Research map / Vision-language and multimodal / Robot action and navigation
Embodied multimodal LLM agents
Applies pretrained multimodal LLMs and embodied chain-of-thought to interactive agents, with benchmarks.
70 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2018–2026) · an approach in Robot action and navigation, keyword cluster “vla · language action · robotic”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
13 papers in 2023–24 and 51 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 17.1% to 10.3%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2018 · Interactive Grounded Language Acquisition and Generalization in a 2D World (ICLR 2018 · 29 citations)
- 2020 · ALFRED: A Benchmark for Interpreting Grounded Instructions for Everyday Tasks (CVPR 2020 · 533 citations)
- 2021 · SILG: The Multi-domain Symbolic Interactive Language Grounding Benchmark (NeurIPS 2021 · 5 citations)
- 2022 · VLMbench: A Compositional Benchmark for Vision-and-Language Manipulation (NeurIPS 2022 · 15 citations)
- 2023 · PaLM-E: An Embodied Multimodal Language Model (ICML 2023 · 349 citations)
- 2024 · Unified-IO 2: Scaling Autoregressive Multimodal Models with Vision Language Audio and Action (CVPR 2024 · 77 citations)
- 2025 · Magma: A Foundation Model for Multimodal AI Agents (CVPR 2025 · 26 citations)
- 2026 · See, Act, Adapt: Active Perception for Unsupervised Cross-Domain Visual Adaptation via Personalized VLM-Guided Agent (ICML 2026 · 0 citations)
Earliest papers here
- Interactive Grounded Language Acquisition and Generalization in a 2D World (ICLR 2018 · 29 citations)
- ALFRED: A Benchmark for Interpreting Grounded Instructions for Everyday Tasks (CVPR 2020 · 533 citations)
- RoboTHOR: An Open Simulation-to-Real Embodied AI Platform (CVPR 2020 · 214 citations)
What has been tried
31 papers here report 336 measured changes (a component added, removed or swapped, compared with a baseline): helped 65%, no clear effect 7%, hurt 27%, mixed 1%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| increase model size from 1.1B to 6.8B parameters | GRIT ablation set · All | 60.8 -> 66.9 | Helped | Unified-IO 2: Scaling Autoregressive Multimodal Models with Vision Language Audio and Action (CVPR 2024 · 77 citations) |
| Train with retrospective LLM feedback rather than using only a plain LLM actor | ALFWorld visual tasks across training trials · success rate | Helped | Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld (CVPR 2024 · 32 citations) | |
| add SoM and ToM to full multimodal pretraining | ScreenSpot Overall · score | 57.4 -> 61.4 | Helped | Magma: A Foundation Model for Multimodal AI Agents (CVPR 2025 · 26 citations) |
| Add UI-Graph-guided visual token selection | Training efficiency · Training speedup | 1.4× | Helped | ShowUI: One Vision-Language-Action Model for GUI Visual Agent (CVPR 2025 · 24 citations) |
| use CLIP instead of MineCLIP as MineLLM's visual encoder | Context-Dependent Tasks · success rate | 95.00 vs 98.50; 90.00 vs 94.50; 87.00 vs 93.00; 80.00 vs 91.00 | Hurt | MP5: A Multi-modal Open-ended Embodied System in Minecraft via Active Perception (CVPR 2024 · 18 citations) |
| Add the 3D vision-language alignment stage | Scan2Cap (ScanNet) · sentence similarity | 62.8 -> 65.4 | Helped | An Embodied Generalist Agent in 3D World (ICML 2024 · 15 citations) |
| Condition on 4 recent frames instead of 1 frame | Ego4D validation · FVD | 315.69 -> 211.3 | Helped | Learning Interactive Real-World Simulators (ICLR 2024 · 12 citations) |
| Add the current UAV state as an encoded input | UAV-Flow · trajectory performance | Hurt | UAV-Flow Colosseo: A Real-World Benchmark for Flying-on-a-Word UAV Imitation Learning (NeurIPS 2025 · 4 citations) | |
| Fine-tune SCR on general robotics datasets | Franka-Kitchen · success rate | 45.0 -> 49.9 | Helped | Pre-trained Text-to-Image Diffusion Models Are Versatile Representation Learners for Control (NeurIPS 2024 · 4 citations) |
| Add the aligner before feeding language embeddings to the connector | Overall · normalized episodic reward | 0.17 ± -> 0.76 ± | Helped | GenRL: Multimodal-foundation world models for generalization in embodied agents (NeurIPS 2024 · 3 citations) |
| Add MLLM-based affordance maps | GRAB seen · MPJPE | 56.04 ± 1.22 -> 47.64±1.03 | Helped | OpenHOI: Open-World Hand-Object Interaction Synthesis with Multimodal Large Language Model (NeurIPS 2025 · 2 citations) |
| add future-view imagination | ObjectNav · SR | 43.0 -> 55.0 | Helped | ImagineNav: Prompting Vision-Language Models as Embodied Navigator through Scene Imagination (ICLR 2025 · 2 citations) |
What reviewers will expect
Datasets and baselines named in the 21 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Mind2Web · 2 of 21 papers
- AITW · 2 of 21 papers
- ALFRED · 2 of 21 papers
- RoboVQA · 2 of 21 papers
- UAV-Flow · 2 of 21 papers
- AndroidControl · 2 of 21 papers
Baselines
- GPT-4o · 8 of 21 papers
- Gemini-1.5-Pro · 4 of 21 papers
- Claude-3.5-Sonnet · 3 of 21 papers
- Gemini-2.0-flash · 3 of 21 papers
- Qwen-VL · 2 of 21 papers
- SeeClick · 2 of 21 papers
- OpenVLA · 2 of 21 papers
- CogAgent · 2 of 21 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 64 papers; from OpenAlex and the papers' affiliations).
- Google DeepMind · 5 papers
- Tsinghua University · 5 papers
- Stanford University · 3 papers
- Peking University · 3 papers
- Northwestern University · 2 papers
- University of Illinois, Urbana Champaign · 2 papers
- SenseTime · 2 papers
- National University of Singapore · 2 papers
- Microsoft · 2 papers
- University of California, Berkeley · 2 papers
Use it
Download BibTeX of 28 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Vision-language-action foundation models (264 papers)
- Memory and subgoal-based manipulation (103 papers)
- Multimodal prompt-driven manipulation (50 papers)
Elsewhere in the topic
- Pretraining and data augmentation for VLN (97 papers)
- Reasoning-enhanced vision-language navigation (46 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).