AI summaryⓘ
The authors studied how much useful information about an AI's correctness is shown when readers see different parts of the AI's reasoning: just the final answer, a summary of the reasoning, the full reasoning steps, or internal signals. They found that when the original question (prompt) is not shown, the summary helps predict correctness better than just the answer, but when the prompt is visible, the full reasoning trace gives more useful clues than the summary. The study also shows that simple methods to read summaries or traces have limited success, and the usefulness depends on what information the user already has. Overall, the authors argue that how well one can monitor AI’s correctness depends both on what reasoning parts are displayed and how the reader interprets them.
Large Language ModelsPromptReasoning TraceSelf-SummaryCorrectness PredictionAUROCLinear PredictorMonitorabilityFaithfulnessMMLU-Pro Benchmark
Authors
Andres Algaba, Francesca Carlon, Lynn Delcon, Marthe Ballon, Bert Verbruggen, Vincent Ginis
Abstract
Large language models often show users a final response and a short reasoning summary while the full reasoning trace stays hidden. We introduce an observability ladder that holds each completed run fixed and varies only what a reader inspects to judge whether the answer is correct: the response, a self-summary the model writes from the trace, the trace itself, and internal signals, each with and without the prompt. Across three benchmarks and five open-weight Qwen3 and gpt-oss models, we train matched linear correctness predictors on each access level. Without the prompt, summaries carry most of the trace's ranking signal (mean AUROC 0.774 versus 0.813) and add +0.156 over the response alone. With the prompt visible, the summary's gain collapses to +0.019, while the trace still adds +0.041. Even at equal length, the trace's last words predict correctness as well as summaries, or slightly better, and carry denser and more discriminative uncertainty and self-correction cues. On MMLU-Pro questions with both correct and incorrect runs, linear summary readers are near chance and trace readers retain only modest signal, both with and without the prompt (prompt-withheld AUROC 0.503-0.545 versus 0.544-0.590). With the prompt withheld, a GPT-5-mini reader recovers substantially more signal from both summaries and traces on gpt-oss-20b, and even then the trace keeps a small +0.034 advantage. Much of the linear readers' trace signal is associated with length. In the common case where users already hold the prompt, summaries are less helpful than the full trace for monitoring correctness. Monitorability is thus a joint property of the display and the reader, so any monitorability claim, including for faithfulness, should specify both.