Alibaba's Qwen Team Builds HopChain to Fix How Vision Models Fall Apart During Multi-Step Reasoning
Alibaba's Qwen team and Tsinghua University have released HopChain, a training framework that forces vision-language models to verify intermediate reasoning steps before proceeding. The result: improvements on 20 of 24 benchmarks tested, with some scores more than doubling on hard visual reasoning tasks.
D.O.T.S AI Newsroom
AI News Desk
Vision-language models have a well-documented weakness: when a task requires more than one or two reasoning steps over an image, small errors compound rapidly. Miscounting objects, misreading spatial relationships, or hallucinating a detail early in the reasoning chain contaminates every conclusion that follows. Alibaba's Qwen team, working with researchers from Tsinghua University, has published a framework called HopChain that directly targets this failure mode — and the benchmark results are notable.
What HopChain Does
HopChain automatically generates multi-stage visual questions that chain 3 to 6 objects together in dependency sequences. A model cannot answer the final question without correctly resolving each intermediate step. The pipeline runs through four stages: Alibaba's Qwen3-VL model first identifies object categories; Meta's SAM3 segments individual instances; the language model generates multi-level questions with linked dependencies; and four human annotators independently verify every answer. Only questions with unanimous human agreement enter the training set.
The forcing function is simple but effective: each step in a HopChain question requires the model to return to the image and visually verify a claim before proceeding. Models trained on this data learn to treat visual verification as mandatory rather than optional.
Benchmark Results
HopChain improved 20 of 24 benchmarks across STEM, document comprehension, and video reasoning tasks. On the 35B model, EMMA rose from 53 to 58 and CharXiv from 69 to 73.1. On the 397B model, ZeroBench — a particularly hard visual reasoning benchmark — doubled from 4 to 8 points. The framework generalized to video tasks despite being trained on still images, suggesting it captures something structural about multi-step visual attention rather than dataset-specific patterns.
Full reasoning chains proved essential — partial chains degraded performance. The benefits scaled with chain length, reaching 50+ point improvements for the longest responses.
Why This Matters
Multi-step visual reasoning is the gateway to reliable AI for domains like medical imaging, scientific analysis, and robotics. Today's models are brittle exactly where precision is most required. HopChain's approach — building training data that enforces step-by-step verification — is a data-centric solution to a problem that extra model capacity alone has not solved. The code and training data are expected to be released alongside the paper.
