Research map / Label-efficient and robust learning / Recognition and long-tail
Source-Target Inference Models for Spatial Instruction Understanding
Hao Tan, Mohit Bansal
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add relative-position and stack features to block coordinates (ablation) | Full Model (Expectation) · Blank-labeled blocks validation | source accuracy: 50.0% -> 52.2% | Improved |
| add relative-position and stack features to block coordinates (ablation) | Full Model (Expectation) · Blank-labeled blocks validation | target mean distance: 3.41 -> 3.23 | Improved |
| replace last-hidden-state attention with CNN attention (ablation) | Full Model (Expectation) · Blank-labeled blocks validation | source accuracy: 47.9% -> 52.2% | Improved |
| replace last-hidden-state attention with CNN attention (ablation) | Full Model (Expectation) · Blank-labeled blocks validation | target mean distance: 3.48 -> 3.23 | Improved |
| replace CNN attention with dual attention (ablation) | Full Model (Expectation) · Blank-labeled blocks validation | source accuracy: 51.4% -> 52.2% | No clear effect |
| replace CNN attention with dual attention (ablation) | Full Model (Expectation) · Blank-labeled blocks validation | target mean distance: 3.32 -> 3.23 | No clear effect |
| use expectation inference and loss instead of annealed sampling (ablation) | Target position prediction · Blank-labeled blocks validation | target mean distance: 3.46 -> 3.23 | Improved |
| share representations through joint source-target training (ablation) | Source block prediction · Blank-labeled blocks validation | source accuracy: 51.3% -> 52.2% | Improved |
| share representations through joint source-target training (ablation) | Target position prediction · Blank-labeled blocks validation | target mean distance: 3.46 -> 3.23 | Improved |
| add an 8-sized ensemble to the expectation model (ablation) | Full Model (Expectation) · Blank-labeled blocks validation | target mean distance: 3.23 -> 3.14 | Improved |
| combine CNN and dual-attention expectation ensembles | Expectation Model · Blank-labeled blocks test | source accuracy: 56.6% -> 57.70% | Improved |
| initialize the instruction encoder with pretrained GloVe embeddings | Instruction LSTM-RNN · Blank-labeled blocks | performance | No clear effect |
Closest papers by meaning
- RoboRefer: Towards Spatial Referring with Reasoning in Vision-Language Models for Robotics (NeurIPS 2025 · similarity 0.88)
- OntoPlan: An Ontology-Grounded Scene Representation and Agentic Framework for Scalable Robot Task Planning (NeurIPS 2026 · similarity 0.88)
- SoFar: Language-Grounded Orientation Bridges Spatial Reasoning and Object Manipulation (NeurIPS 2025 · similarity 0.87)
- LINGO-Space: Language-Conditioned Incremental Grounding for Space (AAAI 2024 · similarity 0.87)
- InternSpatial: A Comprehensive Dataset for Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.87)
- SpatialRGPT: Grounded Spatial Reasoning in Vision-Language Models (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
