RUC and Kling AI propose EventLens: event-structure RL to fix video LLMs' temporal understanding

Guowen Zhang · hf · 2026-10-14

Researchers from Renmin University's AIM3 Lab and Kuaishou's Kling AI released EventLens, a vision-centric reinforcement learning framework targeting a known weakness of multimodal LLMs in video QA: models merge neighboring events, miss single-event progression, or infer implausible cross-event relations.

Key ideas

The authors attribute the gap to a mismatch between post-training objectives and temporal reasoning—caption-style supervision lets models exploit language shortcuts instead of grounding temporal structure visually.
EventLens trains models to recover event structure under controlled perturbations (altered temporal granularity, reversed progression, shuffled event order) rather than reconstruct captions.

Method

Built on a three-level temporal ontology, it derives three verifiable task families—event segmentation, progression discrimination, and multi-event relation reasoning—with deterministic rewards, enabling task-specialized GRPO training without human preference labels or learned reward models.
A multi-teacher on-policy distillation (MT-OPD) stage consolidates the specialized policies into one unified model.

Results

On Qwen3-VL backbones, EventLens yields consistent gains on temporal grounding and event-sensitive reasoning benchmarks while preserving general video-QA performance, with notably lower training cost than direct mixed RL optimization.

Original post →

RUC and Kling AI propose EventLens: event-structure RL to fix video LLMs' temporal understanding

More from Multimodal