🤖 AI

Measuring Black-Box Confidence via Reasoning Trajectories: Geometry, Coverage, and Verbalization

This paper proposes a novel black-box confidence estimation method that embeds chain-of-thought reasoning trajectories to measure geometric convergence toward answer anchors, demonstrating that fusing this trajectory-based score with coverage and verbalization channels significantly outperforms traditional self-consistency baselines across diverse benchmarks and large language models without requiring access to internal model states.

Original authors: Marc Boubnovski Martell, Josefa Lia Stoisser, Kaspar Märtens, Jialin Yu, Robert Kitchen, Philip Torr, Jesper Ferkinghoff-Borg

Published 2026-05-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Marc Boubnovski Martell, Josefa Lia Stoisser, Kaspar Märtens, Jialin Yu, Robert Kitchen, Philip Torr, Jesper Ferkinghoff-Borg

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are hiring a brilliant but mysterious consultant to solve a difficult puzzle. You can only see their written notes (the "Chain of Thought"), but you cannot see their internal brainwaves, their confidence scores, or the raw data they are processing. You need to know: Can we trust their answer?

Usually, to check if this consultant is confident, you ask them to solve the same puzzle 10 times and see if they give the same answer 10 times. This is called "Self-Consistency." It works, but it's expensive (you have to pay for 10 answers) and it's a bit clumsy—it's like asking, "Did you get the same result 10 times?" rather than looking at how they got there.

This paper introduces a new, cheaper, and smarter way to measure confidence by looking at the geometry (the shape and path) of the consultant's notes.

Here is the breakdown of their discovery using simple analogies:

1. The "Walking Path" Analogy (Geometry)

Imagine the consultant's reasoning process is a person walking through a foggy forest toward a specific destination (the correct answer).

The Old Way: You just wait until they reach the end and ask, "Are you sure?"
The New Way: You watch their path. As they walk, do they start drifting closer and closer to the correct destination, even before they say the destination's name out loud?

The researchers found that if you look at the consultant's notes just before they write the final answer (the "penultimate" step), their words naturally cluster closer to the correct answer in a mathematical space. It's like seeing the walker's footsteps get tighter and tighter around the right tree before they finally point at it. This "geometric convergence" is a strong signal that they are confident and correct, even if you can't see their internal brain.

2. The Three-Part Confidence Check

The paper argues that you can't rely on just one signal. They break confidence down into three distinct channels, like checking a car before a long trip:

Coverage (The Map Check): Did the consultant even look at the right map? Did they consider the correct answer as a possibility at all? If they never thought of the right answer, no amount of confidence matters. This is a "reachability" check.
Geometry (The Path Check): Once they are looking at the right map, are they walking steadily toward the right spot? This is the "path" signal described above. It measures how tightly their reasoning locks onto a specific option.
Verbalization (The Self-Report): Finally, you ask the consultant, "On a scale of 0 to 100, how sure are you?"
- Surprise Finding: This last part is the least reliable on its own. Sometimes the consultant says they are 100% sure, but they are actually walking in circles. It only helps when the other two checks are already looking good.

3. The "Penultimate" Secret

One of the coolest discoveries is when to look.

If you look at the very last sentence where they say the answer, the signal gets messy. It's like the walker stopping and shouting, "I'm here!" even if they are actually standing next to the wrong tree.
The researchers found the sweet spot is the sentence right before the final answer. This is where the consultant's internal logic is most committed to the truth, before they get distracted by the act of writing the final word.

4. The Result: Smarter and Cheaper

By combining these three signals (Map, Path, and Self-Report), the researchers created a system that:

Costs less: It only needs 4 attempts to get a better confidence score than the old method's 8 attempts.
Works better: It predicts the correct answer more accurately (higher "AUC" scores) across medical and science exams.
Is robust: It works even if you use different AI models or different ways of measuring the "distance" between words.

The Bottom Line

You don't need to be a mind-reader to know if an AI is confident. You just need to watch how it thinks. If its reasoning path naturally tightens around the right answer just before it speaks, and if it actually considered the right answer in the first place, you can trust it more than if you just asked it to repeat the answer ten times.

Important Limitation: The paper notes that if the AI never thinks of the correct answer in the first place, no amount of "path watching" can fix that. The system can only tell you how confident the AI is in the options it actually considered.

Technical Summary: Measuring Black-Box Confidence via Reasoning Trajectories

Problem Statement

Reliable confidence estimation is a prerequisite for the safe deployment of Chain-of-Thought (CoT) reasoning through text-only APIs, where access to logits, hidden states, or internal retrieval mechanisms is unavailable. The dominant black-box baseline, self-consistency (sampling $K$ CoTs and using vote concentration as a confidence signal), suffers from three primary limitations:

Linear Cost: It requires sampling $K$ traces, making it linearly expensive in $K$ .
Geometric Blindness: It ignores the internal geometry of the reasoning trace, treating the trace merely as a means to a final answer.
Discrete Resolution: It produces a discrete confidence score ( $k/K$ ), which lacks resolution at small $K$ and cannot effectively rank ties.

This paper investigates an alternative signal: the trajectory of the CoT itself. The authors propose measuring how the embedding of a reasoning trace converges toward external answer anchors, aiming to derive a continuous confidence score from text alone without access to model internals.

Methodology

The proposed framework decomposes black-box confidence into three complementary channels: Coverage (C), Geometry (G), and Verbalization (V).

1. The Geometric Channel (G)

This channel measures the "within-trace spatial commitment" of a CoT toward an external answer anchor.

