Source-linked AI summary

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

Shuting Xie, Nathaniel Lesperance, Graham W. Taylor

arXiv:2608.22839v1cs.LGcs.AI

TL;DR

Reliable confidence estimation is difficult for black-box LLMs, especially when biodiversity predictions are hierarchical, long-tailed, and partially abstained. The paper trains lightweight estimators from proxy-LLM features and hierarchy-aware supervision, finding that supervised methods outperform a token-likelihood baseline and that rank-specific multi-head outputs perform best overall. The study is limited by sparse fine-rank supervision, a single taxonomic pipeline, and a global abstention threshold.

  • Problem

    Black-box LLMs lack reliable confidence estimates for hierarchical taxonomic reasoning, where long-tailed biodiversity predictions and partial abstentions complicate principled rejection.

  • Method

    The paper trains lightweight supervised estimators using proxy-LLM features, rank-level accuracy and hierarchical F1 supervision, and rank-aware output heads to predict rank-wise correctness.

  • Results

    Supervised estimators outperform a token-likelihood baseline across tool LLMs for micro discrimination and global-threshold selective prediction, with H3 performing best overall.

  • Takeaways & Limitations

    Estimator design that accounts for hierarchical outputs is important when a unified abstention rule is required.

  • Takeaways & Limitations

    The study is limited by sparse fine-rank supervision, a single taxonomic reasoning pipeline, and a global abstention threshold.

Abstract

from arXiv · show

Large language models (LLMs) are increasingly used for scientific decision support, yet reliable confidence estimation remains difficult in black-box settings. We study uncertainty estimation for hierarchical taxonomic reasoning generated by a black-box LLM in a long-tailed biodiversity monitoring pipeline. Using proxy features extracted by an open-source tool LLM, we train lightweight supervised estimators with hierarchy-aware supervision to predict rank-wise correctness. Across three tool LLMs, the supervised estimators consistently outperform a token-likelihood baseline for micro discrimination and selective prediction under a single global rejection threshold, improving micro AUROC from 0.57 to 0.75--0.80. The best results are achieved by a rank-specific multi-head design (H3), suggesting that accounting for hierarchical output structure is important when a unified abstention rule is required. Our code is publicly available at https://github.com/uoguelph-mlrg/hierarchy-aware-llm-uq

1 Introduction

The paper addresses confidence estimation for black-box LLMs in hierarchical taxonomic reasoning, motivated by biodiversity monitoring where long-tailed predictions and partial abstentions complicate principled rejection. It formulates rank-level uncertainty estimation and develops hierarchy-aware supervised estimators, with rank-specific multi-head outputs performing best overall.

  • Motivation: Black-box LLM confidence estimation remains difficult when hierarchical outputs span multiple semantic levels and abstention produces partial outputs.Reliable uncertainty quantification is especially important in high-stakes decision-support settings.
  • Motivation: Long-tailed biodiversity taxonomies make rare and endangered species difficult to monitor, while qualitative abstentions lack numerical confidence for principled rejection thresholds.The motivating pipeline combines VLM image captioning, retrieval-augmented generation, and LLM reasoning across taxonomic ranks.
  • Study focus: The study investigates uncertainty estimation for hierarchical taxonomic reasoning using only externally accessible signals from black-box LLMs.This frames the work around uncertainty estimation rather than improving the upstream reasoning pipeline.
  • Formulation: The paper casts black-box LLM taxonomic uncertainty estimation as rank-level selective prediction under partial-path abstention.The formulation reflects variable-depth outputs in which the system may abstain at an uncertain rank.
  • Method: Lightweight supervised estimators use proxy-LLM features, hierarchy-aware supervision, and rank-aware output parameterizations to predict rank-wise correctness.These design choices align the estimator with the hierarchical structure of the outputs.
  • Result: Supervised estimators outperform a token-likelihood baseline for global-threshold selective prediction, with a rank-specific multi-head design performing best overall.The comparison is conducted on a realistic long-tailed arthropod pipeline.

2 Related Work

Prior uncertainty-estimation methods use internal signals, self-verbalized confidence, or repeated-response consistency, while taxonomic prediction adds hierarchical, long-tailed, and partial-output challenges. This paper studies uncertainty estimation for such hierarchical outputs rather than improving the upstream reasoning pipeline.

  • LLM uncertainty estimation: White-box uncertainty methods use logits or hidden states, but these signals are unavailable for proprietary API models.Black-box alternatives commonly use self-verbalized confidence or consistency and semantic clustering across sampled responses.
  • LLM uncertainty estimation: Self-verbalized confidence may not faithfully reflect intrinsic uncertainty, while repeated-response consistency or semantic clustering can be prohibitively expensive.These constraints motivate uncertainty estimation methods that operate with externally accessible signals.
  • Hierarchical taxonomic prediction: Taxonomic reasoning complicates uncertainty estimation because predictions are hierarchical, long-tailed, and often partially specified through abstention.Prior hierarchical-classification work uses ancestor overlap and partial credit, and recent work extends this perspective to taxonomy-aware evaluation.
  • Hierarchical taxonomic prediction: The paper complements upstream taxonomic-reasoning work by estimating uncertainty for hierarchical outputs with hierarchy-aware signals as auxiliary supervision.Its focus is uncertainty estimation rather than improving the upstream reasoning pipeline.

