Forest Before Trees: Latent Superposition for Efficient Visual Reasoning

TL;DR

Laser reframes visual reasoning as latent-space dynamic windowed alignment, replacing point-wise next-token prediction with a probabilistic superposition over future semantic concepts. The Dynamic Windowed Alignment Learning (DWAL) framework maintains global semantic representations through a shrinking validity window $\mathcal{W}_t = \{ c_k \mid t \le k \le T \}$ and uses Self-Refined Superposition with an Entropy-Regularized Intervention to stabilize learning. A synthetic ScanPath dataset is generated via a Global-to-Local prompting strategy to train the latent reasoning trajectories without explicit ROI supervision. Empirically, Laser achieves state-of-the-art results among latent-methods across 6 benchmarks, with extreme inference efficiency (average tokens $\approx 6.0$ on BLINK, a $97\%$ reduction) and strong generalization to out-of-distribution domains, while preserving interpretability through decodable latent trajectories.

Abstract

While Chain-of-Thought empowers Large Vision-Language Models with multi-step reasoning, explicit textual rationales suffer from an information bandwidth bottleneck, where continuous visual details are discarded during discrete tokenization. Recent latent reasoning methods attempt to address this challenge, but often fall prey to premature semantic collapse due to rigid autoregressive objectives. In this paper, we propose Laser, a novel paradigm that reformulates visual deduction via Dynamic Windowed Alignment Learning (DWAL). Instead of forcing a point-wise prediction, Laser aligns the latent state with a dynamic validity window of future semantics. This mechanism enforces a "Forest-before-Trees" cognitive hierarchy, enabling the model to maintain a probabilistic superposition of global features before narrowing down to local details. Crucially, Laser maintains interpretability via decodable trajectories while stabilizing unconstrained learning via Self-Refined Superposition. Extensive experiments on 6 benchmarks demonstrate that Laser achieves state-of-the-art performance among latent reasoning methods, surpassing the strong baseline Monet by 5.03% on average. Notably, it achieves these gains with extreme efficiency, reducing inference tokens by more than 97%, while demonstrating robust generalization to out-of-distribution domains.

Figures (7)

• Figure 1: Laser replaces verbose textual rationales (a) with efficient latent superpositions (b).
• Figure 2: Overview of the Laser. Laser employs DWAL. At each step $t$, a dynamic validity window $\mathcal{W}_t$ is defined over future semantic tokens to construct a Reference Superposition Distribution. The latent state is then optimized to align with this distribution via $\mathcal{L}_{DWAL}$. The final answer is generated explicitly after the reasoning using $\mathcal{L}_{CE}$.
• Figure 3: Fine-grained comparison across 14 distinct categories. Laser outperforms Qwen2.5-VL-7B and Monet in 11 tasks, highlighting superior high-level semantic and spatial reasoning.
• Figure 4: Visualization of the latent cognitive trajectory. The decoded tokens reveal a structured multi-hop reasoning path, evolving from entity localization (Step 0: Seats) to spatial analysis (Step 1: Fence) and final deduction.
• Figure 5: Ablation study. We contrast the full Laser model with variants lacking the DWAL (w/o DWAL) and the dynamic windows (w/o Windows). The consistent performance gap across six benchmarks highlights the necessity of the proposed Dynamic Windowed Alignment Learning for effective visual reasoning.