# Distill What RGB Can Recover: Privileged 3D Evidence for RGB-Only Vision-Language Models

arXiv: https://arxiv.org/abs/2608.00110 · HTML: https://mlsanity.com/paper/2608.00110/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Embodied and 3D vision-language models, one of 16 directions in Multimodal LLMs / vision-language models.
- Builds on 27 of its 33 references that are in the map.
- First posted on arXiv 31 Jul 2026.

## Placed in

- Multimodal LLMs / vision-language models / [Embodied and 3D vision-language models](https://mlsanity.com/mllm/embodied-and-3d-vision-language-models/)
- Knowledge distillation / [Detection and dense prediction distillation](https://mlsanity.com/distill/detection-and-dense-prediction-distillation/)

## Builds on (most cited first, 27 of 27)

- [Distilling the Knowledge in a Neural Network](https://mlsanity.com/paper/1503.02531/) (2015)
- [LLaVA-Video: Video Instruction Tuning With Synthetic Data](https://mlsanity.com/paper/2410.02713/) (2024)
- [ScanQA: 3D Question Answering for Spatial Scene Understanding](https://mlsanity.com/paper/2112.10482/) (2021)
- [SQA3D: Situated Question Answering in 3D Scenes](https://mlsanity.com/paper/2210.07474/) (2022)
- [Unifying distillation and privileged information](https://mlsanity.com/paper/1511.03643/) (2015)
- [Scan2Cap: Context-aware Dense Captioning in RGB-D Scans](https://mlsanity.com/paper/2012.02206/) (2020)
- [LL3DA: Visual Interactive Instruction Tuning for Omni-3D Understanding, Reasoning, and Planning](https://mlsanity.com/paper/2311.18651/) (2023)
- [3D-VisTA: Pre-trained Transformer for 3D Vision and Text Alignment](https://mlsanity.com/paper/2308.04352/) (2023)
- [LLaVA-3D: A Simple yet Effective Pathway to Empowering LMMs with 3D-awareness](https://mlsanity.com/paper/2409.18125/) (2024)
- [Chat-Scene: Bridging 3D Scene and Large Language Models with Object Identifiers](https://mlsanity.com/paper/2312.08168/) (2023)
- [VLM-3R: Vision-Language Models Augmented with Instruction-Aligned 3D Reconstruction](https://mlsanity.com/paper/2505.20279/) (2025)
- [Bottom Up Top Down Detection Transformers for Language Grounding in Images and Point Clouds](https://mlsanity.com/paper/2112.08879/) (2021)
- [Modality Distillation with Multiple Stream Networks for Action Recognition](https://mlsanity.com/paper/1806.07110/) (2018)
- [End-to-End 3D Dense Captioning with Vote2Cap-DETR](https://mlsanity.com/paper/2301.02508/) (2023)
- [3DRS: MLLMs Need 3D-Aware Representation Supervision for Scene Understanding](https://mlsanity.com/paper/2506.01946/) (2025)
- [3D-LLaVA: Towards Generalist 3D LMMs with Omni Superpoint Transformer](https://mlsanity.com/paper/2501.01163/) (2025)
- [How Large Language Models Encode Context Knowledge? A Layer-Wise Probing Study](https://mlsanity.com/paper/2402.16061/) (2024)
- [3D-to-2D Distillation for Indoor Scene Parsing](https://mlsanity.com/paper/2104.02243/) (2021)
- [3D-GRAND: A Million-Scale Dataset for 3D-LLMs with Better Grounding and Less Hallucination](https://mlsanity.com/paper/2406.05132/) (2024)
- [Graph Distillation for Action Detection with Privileged Modalities](https://mlsanity.com/paper/1712.00108/) (2017)
- [Learning an Augmented RGB Representation with Cross-Modal Knowledge Distillation for Action Detection](https://mlsanity.com/paper/2108.03619/) (2021)
- [SpatialStack: Layered Geometry-Language Fusion for 3D VLM Spatial Reasoning](https://mlsanity.com/paper/2603.27437/) (2026)
- [Generation Models Know Space: Unleashing Implicit 3D Priors for Scene Understanding](https://mlsanity.com/paper/2603.19235/) (2026)
- [CVP: Central-Peripheral Vision-Inspired Multimodal Model for Spatial Reasoning](https://mlsanity.com/paper/2512.08135/) (2025)
- [Attentive multilayer fusion for vision transformers](https://mlsanity.com/paper/2601.09322/) (2026)
- [Beyond 3D VQAs: Injecting 3D Spatial Priors into Vision-Language Models for Enhanced Geometric Reasoning](https://mlsanity.com/paper/2605.30231/) (2026)
- [PAR3D: A Unified 3D-MLLM with Part-Aware Representation for Scene Understanding](https://mlsanity.com/paper/2606.06485/) (2026)

## Abstract (arXiv)

3D scene understanding requires reasoning about entity existence, spatial layout, and object relations, yet RGB images alone often provide insufficient 3D cues. Existing 3D-VLMs commonly rely on depth or 3D-position-aware inputs at inference time, introducing additional acquisition, reconstruction, or annotation costs that limit RGB-only deployment. We therefore study how training-time 3D evidence can be converted into spatial reasoning capabilities retained under RGB-only inference. We propose a privileged-evidence distillation framework that constructs a distillable teacher through a unified evidence interface and controlled residual injection, and transfers its knowledge to a deployable student receiving only RGB images and questions through logit and structured representation distillation. To avoid imitating teacher signals unsupported by RGB, we further introduce evidence-sensitivity-guided distillation, which uses corrupted evidence to identify highly evidence-dependent targets and down-weight their supervision. We also define a recoverability decomposition based on the matched baseline, teacher, and student, separating privileged gains into RGB-recoverable improvements and residual teacher advantages. Across four benchmarks, the teacher achieves the best result on 7 of 11 reported metrics among the compared methods. The RGB-only student outperforms its matched baseline on all 11 metrics, including gains of 10.4 ScanQA CIDEr and 19.1 Scan2Cap CIDEr@0.5, without additional inference-time inputs. These results validate the effectiveness of training-time privileged 3D evidence distillation for both teacher performance and deployable RGB-only spatial reasoning. Separately, our matched baseline-teacher-student analysis characterizes privileged-gain transfer across evidence types and spatial skills.