IC-1331SIFT and RLHF variants of CodeLlama and Llama-2 perform worse than their base counterparts in multi-turn interaction

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

Comparing base, SIFT, and RLHF variants within the CodeLlama and Llama-2 families, the paper finds that instruction fine-tuning and RLHF alignment generally degrade multi-turn tool-use performance. SIFT hurts CodeLlama-34b by 11.1% in absolute success rate (no feedback) and 15.4% (with feedback); RLHF hurts Llama-2-70b by 8.5% and 8.7% respectively. Notable exceptions are Vicuna-v1.5-7b and Lemur-v1-70b SIFT, where SIFT on multi-turn ShareGPT data improves performance.

Evidence
correlational
Key metric
SIFT hurts CodeLlama-34b by 11.1% (sr5: 28.2 base vs 17.1 SIFT) and 15.4% (sr5_feedback: 42.7 vs 27.3); RLHF hurts Llama-2-70b by 8.5% (sr5: 26.4 base vs 17.9 RLHF) and 8.7% (sr5_feedback: 35.3 vs 26.6)
Caveat
The authors note it is hard to conclude that RLHF in general hurts model performance, and leave exploring the role of RLHF in multi-turn interaction for future work. Vicuna-v1.5 and Lemur-v1 are exceptions where SIFT helps.
Model
CodeLlama-13B CodeLlama-7B, CodeLlama-34B, Llama 2 / Llama 2 base Llama 2 7B, Llama 2 13B, Llama 2 70B, Vicuna Vicuna-v1.5-7B, Lemur-v1-70B / Lemur-70B-Chat-V1
Related work
ReAct [builds-on]
Related findings
IC-1330, IC-1332
Extraction
automatic-extraction