MINT: Evaluating LLMs in Multi-turn Interaction with Tools and Language Feedback
MINT evaluates LLMs' multi-turn tool use and feedback leveraging, with performance gains of 2-17% across 20 models, revealing training impacts and evaluation gaps.
Key Findings
Methodology
The framework combines Python code execution for tool calls and GPT-4 simulated user feedback to create realistic multi-turn scenarios. It integrates diverse datasets—reasoning, coding, decision-making—curated into 586 challenging instances. Models are evaluated across 1-5 interaction turns, analyzing performance improvements via linear regression. The setup assesses models' abilities to utilize tools and feedback, comparing open- and closed-source architectures, and examining training effects like SIFT and RLHF.
Key Results
- All models benefit from multi-turn interactions, with tool use improving success rates by 1-8% per turn and natural language feedback boosting 2-17%.
- Superior single-turn performance does not guarantee multi-turn success; some models excel in multi-turn despite weaker single-turn results.
- Open-source models lag behind closed-source counterparts, with the best open-source model still 8.7% below the top closed-source model in success rate.
- Training methods like SIFT and RLHF generally harm multi-turn capabilities, though some models like Vicuna-7B and Lemur-70B show improvements in multi-turn scenarios.
Significance
This work addresses the gap in multi-turn evaluation, emphasizing the importance of tool and feedback integration for real-world applications. It provides a standardized, scalable benchmark, especially valuable for open-source communities lacking extensive human evaluation resources. The insights into training impacts inform future model development, aiming for more robust multi-turn interaction abilities in complex tasks.
Technical Contribution
The paper introduces a novel multi-turn evaluation framework combining Python tool invocation and GPT-4 simulated feedback, with datasets designed for multi-step reasoning, coding, and decision-making. It employs regression analysis to quantify performance gains, revealing how training strategies influence multi-turn capabilities. The approach offers a systematic, reproducible method for benchmarking and improving LLMs' interactive skills.
Novelty
This is the first comprehensive benchmark explicitly assessing LLMs' multi-turn interaction with tools and natural language feedback. It innovates by integrating simulated user feedback, curated challenging datasets, and a regression-based performance metric, filling a critical gap in current evaluation standards and revealing unexpected effects of training methods.
Limitations
- The reliance on GPT-4 simulated feedback may not fully capture real user interactions, affecting external validity.
- The curated dataset, while diverse, cannot encompass all real-world complexities, limiting scope.
- Evaluation costs and API limitations restrict large-scale, real-time assessments.
Future Work
Future research should incorporate real user feedback, expand task diversity, and explore multi-modal interactions. Improving training strategies to enhance multi-turn robustness and developing standardized multi-turn benchmarks will further advance the field. Collaboration with industry to deploy and validate models in practical settings is also anticipated.
AI Executive Summary
The rapid evolution of large language models (LLMs) has transformed natural language processing, yet their evaluation remains predominantly focused on single-turn tasks. This approach overlooks the nuanced dynamics of real-world interactions, where models must engage in multiple rounds of reasoning, tool utilization, and feedback integration. Recognizing this gap, Wang et al. introduce MINT, a comprehensive benchmark designed to assess LLMs' multi-turn capabilities in a realistic setting.
MINT employs a novel evaluation framework combining Python code execution for tool calls and GPT-4 simulated natural language feedback, creating a scalable and reproducible environment. By curating a diverse set of 586 challenging instances from reasoning, coding, and decision-making datasets, the benchmark captures the complexity of multi-step problem solving. The evaluation spans 20 models, including both open-source and commercial closed-source architectures, across interaction limits from one to five turns.
Experimental results reveal that all models benefit from multi-turn interactions, with performance gains ranging from 2% to 17%. Interestingly, models with strong single-turn performance do not necessarily excel in multi-turn settings, highlighting the importance of interaction-specific training and evaluation. Open-source models generally lag behind their closed-source counterparts, with the best open-source model still trailing by 8.7% in success rate. Moreover, training strategies like SIFT and RLHF, while effective for single-turn tasks, tend to impair multi-turn capabilities, suggesting a need for rethinking training paradigms.
This work significantly advances the understanding of multi-turn interaction in LLMs, providing a standardized benchmark and insights into training impacts. It paves the way for future research focused on enhancing models' ability to collaborate, reason, and adapt in complex, real-world scenarios. The findings have broad implications for deploying more robust, interactive AI systems across industries, from customer service to autonomous decision-making. Despite some limitations, such as reliance on simulated feedback, the study offers a vital step toward more human-like, versatile language models.
