AdaThink-Med: Adaptive Medical Reasoning

• AdaThink-Med is an adaptive reasoning framework for medical QA that uses uncertainty-guided calibration to adjust output length based on question difficulty.
• It employs a two-stage training process where initial medical reasoning competence is enhanced with reinforcement learning using uncertainty and correctness signals.
• Empirical evaluations report up to 6.4x reduction in output length with minimal performance degradation, optimizing the trade-off between efficiency and accuracy.

Searching arXiv for AdaThink-Med and closely related adaptive reasoning medical LLM papers. "AdaThink-Med" denotes an end-to-end framework for medical reasoning LLMs that targets adaptive thinking through uncertainty-guided length calibration (Rui et al., 29 Sep 2025). It is motivated by a practical inefficiency in long chain-of-thought inference: contemporary medical reasoning models often produce lengthy reasoning traces regardless of input difficulty, increasing latency, token cost, and susceptibility to overthinking. AdaThink-Med addresses this by learning when to suppress redundant reasoning and when to permit extended reasoning, with the stated goal that models think less for simpler questions and think more for complex ones (Rui et al., 29 Sep 2025). On six public medical QA benchmarks, it is reported to achieve up to 6.4x length reduction on average while retaining performance with only minimal degradation, and to exhibit two emergent reasoning modes characterized as "non-thinking" and "thinking" (Rui et al., 29 Sep 2025).

1. Definition and problem setting

AdaThink-Med is presented as the first end-to-end framework designed to enhance adaptive thinking ability in medical reasoning models with uncertainty-guided length calibration (Rui et al., 29 Sep 2025). Its problem setting is medical QA under inference-time scaling, where extended long chain-of-thought has improved reasoning performance but introduces a deployment trade-off between answer quality and computational cost.

The framework is premised on four observations stated in the source description. First, many medical questions are solvable with short outputs or no explicit reasoning. Second, forcing long reasoning on every question creates unnecessary compute cost. Third, overthinking can degrade performance on easy tasks. Fourth, medical deployment requires both accuracy and efficiency (Rui et al., 29 Sep 2025). The method is therefore not a generic length penalty; it is a difficulty-aware calibration scheme intended to make output length conditional on whether the question appears easy or hard.

A common misconception is that AdaThink-Med is simply a compression method for chain-of-thought. The description instead frames it as an adaptive reasoning framework: it does not reward shorter outputs uniformly, but calibrates length differently for easy correct cases and hard incorrect cases (Rui et al., 29 Sep 2025). This distinction is central to its design, because naive shortening can induce what the paper calls a "length reward hack," in which models become unnaturally short, repetitive, or unhelpful.

2. Core framework and training stages

AdaThink-Med consists of two training stages and an inference-time behavior that emerges from the learned policy (Rui et al., 29 Sep 2025). In Stage 1, the model is trained normally, without length calibration, so that it learns medical reasoning competence. In Stage 2, adaptive thinking is introduced through uncertainty-guided length calibration.

For each prompt $\mathbf{x}$, the model samples a group of $G$ candidate outputs,

$\{o_i\}_{i=1}^G \sim \pi_\theta(\cdot|\mathbf{x}),$

where each output contains a full reasoning trace plus final answer (Rui et al., 29 Sep 2025). These candidates are then evaluated against the ground-truth answer $y^*$ to determine correctness. The framework also computes token-level entropy during decoding and aggregates the highest-entropy tokens to obtain an uncertainty signal for each output. Correctness and uncertainty are then combined into a question-level difficulty estimate $\mathcal{D}_q$, which drives a difficulty-aware reward on output length (Rui et al., 29 Sep 2025).

The resulting policy is optimized with GRPO plus KL regularization to remain close to a reference model (Rui et al., 29 Sep 2025). This design places AdaThink-Med in the broader class of RL-based reasoning calibration methods, but the distinctive claim is that it explicitly models problem difficulty rather than relying only on correctness or a global output-length penalty.

