IC-1330All 20 evaluated LLMs improve in multi-turn task-solving with additional tool-use turns and GPT-4-simulated language feedback

Xingyao Wang, Zihan Wang, Jiateng Liu, Yangyi Chen, Lifan Yuan, Hao Peng, Heng Ji

SourceMINT: Evaluating LLMs in Multi-turn Interaction with Tools and Language Feedback

The paper evaluates 20 open- and closed-source LLMs on 586 multi-turn interaction tasks spanning reasoning, code generation, and decision-making. Every model shows absolute performance gains of 1-8% per additional tool-use turn (k=1 to k=5) and 2-17% from GPT-4-simulated natural language feedback. Closed-source models (GPT-3.5-turbo-0613, Claude-instant-1) show the steepest per-turn improvement slopes (+8.2 and +8.0), while the best open-source model (Lemur-v1-70b SIFT) still lags the best closed-source model by 8.7% in absolute success rate with feedback.

Evidence
correlational
Key metric
performance gains of 1-8% for each turn of tool use and 2-17% with natural language feedback; slopes range +1.1 (CodeLlama-7b base) to +8.2 (GPT-3.5-turbo-0613); delta_feedback ranges +1.7 (Llama-2-7b RLHF) to +17.2 (CodeLlama-7b SIFT)
Caveat
GPT-4-simulated feedback may not cover all possible human responses, particularly in tasks involving nuanced judgments of human values. Metrics may not fully assess interaction quality beyond outcomes.
Model
GPT-3.5 / ChatGPT-3.5 GPT-3.5-turbo-0613, GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report GPT-4 0613, Claude Instant 1, Chat-bison-001, Llama 2 / Llama 2 base Llama 2 7B, Llama 2 13B, Llama 2 70B, Vicuna Vicuna-v1.5-7B, Vicuna-v1.5-13B, CodeLlama-13B CodeLlama-7B, CodeLlama-34B, Lemur-v1-70B / Lemur-70B-Chat-V1
Datasets
HumanEval [eval], MBPP [eval], ALFWorld [eval], GSM8K [eval], HotpotQA [eval], MATH [eval], MMLU / MMLU-Math [eval], TheoremQA [eval]
Related work
ReAct [builds-on], Reflexion [context]
Related findings
IC-1331, IC-1332
Extraction
automatic-extraction