Research map / Vision-language and multimodal / Multimodal LLMs

HopChain: Multi-Hop Data Synthesis for Generalizable Vision-Language Reasoning

NeurIPS 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.

Synthesizes multi-hop, instance-grounded vision-language queries with a four-stage pipeline (VLM categories, SAM3 segmentation, VLM query generation, human verification) and adds them to RLVR data, improving Qwen3.5 models on 20 of 24 benchmarks.

Problem

VLMs struggle with fine-grained multi-step vision-language reasoning, and existing RLVR data does not require long reasoning chains that depend on visual evidence at every step, so long-CoT errors (perception, reasoning, knowledge, hallucination) go unexposed during training.

Approach

Generate benchmark-agnostic multi-hop queries over segmented instances in real images, where each hop depends on earlier hops (perception-level and instance-chain hops) and ends in a single verifiable numerical answer, and mix them into standard RLVR data with SAPO.

Key results

• RLVR w/ Multi-Hop improves 20 of 24 benchmarks over RLVR w/o Multi-Hop on both Qwen3.5-35B-A3B and Qwen3.5-397B-A17B.
• On five representative benchmarks, average score drops from 70.4 (full multi-hop) to 66.7 (half-multi-hop) and 64.3 (single-hop) on Qwen3.5-35B-A3B.
• Multi-hop gains exceed 50 points in the ultra-long-CoT response regime on Qwen3.5-397B-A17B, per the paper's analysis.

Surprising findings

• Multi-hop image-derived training transfers to video understanding, improving 5 of 6 video benchmarks on both model scales.
• Gains are reported on benchmarks not targeted by the synthesized data, which is benchmark-agnostic.

Limitations

• Pipeline depends on successful SAM3 instance segmentation, so images without detectable objects are excluded.
• Relies on internal pre-final SFT and RLVR data, limiting reproducibility.
• Some benchmarks show regressions (e.g., HallusionBench, LVBench on 35B model, MMBench-EN on 397B).

Experimental setup

Datasets: original RLVR data (internal, pre-final) math RLVR data HopChain multi-hop data synthesized from filtered real images

Benchmarks: MathVision MMMU Pro MMMU MathVista BabyVision ZeroBench EMMA LogicVista MMBench-CN MMBench-EN RealWorldQA MMStar

Metrics: accuracy

Baselines: RLVR w/o Multi-Hop RLVR w/ Single Hop RLVR w/ Half-Multi-Hop Before RLVR

Builds on: SAPO GSPO GRPO SAM3 Qwen3.5

Closest papers by meaning

• Unveiling Chain of Step Reasoning for Vision-Language Models with Fine-grained Rewards (NeurIPS 2025 · similarity 0.92)
• Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.91)
• Latent Chain-of-Thought for Visual Reasoning (NeurIPS 2025 · similarity 0.91)
• OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles (NeurIPS 2025 · similarity 0.91)
• ReaGEN: Adaptive Generation of Structured Chains-of-Thought for Efficient Multimodal Reasoning (CVPR 2026 · similarity 0.91)
• Revisiting the Necessity of Lengthy Chain-of-Thought in Vision-centric Reasoning Generalization (CVPR 2026 · similarity 0.91)