolated from evaluation logic, allowing iteration without changes to core inference or benchmark code.Source: §A.3.3 · p. 17

A.4 Extensibility

  • TorchUMM extensions require component-specific additions: register a BackboneAdapter for a model, add a handler and CLI route for a benchmark, or add a training loop and entry point for a method.New models need no evaluation-script or infrastructure changes, and trained models can use existing evaluation configs by changing the weights path.Source: §A.4 · pp. 17–18

B.1 Supported Datasets and Benchmarks

TorchUMM covers understanding, generation, editing, and cross-task benchmarks, using protocols for both direct evaluation and post-training followed by evaluation. Together, these benchmarks assess task-specific abilities and integrated multimodal capabilities.

Generation Benchmarks

  • Generation benchmarks assess prompt fidelity, compositional generalization, and world-knowledge integration in text-to-image models.DPG-Bench [14] focuses on fine-grained detail and prompt fidelity; GenEval [20] tests object, attribute, and relation combinations; WISE [26] evaluates commonsense and factual alignment.Source: §B.1 · p. 18

Understanding Benchmarks

  • Understanding benchmarks span expert reasoning, general multimodal skills, diagnostic perception and cognition, complex inference, visual mathematics, and vision-dependent reasoning.The suite includes MMMU [56], MMBench, MME [10], MM-Vet [55], MathVista [22], and MMStar [2].Source: §B.1 · p. 18

Unified and Cross-Task Benchmarks

  • Cross-task benchmarks evaluate unified capabilities, cross-task generalization, and bidirectional synergy between understanding and generation.UEval [18] spans generation and reasoning; Uni-MMMU [60] jointly assesses understanding, generation, and editing; RealUnify [32] tests how understanding guides synthesis and generation supports problem solving.Source: §B.1 · p. 18

Editing Benchmarks

  • Editing benchmarks assess instruction alignment and precise, consistent image modifications.GEdit-Bench [20] targets attribute manipulation and multi-step edits; ImgEdit [53] tests region or attribute changes while preserving structure, identity, and visual coherence.Source: §B.1 · p. 18

Evaluation Protocols

  • TorchUMM supports direct inference-and-scoring evaluation as well as multi-stage post-training followed by evaluation.These options enable flexible assessment protocols.Source: §B.1 · p. 18

C Detailed Results · D Detailed Unification Analysis

TorchUMM’s detailed results report fine-grained subscores across benchmarks, while the supplied passages do not describe the Detailed Unification Analysis section.

C Detailed Results

  • The detailed results provide per-dimension subscores for generation, understanding, and editing across diverse benchmarks.The passages describe this as a finer-grained view of model performance under TorchUMM’s unified evaluation framework.Source: C Detailed Results · p. 19
  • Generation subscores are reported for GenEval and WISE in Tables 8 and 9.Table 8 is identified as GenEval subscores, with categories including single-object, two-object, counting, colors, position, and color attribute; no values are provided in the excerpt.Source: C Detailed Results · p. 19, Table 8 · p. 19
  • Understanding subscores are reported for MathVista, MMMU, and uni-MMMU in Tables 10, 11, and 15.Source: C Detailed Results · p. 19
  • Editing subscores are reported for GEdit-EN and RealUnify UEG score in Tables 12–14.Source: C Detailed Results · p. 19

D.1 Unification Degree is Not a Reliable Capability Proxy

The UEval case study finds that the least architecturally unified model, OmniGen2, performs best on two tightly coupled understanding–generation prompts. The results suggest unification degree is not a reliable capability proxy, but do not isolate its causal effect.

D.1 Unification Degree is Not a Reliable Capability Proxy

  • On Sample 4, OmniGen2 scored 0.79 vs. 0.46 for Show-o2 and 0.29 for MMaDA; on Sample 481, scores were 0.13 vs. 0.07 and 0.00, respectively.OmniGen2 alone consistently attempted the requested structure: progressive tutorial panels for Sample 4 and a structured Transformer diagram for Sample 481.Source: Figure 4 · p. 23, §D.1 · p. 24
  • The architectural prior ranks unification MMaDA > Show-o2 > OmniGen2, but observed capability ordering nearly reverses it, with OmniGen2 outperforming both on cross-modal understanding and final visual presentation.MMaDA uses a single token sequence and diffusion language model; Show-o2 shares a token space but uses separate generation heads; OmniGen2 conditions a separate visual generator from a VLM.Source: §D.1 · pp. 19, 24
  • The comparison does not causally isolate unification because the models differ in backbone inheritance, generative formulation, and uncontrolled training-data composition and quality.MMaDA builds on LLaDA-8B-Instruct, Show-o2 on Qwen2.5-7B-Instruct, and OmniGen2 uses a Qwen2.5-3B-Instruct VLM with a separate visual generator.Source: §D.1 · p. 24
  • One proposed explanation is that tighter unification can increase cross-modal and cross-task interference, making optimization burdens outweigh sharing benefits when representations and data are insufficient.The authors call for controlled comparisons varying unification while holding backbone architecture, initialization, tokenizer, and data recipe fixed.Source: §D.1 · p. 24

D.2 Heterogeneous Effects of Unification Training · D.3 Reasoning Traces Expose a Test-Time Scaling Gap

Unification training can preserve or substantially alter a backbone’s behavior, and stronger unification does not guarantee better consistency or the ability to use visible chain-of-thought reasoning at test time.

