New Benchmark Evaluates LLM Uncertainty in Long-Form Generation

As Large Language Models (LLMs) produce increasingly lengthy outputs, accurately estimating their uncertainty at a fine-grained level becomes crucial for identifying specific errors rather than dismissing entire responses. Current evaluation methods for uncertainty are often hampered by imperfect labels, highlighting the need for benchmarks with zero-noise, deterministic ground truth. This paper introduces the Single-answer Atomic Long-form Target (SALT) benchmark, comprising six procedurally generated tasks. SALT provides single, deterministic long textual ground truths, allowing for unit-level evaluation of correctness, calibration, and ranking without relying on external human judges. Using SALT, an analysis of over 50 LLMs yielded significant insights: confidence functions vary in their effectiveness across different uncertainty aspects, and confidence ranking largely fails at the atomic resolution, though it improves at coarser line-level units. The study also identified two distinct drivers of future errors: propagation from corrupted prefixes, primarily influenced by global context correctness, and bounded degradation due to increasing answer-context length. Furthermore, reasoning techniques like Chain-of-Thought prompting or internalized training improve accuracy but can degrade confidence ranking, presenting a trade-off for risk-critical applications.