Research map / Large language models / Agents, code and math
Distilling Neuro-Symbolic Programs into 3D Multi-modal LLMs
Wentao Mo, Yang Liu
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 CoT-RL stage (Stage 3) instead of direct-answer SFT (ablation) | APEIRIA · ScanRefer Acc@0.25 | accuracy: 51.5 -> 58.4 | Improved |
| add CoT-SFT symbolic reasoning injection (Stage 2) before RL (ablation) | APEIRIA · Multi3DRefer F1@0.25 | F1: 36.7 -> 59.2 | Improved |
| add format reward in RL (ablation) | APEIRIA · ScanRefer Acc@0.25 | accuracy: 55.7 -> 58.4 | Improved |
| use soft grounding reward instead of IoU reward (ablation) | APEIRIA · ScanRefer Acc@0.25 | accuracy: 57.7 -> 58.4 | Improved |
| generate thinking trace at inference instead of direct answer (ablation) | APEIRIA · ScanRefer Acc@0.25 | accuracy: 56.8 -> 58.4 | Improved |
| CoT-RL on short reasoning chains (<=4 steps) (ablation) | APEIRIA · ScanRefer Acc@0.5 | accuracy: 47.2 -> 45.4 | Hurt |
| CoT-RL on medium chains (5 steps) (ablation) | APEIRIA · ScanRefer Acc@0.5 | accuracy: 50.7 -> 52.2 | Improved |
| CoT-RL on long chains (>=6 steps) (ablation) | APEIRIA · ScanRefer Acc@0.5 | accuracy: 45.2 -> 47.9 | Improved |
| inject plans from Claude 4.5 Opus instead of self-generated plans (ablation) | APEIRIA · ScanRefer Acc@0.25 | accuracy: 58.4 -> 58.6 | No clear effect |
| replace perception with SegDINO3D outputs (ablation) | APEIRIA · ScanRefer Acc@0.25 | accuracy: 58.4 -> 60.4 | Improved |
| replace perception with Mask3D outputs (ablation) | APEIRIA · ScanRefer Acc@0.25 / Multi3DRefer F1@0.25 | accuracy / F1: 58.4 -> 58.5; 59.2 -> 58.3 | Mixed |
| train on Sr3D only and test zero-shot on Nr3D | APEIRIA vs supervised NS3D · Nr3D | accuracy: 33.9 -> 36.5 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Chat-Scene — ChatScene: Knowledge-Enabled Safety-Critical Scenario Generation for Autonomous Vehicles (CVPR 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Struct2D: A Perception-Guided Framework for Spatial Reasoning in MLLMs (NeurIPS 2025 · similarity 0.90)
- EgoMind: Activating Spatial Cognition through Linguistic Reasoning in MLLMs (CVPR 2026 · similarity 0.89)
- Think with 3D: Geometric Imagination Grounded Spatial Reasoning from Limited Views (CVPR 2026 · similarity 0.89)
- S$^2$-MLLM: Boosting Spatial Reasoning Capability of MLLMs for 3D Visual Grounding with Structural Guidance (CVPR 2026 · similarity 0.89)
- Can Large Language Models Understand Symbolic Graphics Programs? (ICLR 2025 · similarity 0.89)
- SpatialLLM: A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models (CVPR 2025 · similarity 0.89)
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.
