Uncertainty-Aware Dynamic Thinking (UADT)

• UADT is a framework that uses entropy-derived uncertainty estimates to flexibly adjust exploration depth and reasoning policies.
• It employs techniques like exponential moving averages and hyperbolic tangent mappings to shape advantage in reinforcement learning pipelines.
• By dynamically modulating decision speed and depth, UADT enhances accuracy and efficiency in handling diverse, multimodal inputs.

Uncertainty-Aware Dynamic Thinking (UADT) denotes a class of mechanisms in which uncertainty estimates are used to modulate reasoning, exploration, supervision, information acquisition, or decision policy during learning or inference. In the explicit nomenclature of ophthalmic multimodal reasoning, UADT is introduced as an advantage-shaping mechanism inside a reinforcement-learning pipeline: it estimates sample-level uncertainty via entropy, constructs an adaptive threshold with an exponential moving average (EMA), maps entropy deviations through a hyperbolic tangent, and reshapes the GRPO advantage so that high-uncertainty samples receive a bonus while low-uncertainty samples are down-weighted (Wu et al., 22 Aug 2025). In a closely related inference-time formulation, dynamic control of thinking path length is achieved by making the overclocking hyperparameter $\alpha$ context-aware and uncertainty-aware, so that reasoning speed changes token by token according to normalized entropy and input complexity (Jiang et al., 21 Sep 2025). Taken together, these formulations define UADT as a family of uncertainty-conditioned control schemes rather than a single architecture.

1. Definition and conceptual scope

The named method “Uncertainty-Aware Dynamic Thinking (UADT)” is introduced in OphthaReason as a mechanism that “estimates sample-level uncertainty via entropy and dynamically modulates the model’s exploration depth using a shaped advantage mechanism” (Wu et al., 22 Aug 2025). The immediate motivation is heterogeneity across reasoning regimes: the MM-Retinal-Reason dataset spans both basic reasoning, characterized as single-image and visual feature–matching, and complex reasoning, characterized as multi-image, multimodal inputs plus heterogeneous clinical text and clinical context. Under that heterogeneity, a fixed exploration policy is described as suboptimal because it can under-explore ambiguous or rare cases and over-explore trivial cases.

A parallel motivation appears in “Adaptive Overclocking,” where Large Reasoning Models are described as suffering from computational inefficiency due to overthinking when a fixed reasoning budget is applied uniformly across tasks (Jiang et al., 21 Sep 2025). There, a fixed scalar $\alpha$ shortens the reasoning path by pushing hidden states along a Thinking Progress Vector, but the same fixed $\alpha$ can cause over-acceleration on hard or ambiguous steps and under-acceleration on easy tasks. This establishes a second, complementary formulation of UADT: uncertainty should not merely be estimated, but should directly govern how much deliberation is performed at each step.

Across these formulations, UADT is not limited to one operational meaning of “thinking.” In OphthaReason, “exploration depth” is defined in a policy/learning sense rather than as an explicit tree-search depth or explicit sequence-length controller (Wu et al., 22 Aug 2025). In Adaptive Overclocking, by contrast, higher $\alpha$ means faster progression and shorter paths, so dynamic uncertainty control directly affects the length of the internal Chain-of-Thought trajectory (Jiang et al., 21 Sep 2025). This distinction is central to the literature: UADT can operate at the level of policy optimization, token generation, information seeking, or decision gating.

2. Mathematical structure of uncertainty-conditioned control

A recurring technical pattern in this literature is the use of an uncertainty signal, a reference or threshold, and a control law that alters either optimization pressure or inference dynamics. In OphthaReason, token-level entropy is defined as

$$\mathcal{H}_t(y_t \mid X, q, y_{<t}) = - \sum_{v \in V} \pi_{\theta}(y_t = v \mid X, q, y_{<t}) \log \pi_{\theta}(y_t = v \mid X, q, y_{<t}),$$

and the sample-level uncertainty is the sequence average

$$\mathcal{H}_{\text{seq}}(Y \mid X, q) = \frac{1}{T} \sum_{t=1}^{T} \mathcal{H}_t(y_t \mid X, q, y_{<t}).$$

The adaptive threshold is

$$

\tau_s = \alpha \cdot \