RLER: A New Paradigm for Video Reasoning Combining Reinforcement Learning and Evidence Election

This paper proposes the RLER dual-paradigm framework, which uses reinforcement learning to train models to generate structured evidence, then selects reliable answers via a training-free evidence weighted election mechanism. It achieves SOTA on 8 video reasoning benchmarks, with an average improvement of 6.3% and only requiring 3.1 candidates.

Video reasoning requires understanding visual content and performing temporal reasoning and causal inference. While large multimodal models (LMMs) bring new hope, existing methods lack an evidence verification mechanism, leading to a disconnect between reasoning and evidence, which causes risks of hallucinations, non-interpretability, and fragility.

The RLER framework decomposes video reasoning into two independent stages: RLER-Training (generating structured evidence via reinforcement learning) and RLER-Inference (obtaining answers through evidence election). The decoupled design allows training to focus on improving evidence quality and reasoning to focus on using evidence for decision-making, jointly enhancing the system's reliability and interpretability.

The training stage uses population-based relative reinforcement learning and designs three task-driven reward functions: frame-sensitive reward (anchoring key frames), transparent thinking reward (structured reasoning trajectory), and anti-redundancy reward (improving information density). It can be applied on a large scale without manual annotation of reasoning processes.

The inference stage completes four steps via an orchestrator: generating 3.1 diverse candidates, parsing candidate answers and referenced frames, scoring from four dimensions (evidence consistency, confidence, transparency, non-redundancy), and selecting the most reliable answer through weighted election to reduce misjudgments.

RLER achieves SOTA on 8 video reasoning benchmarks with an average improvement of 6.3%; the average number of candidates is only 3.1, with a computational overhead increase of about 2x but a significant performance improvement; the generated reasoning trajectory has high transparency and strong verifiability, making it suitable for high-risk scenarios.

Insights from RLER: Explicit evidence improves interpretability and accuracy; performance can be enhanced without expanding model parameters; the collaborative design of training and inference stages forms a closed loop to tap the potential of existing models.

Current limitations: Only focuses on question-answering tasks, and real-time scenarios need optimization; future directions: exploring efficient candidate generation, cross-video evidence transfer, and expanding to multi-modal reasoning such as audio-text.