🤖 machine learning

Hierarchy-Aware Supervised Uncertainty Estimation for Black-box LLM Taxonomic Reasoning

This paper proposes a hierarchy-aware supervised uncertainty estimation framework that leverages proxy features from open-source LLMs to train lightweight estimators, significantly outperforming token-likelihood baselines in predicting rank-wise correctness for black-box LLM taxonomic reasoning within biodiversity monitoring.

Original authors: Shuting Xie, Nathaniel Lesperance, Graham W. Taylor

Published 2026-08-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shuting Xie, Nathaniel Lesperance, Graham W. Taylor

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

In the high-stakes world of scientific discovery, from tracking endangered species to diagnosing rare diseases, computers are increasingly asked to make critical judgments. These machines, known as large language models, are powerful tools that can read vast amounts of text and generate intelligent-sounding answers. However, a significant problem remains: when these models are used as "black boxes"—systems where we can see the input and the output but cannot peek inside to understand how they reached a conclusion—it is incredibly difficult to know when to trust them. In fields like ecology, where a wrong guess about a species could lead to wasted conservation efforts or missed protection for a rare animal, knowing how confident a computer is in its answer is just as important as the answer itself. This uncertainty becomes even harder to measure when the task involves complex hierarchies, such as the biological classification of life, where a mistake at a broad level (like the family of an animal) might be less severe than a mistake at a specific level (like the exact species), yet both matter.

A team of researchers set out to solve this problem for a specific, challenging scenario: identifying arthropods, a group of invertebrates that includes insects and spiders, many of which are rare or endangered. They worked with a system designed to identify these creatures from images, a process that involves several steps. First, a vision model describes the image in words. Then, a retrieval system gathers relevant facts from online databases. Finally, a large language model uses this information to propose a taxonomic path, which is a list of names ranging from broad categories like "insect" down to the specific species. The system is designed to be cautious; if it is unsure at any step, it stops and admits it doesn't know, rather than guessing and potentially leading a researcher astray. The researchers' goal was to build a way to assign a numerical confidence score to these predictions, allowing scientists to automatically filter out uncertain cases and send them to human experts for review.

The challenge was that the main language model used for these predictions was a black box, meaning the researchers could not access its internal calculations to measure uncertainty directly. To get around this, they developed a clever workaround. They used a separate, open-source language model as a "tool" to analyze the conversation between the image description and the final prediction. This tool model acted like a translator, extracting specific clues and patterns from the text that hinted at whether the prediction was likely correct or incorrect. These clues served as a proxy, a stand-in signal for the hidden confidence of the main model.

Using these extracted clues, the researchers trained a small, lightweight computer program to act as a supervisor. This supervisor learned to look at the text patterns and predict, for each level of the biological hierarchy, whether the main model's guess was right or wrong. They tested three different designs for this supervisor. One design treated every level of the hierarchy the same. A second design added a helper task to teach the system about the overall structure of the classification. The third design was more specialized, giving the system a unique "head" or dedicated processor for each specific level of the hierarchy, from broad categories down to specific species.

The results showed that this supervised approach was far more effective than simply looking at how likely the model thought its own words were to be correct. In a direct comparison, the new method consistently did a better job of distinguishing between correct and incorrect answers. The most successful design was the specialized one with unique heads for each rank. This suggests that when a system needs to make a single, unified decision about whether to accept or reject a whole chain of reasoning, it is crucial to account for the fact that different levels of the hierarchy have different structures and difficulties. The specialized supervisor learned that a mistake at the species level is different from a mistake at the family level, and it adjusted its confidence scores accordingly.

This work demonstrates that even when we cannot see inside a powerful artificial intelligence, we can still build reliable safety nets by using other tools to interpret its behavior. By training a simple estimator on the patterns of the model's output, the researchers created a system that can reliably flag uncertain predictions in a long-tailed, hierarchical setting. This means that in real-world applications, such as monitoring biodiversity, scientists can now rely on a clear, numerical threshold to decide when to trust a computer's identification and when to call in a human expert, making the use of artificial intelligence in conservation both more powerful and more responsible.

Technical Summary: Hierarchy-Aware Supervised Uncertainty Estimation for Black-Box LLM Taxonomic Reasoning

Problem Statement

The paper addresses the challenge of reliable uncertainty quantification (UQ) for Large Language Models (LLMs) operating in black-box settings (accessible only via APIs) within hierarchical taxonomic reasoning tasks. This is particularly critical in biodiversity monitoring, where predictions span multiple semantic levels (taxonomic ranks) and the data distribution is strongly long-tailed.

Existing solutions face two primary limitations in this context:

Black-box constraints: White-box methods relying on logits or hidden states are inapplicable to proprietary API models.
Hierarchical complexity: Standard uncertainty metrics often fail to account for the structure of hierarchical outputs, where models may abstain at uncertain ranks, producing partial paths rather than flat labels. Furthermore, abstention in hierarchical settings yields partial outputs rather than a single rejection, complicating the definition of confidence thresholds.

