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

TL;DR

OG-Rank presents a practical, single-decoder ranking approach that achieves encoder-like scoring efficiency by using a pooled first-token yes/no signal to rank candidates, and selectively generates a structured rationale only when uncertainty is high. A listwise uncertainty gate (threshold fixed at $T=0.9$) routes ambiguous queries to a slow, one-shot JSON ranking with a brief rationale, balancing fast response times with justification where it matters. The model is trained with a curriculum that allocates more rollout and rationales to hard cases, guided by a multi-axis judge and KL-regularized policy optimization, enabling safer and domain-aligned decisions. Across encounter-scoped ranking tasks, OG-Rank attains strong fast-path performance (Recall@1 around $0.45$ and nDCG@20 around $0.625$) and further gains when gating activates (Recall@1 about $0.56$ and nDCG@20 about $0.699$ at a $45egin{document}" gate rate), while maintaining predictable budgets and simpler deployment compared with multi-model stacks. The approach demonstrates substantial compute savings through bucketing and curriculum adjustments, and its general two-speed, uncertainty-guided paradigm is transferable to other decision tasks requiring selective generation and reliable latency budgets.

Abstract

Clinicians need ranking systems that work in real time and still justify their choices. Motivated by the need for a low-latency, decoder-based reranker, we present OG-Rank, a single-decoder approach that pairs a pooled first-token scoring signal with an uncertainty-gated explanation step. The model scores all candidates in one pass and generates a brief, structured rationale only when the list is genuinely ambiguous, keeping latency predictable. Trained with a curriculum that concentrates effort on hard cases, OG-Rank delivers strong effectiveness on encounter-scoped order selection (fast path: Recall@1~0.45, nDCG@20~0.625) and improves further when the gate activates (Recall@1~0.56, nDCG@20~0.699 at a 45\% gate rate), while compact backbones show similar gains under the same policy. Encoder baselines trail in both effectiveness and flexibility. The result is a practical recipe: rank fast by default and explain when it helps, a pattern that applies broadly to decision tasks where selective generation buys accuracy at acceptable cost. The single-policy design simplifies deployment and budget planning, and the curriculum principle (spend more on the hard cases, less on the easy ones) readily transfers beyond clinical order selection.

Figures (3)

• Figure 1: GRPO with uncertainty- and reward-trend-guided curriculum. First-step uncertainty determines the curriculum bucket for the next epoch together with the previous epoch’s reward trend. Each bucket uses a fixed rollout configuration; group-relative centering and KL regularization to a pre-epoch reference stabilize updates under a multi-axis clinical judge.
• Figure 2: Rebucketing across epochs. At the start of epoch $e{+}1$, each pointwise sample is reassigned to hard, medium, or easy using both uncertainty $q_e(u_j)$ and the previous epoch’s mean reward $\bar{r}_e(u_j)$.
• Figure 3: Two-speed inference. The fast path reads the pooled first-step signal (via log-odds) as a score with no decoding. The slow path is triggered by listwise uncertainty and decodes a concise rationale under a fixed budget.