How Much Does a Reasoning Summary Reveal? An Observability Ladder for Large Language Models

2026-08-03

AI Summary

In large language model (LLM) inference, full reasoning trajectories are often concealed, with only final answers or summaries provided, impeding users’ ability to accurately assess response correctness. This work proposes an “observability ladder” framework to systematically evaluate the utility of different information granularities—ranging from answers and summaries to full trajectories and internal signals, with or without access to the original prompt—for correctness judgment. Using linear predictors and AUROC as a metric of monitoring capability, experiments reveal that when the prompt is known, summaries offer negligible improvement over answers alone (+0.019 AUROC), whereas full trajectories substantially enhance judgment accuracy (+0.041). Moreover, terminal tokens of trajectories contain richer uncertainty and self-correction cues than equally long summaries. The study demonstrates that monitoring efficacy is a joint function of the presented content and the observer’s contextual knowledge.

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.

Problem

• observability
• reasoning summary
• large language models
• correctness prediction
• monitorability

Innovation

• observability ladder
• reasoning trace
• correctness monitoring
• faithfulness evaluation
• large language models