D.2 Heterogeneous Effects of Unification Training

  • The D.2 diagnostic compares response embeddings and response-span hidden states across meaning-preserving query variants on a 200-sample MathVista subset.Two rephrasings per question created three variants per image; hidden states were sampled every five layers.Source: §D.2 · p. 24
  • OmniGen2 remained close to Qwen2.5-VL-3B-Instruct in response consistency and layerwise latent trajectories, suggesting its unification recipe largely preserved backbone behavior.Both response-consistency curves nearly overlapped, and their latent trajectories stayed near the top of the scale with only tiny fluctuations.Source: §D.2 · p. 25, Figure 5 · p. 25
  • Show-o2 diverged from Qwen2.5-VL-7B-Instruct, with broader, lower response consistency and a less consistent latent trajectory that only partially recovered near the final layers.The Qwen backbone’s response consistency remained high and its latent trajectory nearly flat across layers.Source: §D.2 · p. 25, Figure 5 · p. 25
  • Across the two pairings, stronger unification was associated with greater disruption of inherited consistency; tighter sharing may introduce cross-task interference that depends on optimization and data mixture.The passage presents cross-task interference as a natural interpretation, not a demonstrated causal mechanism.Source: §D.2 · p. 25

D.3 Reasoning Traces Expose a Test-Time Scaling Gap

  • 99.5% of samples received explicit CoT traces from OmniGen2, whose median visible length was 220 tokens and accuracy was 58.1%; MMaDA and Show-o2 accuracies were 29.3% and 50.0%.OmniGen2 had the highest open-source accuracy despite the other models’ higher architectural unification.Source: §D.3 · p. 26
  • MMaDA and Show-o2 often produced answer-only responses rather than effective CoT traces, with median lengths of 2 and 5 tokens and answer-only rates of 63.1% and 49.5%.OmniGen2’s entropy started moderately and decreased gradually across its trace, consistent with progressively constraining the answer.Source: §D.3 · p. 26
  • The CoT results expose a trace-formation gap: without preserving or activating backbone reasoning behavior, additional test-time reasoning instructions do not yield visible reasoning that MathVista rewards.This links the accuracy gap to failure to form useful traces, not only to lower final accuracy.Source: §D.3 · p. 26

D.4 Unification Does Not Necessarily Reduce the Modality Gap · E AI Assistants Usage · F Broader Impact

On UEval, the least architecturally unified model has stronger paired text–image alignment than the more aggressively shared models, showing that architectural unification does not necessarily produce representational unification. TorchUMM supports systematic analysis, but its impact depends on use and results inherit limitations from existing datasets and models.

D.4 Unification Does Not Necessarily Reduce the Modality Gap

  • The analysis re-encodes each UEval prompt, generated image, and text answer through the understanding pathway, then compares pooled text-answer and image-answer representations.Selected-layer representations are extracted, and final-layer representations are visualized; paired spans capture both cross-modal alignment and cross-task continuity.Source: §D.4 · p. 26
  • CMAM is 0.126 for OmniGen2, 0.030 for Show-o2, and 0.019 for MMaDA, indicating stronger paired text–image alignment in the least architecturally unified model.A higher CMAM means a sample’s own text–image pair is closer than mismatched pairs; Figure 6 visualizes final-layer representations.Source: Figure 6 · p. 27, §D.4 · p. 27
  • OmniGen2’s smaller effective modality gap is consistent with stronger instruction following, more backbone-like behavior, and better CoT following, while aggressive sharing may intensify cross-modal and cross-task interference.Interference may arise when shared parameters support generation, reasoning, understanding, and trace formation before latent geometry is sufficiently organized.Source: §D.4 · p. 27
  • Before sufficient scale, data, or optimization support, greater architectural sharing can increase the practical modality gap; controlled studies should vary unification while fixing other factors.The proposed controls are backbone, tokenizer, data mixture, and optimization recipe; TorchUMM is intended to enable this analysis.Source: §D.4 · pp. 27–28

E AI Assistants Usage

  • AI assistants supported manuscript language refinement, clarity, and organization, while authors made the experimental design and methodological choices.Source: E AI Assistants Usage · p. 28

F Broader Impact

  • TorchUMM standardizes evaluation pipelines to support more consistent, reproducible comparisons and systematic analysis, though it could indirectly aid stronger generative models that may be misused.Potential misuse includes producing misleading or low-quality content at scale.Source: F Broader Impact · p. 28
  • Because TorchUMM relies on existing datasets and models, it may inherit biases or incomplete real-world coverage, so results warrant caution and additional analysis in practice.The toolkit is research-oriented, and its broader impact depends on how the community uses it.Source: F Broader Impact · p. 28

Related papers

Loading 2604.10784v2…