STEER: Inference-Time Risk Control via Constrained Quality-Diversity Search

TL;DR

STEER addresses the lack of controllable risk in LLM-based decision support for ordinal tasks by introducing a training-free constrained quality-diversity search to build a diverse persona ensemble and an inference-time percentile tuning dial. The method expands the latent risk spectrum while preserving safety on unambiguous cases, outperforming temperature-based sampling and post-training alignment in two clinical triage benchmarks. By being model-agnostic and deployable without weight updates, STEER enables precise, monotone control over the ROC operating point in real time, facilitating context-aware safety. This work highlights the value of explicit diversity governance for safe, flexible AI in high-stakes domains and beyond.

Abstract

Large Language Models (LLMs) trained for average correctness often exhibit mode collapse, producing narrow decision behaviors on tasks where multiple responses may be reasonable. This limitation is particularly problematic in ordinal decision settings such as clinical triage, where standard alignment removes the ability to trade off specificity and sensitivity (the ROC operating point) based on contextual constraints. We propose STEER (Steerable Tuning via Evolutionary Ensemble Refinement), a training-free framework that reintroduces this tunable control. STEER constructs a population of natural-language personas through an offline, constrained quality-diversity search that promotes behavioral coverage while enforcing minimum safety, reasoning, and stability thresholds. At inference time, STEER exposes a single, interpretable control parameter that maps a user-specified risk percentile to a selected persona, yielding a monotonic adjustment of decision conservativeness. On two clinical triage benchmarks, STEER achieves broader behavioral coverage compared to temperature-based sampling and static persona ensembles. Compared to a representative post-training method, STEER maintains substantially higher accuracy on unambiguous urgent cases while providing comparable control over ambiguous decisions. These results demonstrate STEER as a safety-preserving paradigm for risk control, capable of steering behavior without compromising domain competence.

Figures (7)

• Figure 1: STEER Reintroduces the Risk Operating Point. (A) Problem: Standard models exhibit mode collapse, shrinking the distribution of valid disagreement (gray curve). They are also unsteerable, reacting erratically to prompts (blue arrows). (B) The STEER Framework: STEER enhances this distribution via a constrained quality-diversity evolutionary search. This creates a deterministic conservativeness dial ($P$). (C) Context-Aware Control: The system generates a steerable ordinal performance curve (analogous to an ROC), enabling users to select a precise operating point ($P$) that satisfies specific sensitivity/specificity requirements.
• Figure 2: The STEER Framework. (A) Evolution: We iteratively optimize a persona population to maximize bias diversity subject to constraints. (B) Assembly: The evolved pool is distilled into a compact team to ensure uniform spectral coverage. (C) Inference: At deployment, the team acts as a distributional generator. The Inference with Percentile Tuning mechanism acts as a conservativeness dial.
• Figure 3: Evolutionary Steerability vs. Baselines. Ordinal AUC gain of STEER ensembles (varying size $N \in \{5, 10, 20, 30\}$) relative to high-temperature sampling ($N=10$) and static persona baselines ($N=10$) for (A) MIETIC Triage and (B) EHR Triage.
• Figure 4: Deterministic Risk Control via Percentile Tuning. The effect of the Percentile Selection dial on the predicted urgency distribution for STEER ($N=10$) vs. the high-temperature baseline ($N=10$) on (A) MIETIC and (B) EHR Triage. Higher percentiles ($P \to 100$) correspond to strictly more conservative (urgent) predictions.
• Figure 5: Inference-Time Evolution vs. Post-Training Alignment. Comparison between Spectrum Tuning (red) and STEER (blue) on Gemma-3-12B-IT. (A, B) Steerability: Ordinal AUC scores on the MIETIC and EHR Triage datasets. (C, D) Safety: Accuracy on the unambiguous urgent safety sets.