Research map / Vision-language and multimodal / Multimodal LLMs

SeqAfford: Sequential 3D Affordance Reasoning via Multimodal Large Language Model

CVPR 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Instruction-tuned multimodal assistants · 5 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces Sequential 3D Affordance Reasoning and a 180K instruction-point-cloud benchmark, and proposes SeqAfford, which adapts the ShapeLLM 3D MLLM with LoRA and a multi-granular language-point integration module to output sequences of affordance masks.

Problem

Existing 3D affordance segmentation follows a single-object, single-affordance paradigm and cannot reason over complex, long-horizon instructions that imply sequences of affordances across multiple objects.

Approach

Unlock a pretrained 3D multimodal LLM with <SEG> tokens whose embeddings are fused with multi-granular dense point features via cross-attention, so the LLM decomposes an instruction into an ordered series of affordance masks.

Key results

• On the seen single-affordance split, the model reaches 19.5 mIoU and 86.9 AUC versus 16.3 mIoU and 84.3 AUC for PointRefer, the strongest baseline.
• On the unseen single-affordance split, the model reaches 13.8 mIoU and 82.4 AUC, versus 12.4 mIoU and 76.1 AUC for PointRefer.
• On the sequential split, the model reaches 14.6 mIoU and 84.2 AUC, versus 14.3 mIoU and 80.7 AUC for PointRefer given ground-truth sequential order.

Measured effects reported in the paper

Limitations

• Baselines on the sequential task were given ground-truth sequential order, so the comparison is not like-for-like with the model's predicted order.
• Sequential order for multi-object cases is derived by extracting nouns from the generated text, a heuristic not validated in the paper.
• Code and dataset release is promised only after acceptance.

Experimental setup

Datasets: 3D AffordanceNet PartNet IAGNet HOI images LASO

Benchmarks: SeqAfford single-affordance seen split SeqAfford single-affordance unseen split SeqAfford sequential split

Metrics: mIoU AUC SIM MAE

Baselines: ReferTrans ReLA 3D-SPS IAGNet PointRefer LASO

Builds on: ShapeLLM LASO 3D AffordanceNet LLaVA PointNet++

Closest papers by meaning

• 3D-AffordanceLLM: Harnessing Large Language Models for Open-Vocabulary Affordance Detection in 3D Worlds (ICLR 2025 · similarity 0.93)
• AffordBot: 3D Fine-grained Embodied Reasoning via Multimodal Large Language Models (NeurIPS 2025 · similarity 0.93)
• Grounding 3D Object Affordance with Language Instructions, Visual Observations and Interactions (CVPR 2025 · similarity 0.91)
• UniFunc3D: Unified Active Spatial-Temporal Grounding for 3D Affordance Segmentation (NeurIPS 2026 · similarity 0.90)
• GEAL: Generalizable 3D Affordance Learning with Cross-Modal Consistency (CVPR 2025 · similarity 0.90)
• GREAT: Geometry-Intention Collaborative Inference for Open-Vocabulary 3D Object Affordance Grounding (CVPR 2025 · similarity 0.90)