PaperlayerView exact revision on arXiv

Source-linked AI summary

T-Eval: Evaluating the Tool Utilization Capability of Large Language Models Step by Step

Zehui Chen, Weihua Du, Wenwei Zhang, Kuikun Liu, Jiangning Liu, Miao Zheng, Jingming Zhuo, Songyang Zhang, Dahua Lin, Kai Chen, Feng Zhao

arXiv:2312.14033v3cs.CL

TL;DR

Evaluation of tool use remains limited by holistic final-output or single-step assessments, despite the multi-step nature of tool use. T-Eval evaluates six capabilities separately using a step-by-step benchmark. Its evaluations are reported to align with overall performance while providing finer-grained analysis of model abilities.

  • Problem

    Existing evaluations often judge only final outputs or single-step calls, omitting intermediate steps in complex multi-tool tasks and potentially varying with real-time tool factors.

    Source: §1 · p. 1

  • Method

    T-Eval uses human-verified annotations and tailored protocols to assess planning, reasoning, retrieval, understanding, instruction following, and review separately.

    Source: §1 · pp. 1–2

  • Results

    T-Eval's individual and comprehensive ability evaluations are reported as consistent, and its single-index evaluation is consistent with end-to-end performance.

    Source: §1 · p. 1

  • Takeaways & Limitations

    Higher individual ability scores are linked to better performance on complex downstream tasks, while T-Eval offers finer-grained analysis of tool-use abilities.

    Source: §1 · p. 2

Abstract

from arXiv · show

Large language models (LLM) have achieved remarkable performance on various NLP tasks and are augmented by tools for broader applications. Yet, how to evaluate and analyze the tool-utilization capability of LLMs is still under-explored. In contrast to previous works that evaluate models holistically, we comprehensively decompose the tool utilization into multiple sub-processes, including instruction following, planning, reasoning, retrieval, understanding, and review. Based on that, we further introduce T-Eval to evaluate the tool utilization capability step by step. T-Eval disentangles the tool utilization evaluation into several sub-domains along model capabilities, facilitating the inner understanding of both holistic and isolated competency of LLMs. We conduct extensive experiments on T-Eval and in-depth analysis of various LLMs. T-Eval not only exhibits consistency with the outcome-oriented evaluation but also provides a more fine-grained analysis of the capabilities of LLMs, providing a new perspective in LLM evaluation on tool-utilization ability. The benchmark will be available at https://github.com/open-compass/T-Eval.

1 Introduction

T-Eval addresses limits in evaluating tool use from final answers or single-step calls by assessing intermediate capabilities separately. Experiments report consistency with overall evaluation while offering finer-grained analysis of model abilities and bottlenecks.

1 Introduction

  • Final-answer evaluation omits intermediate steps, while single-step benchmarks miss the multi-tool planning and execution common in real-world tasks.Reliance on real-time tools can also introduce evaluation variance from API instability or temporal shifts.Source: §1 · p. 1
  • T-Eval evaluates planning, reasoning, retrieval, understanding, instruction following, and review as separate tool-use capabilities.Its step-by-step protocols are based on human-verified tool-use annotations.Source: §1 · pp. 1–2
  • Experiments report that T-Eval's individual and comprehensive ability evaluations are consistent, with higher individual scores linked to better complex-task performance.The paper also reports consistency with outcome-oriented evaluation.Source: §1 · pp. 1–2
  • The benchmark uses a multi-agent data-generation pipeline with human verification to reduce external influences and support more stable, fair assessment.The process is presented as a way to mitigate evaluation effects from external factors.Source: §1 · p. 2
  • T-Eval's decomposition provides finer-grained insight into current LLMs' tool-use capabilities and bottlenecks.The authors report extensive experiments across various LLMs.Source: §1 · p. 2

2 T-Eval

T-Eval decomposes tool use into abilities and supplies tailored evaluation protocols, supported by human-in-the-loop benchmark construction. Its single-index evaluation is reported to be consistent with end-to-end performance evaluation.

