Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Ambiguity in LLMs is a concept missing problem

Zhibo Hu, Chen Wang, Yanfeng Shu, Hye-Young Paik, Liming Zhu

TL;DR

Ambiguity in natural language presents a major hurdle for mapping to structured data with LLMs. The authors propose a representation-based framework that detects ambiguity by examining latent-space concept differences using a sparse autoencoder (SAE) and a path kernel over concept activations ($K_{ ext{path}}$), with a concept mask to isolate task-relevant notions; they further show that activating missing concepts increases semantic entropy $H_{ ext{sem}}$ and improves discriminability via distances $D_1$ and $D_2$. They extend this approach to mitigate ambiguity in agentic tool calling by predicting missing concepts with a LightGBM-based predictor to guide API retrieval through a union-joint ranking of calls. Empirical results on AMBROSIA for text-to-SQL and Gorilla API bench demonstrate state-of-the-art performance, outperforming dense embedding baselines and certain fine-tuned models. Overall, the work provides a scalable, concept-centered method for detecting and resolving ambiguity in LLM-driven structured data tasks, with broad implications for robust NLU and tool integration.

Abstract

Ambiguity in natural language is a significant obstacle for achieving accurate text to structured data mapping through large language models (LLMs), which affects the performance of tasks such as mapping text to agentic tool calling and text-to-SQL queries. Existing methods to ambiguity handling either rely on the ReACT framework to obtain correct mappings through trial and error, or on supervised fine-tuning to bias models toward specific tasks. In this paper, we adopt a different approach that characterizes representation differences of ambiguous text in the latent space and leverages these differences to identify ambiguity before mapping them to structured data. To detect sentence-level ambiguity, we focus on the relationship between ambiguous questions and their interpretations. Unlike distances calculated by dense embeddings, we introduce a new distance measure based on a path kernel over concepts. With this measurement, we identify patterns to distinguish ambiguous from unambiguous questions. Furthermore, we propose a method for improving LLM performance on ambiguous agentic tool calling through missing concept prediction. Both achieve state-of-the-art results.

Ambiguity in LLMs is a concept missing problem

TL;DR

Ambiguity in natural language presents a major hurdle for mapping to structured data with LLMs. The authors propose a representation-based framework that detects ambiguity by examining latent-space concept differences using a sparse autoencoder (SAE) and a path kernel over concept activations (), with a concept mask to isolate task-relevant notions; they further show that activating missing concepts increases semantic entropy and improves discriminability via distances and . They extend this approach to mitigate ambiguity in agentic tool calling by predicting missing concepts with a LightGBM-based predictor to guide API retrieval through a union-joint ranking of calls. Empirical results on AMBROSIA for text-to-SQL and Gorilla API bench demonstrate state-of-the-art performance, outperforming dense embedding baselines and certain fine-tuned models. Overall, the work provides a scalable, concept-centered method for detecting and resolving ambiguity in LLM-driven structured data tasks, with broad implications for robust NLU and tool integration.

Abstract

Ambiguity in natural language is a significant obstacle for achieving accurate text to structured data mapping through large language models (LLMs), which affects the performance of tasks such as mapping text to agentic tool calling and text-to-SQL queries. Existing methods to ambiguity handling either rely on the ReACT framework to obtain correct mappings through trial and error, or on supervised fine-tuning to bias models toward specific tasks. In this paper, we adopt a different approach that characterizes representation differences of ambiguous text in the latent space and leverages these differences to identify ambiguity before mapping them to structured data. To detect sentence-level ambiguity, we focus on the relationship between ambiguous questions and their interpretations. Unlike distances calculated by dense embeddings, we introduce a new distance measure based on a path kernel over concepts. With this measurement, we identify patterns to distinguish ambiguous from unambiguous questions. Furthermore, we propose a method for improving LLM performance on ambiguous agentic tool calling through missing concept prediction. Both achieve state-of-the-art results.
Paper Structure (21 sections, 14 equations, 13 figures, 1 table, 1 algorithm)

This paper contains 21 sections, 14 equations, 13 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminary
  3. Methodology
  4. Ambiguity Resolution as a Missing Concept Problem
  5. Effect of adding missing concepts into the latent space
  6. Representation-based Ambiguity Detection
  7. Predicting Missing Concepts to Mitigate Ambiguity
  8. Experiments
  9. Ambiguity Detection
  10. Agentic Tool Calling
  11. Conclusion
  12. Algorithm: Measuring Semantic Entropy by LLM
  13. Ambiguity Detection: Prompt examples and Additional Results
  14. Few-shot Examples used in Prompts to Inject Concepts
  15. Distance Relationship
  16. ...and 6 more sections

Figures (13)

  • Figure 1: An example to show the difference of the distance measurement on triplets for ambiguous question ($q$, $i_1$, $i_2$) and two kind of unambiguous question ($q'$, $i'_1$, $i'_2$), ($q"$, $i"_1$, $i"_2$). For similar queries ambiguous $q$ and unambiguous $q'$, $q"$, we expect $\overline{(\mathrm{D}(q, i_1), \mathrm{D}(q, i_2), \mathrm{D}(i_1, i_2))}$$>$$\overline{(\mathrm{D}(q', i'_1), \mathrm{D}(q', i'_2), \mathrm{D}(i'_1, i'_2))}$ and $|\mathrm{D}(q, i_1) - \mathrm{D}(q, i_2)| \ll |\mathrm{D}(q", i"_1) - \mathrm{D}(q", i"_2)|$ (the overline means average).
  • Figure 2: Entropy/ambiguity change with missing concepts added.
  • Figure 3: The workflow of the path kernel calculation with SAE.
  • Figure 4: Tool calling framework based on missing concept prediction in ambiguous questions.
  • Figure 5: Distribution of average distances calculated using the path kernel method with SAE, and dense vector-based methods with SFR-Embedding-Mistral and LLama-3.3-70B-Instruct. A smaller overlapping area indicates a stronger ability to distinguish ambiguous from unambiguous questions.
  • ...and 8 more figures

Theorems & Definitions (1)

  • Definition A.1: Semantic Entropy