LLMs Struggle to Learn Per-Instance Acquisition Policies.

Many AI systems can acquire additional, often expensive, information—like an LLM's structured reasoning or a slow oracle—to improve decisions. This research highlights a crucial difference: merely observing that such an auxiliary signal helps on average does not mean a system can learn to effectively decide when to acquire it for individual instances. The authors propose a "reward-SNR floor," a minimum signal-to-noise ratio, below which learning an effective per-instance acquisition policy becomes impossible.The study demonstrates that even when an auxiliary signal shows a clear average benefit, and an ideal "oracle" could pick the best instances, learned policies often perform no better than random acquisition. This failure occurs across various granularities (per-impression, cluster, etc.) because the apparent "learnable structure" is often just noise. The paper introduces Structured Hypothesis Embeddings (SHE) as a concrete example, showing that while SHE can be faithful, its value is conditional, and learned acquisition policies collapse on datasets that fall below the reward-SNR floor. The conclusion is that for low-SNR scenarios, a design-time regime gate is more effective than attempting a per-instance policy.