IC-190LLMs show progressive degradation in system message constraint following across multi-turn conversations, with dependent conversations degrading faster than parallel ones

Yanzhao Qin, Tao Zhang, Yanjun Shen, Wenjing Luo, sunhaoze, Yan Zhang, Yujing Qiao, weipeng chen, Zenan Zhou, Wentao Zhang, Bin CUI

SourceSysBench: Can LLMs Follow System Message?

Using the session stability rate (SSR) and per-round satisfaction (Rn), the paper shows all 16 models degrade over 5 turns. The best SSR is 54.4% (GPT-4o, parallel). Multi-turn dependent conversations show steeper decline than parallel ones for most models, as evidenced by higher linear regression slopes (k). For example, Qwen2-7b drops to 0.0% by round 5 in parallel conversations and 1.1% in dependent conversations. Replacing model-generated history with ground truth improves ISR, confirming that prior errors compound and degrade subsequent performance.

Evidence

correlational

Key metric

GPT-4o SSR 54.4% (parallel), 56.7% (dependent); R5: GPT-4o 26.4% (parallel) / 33.7% (dependent), Qwen2-7b 0.0% (parallel) / 1.1% (dependent); best overall SSR 54.4%

Caveat

Conversations are limited to 5 turns; the paper does not test longer horizons. The ground-truth history experiment uses only 4 models.

Model

GPT-4o, GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report GPT-4-turbo-20240409, Claude 3 Opus, Llama 3.1 70B Instruct, 8B Instruct, Mixtral 8x22B Instruct, GPT-3.5 / ChatGPT-3.5 GPT-3.5-turbo-20231106, Qwen 2.5 72B Instruct, Qwen 2 Qwen2-72B-Instruct, 7B Instruct, GLM-4 GLM-4-0520, GLM-4-9B-Chat, DeepSeek-V2-0628, Moonshot-v1-8k

Concepts

Failure mode

Related work

Large Language Models Can Be Easily Distracted by Irrelevant Context [context]

Related findings

IC-188, IC-189, IC-191

Extraction

automatic-extraction