OG-Rank: Learning to Rank Fast and Slow with Uncertainty and Reward-Trend Guided Adaptive Exploration

Praphul Singh, Corey D. Barrett, Sumana Srivasta, Irfan Ullah, Sri Gadde, Krishnaram KenthapadiarXiv 2025 · DOI · arXiv

In high-stakes clinical environments, decision-support systems face a dual challenge: they must be nearly instantaneous to maintain workflow efficiency, yet sufficiently transparent to earn clinician trust. Traditional Information Retrieval (IR) systems offer speed but lack interpretability, while Large Language Models (LLMs) provide sophisticated reasoning at the cost of high latency and compute.

OG-Rank introduces a novel reranking architecture that bridges this gap by mimicking human cognitive processes—specifically, the "fast and slow" thinking paradigm. By utilizing a single-decoder model with uncertainty-gated routing and an adaptive curriculum, OG-Rank achieves state-of-the-art performance in clinical order selection while optimizing computational resources.

Two-Speed Inference Architecture

OG-Rank is built upon a single-decoder model (initialized from Llama 3.2 3B) that supports two distinct inference paths. This "two-speed" approach allows the system to handle routine cases rapidly while reserving deep reasoning for complex scenarios.

  1. The Fast Path: The model calculates the log-odds of the initial generated token (e.g., the probability of a "yes" vs. "no" response for a candidate). This allows the system to rank an entire list of candidates in a single pass without full text decoding.
  2. Uncertainty Gating: The system monitors the "normalized entropy" ($q$) of the candidate list. If the model is confident, it proceeds with the fast-path ranking. If entropy exceeds a predefined threshold ($T=0.9$), the system triggers the slow path.
  3. The Slow Path: For high-uncertainty cases, the model performs a full listwise generation, producing a structured JSON output that includes both a reranked list and a clinical rationale to justify its decision.
Figure 2: The OG-Rank inference pipeline, illustrating the transition from fast ranking to slow, rationale-based ranking based on uncertainty gating (q > T).

Adaptive Curriculum Learning

To optimize training, the authors employed Group Relative Policy Optimization (GRPO) combined with a Reward-Trend Guided Curriculum. Rather than uniform training, examples are bucketized into "easy," "medium," and "hard" categories based on real-time uncertainty and historical reward trends.

This curriculum ensures that "hard" cases receive a higher allocation of training rollouts and longer rationale budgets. This strategy forces the model to focus its learning capacity on the most challenging clinical scenarios. Furthermore, the training incorporates "hard vetoes" for safety; any suggestion that violates critical clinical constraints (e.g., prescribing a known allergen) results in a minimum reward, prioritizing safety over formatting or fluency.

Experimental Results

OG-Rank was evaluated on encounter-scoped clinical order selection tasks, where the goal is to select the correct medical order based on patient transcripts. Performance was measured using Recall@1 (R@1) and Normalized Discounted Cumulative Gain (nDCG@20).

Figure 4: Comparative performance of OG-Rank against various baseline models. OG-Rank significantly outperforms larger models like Llama 3.1 8B in both fast and two-speed modes.

As shown in the results, OG-Rank's fast path alone (R@1: 0.45) significantly outperforms standard compact models and even larger models like Llama 3.1 8B (R@1: 0.21). When the uncertainty gate is active (triggering the slow path in 45% of cases), performance improves further to a Recall@1 of 0.56.

Beyond accuracy, the curriculum-based approach demonstrated significant efficiency gains, reducing the total decoding budget by 8.6% and average rollouts by 6.8% compared to uniform training. This confirms that targeted exploration of difficult cases is more effective than exhaustive training across all data points.

Conclusion

OG-Rank demonstrates that a single-policy model can effectively manage the trade-off between speed and reasoning in clinical settings. By leveraging internal probability signals for fast ranking and reserving full-text rationales for uncertain cases, the system provides a scalable, safe, and high-performance solution for real-time medical decision support. Future work will explore dynamic, budget-aware gating to further optimize performance under varying server loads.

Summary updated

BibTeX
@misc{singh2025og,
  author = {Praphul Singh and Corey D. Barrett and Sumana Srivasta and Irfan Ullah and Sri Gadde and Krishnaram Kenthapadi},
  title = {{OG-Rank: Learning to Rank Fast and Slow with Uncertainty and Reward-Trend Guided Adaptive Exploration}},
  year = {2025},
  eprint = {2510.17614},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.17614},
}
Figure