This suggests that AdaThink-Med should be understood as a training-time mechanism for acquiring a conditional inference policy, rather than as a separate routing module. The source description explicitly contrasts this with approaches requiring manual labeling and extra routing or classifier modules, noting that AdaThink-Med produces two reasoning modes end-to-end, without extra supervision (Rui et al., 29 Sep 2025).

3. Uncertainty-guided difficulty estimation and calibrated rewards

The mathematical core of AdaThink-Med is an efficient reasoning objective that balances task reward, computational cost, and KL regularization:

$\max_{\theta} \ \mathbb{E}_{\mathbf{x} \sim \mathcal{D}, \mathbf{y} \sim \pi_\theta(\cdot|\mathbf{x})} \Big[ R(\mathbf{y}, \mathbf{y}^*) \ - \ \lambda \cdot \mathcal{C}(\pi_\theta, \mathbf{x}) \ - \ \beta \cdot D_{\mathrm{KL}\!\big(\pi_\theta(\mathbf{y}|\mathbf{x}) \parallel \pi_{\mathrm{ref}(\mathbf{y}|\mathbf{x})\big) \Big]$

(Rui et al., 29 Sep 2025).

Uncertainty is computed from token-level entropy. At decoding step $t$,

$$H_t := - \sum_{j=1}^{V} p_{t,j} \log p_{t,j}, \quad \text{where } p_t = (p_{t,1}, \cdots, p_{t,V}) = \pi_\theta(\cdot \mid \mathbf{x}, s_{<t}) = \text{Softmax} \left( \frac{z_t}{T} \right).$$

Output-level uncertainty is then defined as the mean entropy of the top-$K$ most uncertain tokens:

$$\mathcal{H}_i := \frac{1}{K} \sum_{t \in \mathcal{T}_K} H_t.$$

These values are min-max normalized within the batch to yield $G$0 (Rui et al., 29 Sep 2025).

A naive difficulty score based only on sampling accuracy is

$G$1

AdaThink-Med instead defines uncertainty-guided difficulty as

$G$2

The intended interpretation is that correct answers with higher uncertainty provide weaker evidence that a question is easy (Rui et al., 29 Sep 2025). The batch threshold separating easy and hard questions is the $G$3-quantile,

$G$4

The calibrated length reward is asymmetric:

$G$5

with

$G$6

and $G$7 and $G$8 defined as EMA-smoothed average lengths for simple and hard questions (Rui et al., 29 Sep 2025). The full per-output reward is

$G$9

The asymmetry is the method’s main safeguard against indiscriminate compression. Easy correct outputs are pushed toward brevity, whereas hard incorrect outputs are encouraged to extend reasoning in order to explore alternative solutions (Rui et al., 29 Sep 2025). A plausible implication is that the framework operationalizes "thinking budget" as a reward-shaped function of estimated epistemic difficulty rather than as a fixed prompt instruction.

4. Optimization, backbones, and evaluation protocol

AdaThink-Med uses GRPO as its RL backbone. The appendix gives the group loss:

$\{o_i\}_{i=1}^G \sim \pi_\theta(\cdot|\mathbf{x}),$0

with the full update also including the KL penalty from the main objective (Rui et al., 29 Sep 2025).

The primary training corpus is AlphaMed19k, which combines the official training splits of MedQA and MedMCQA (Rui et al., 29 Sep 2025). Evaluation is conducted on six public medical QA benchmarks: MedQA, MedMCQA, PubMedQA, MMLU-ProM, GPQA-M, and MedXpertQA / MedXpert (Rui et al., 29 Sep 2025). The main backbones are Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct, with additional comparisons to MedLLama3, MMed, Med42, OpenBioLLM, UltraMedical, HuatuoGPT-o1, and m1 (Rui et al., 29 Sep 2025).

