How Laser Cuts Token Use by 97% with Probabilistic Superposition for Implicit Multimodal Reasoning

1. The Bottleneck of Explicit Chain‑of‑Thought in Multimodal Models

Chain‑of‑Thought (CoT) techniques have boosted multimodal large model (VLM) reasoning, but relying on explicit textual tokens creates an "information bandwidth bottleneck" that discards rich visual details and introduces language priors that cause hallucinations.

2. Laser: Latent Superposition for Efficient Visual Reasoning

The research team proposes Laser (Latent Superposition for Effective Visual Reasoning), inspired by the "Forest‑before‑Trees" cognitive mechanism. Its core innovation is Dynamic Windowed Alignment Learning (DWAL) , which abandons point‑wise next‑token prediction in favor of aligning the latent state with a dynamic semantic window that contains future potential tokens.

Dynamic Semantic Window : At each reasoning step the latent state is forced to cover all valid semantics inside the window, which gradually shrinks as reasoning proceeds, mirroring human global‑first then local‑focus perception.

Self‑Refined Superposition : The model extracts its own next‑token distribution, smooths it with a temperature‑scaled Softmax, and uses this distribution as a soft target, allowing the latent state to maintain a probabilistic superposition of multiple future tokens.

Entropy‑Regularized Intervention : Normalized entropy of the soft target measures uncertainty. When entropy exceeds a threshold η, a hard label weighted by α is injected; otherwise the model continues with the soft superposition, forming an implicit curriculum.

3. ScanPath: A Cognitive Trajectory Dataset

To train DWAL, the authors built the ScanPath dataset (≈270 k samples) that follows a strict "global‑to‑local" scanning logic, provides atomic visual concepts without grammatical fillers, and achieves 91.5% logical validity in human evaluation.

4. Optimization Objective

The overall loss combines the DWAL superposition loss with the standard cross‑entropy loss for the final answer generation, balancing global visual semantics and precise local grounding.

5. Experimental Results

Laser was evaluated on six challenging visual benchmarks. Compared with implicit‑reasoning baselines, Laser improves average performance by 5.03%, with 11.36% gain on HallusionBench and 6.21% on BLINK. Token consumption drops by more than 97%; on BLINK the average token count falls to 6.0, far below explicit‑CoT methods.

6. Ablation Studies

Probability Superposition : Removing DWAL’s target collapses performance on fine‑grained benchmarks, confirming its role in preventing premature semantic collapse.

Dynamic Window : Fixing the window size harms results on complex logical tasks (e.g., MMStar), demonstrating the necessity of progressive window shrinkage for the "Forest‑before‑Trees" hierarchy.

Entropy Threshold : An η of 0.6 yields the best trade‑off, triggering hard interventions in ~10% of tokens; lower η (0.5) causes over‑intervention and performance drop, while higher η (0.8–1.0) leads to under‑intervention and degraded reasoning.

7. Interpretability

Because the latent state retains a superposition of semantic tokens, it can be projected onto the language model vocabulary, allowing researchers to visualize the model’s internal "cognitive trajectory".

8. Conclusion

Laser demonstrates that moving reasoning from discrete text to a compact latent space with probabilistic superposition and dynamic alignment yields both high efficiency (97% token reduction) and strong performance, offering a new direction for multimodal large‑model research.