Current black-box approaches, such as self-verbalized confidence or semantic clustering of multiple sampled responses, are either unfaithful to intrinsic uncertainty or prohibitively expensive in API-based pipelines.

Methodology

The authors formulate uncertainty estimation as a supervised prediction problem using proxy features extracted from prompt-response pairs. The core pipeline involves:

Upstream System: A target LLM (GPT-4o) generates taxonomic predictions from image captions and retrieved context. It employs a coarse-to-fine reasoning strategy, abstaining at the first uncertain rank to produce variable-depth partial paths.
Proxy Feature Extraction: An open-source "tool LLM" (e.g., Gemma 7B, GPT-OSS 20B, or Qwen3 30B) processes the prompt-response pairs to extract two families of features:
- Representation features: Averaged and last-token hidden states from the middle layer.
- Distributional features: Statistics of entropy, negative log-probability, and raw token probabilities.
Supervised Estimators: A lightweight Multi-Layer Perceptron (MLP) is trained to predict rank-wise correctness using the extracted features. The architecture includes:
- Shared Trunk: A two-layer shared backbone (hidden size 16, ReLU, dropout 0.3).
- Heads: Task-specific output heads. The primary head is a sigmoid classifier predicting binary correctness ( $y \in \{0, 1\}$ ).
- Hierarchy-Aware Supervision: To leverage the hierarchical structure, an auxiliary regression head predicts a hierarchical F1 score ( $h$ ) based on taxonomic overlap (ancestor intersection). The loss function combines Binary Cross-Entropy (BCE) for correctness and Mean Squared Error (MSE) for the hierarchical score:
  $L = BCE(\hat{p}, y) + \lambda \|\hat{h} - h\|_2^2$
Architectural Variants:
- H1: Single binary head trained only on BCE.
- H2: Single binary head + auxiliary regression head trained on combined loss.
- H3: Rank-specific multi-head design with five binary heads (one per rank) plus one auxiliary regression head. This allows the model to output correctness probabilities for all ranks in a single forward pass.

Key Contributions

Formulation: The authors cast black-box LLM taxonomic UQ as a rank-level selective prediction problem under partial-path abstention, explicitly modeling uncertainty at the rank level to capture error propagation.
Method: They develop lightweight supervised estimators that utilize proxy-LLM features and hierarchy-aware supervision (via auxiliary hierarchical F1 targets) to predict rank-wise correctness without access to the target model's internal states.
Design Insight: Through empirical evaluation, they demonstrate that a rank-specific multi-head design (H3) yields the best performance, suggesting that accounting for the hierarchical output structure is crucial when a unified abstention rule is required.

Results

The study was conducted on a dataset of 829 arthropod images from the Rare Species dataset, covering seven taxonomic ranks with a focus on conservation-biased (endangered) species.

Performance vs. Baseline: Across three different tool LLMs, the supervised estimators consistently outperformed a token-likelihood baseline (Negative Mean Token NLL).
- Micro AUROC: Improved from a baseline of 0.573 to a range of 0.75–0.80.
- Selectivity: Under a single global rejection threshold, the supervised estimators achieved significantly lower selective risk (error rate among accepted examples) across a broad coverage range compared to the baseline.
Architecture Comparison:
- H3 (Rank-specific multi-head) achieved the highest mean micro AUROC overall (e.g., 0.796 with Gemma 7B).
- While H2 (auxiliary supervision) showed modest gains over H1, the performance gap between H1 and H3 was more pronounced, indicating that rank-specific output heads are more impactful than auxiliary hierarchical targets alone for unified-threshold deployment.
Metric Divergence: The paper notes a divergence between Micro AUROC (dominated by high-coverage coarse ranks) and Macro AUROC (rank-balanced). The supervised estimators performed well on Micro AUROC but showed less consistent gains on Macro AUROC, particularly at deeper ranks where sparse supervision and frequent abstentions increase variance.

Significance and Claims

The paper claims that for unified-threshold deployment in hierarchical settings, the design of the uncertainty estimator is as critical as the uncertainty signal itself. Specifically:

Practicality: The proposed method offers a practical solution for black-box LLMs where internal signals are unavailable and multi-sampling is too costly.
Hierarchy Awareness: Simply treating hierarchical outputs as flat labels is suboptimal. The results suggest that explicitly modeling the rank-specific nature of the output (via H3) leads to better discrimination and risk-coverage trade-offs.
Deployment Readiness: The approach supports principled rejection thresholds for routing uncertain cases to domain experts, a critical requirement for high-stakes domains like ecology and conservation.

The authors remain modest regarding generalizability, noting that their conclusions are based on a single taxonomic reasoning pipeline and a specific global abstention threshold. They acknowledge that sparse fine-rank supervision remains a limitation and that future work is needed to assess generalization across wider distribution shifts and rank-specific decision rules.