The baselines include general zero-shot LLMs, medical fine-tuned LLMs, and length-calibration methods Kimi1.5, ShortBetter, DAST, and CosFn (Rui et al., 29 Sep 2025). Training details reported in the source include 8× H100 GPUs, PyTorch, FlashAttention-2, GRPO via verl, batch size 256, learning rate $\{o_i\}_{i=1}^G \sim \pi_\theta(\cdot|\mathbf{x}),$1, KL coefficient $\{o_i\}_{i=1}^G \sim \pi_\theta(\cdot|\mathbf{x}),$2, inference with vllm, 8 completions per sample ($\{o_i\}_{i=1}^G \sim \pi_\theta(\cdot|\mathbf{x}),$3), temperature 1.0, 300 GRPO steps without length calibration followed by 200 steps with adaptive length calibration (Rui et al., 29 Sep 2025). For LLaMA-based experiments, an n-gram repetition penalty is added because repetition-based length hacking was observed (Rui et al., 29 Sep 2025).

Three evaluation metrics are used: Accuracy, Output length, and AES (Accuracy-Efficiency Score) (Rui et al., 29 Sep 2025). AES is defined by

$\{o_i\}_{i=1}^G \sim \pi_\theta(\cdot|\mathbf{x}),$4

with default weights

$\{o_i\}_{i=1}^G \sim \pi_\theta(\cdot|\mathbf{x}),$5

This metric formalizes the efficiency-performance trade-off that the framework is designed to optimize (Rui et al., 29 Sep 2025).

5. Reported empirical behavior

The main quantitative claim is that AdaThink-Med substantially reduces output length while preserving accuracy with only minimal degradation (Rui et al., 29 Sep 2025). For AdaThink-Med-Llama, the reported average length is 64 tokens, average length reduction is 84.39%, average accuracy is 55.59%, and AES is 0.92. For AdaThink-Med-Qwen, the reported average length is 106 tokens, average length reduction is 78.67%, average accuracy is 54.99%, and AES is 0.93 (Rui et al., 29 Sep 2025). The paper also states that the method can achieve up to 6.4x average length reduction.

Against length-calibration baselines, the reported comparison is explicitly framed in terms of trade-off rather than absolute accuracy. On LLaMA, AdaThink-Med has average length 64 versus 125 for Kimi1.5, with comparable accuracy 55.59% versus 55.65%, and AES 0.92 versus 0.78 (Rui et al., 29 Sep 2025). On Qwen, AdaThink-Med has average length 106 versus 298 for Kimi1.5, accuracy 54.99% versus 55.41%, and AES 0.93 versus 0.55 (Rui et al., 29 Sep 2025). The interpretation given in the source is that some baselines attain slightly higher accuracy but at much larger length cost, while more aggressive shortening methods may collapse in performance.

An especially notable empirical observation is the spontaneous emergence of two reasoning modes, "non-thinking" and "thinking" (Rui et al., 29 Sep 2025). In non-thinking mode, the model outputs a direct answer with little or no intermediate reasoning. In thinking mode, it produces concise but necessary reasoning steps. Output-length distributions and case studies are reported to show that easy questions receive very short outputs, hard questions trigger longer reasoning, and redundant reasoning is dynamically suppressed (Rui et al., 29 Sep 2025).

The source also describes a data-selection use case. By partitioning samples according to output length from the trained model and balancing easy and hard samples, a 40% subset can outperform full-data training on some benchmarks, and 20% retains 98.8% of full-set performance (Rui et al., 29 Sep 2025). This suggests that the learned adaptive-thinking signal may function as a proxy for curriculum structure or sample utility, though the description does not generalize this beyond the reported use case.

