Perceptual Time Expansion (PTS) Technology: Enabling Multimodal Reasoning Models to "Think Deliberately"

The latest ICLR 2026 research proposes the Perceptual Time Expansion (PTS) technology. By introducing a computational expansion strategy in the visual perception phase, it shifts the focus of computational expansion from "generating answers" to "understanding problems", significantly improving the performance of large multimodal models on complex reasoning tasks and providing a new idea for balancing reasoning efficiency and quality.

In recent years, large language models have made progress in pure text reasoning through "test-time computational expansion" technology, but in multimodal tasks, the visual perception phase is the key to reasoning quality. When faced with complex visual problems, humans repeatedly observe and focus on details, but current multimodal models lack this "deliberate" perception process, and the traditional fixed visual encoding process is difficult to meet the needs of fine-grained understanding.

The core innovation of the PTS framework lies in shifting the focus of computational expansion from the generation phase to the perception phase. It introduces an extensible computational strategy at the visual encoder level, allowing the model to dynamically adjust the depth of perception according to task complexity, breaking the fixed encoding process, and gradually building a deep understanding of visual content through multiple rounds of iteration.

PTS includes three key components:

1. Adaptive Perception Iteration Mechanism: Determines the number of perception iterations based on input complexity—fewer iterations for simple tasks and multiple rounds of processing for complex tasks;
2. Perception State Caching and Reuse: Reuses low-level visual features and only optimizes high-level semantic representations to control computational overhead;
3. Perception-Reasoning Collaborative Scheduling: Unifies visual perception and text reasoning to form a positive cycle of "perception guides reasoning, reasoning feeds back to perception".

The research team verified the effectiveness of PTS on multimodal reasoning benchmarks such as MathVista and MMMU. The model's performance was significantly improved while maintaining performance on general visual question answering tasks. PTS has excellent computational-performance trade-off characteristics; it can flexibly balance reasoning speed and answer quality by adjusting the number of perception iterations, adapting to different application scenarios.

PTS proves that there is huge optimization space for test-time computational expansion in the visual perception phase, and its modular design is easy to integrate into existing multimodal models. In the future, it can be combined with chain-of-thought technology to improve complex reasoning, with efficient attention mechanisms to reduce computational overhead, and with model compression technology to achieve high-quality reasoning on edge devices.

Building an intelligent multimodal system requires smarter computing methods. PTS enables models to "think deliberately" about the world, which is an important step towards more reliable AI. With the development and popularization of technology, multimodal AI is expected to play a greater role in fields such as scientific research, educational tutoring, and medical diagnosis.