Research map / Vision-language and multimodal / Multimodal LLMs

Discriminative Perception via Anchored Description for Reasoning Segmentation

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 0 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.

Proposes DPAD, a discriminative reward for RL-trained reasoning segmentation MLLMs that scores whether reasoning stays anchored on the referred region, using a frozen CLIP scorer, to prune verbose reasoning while improving ReasonSeg accuracy.

Problem

Reasoning segmentation RL uses geometric (mask/IoU) rewards that only guide final localization and cannot tell whether the reasoning chain stays anchored on the referred region or drifts into irrelevant context.

Approach

Adds an anchored-description discriminative reward: a frozen CLIP model scores how well the generated reasoning matches the referred region (cropped by GT box) relative to the wider context (full image), giving a binary/margin reward that suppresses verbose, off-target reasoning and improves segmentation.

Key results

• On ReasonSeg val, DPAD reaches 61.2 cIoU vs 57.4 for base, while reducing reasoning tokens from ~117.5 to ~68.5.
• On ReasonSeg test, DPAD reaches 57.5 cIoU vs 54.4 for base.
• On MMR multi-target, gIoU improves from 30.22 to 31.28 with fewer tokens (79.04 to 75.42).

Surprising findings

• A simple length penalty reduces tokens to about 36 without improving val cIoU (57.4 to 57.2), suggesting DPAD gains are not from mere truncation.

Limitations

• Gains are smaller on RefCOCO with explicit queries, where the baseline is already accurate (ceiling effect).
• Limited token compression and gains on part-level queries, attributed to object-level granularity.
• Reward weight set to default 1.0 without parameter search.

Experimental setup

Datasets: ReasonSeg RefCOCO MMR

Benchmarks: ReasonSeg RefCOCO MMR

Metrics: cIoU gIoU token count

Baselines: Base (geometric-reward RL) Base + length penalty

Builds on: GRPO

Closest papers by meaning

• DR$^2$Seg: Decomposed Two-Stage Rollouts for Efficient Reasoning Segmentation in Multimodal Large Language Models (ICML 2026 · similarity 0.89)
• CoPRS: Learning Positional Prior from Chain-of-Thought for Reasoning Segmentation (ICLR 2026 · similarity 0.89)
• LISA: Reasoning Segmentation via Large Language Model (CVPR 2024 · similarity 0.88)
• VisuRiddles: Fine-grained Perception is a Primary Bottleneck for Multimodal Large Language Models in Abstract Visual Reasoning (ICLR 2026 · similarity 0.88)
• Activating Visual Context and Commonsense Reasoning Through Masked Prediction in VLMs (AAAI 2026 · similarity 0.88)
• VLM-R³: Region Recognition, Reasoning, and Refinement for Enhanced Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.88)