Chains That See, Answers That Don't: A Multi-Aspect Evaluation Recipe for Forced Chain-of-Thought on Video-MME

We present an efficient method to distill reasoning capabilities into compact video-language models (VLMs) for video question answering (VideoQA). Our approach fine-tunes a 2B-parameter model using only $\sim$900 uncertainty-selected examples, each augmented with synthetic chain-of-thought (CoT) rationales generated by a 4B teacher. Despite its minimal compute cost - under two hours on a single A100 GPU - our method enables the 2B model to outperform VLMs up to 4$\times$ larger, and generalize across CinePile, ActivityNet-QA, and MLVU, approaching the performance of its own 4B teacher. A key finding is that placing CoT rationales after the answer - contrary to standard prompting - substantially improves reasoning in compact models. This insight challenges prevailing CoT conventions and reveals new alignment strategies under limited model capacity. Our findings offer a practical blueprint for training deployable, reasoning-rich VLMs suited for mobile and edge applications.

Recent advancements in chain-of-thought (CoT) reasoning have shown promise in enhancing video understanding and reasoning capabilities of multimodal large language models (MLLMs). However, existing CoT-based MLLMs require labor-intensive CoT annotations and incur substantial training and inference overhead. While visual latent reasoning has emerged as a more efficient alternative, existing methods primarily focus on image tasks and heavily rely on additional supervision signals for visual latent generation (e.g., CoT traces, auxiliary images, or fine-grained annotations), limiting their scalability and transferability to video tasks. To bridge this gap, we introduce VideoLatent, a novel MLLM equipped with a latent injection module tailored for video understanding and reasoning. Specifically, VideoLatent learns to perform visual latent reasoning using a new latent self-forcing training paradigm, which comprises latent alignment and latent diversity objectives, and relies solely on standard video-question-answer triplets. Extensive experiments across 14 benchmarks demonstrate that our model consistently outperforms existing standard and latent MLLMs on general video understanding and complex video reasoning. Compared with Video-R1, our VideoLatent achieves superior computational efficiency, reducing training/inference overhead by $\sim$6$\times$/$\sim$68$\times$. Moreover, experiments demonstrate that our method has strong generalizability to different MLLM backbones and different model scales.

HopChain has shown on still images that multi-hop data synthesis improves vision-language reasoning, because long chain-of-thought reasoning exposes errors that compound across steps, while most data used for reinforcement learning with verifiable rewards (RLVR) rarely demands a chain of visual evidence, so these weaknesses are likely to stay unexposed. We observe the same problem in video, where this framework has not yet been explored. We therefore build Video-HopChain, a dataset of 22,550 multi-hop video questions over 13,378 videos, together with a held-out benchmark of 1,000 questions. Each question chains three to six yes/no questions about moments in one video, and each yields one of two integers depending on its answer. The final answer is the sum of these integers, so an exact match on that sum gives the verifiable reward that RLVR needs. We first train Qwen3-VL-8B with GRPO on a standard video dataset, and a second stage on Video-HopChain then raises the mean over eight video understanding and reasoning benchmarks from 55.4 to 57.9 and improves every one of them. Training on such a dataset, however, exposes a known limitation of GRPO: its learning signal comes from the reward variance within a group, so hard questions whose rollouts are all incorrect and easy questions whose rollouts are all correct both leave the group with no gradient. To recover these groups at the same compute budget, we introduce Confidence-Gated Exploration (CGE). With 8 rollouts per question, CGE samples the first 4 as usual. If these 4 are either all correct or all incorrect, it samples the last 4 with the policy's most confident token masked inside the reasoning span, and removes the masked positions from the loss while all 8 rollouts enter the advantage. With CGE, the mean rises further to 59.3. We release the dataset, the checkpoint, and the data generation and training code.