Trajectory Construction: A CoT is embedded as a trajectory of overlapping sliding windows ( $W=30$ words, stride $S=15$ ) using a fixed embedding model $\phi$ .
External Anchors: In the blinded protocol, the generator reasons without seeing the option list. The option texts are embedded externally as fixed anchors.
Scoring: The method calculates the cosine distance from the penultimate sliding window of the trace to each option anchor. The "pick" is the anchor with the minimum mean distance.
Calibration: A one-parameter softmax model maps these distances to a confidence probability:
$\text{GEO}(q) = \frac{\exp(\hat{\beta}_{\text{all}} D_{\hat{y}})}{\sum_{j'} \exp(\hat{\beta}_{\text{all}} D_{j'})}$
where $\hat{\beta}_{\text{all}}$ is fitted via maximum likelihood. A strongly negative $\hat{\beta}$ indicates that closer anchors are more likely to be correct.
Penultimate Window Selection: Empirical analysis reveals that the calibration signal peaks in the penultimate window. The final window often contains literal answer verbalization, which aligns with the model's picked answer rather than the gold answer, causing the signal to invert (positive $\hat{\beta}$ ) in specific benchmarks like GPQA Diamond.

2. The Three-Channel Decomposition

To address the structural ceiling where a trace might never reach the correct hypothesis, the authors introduce an open-ended regime with a Proposer/Judge setup:

Coverage (C): A judge-mediated prior measuring whether the correct hypothesis is reachable at all. It uses a proposer LLM to generate a hypothesis pool and a judge LLM to evaluate reachability. This acts as a question-difficulty signal.
Geometry (G): As defined above, but using proposer-generated hypotheses as anchors instead of fixed dataset options.
Verbalization (V): The generator is shown its own CoT and asked to provide a numeric self-rating (0–100) of its confidence.

Fusion: The three channels are fused via a logistic regression to produce a final confidence score.

Key Contributions

Black-Box Trajectory-Confidence Score: A method to map $(\text{text}, \text{embedder}, \text{anchor set})$ to a continuous confidence score using a one-parameter softmax on sliding-window embeddings. The sign and significance of the fitted slope ( $\hat{\beta}$ ) replicate across three different embedders (Gemini, OpenAI, E5), three reasoners, and three benchmarks.
Late-Trace Commitment Signature: The authors identify a mechanistic signature where the geometry signal is deepest (most negative $\hat{\beta}$ ) in the penultimate window and inverts at the terminal window on GPQA Diamond. This validates the penultimate window as a robust scoring position that captures answer commitment before literal verbalization.
Asymmetric Three-Channel Decomposition:
- Coverage (C) and Geometry (G) carry independent signal in nearly all settings (18/18 and 16/18 combinations, respectively).
- Verbalization (V) contributes unique residual signal in only 6/18 settings, acting as a gated metacognitive channel rather than a universal pillar.
- The fusion of C+G+V at $K=4$ Pareto-improves self-consistency at $K=8$ in all six closed-set settings (median $\Delta \text{AUC} = +0.075$ ).

Results

The paper evaluates the method on MedQA-USMLE, GPQA Diamond, and MMLU-Pro using Gemini 3.1 Pro and Claude Sonnet 4.6.

Geometric Signal: The single geometric feature (G) achieves a selective-prediction AUC of 0.71–0.82 in blinded protocols, significantly outperforming self-consistency in 9/10 benchmark-model combinations when budget-matched.
Fusion Gain: The fused C+G+V score at $K=4$ $K = 4$ outperforms self-consistency at $K=8$ $K = 8$ (SC@8) across all six closed-set settings.
- Median $\Delta \text{AUC} = +0.075$ (Stouffer $z = 5.67, p = 7.1 \times 10^{-9}$ ).
- The gain persists even when the chosen answer is fixed to the SC@8 plurality pick, confirming the improvement is due to better calibration, not answer switching.
Robustness:
- The results hold across different embedders (including E5-large-v2).
- Swapping the judge LLM (GPT-5-mini to Claude Sonnet 4.6) leaves the Geometry-only AUC unchanged ( $|\Delta| \le 0.013$ ) and shifts Coverage-only AUC by at most $\pm 0.02$ , indicating the signals are not single-vendor artifacts.
- The method is robust to anchor design; using dataset-option anchors is critical, while anchors derived from the model's own output collapse to near-chance performance.

Significance and Claims

The paper claims that black-box CoT confidence can be read more reliably from a trace's geometric convergence in embedding space than from sample-vote agreement, achieving this at lower sampling cost and without access to logits or hidden states.

Key takeaways emphasized by the authors:

Structure over Sampling: Confidence is a structured signal composed of distinct channels (reachability, commitment, self-report) rather than a single scalar.
Mechanistic Insight: The geometry signal peaks before the model literally verbalizes its answer, suggesting that "answer commitment" is a distinct phase in the reasoning trajectory that can be detected via embedding space dynamics.
Practical Efficiency: Fusing C+G+V at $K=4$ provides better calibration than SC@8 while reducing total API cost by approximately 51% (due to fewer generation calls and the efficiency of the embedding pass).
Limitations: The authors modestly note that the Coverage channel is mediated by an LLM judge (introducing potential judge priors), Verbalization is not universally useful, and no within-trace score can recover a correct hypothesis if the model never proposed it (the "structural ceiling").

The work provides a trace-level confidence primitive for pipelines requiring black-box access, enabling safer abstention and selective prediction in high-stakes domains like clinical decision support.