What happened
The paper, *Mixture-of-Visual-Thoughts: Exploring Context-Adaptive Reasoning Mode Selection for General Visual Reasoning*, introduces a paradigm called Mixture-of-Visual-Thoughts (MoVT) and a two-stage training framework called AdaVaR. Its authors are affiliated with Fudan University, Alibaba Group Holding Limited, the University of Southern California, and Shanghai Innovation Institute. The arXiv page identifies the paper as version 2, dated 14 May 2026. The researchers start from Qwen2.5-VL and focus on two forms of visual reasoning: text-based reasoning, expressed in natural language, and visually grounded reasoning, which aligns concepts with image regions using structured information such as bounding boxes.
AdaVaR first combines these modes in a supervised fine-tuning cold-start stage. Each reasoning path begins with a mode-specific prefix, allowing one model to learn both formats in a shared sequence representation. The authors say they balance the two data sources at a 1:1 ratio. Text-based trajectories are constructed by distilling a text-based visual reasoning model and applying rejection sampling; grounded trajectories use existing high-quality supervised data. The paper argues that supervised examples alone cannot reliably teach which mode to choose for a particular question, because the preferred mode depends on the question and image.
The second stage applies reinforcement learning with the authors’ AdaGRPO algorithm. For each sample, prefix-guided exploration makes the policy generate rollouts in multiple modes rather than relying on a single mode. AdaGRPO combines rollout-level advantages with a mode-relative advantage, intended to reward the mode more likely to produce the correct answer. A curriculum-based schedule begins with easier data and moves to harder questions. The model generates a mode prefix and then its reasoning and answer in one autoregressive sequence.
The paper reports that AdaVaR-3B matches Qwen2.5-VL-7B on average across eight benchmarks, while AdaVaR-7B surpasses GPT-4o on that average. The authors also report consistent improvements across scenarios. The excerpt describes an evaluation setup involving benchmarks including CHAIR, POPE, AMBER, and GPT-4 assessment in parts of the paper; the reported headline comparison is an average over eight benchmarks. The authors say code, models, and data will be open-sourced. The paper does not present this as a priced or hosted service.
What the archive adds
New Runtime's earlier coverage of C3PO reported a paper aiming to reduce hallucinations in multimodal reasoning models, with results across five 7B models and multiple benchmarks (the earlier note, 27 October 2026). This paper is a new event in the broader research thread: rather than focusing on hallucination reduction, it trains a model to choose between complementary reasoning modes.
Trend signal
This supports the idea that model answers are becoming task interfaces: the model can select a bounded response mode for the task rather than using one general format. That connection is plausible but early; this paper describes a research method, not a deployed interface (Model answers become task interfaces).
Why now
The paper frames its work against recent progress in language-model reasoning and reinforcement learning, and against visual reasoning methods that tend to specialize in one mode or a narrower task domain. Its proposed response is to train for mode selection as a capability, rather than assume that more reasoning in a single format will generalize across visual questions.
Why it matters
For teams building multimodal systems, the useful idea is not simply to add longer chains of thought. Text-based reasoning and visual grounding carry different biases: the former can help with abstract questions, while the latter can anchor claims to image regions. A model that learns when to use each could avoid forcing every task through the same reasoning path.
The paper also makes the training objective more explicit: evaluate alternative modes on the same input and use their relative outcomes to teach selection. That is a research direction, not yet a recipe teams can assume will transfer to their data. Still, builders can use the distinction to design evaluations that separate abstract reasoning from object- or region-centered questions, then check whether a model's chosen approach improves accuracy without increasing unsupported claims.
Limits and uncertainty
The reported comparisons come from the authors' experiments, and this account has not independently reproduced them. The headline claims depend on the paper's benchmark selection, data construction, implementation, and evaluation setup; the reported average does not establish performance on every visual task or in production. The paper says code, models, and data will be released, but those releases and their practical usability are not established here. It also does not demonstrate that the learned mode selection transfers to arbitrary domains or user-provided datasets.
Reader takeaway
AdaVaR treats visual reasoning style as something a model can learn to choose, not a fixed property of the system. Its benchmark results are promising research claims, but the practical test for builders is whether mode selection improves performance on their own mix of abstract and visually grounded tasks.