Conformal Language Modeling via Posterior Sampling

Researchers have proposed new methods to improve Large Language Models (LLMs) and neural operator surrogates. The LLM method samples from approximations to the LLM posterior, achieving target risk control and higher downstream utility. Neural operator surrogates were improved using a diffusion prior with diffusion posterior sampling.

A new method for LLMs uses posterior sampling to reduce hallucinations, achieving target risk control and higher downstream utility compared to prior work^[1]. This approach samples from approximations to the LLM posterior, where the conditioning event corresponds to a calibrated, high-scoring region. In a separate development, researchers have proposed the FreqNO-DPS method to improve neural operator surrogates, which approximate PDE solutions orders of magnitude faster than numerical solvers but suffer from spectral bias^[2]. The FreqNO-DPS method combines an unconditional score-based diffusion prior with diffusion posterior sampling conditioned on sparse observations. It uses a closed-form, spectrally shaped guidance score that weights the surrogate by its frequency-dependent accuracy. The method has been tested on 3D elastic wavefield prediction at 5% and 2% sensor coverage. While LLMs remain plagued by hallucinations, recent work has shown empirical success in reducing their prevalence using statistical techniques based on conformal prediction.