6. Hyperparameters, ablations, and limitations

The ablation study centers on $\{o_i\}_{i=1}^G \sim \pi_\theta(\cdot|\mathbf{x}),$6 and $\{o_i\}_{i=1}^G \sim \pi_\theta(\cdot|\mathbf{x}),$7 (Rui et al., 29 Sep 2025). The threshold parameter $\{o_i\}_{i=1}^G \sim \pi_\theta(\cdot|\mathbf{x}),$8 controls the easy-hard split. When $\{o_i\}_{i=1}^G \sim \pi_\theta(\cdot|\mathbf{x}),$9, the model produces longer reasoning overall with some redundant overthinking. When $y^*$0, outputs become extremely short, but accuracy drops and reward hacking appears. The best reported setting is $y^*$1 (Rui et al., 29 Sep 2025).

The parameter $y^*$2 controls the balance between uncertainty and correctness in difficulty estimation. With $y^*$3, performance is weaker and AES is 0.82. With $y^*$4, performance is slightly worse than the best setting. The best reported setting is $y^*$5 (Rui et al., 29 Sep 2025). These results are consistent with the framework’s conceptual claim that uncertainty should modulate, but not dominate, correctness-based difficulty estimation.

A staged-training ablation further indicates that the two-stage procedure is important. Without staged training, RL begins from a state of low accuracy and high uncertainty, almost all samples appear hard, optimization becomes imbalanced, and length compression weakens while performance degrades (Rui et al., 29 Sep 2025). This reinforces the interpretation that adaptive length calibration presupposes an already competent reasoning policy.

The source identifies one main limitation: the uncertainty signal depends on the quality and diversity of sampled outputs, so entropy-based difficulty estimation may be sensitive to sampling behavior (Rui et al., 29 Sep 2025). The stated future direction is to explore more robust uncertainty estimators and broader clinical validation. A related misconception would be to treat the entropy signal as a stable measure of intrinsic medical difficulty. The paper does not make that stronger claim; it uses entropy as a practical proxy within a sampling-based RL pipeline (Rui et al., 29 Sep 2025).

7. Position within medical AI research

AdaThink-Med belongs to a line of work concerned with interpretability, efficiency, and adaptive reasoning in medical AI, but its contribution is specific to inference-time reasoning length in medical LLMs (Rui et al., 29 Sep 2025). In this respect, it differs from image-centric teacher-student-explainer frameworks such as "MED-TEX" (Nguyen-Duc et al., 2020), which combine knowledge distillation and model interpretation for medical image classification, and from rationale-aware multimodal generation frameworks such as "MedThink" (Gai et al., 2024), which target MedVQA by generating answers together with medical decision-making rationales. AdaThink-Med is not a pixel-level explanation model and not a MedVQA rationale generator; it is a medical QA framework that calibrates how much reasoning is produced.

The comparison with "MedThink" is particularly instructive. MedThink studies generation order for answer-rationale production in multimodal QA, including Explanation, Reasoning, and Two-Stage Reasoning strategies (Gai et al., 2024). AdaThink-Med instead learns whether a case warrants extended reasoning at all, using correctness and uncertainty to modulate output length (Rui et al., 29 Sep 2025). This suggests a broader research distinction between rationale generation as an interpretability artifact and adaptive thinking as a compute-allocation policy.

Within the wider medical AI landscape represented by the cited works, AdaThink-Med is aligned with a recurring practical concern: deploying high-performing systems under resource constraints without assuming that more computation is always better. MED-TEX addresses this through small-student distillation with pixel-level explanations (Nguyen-Duc et al., 2020); Med-NCA does so through lightweight segmentation with strong invariance properties (Kalkhof et al., 2023); AdaViT addresses variable-modality transfer in 3D MRI (Das et al., 4 Apr 2025); Med-Art focuses on data-efficient diffusion adaptation under limited text supervision (Guo et al., 25 Jun 2025). AdaThink-Med extends that resource-sensitive perspective to medical language reasoning, where the constrained resource is inference-time token budget rather than memory, modality availability, or paired-data scale.

In that sense, AdaThink-Med can be situated as a medical reasoning analogue of adaptive computation: it does not reject long reasoning, but seeks to reserve it for cases where it is most needed (Rui et al., 29 Sep 2025).