Morning Edition · Monday, August 10, 2026Published at 2:21 AM EDT · New York

Paper Finds Bundling Many Verdicts Into One Judge Call Degrades Oversight, Regardless of Compute

Splitting the work into separate calls, each returning fewer decisions, keeps verdicts grounded in evidence where added tokens and tools do not.

A preprint titled Sharding Prevents LLM Oversight Failures and Adversarial Exploitation makes a claim that is simple to state and difficult to act on. Giving a large language model (LLM) acting as a judge more compute does not necessarily m…