2.1 Evaluation Decomposition

  • The decomposition treats planning, reasoning, retrieval, understanding, instruction following, and review as distinct abilities in multi-step tool use.Planning guides actions; reasoning handles context; retrieval selects tools; understanding interprets documentation and parameters; instruction following formats calls; review assesses responses.Source: §2.1 · pp. 2–3

2.2.1 Definition

  • A query datum is a tool list and query, while its solution path records per-step thought, action, observation, and review alongside a final answer.A plan is the proposed sequence of actions.Source: §2.2.1 · p. 3

2.2.2 Single-Index Evaluation

  • Planning compares predicted and gold action sequences by semantic similarity and matching, scoring the longest ordered sequence of paired actions.Reasoning generates the next thought from a solution-path prefix, while retrieval selects the next tool against the gold choice.Source: §2.2.2 · p. 3
  • Understanding evaluates predicted parameters against gold parameters, instruction following scores tool names and parameter values in formatted calls, and review classifies tool-response success or error.Review uses five choices: Success, Internal Error, Input Error, Irrelevant Response, and Unable to Accomplish.Source: §2.2.2 · p. 3

2.2.3 End-to-End Evaluation

  • Single-index evaluation is reported to be consistent with end-to-end performance measured by win rate against GPT-3.5.The end-to-end test asks models to generate a complete solution path and final answer.Source: §2.2.3 · pp. 3–4

2.3 Dataset Construction

  • Construction combines tool selection and documentation, GPT-3.5 query generation refined by GPT-4, and planner-executor-reviewer annotation with human checking.The annotation roles divide next-step planning, tool execution, and response review; the authors say this decomposition reduces errors compared with conventional CoT or ReAct approaches.Source: §2.3.1 · p. 4, §2.3.2 · p. 4, §2.3.3 · pp. 4–5

2.3.4 Inclusive Difficulty Evaluation

  • The instruction-following evaluation offers an easy semantic-text level and a stricter JSON level that checks exact tool names and parameters.The two levels address parse failures that can distort scores, particularly for some small-scale models.Source: §2.3.4 · p. 5

2.4 Dataset Summary

  • The benchmark contains 23,305 test cases across six ability subsets, selected from 1,500 initial pairs after two human-verification rounds yielded 553 pairs.The subsets are INSTRUCT, RETRIEVE, PLAN, REASON, UNDERSTAND, and REVIEW.Source: §2.4 · p. 5

3 Experiments

The experiments compare commercial and open-source LLMs on T-Eval, where GPT-4 achieves the highest reported overall score. Results also show scaling gains alongside persistent gaps and weaknesses in format following, retrieval, and review.

3.1 Experimental Setup

  • The evaluation covers 20 API-based commercial and open-source LLMs, including three representative commercial models and a range of open-source model families.The supplied setup names GPT-3.5, GPT-4, and Claude2 among commercial models and lists open-source families including LLaMA2, Qwen, and Mistral.Source: §3.1 · p. 5

3.2 Main Results

  • As LLaMA2 and Qwen scale up, their tool-utilization scores rise, but a clear gap from GPT-4 remains.Qwen-72B reaches 71.4% overall, reducing the gap to API-based models.Source: Figure 3 · p. 5, §3.2 · pp. 5–6
  • 86.4 overall score gives GPT-4 the highest result among the evaluated models.Table 1 defines Overall as the average of metrics across all subsets.Source: §3.2 · p. 5, Table 1 · p. 6
  • Open-source models often struggle with JSON protocols, while Qwen-72B's string-format understanding score is 84.5 versus GPT-4's 83.2, and its JSON result is over 20 points lower.The text also notes that few models score highly on INSTRUCT under both string and JSON formats.Source: §3.2 · p. 6
  • Tool retrieval remains difficult: Qwen-72B reaches 65.0%, more than 20 points below GPT-3.5.Source: §3.2 · p. 6
  • Most models score 50%–60% on review versus GPT-4's 95%, and the training-data comparisons suggest that instruction quality matters alongside model scale.At 7B, high-quality general instructions substantially help, but gains diminish or reverse at larger scales; AgentLM outscores CodeLLaMA, yet neither significantly surpasses Vicuna.Source: §3.2 · pp. 6–7

