Research map / Vision-language and multimodal / Multimodal LLMs

PointThinker: Point-Incentivized Parallel Thinking for Multimodal Large Language Model

CVPR 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 PointThinker, which makes an MLLM list visual key points each starting an independent reasoning path, trained with GPPO (a point-level and rollout-level verification RL scheme giving dense rewards) on Qwen2.5-VL-7B.

Problem

Parallel thinking for multimodal LLMs suffers from redundant, low-diversity reasoning paths, and sparse outcome rewards (GRPO) assign identical credit to helpful and invalid reasoning points within a chain.

Approach

Require the model to enumerate key points (<p>) each followed by its own reasoning (<r>), then use Group Points Policy Optimization: re-reason each point in isolation to get point-level validity, and only credit tokens of points whose validity agrees with the rollout's outcome.

Key results

• PointThinker-7B reaches 58.7% on HallusionBench (+5.8 over Qwen2.5-VL-7B) and 58.1% average across the 7 image benchmarks, the best in Table 1.
• On video benchmarks PointThinker-7B with 64 frames reaches 62.5% average-like score on VideoMME-style comparisons, vs 61.4 for Video-R1-7B at 64 frames; it reaches 36.4% on VSI-Bench at 16 frames vs 34.6 for Video-R1.
• Ablation on HallusionBench: point structure alone gives 56.8% (+3.9), adding GPPO gives 58.7%; standard parallel thinking gives 55.6%.

Surprising findings

• SFT cold-start alone degraded performance on some benchmarks (e.g., HallusionBench 52.9 to 50.0), so the authors skip SFT and use RL only.

Limitations

• Point-level verification requires extra single-point re-reasoning rollouts (G x K extra generations), adding training cost not quantified.
• Gains on some benchmarks are modest relative to the base model and rely on a single base model family (Qwen2.5-VL).

Experimental setup

Datasets: MM-Eureka WeThink Video-R1-260k

Benchmarks: MMStar MMMU HallusionBench MMVet MathVista MathVision MathVerse VSI-Bench VideoMMMU MMVU MVBench TempCompass

Metrics: accuracy

Baselines: Qwen2.5-VL-7B WeThink-VL-7B Ovis2-8B MUG-U-7B Video-R1-7B LLaVA-OneVision-7B InternVL2.5-8B Qwen2-VL-7B

Builds on: GRPO Parallel-R1 Self-Consistency Easy-R1 Video-R1

Closest papers by meaning

• Parallel-R1: Towards Parallel Thinking via Reinforcement Learning (ICLR 2026 · similarity 0.90)
• Unveiling Chain of Step Reasoning for Vision-Language Models with Fine-grained Rewards (NeurIPS 2025 · similarity 0.89)
• SophiaVL-R1: Reinforcing MLLMs Reasoning with Thinking Reward (ICLR 2026 · similarity 0.88)
• ThreadWeaver: Adaptive Threading for Efficient Parallel Reasoning in Language Models (ICML 2026 · similarity 0.88)
• Prioritize the Process, Not Just the Outcome: Rewarding Latent Thought Trajectories Improves Reasoning in Looped Language Models (ICML 2026 · similarity 0.88)
• More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.88)