Research map / Vision-language and multimodal / Multimodal LLMs

CoPRS: Learning Positional Prior from Chain-of-Thought for Reasoning Segmentation

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning

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.

Proposes CoPRS, an end-to-end MLLM that reasons with multimodal chain-of-thought, emits a concentration token whose query produces a heatmap positional prior, and decodes it to masks, trained jointly with GRPO and segmentation losses.

Problem

Reasoning segmentation either decodes hidden LLM features directly into masks (opaque) or emits discrete text coordinates (coarse, format-fragile); a better interpretable and fine-grained interface between reasoning and segmentation is needed.

Approach

A learnable concentration token after MCoT yields a query that attends over SAM image keys to produce a dense heatmap positional prior, which a lightweight SAM-style decoder turns into a mask; GRPO on the language path is trained jointly with BCE/dice/focal segmentation losses in one backward pass.

Key results

• CoPRS-7B reaches 81.6 / 85.3 / 79.5 cIoU on RefCOCO val/testA/testB and trails RAS-13B on only 2 of 8 RefCOCO-series splits.
• Zero-shot on ReasonSeg, CoPRS-7B gets 65.2 gIoU / 64.5 cIoU on val and 59.8 gIoU / 55.1 cIoU on test, the best among compared methods.
• Correlation analysis reports R > 0.7 between heatmap and mask quality across training and test splits.

Surprising findings

• CoPRS-3B surpasses the 7B Seg-R1 and Seg-Zero variants on RefCOCO-series splits.

Limitations

• Fails on very small objects due to 256x256 prior and 64x64 SAM embedding resolution.
• Struggles with dense groups of similar instances where text alone cannot disambiguate.
• Correlation analysis between CoT quality and masks relies on an LLM judge (Gemini-2.5-Flash) with no human validation reported.

Experimental setup

Datasets: RefCOCO RefCOCO+ RefCOCOg

Benchmarks: RefCOCO RefCOCO+ RefCOCOg ReasonSeg

Metrics: cIoU gIoU

Baselines: LISA PerceptionGPT PixelLM LaSagnA SegLLM OMG-LLaVA GroundHog GLaMM RAS SAM4MLLM Seg-R1 Seg-Zero

Builds on: GRPO LISA Seg-Zero SAM Qwen2.5-VL

Closest papers by meaning

• LISA: Reasoning Segmentation via Large Language Model (CVPR 2024 · similarity 0.89)
• Discriminative Perception via Anchored Description for Reasoning Segmentation (CVPR 2026 · similarity 0.89)
• SegCompass: Exploring Interpretable Alignment with Sparse Autoencoders for Enhanced Reasoning Segmentation (CVPR 2026 · similarity 0.88)
• DR$^2$Seg: Decomposed Two-Stage Rollouts for Efficient Reasoning Segmentation in Multimodal Large Language Models (ICML 2026 · similarity 0.88)
• SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.88)
• To Think or Not to Think: Pre-Decisional Reasoning Budgets for Referring Audio-Visual Segmentation (NeurIPS 2026 · similarity 0.88)