LLM Output Correctness Estimation
Methods for estimating the correctness of LLM outputs at the step or span level without additional training, often using internal model signals.
6 papers
Where this stands
The written synthesis of this thread is for subscribers. Subscribe.
How this thread developed
Oct 2026 · Technical University of Munich, Huawei Heisenberg Research Center
Language models’ reasoning steps reveal mistakes better than final answers alone
Introduces STEP, a training-free framework for step-level correctness estimation using soft probability distributions from partial trajectories.
Oct 2026 · TokenRhythm Technologies
AI services secretly using cheaper models can be detected with few queries
Proposes API-TAIL, a method to detect LLM model substitution in APIs by estimating correctness through temperature alignment, extending output correctness estimation to API audits.
Oct 2026 · Anthropic Fellows Program, Anthropic
AI auditors find more problematic behavior by reading models’ internal signals
Uses internal activity to estimate likelihood of harmful behavior, extending correctness estimation to safety-flagged outputs.
2 further papers
Nov 2026 · Indian Institute of Technology Roorkee
Language models can improve answers without better internal error recognition
Shows RL improves arithmetic outputs without improving the model's internal error detection, exposing a limit of self-monitoring signals.
4 of 6 papers shown