Resource-Constrained Adaptive Inference for Sequential Pricing

Researchers have proposed two new frameworks, one for target-aware pricing and another for long-context LLM inference, to address specific challenges in their respective fields.

A team of researchers has designed a target-aware pricing controller that certifies feasible target bands and logs continuous local densities, addressing the issue of support-exclusion failure in resource-constrained pricing controllers^[1]. The controller's resource state may remove the target price neighborhood from the feasible set, even when every realized action has a known positive density. Experiments show calibration in certified bands and diagnostic abstention when the resource state collapses target support. Meanwhile, another research team proposed a framework called EntropyInfer for long-context LLM inference, which adaptively allocates compute at the granularity of individual heads and segments^[2]. The framework uses attention entropy to adaptively allocate compute during prefilling and introduces a latent KV cache compression scheme. Experiments on Llama, Qwen, and openPangu model series showed that EntropyInfer outperformed baselines with up to 2.39$ imes$ end-to-end speedup beyond 100k tokens^[2].