Conan: Progressive Learning to Reason Like a Detective over Multi-Scale Visual Evidence

TL;DR

Conan addresses the challenge of multi-step video reasoning by grounding reasoning in explicitly identified visual evidence across frames. It introduces Conan-91k, a dataset of Conan-style traces (frame identification, evidence reasoning, action decision) built from GenS-Video-150K, and a two-phase training curriculum comprising a multi-stage progressive cold-start and an AIR RLVR framework with carefully designed rewards. Experiments on six multi-step benchmarks and four long-video benchmarks show Conan consistently outperforms the base model and many state-of-the-art baselines, demonstrating strong generalization to long videos. The combination of multi-scale evidence, progressive learning, and reinforcement learning rewards yields robust, grounded, and scalable video reasoning capabilities with practical implications for complex real-world reasoning over video streams.

Abstract

Video reasoning, which requires multi-step deduction across frames, remains a major challenge for multimodal large language models (MLLMs). While reinforcement learning (RL)-based methods enhance reasoning capabilities, they often rely on text-only chains that yield ungrounded or hallucinated conclusions. Conversely, frame-retrieval approaches introduce visual grounding, yet still struggle with inaccurate evidence localization. To address these limitations, we present Conan, a framework for evidence-grounded multi-step video reasoning. Conan identifies context and evidence frames, reasons over cross-frame clues, and adaptively decides when to conclude or explore further. To achieve this, we 1) construct Conan-91K, a large-scale dataset of automatically generated reasoning traces that include frame identification, evidence reasoning, and action decision, and 2) design a multi-stage progressive cold-start strategy combined with an Identification-Reasoning-Action (AIR) RLVR training framework to progressively incentivize multi-step visual reasoning. Extensive experiments on six multi-step reasoning benchmarks demonstrate that Conan surpasses the baseline Qwen2.5-VL-7B-Instruct by an average of over 10% in accuracy, achieving state-of-the-art performance. Furthermore, Conan generalizes effectively to long video understanding tasks, validating its strong scalability and robustness.

Figures (11)

• Figure 1: Top: The reasoning process of our Conan on an example question. Bottom: Performance comparison across six multi-step reasoning benchmarks.
• Figure 2: a) Reasoning Trace Construction. b) Data Example. c) Multi-stage Progressive Cold-start, including textual, multimodal alignment, and vision-centric reasoning stages. d) The Joint Identification-Reasoning-Action RLVR.
• Figure 3: Performance comparison across different evidence retrievers, with Qwen2.5-VL-7B-Instruct as the question-answering model.
• Figure 4: A qualitative example from VRBench, including the QA pair and the reasoning traces of Video-R1 (Text CoT), Video-MTR (Video CoT), and Conan.
• Figure 5: Prompt templates for reasoning trace construction.