4 Discussion

T-Eval separates format compliance from problem solving and offers an inclusive evaluation protocol for models that struggle with strict JSON. Its scores show trends similar to holistic ToolBench win rates while exposing ability-domain gaps those evaluations may miss.

4.1 Format Following v.s Problem Solving

  • ChatGLM3-6B and Baichuan2-7B score roughly 80% on JSON INSTRUCT, yet struggle to produce valid JSON on PLAN and REASON.The results suggest format following may need to be learned jointly with problem-solving tasks.Source: §4.1 · p. 7

4.2 Inclusive Evaluation Protocol

  • Baichuan2-13B scores 65.6% on PLAN with string evaluation, 25 points above JSON evaluation, showing strict-format scores can understate planning ability.The inclusive protocol uses semantic sentence analysis rather than strict format matching.Source: §4.2 · p. 7

4.3 Comparison to Other Benchmarks

  • T-Eval scores show trends similar to ToolBench win rates, but Qwen-7B’s 52% win rate over GPT-3.5-turbo coexists with ability-domain gaps under human judgment.The authors argue that T-Eval reveals these divergences and provides more detail about tool-utilization abilities.Source: Figure 4 · p. 8, §4.3 · p. 8

5 Related Work

Related work covers tool augmentation, LLM evaluation, and prompting LLMs as agents. T-Eval is positioned as a fine-grained benchmark for tool-utilization capabilities.

Augmenting LLMs with Tools

  • Tool augmentation ranges from fine-tuning models with tool-specific tokens to in-context tool descriptions and demonstrations, the latter paradigm being T-Eval’s focus.The token-based approach requires substantial tool-relevant data and can struggle to adapt to newly appearing tools.Source: §5 · p. 8

Evaluating LLMs

  • Tool-use benchmarks assess response quality, call accuracy, or solution paths; T-Eval is presented as the first dedicated benchmark for fine-grained tool-utilization capabilities.Examples include ToolQA, Gorilla, API-Bank, and ToolBench.Source: §5 · p. 8

Prompting LLMs as Agents

  • Agent-prompting work uses methods such as Chain of Thought, Tree of Thought, and ReAct to guide LLM reasoning and agent behavior.The related work also names ReWOO, SwiftSage, DyLAN, and DP-LLM.Source: §5 · p. 8

6 Conclusion

The paper introduces T-Eval as a fine-grained benchmark that evaluates tool utilization step by step along model abilities. Its analysis identifies bottlenecks in current LLMs’ tool learning.

6 Conclusion

  • T-Eval is a comprehensive, fine-grained benchmark that disentangles tool-utilization tasks by model ability and uses dedicated protocols to evaluate them.The conclusion describes this step-wise evaluation as a way to uncover the evaluated models’ abilities.Source: §6 · p. 8
  • The authors report that T-Eval’s analysis pinpoints current LLMs’ tool-learning bottlenecks and offers insights for developing tool agents.Source: §6 · p. 9

A.1 Dataset Statistics

T-Eval’s dataset comprises 23,305 test cases across six capability subsets, and its annotation paths average 5.8 tool-calling steps per query. The appendix also points readers to implementation details and prompt demonstrations.

