cs.CLMar 2, 2026

Quantifying Conversational Reliability of Large Language Models under Multi-Turn Interaction

Authors: Jiyoon Myung

Organizations: Samsung SDS

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Large Language Models (LLMs) are increasingly deployed in real-world applications where users engage in extended, mixed-topic conversations that depend on prior context. Yet, their reliability under realistic multi-turn interactions remains poorly understood. We conduct a systematic evaluation of conversational reliability through three representative tasks that reflect practical interaction challenges: (1) maintaining global constraints across topic shifts, (2) selecting the correct tool or agent amid interleaved intents, and (3) tracking structured entities under revisions and distractions. Each task pairs single-turn and multi-turn settings, allowing us to quantify reliability degradation under extended dialogue. Across both commercial and open-source models, we observe substantial declines in reliability, particularly for smaller models. Error analyses reveal recurring failure modes such as instruction drift, intent confusion, and contextual overwriting, which compromise dependable behavior in operational systems. Our findings highlight the need for stress-testing LLMs for conversational reliability and developing more robust evaluation methods for trustworthy deployment.

Figures & tables

Figure 1: Single-turn vs Multi-turn accuracy across three evaluation tasks. Each panel shows a task example on the left and model accuracy on the right. Performance drops most severely in Instruction Following , while Entity Extraction remains relatively robust, and Tool Selection shows mixed degradation depending on model size.
 GPT-4oGPT-4o-miniGemini-2.5-FlashGemma-3-12B
Single-turn
Instruction Following96939692
Tool Selection100100100100
Entity Extraction1009610092
Multi-turn
Instruction Following63248933
Table 1: Accuracy (%) of all models across tasks in single-turn and multi-turn settings.
Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

TurnsAccuracy
50.40
60.28
70.38
80.15
90.29
110.25
Appendix
Table 2: Instruction Following accuracy by number of dialogue turns.
TurnsAccuracy
60.76
70.74
80.82
90.74
100.96
110.78
Appendix
Table 3: Tool Selection accuracy by dialogue length.
# ToolsAccuracy
20.98
30.82
40.74
50.64
60.71
Appendix
Table 4: Tool Selection accuracy by number of available tools.
Scenario TypeDateTimePeopleAll Match
Change in Mind0.910.960.990.85
Intermediate Chit Chat0.910.970.970.86
Multiple Mention0.940.980.990.91
Appendix
Table 5: Entity Extraction accuracy by scenario type.
Context (excerpt)Model OutputGround Truth
Instruction Following User: From now on, always answer in at most 5 sentences. (after 10 turns of unrelated chit-chat) User: Could you walk me through the entire rise and fall of the Byzantine Empire?Provides a long, detailed explanation in 13 sentences.Concise 4-5 sentence summary.
Tool Selection User: What is the stock price of AAPL today? User: And what is the weather in Paris right now?[Stock][Weather]
Entity Extraction User: I’d like a table for 3 on April 18 at 7 pm. User: By the way, my friend Sarah booked for 6pm yesterday. User: Oh actually, make it 4 people.{”date”: ”04/18”, ”time”: ” 7pm ”, ”people”: ”4”}{”date”: ”04/18”, ”time”: ”6pm”, ”people”: ”4”}
Appendix
Table 6: Representative qualitative error cases.
TaskSingle-turn Acc.Multi-turn Acc.Exact ppχ2\chi^{2} (cc)pp
Instruction Following0.7640.3961.7×10−611.7\times 10^{-61}244.563.99×10−553.99\times 10^{-55}
Tool Selection0.9990.7633.0×10−643.0\times 10^{-64}210.001.37×10−471.37\times 10^{-47}
Entity Extraction0.9810.9245.2×10−95.2\times 10^{-9}31.651.85×10−81.85\times 10^{-8}
Appendix
Table 8: McNemar test results on 900 paired samples per task. All differences between single-turn and multi-turn settings are statistically significant.
More in this topic

Tool Selection for Language Model Agents

See all 27 papers

Explore similar work

CardsList

  1. Stop Listening to Me! How Multi-turn Conversations Can Degrade LLM Reliability

    Mar 12, 2026Kevin H. Guo, Chao Yan, Avinash Baidya +5LLM ReliabilityMulti-Turn Dialogue Evaluation

    Vanderbilt University · Vanderbilt University Medical Center · Intuit AI Research

  2. Confidence Should Be Calibrated More Than One Turn Deep

    Apr 7, 2026Zhaohan Zhang, Chengzhengxu Li, Xiaoming Liu +3Confidence Estimation in Language ModelsLanguage Model Calibration

    Queen Mary University of London · Xi’an Jiaotong University

  3. Toward a Graded Measure of Belief Stability in Large Language Models

    Sep 28, 2026Samantha Dies, Branden Fitelson, Tina Eliassi-RadLLM ReliabilityLLM Uncertainty Estimation

    Northeastern University, 360 Huntington Ave, Boston, MA 02115 USA · Santa Fe Institute, 1399 Hyde Park Road, Santa Fe, NM 87501 USA