3 Methodology

The study formulates black-box taxonomic uncertainty estimation as rank-level selective prediction and trains lightweight estimators from tool-LLM proxy features with hierarchy-aware supervision. It evaluates rank-wise correctness using AUROC and global-threshold risk–coverage metrics.

  • Pipeline: The pipeline maps target-LLM taxonomic outputs and tool-LLM proxy features to confidence scores for uncertainty estimation.
  • Supervision labels: Each instance receives binary correctness supervision and hierarchical accuracy based on taxonomic overlap, with constant kingdom and phylum ranks excluded.
  • Proxy features: Tool LLMs provide representation features from hidden states and distributional features based on entropy, negative log-probability, and token probability.
  • Supervised estimators: H3 uses five rank-specific binary heads plus an auxiliary regression head, whereas H1 and H2 use shared binary-head designs with progressively richer supervision.
  • Evaluation protocol: Rank-level evaluation treats each image–rank pair as one example with a correctness label and confidence score.
  • Evaluation protocol: Global-threshold risk–coverage pools validation examples across ranks, while AUROC is reported as both pooled micro AUROC and per-rank-average macro AUROC.

4 Results

Supervised uncertainty estimators outperform the mean-NLL baseline across tool LLMs for micro discrimination and global-threshold selective prediction. H3 provides the strongest overall risk–coverage trade-off, while gains are less consistent under macro AUROC.

  • Supervised estimators consistently and substantially outperform the mean-NLL baseline in micro AUROC across all three tool LLMs.
  • H3 achieves the highest mean micro AUROC overall, with a small margin over H1 for Qwen3 and clearer gains for Gemma 7B and GPT-OSS 20B.
  • Under a single global rejection threshold, supervised estimators achieve lower selective risk than the mean-NLL baseline across a broad coverage range.The largest gains occur at low-to-mid coverage.
  • H3 consistently yields the best overall risk–coverage trade-off, while H1 and H2 remain better than the baseline throughout.
  • Gains are less consistent under rank-balanced macro AUROC, although calibration results generally favour supervised estimators over the mean-NLL baseline.

5 Discussion

Micro and macro AUROC diverge because rank imbalance favors coarse, high-coverage ranks, while deeper ranks have weaker training signals. For unified-threshold deployment, H3’s rank-specific heads perform best, but sparse fine-rank supervision and limited evaluation scope constrain the conclusions.

  • Rank imbalance makes micro AUROC favor high-coverage coarse ranks, whereas macro AUROC weights ranks equally and exposes weaknesses at deeper ranks.Frequent abstentions reduce fine-rank training signal and increase variance.
  • H3 performs best overall for unified-threshold deployment, while H2 provides modest gains over H1 for most tool LLMs except GPT-OSS 20B.
  • Because H2 and H3 share auxiliary supervision, their pattern suggests rank-specific output heads matter more than the auxiliary hierarchical target alone.
  • The study is limited by sparse fine-rank supervision, a single taxonomic reasoning pipeline, and a global abstention threshold.Future work should test richer fine-rank coverage, wider distribution shifts, and rank-specific decision rules.

6 Conclusion

The paper studies uncertainty estimation for black-box LLMs performing hierarchical taxonomic reasoning under long-tailed data and partial-path abstention. Lightweight supervised estimators using proxy features outperform an NLL baseline, with rank-specific multi-head output performing best overall.

  • The study addresses UQ for black-box LLMs in hierarchical taxonomic reasoning with long-tailed data and partial-path abstention.
  • Lightweight supervised estimators use proxy features from a tool LLM and hierarchy-aware signals to predict rank-wise correctness.
  • Across tool LLMs, supervised estimators outperform an NLL baseline in micro discrimination and global-threshold selective prediction.
  • A rank-specific multi-head design performs best overall, highlighting estimator design for hierarchical outputs in unified-threshold deployment.

A Implementation details

The implementation trains lightweight MLP estimators with Adam, early stopping, hierarchical supervision, and an L1 penalty encouraging sparse input-feature use. Feature selection is applied only to hidden-state representation features.

  • MLPs use Adam with learning rate 10^-3, weight decay 3 × 10^-4, batch size 128, and up to 200 epochs.
  • Early stopping monitors validation micro AUROC with patience 5 and minimum improvement 10^-4.
  • Hierarchical supervision uses λ = 0.1, while an L1 penalty of 10^-4 encourages sparse input-to-hidden weights.
  • Feature selection is applied only to hidden-state representation features, combining the top-100 features from each listed selection method.

B Upstream taxonomic pipeline statistics

Table 2 reports rank coverage and prediction accuracy in the upstream system before filtering.

  • Table 2 reports rank coverage and prediction accuracy before filtering.

C Tool LLM configurations and extracted features

The paper describes tool LLM configurations and feature types used for uncertainty estimation, including representation and distributional features.

  • Table 3 lists each tool LLM and its exact checkpoint identifier.
  • Tool LLMs extract representation, distribution, and hybrid feature types.

D Additional UQ results

Additional results cover macro AUROC, per-rank AUROC, and calibration before and after temperature scaling.

  • Table 4 reports all-rank aggregated macro AUROC across tool LLMs.Results are reported as mean ± std over 5 random seeds.
  • Figure 4 reports per-rank AUROC before calibration across 5 random seeds.Horizontal bars indicate ±1 standard deviation.
  • Table 5 reports calibration metrics before and after temperature scaling.Entries are mean with standard deviation across seeds, and arrows indicate whether lower or higher values are preferred.
Loading 2608.22839v1…