A.1 Dataset Statistics

  • T-Eval contains 23,305 test cases across INSTRUCT, PLAN, REASON, RETRIEVE, UNDERSTAND, and REVIEW, originating from 533 high-quality query-solution annotation pairs.Source: §A.1 · p. 11
  • The annotation paths cover all tool sets and average 5.8 tool-calling steps per query.Figure 5 depicts tool-call category and step distributions across the paths.Source: §A.1 · p. 11, Figure 5 · p. 11

B Implementation Details

  • End-to-end evaluation uses ReAct with a maximum of 20 action steps; single-index evaluation uses multi-turn conversations.Unless specified otherwise, the evaluated open-source models are chat or instruct versions.Source: B Implementation Details · p. 11

B.1 Prompts Demonstration

C Detailed Evaluation Metrics

T-Eval evaluates six tool-use dimensions with dedicated metrics and JSON or string response formats. The section describes how each dimension is scored and notes that its API documentation is manually expanded.

C Detailed Evaluation Metrics

  • T-Eval separates tool utilization into INSTRUCT, PLAN, REASON, RETRIEVE, UNDERSTAND, and REVIEW, with JSON and string difficulty formats.JSON requires standard JSON responses; string allows a looser response format.Source: C Detailed Evaluation Metrics · p. 12

C.1 INSTRUCT

  • INSTRUCT awards 0.5 for format compliance and up to 0.5 according to the proportion of correctly matched parameters.The evaluator checks format first, then parameter accuracy.Source: §C.1 · p. 12

C.2 PLAN

  • PLAN scores match predicted and gold actions by semantic similarity, maximum-weight matching, and the longest increasing subsequence.Sentence-BERT measures similarity; the implementation uses a 0.7 link threshold and sets the tool-name weight β to 0.75.Source: §C.2 · pp. 12–13

C.3 REASON

  • REASON scores the next thought against the gold thought using Sentence-BERT; JSON output also supplies the next tool and parameters.In string format, the model outputs only the next thought.Source: §C.3 · p. 13

C.4 RETRIEVE

  • RETRIEVE receives a binary score for whether the predicted next tool name matches the gold tool name.In string format, the gold thought is given and only the tool name is generated.Source: §C.4 · p. 13

C.5 UNDERSTAND

  • UNDERSTAND scores the next tool’s predicted parameters by their Sentence-BERT similarity to the gold parameters.In string format, the gold thought and tool name are given, and the model generates the parameters.Source: §C.5 · p. 13

C.6 REVIEW

  • REVIEW classifies each tool response as Success, Internal Error, Input Error, Irrelevant Response, or Unable to Accomplish, scoring correct classifications as 1.The classification is based on the preceding thought and tool response.Source: §C.6 · p. 13

D API Documentation

  • The authors manually curate extensive API documentation, making descriptions more detailed than official RapidAPI documentation to address failures from incomplete benchmark documentation.BINGMap is shown as an example.Source: D API Documentation · p. 13

B.2.3 Dataset Demonstration – REASON

The supplied examples identify prompts for REASON, RETRIEVE, and UNDERSTAND, alongside dataset demonstrations that provide tool access.

B.2.3 Dataset Demonstration – REASON

  • The appendix identifies Figure 11 as an example prompt in the REASON dataset.The demonstration prompt provides access to APIs.Source: Figure 11 · p. 16, §B.2.3 · p. 16

Dataset demonstrations

  • The appendix also labels example prompts for RETRIEVE and UNDERSTAND datasets in Figures 12 and 13.The passages identify these as prompt examples without describing their contents.Source: Figure 12 · p. 17, Figure 13 · p. 17

B.2.6 Dataset Demonstration – REVIEW

The supplied appendix passages identify an example prompt in the REVIEW dataset and an example BINGMap API document.

B.2.6 Dataset Demonstration – REVIEW

  • Figure 14 is identified as an example prompt in the REVIEW dataset.The demonstration passage introduces a task and its corresponding responses.Source: Figure 14 · p. 18, §B.2.6 · p. 18
  • Figure 15 is identified as an example API document for BINGMap.The caption identifies the document example but gives no further details.Source: Figure 15 · p. 19

Related papers

Loading 2312.14033v3…