Deep Analysis
Background
随着GPT、LLaMA等大规模语言模型的出现，模型在自然语言理解和生成方面取得了巨大突破。早期研究主要关注单轮任务，如问答和文本生成，代表工作包括GPT系列、BERT和T5。随后，研究逐步转向多轮交互，旨在模拟真实应用场景，如对话系统和复杂推理。现有评估多偏重单轮指标，缺乏对多轮交互能力的系统衡量。部分研究引入工具调用（如Wang et al., 2024）和用户反馈（如Shridhar et al., 2020），但缺少统一的评估框架，难以量化模型在连续交互中的表现。多轮交互的挑战在于模型如何有效利用上下文信息、调用工具以及从用户反馈中不断优化答案。
Core Problem
当前评估体系多集中于单轮性能，忽视了模型在多轮交互中的能力，尤其是在复杂任务中模型需要多轮推理、工具调用和反馈整合。单纯的单轮指标无法反映模型在实际应用中的表现，导致模型在真实场景中效果不佳。此外，训练策略如SIFT和RLHF虽提升单轮表现，但在多轮场景中效果未明，甚至可能削弱模型的持续交互能力。这限制了模型在复杂、多轮任务中的实际应用，亟需建立系统的多轮评估标准。
Innovation
本文提出结合Python工具调用和GPT-4模拟用户自然语言反馈的多轮评估框架，首次系统性量化模型在多轮交互中的工具利用和反馈能力。设计多样化任务场景，涵盖推理、编码和决策，筛选出具有代表性的实例，确保评估的科学性和实用性。引入回归分析，定量衡量工具和反馈对性能的提升，揭示不同训练策略在多轮场景中的影响，为模型优化提供新思路。这一框架突破了传统单轮评估的局限，为多轮交互研究提供了基础。
Methodology
- �� 构建多轮交互场景：模型在每轮中可调用Python工具或提出解决方案，模拟用户反馈由GPT-4生成。
- �� 数据集整合：从8个公开数据集筛选出586个复杂、多轮依赖实例，确保任务难度和代表性。
- �� 评估流程：在不同交互轮数（k=1-5）下，测定成功率（SR），利用线性回归分析每轮的性能提升。
- �� 反馈机制：模型可接收GPT-4生成的自然语言反馈，评估其利用反馈的能力。
- �� 训练影响分析：比较SIFT、RLHF等训练方式对多轮性能的影响，揭示潜在弊端。
Experiments
采用多源数据集，评估20个模型（包括OpenAI、Anthropic、LLaMA、Vicuna等），在不同交互轮数下测定成功率。通过回归分析计算每轮性能提升率，比较开源与闭源模型差异。引入自然语言反馈，分析模型利用反馈的效果。还验证了GPT-4模拟反馈的有效性，进行人类评估确认其质量。实验设计确保公平性和可重复性，控制交互轮数和模型参数。
Results
所有模型在引入工具和反馈后均表现出性能提升，工具每轮提升1-8%，自然语言反馈提升2-17%。开源模型普遍落后于闭源模型，最高开源模型成功率仍低于闭源8.7%。训练方式如SIFT和RLHF在多轮场景中表现不佳，部分模型如Vicuna-7B在多轮中反而表现更优。实验还揭示了模型在格式遵循和数据Artifact方面的不足，为未来优化提供方向。
Applications
该评估框架适用于模型开发者优化多轮交互能力，特别是开源社区，可借助此工具检测模型在复杂任务中的实际表现。可应用于对话系统、智能助手、自动推理等场景，推动模型在真实环境中的部署。未来还可结合用户反馈持续改进模型交互策略，提升用户体验。
Limitations & Outlook
评估依赖模拟用户反馈，可能与真实用户偏差较大，影响外推性。多轮场景虽丰富，但仍未涵盖所有复杂任务类型，未来需扩展多样性。模型性能受硬件和API调用成本限制，限制大规模实证研究的推广。
Plain Language Accessible to non-experts
想象你在厨房里做饭，厨师（模型）需要按照食谱（任务）一步步完成菜肴。单次尝试可能做得不错，但如果厨师能多次尝试、借助工具（如电饭煲、搅拌机）和听取厨师长（用户）的建议，菜肴会更好。这个研究就像在厨房里测试厨师的多次表现，看看他们是否能通过不断试验和听取建议，做出最美味的菜。模型就像厨师，工具是厨房设备，反馈是厨师长的建议。通过模拟多轮互动，研究评估厨师（模型）是否能借助工具和建议不断改进，最终做出令人满意的菜肴（解决复杂任务）。
ELI14 Explained like you're 14
想象你在学校里学习做手工艺品，老师（模型）需要多次尝试才能做出漂亮的作品。有时候，老师可以用工具（剪刀、胶水）帮助自己，也可以听老师的建议（反馈）来改进。这个研究就像在观察老师是否能通过多次尝试、用工具和听建议，做出更好的作品。科学家们设计了一个模拟场景，让模型像老师一样不断尝试，用工具，听建议，看看能不能做得更好。结果发现，模型用工具和听建议都能让表现提升不少，但用得不好或不听建议也会影响效果。这个研究帮助我们理解，未来让AI更聪明、更会和人合作的方法。
Abstract
To solve complex tasks, large language models (LLMs) often require multiple rounds of interactions with the user, sometimes assisted by external tools. However, current evaluation protocols often emphasize benchmark performance with single-turn exchanges, neglecting the nuanced interactions among the user, LLMs, and external tools, while also underestimating the importance of natural language feedback from users. These oversights contribute to discrepancies between research benchmark evaluations and real-world use cases. We introduce MINT, a benchmark that evaluates LLMs' ability to solve tasks with multi-turn interactions by (1) using tools and (2) leveraging natural language feedback. To ensure reproducibility, we provide an evaluation framework where LLMs can access tools by executing Python code and receive users' natural language feedback simulated by GPT-4. We repurpose a diverse set of established evaluation datasets focusing on reasoning, coding, and decision-making and carefully curate them into a compact subset for efficient evaluation. Our analysis of 20 open- and closed-source LLMs offers intriguing findings. (a) LLMs generally benefit from tools and language feedback, with performance gains (absolute, same below) of 1-8% for each turn of tool use and 2-17% with natural language feedback. (b) Better single-turn performance does not guarantee better multi-turn performance. (c) Surprisingly, on the LLMs evaluated, supervised instruction-finetuning (SIFT) and reinforcement learning from human feedback (RLHF) generally hurt multi-turn capabilities. We expect MINT can help measure progress and incentivize research in improving LLMs' capabilities in multi-turn interactions, especially for open-source communities where multi-turn human evaluation can be less accessible compared to commercial LLMs with a larger user base.