Can Large Language Models Extend Their Agentic Capabilities to Goal Pursuit in Open-Ended Multi-Turn Interaction?

Mingyu Chen; Jingkai Lin; Xiaofen Xing; Liankai Huang; Ruiqi Xiong; Xiangmin Xu

cs.CLNov 1, 2026

Can Large Language Models Extend Their Agentic Capabilities to Goal Pursuit in Open-Ended Multi-Turn Interaction?

Authors: Mingyu Chen, Jingkai Lin, Xiaofen Xing, Liankai Huang, Ruiqi Xiong, Xiangmin Xu

Organizations: School of EE., South China University of Technology, Guangzhou, China · Foshan University, Foshan, China

Abstract

Goal pursuit in open-ended dialogue requires agents to adapt their actions to interaction feedback across successive exchanges. We investigate whether agentic capabilities developed for multi-step tool use carry over to dialogue organized through the model's native tool interface, and how this changes interaction behavior. We define Tool-Mediated Interaction (TMI) as a training-free protocol that represents dialogue actions as tool calls and counterpart responses as tool results, maintaining a continuous reasoning-action-observation loop with retained reasoning and interaction history. We evaluate TMI across social interaction, emotional support, persuasion, and price negotiation using two models and 15,678 main-experiment episodes. Across both models, TMI improves social goal attainment, emotional-support preference, donation amounts, and buyer negotiation margins over conversational baselines with and without thinking. A Qwen ablation further shows that TMI outperforms ordinary conversation with retained reasoning on all four primary outcomes. Behavioral analyses associate the improvements with task-specific patterns in initial proposals and responses to interaction feedback. Alongside these gains, TMI reduces mean reasoning tokens per decision by 42.0-81.5% relative to the thinking baseline and lowers mean cross-turn phrase reuse across all eight model-environment pairs. These findings show that interaction organization shapes both goal attainment and reasoning across dialogue turns.

Figures & tables

Figure 1: Within-turn agentic tool use (a) and goal-directed multi-turn dialogue (b) share a reasoning–action–observation loop. Tool results and counterpart responses inform continued reasoning and subsequent actions toward a goal.

Figure 2: From conversational responses to tool-mediated interaction. (a) The model responds directly; (b) it reasons before each response, with earlier reasoning omitted from subsequent requests. (c) Agentic tool use maintains a reasoning–action–observation loop during tool execution. (d) TMI applies this structure to dialogue: dialogue actions become tool calls and counterpart responses become tool results, preserving reasoning and interaction history across exchanges.

Qwen3.8-Flash
	Sotopia-Hard					ESC-Eval
Setting	Goal $\uparrow$	Rel. $\uparrow$	Know. $\uparrow$	Fin. $\uparrow$	Overall $\uparrow$	Expr. $\uparrow$	Info. $\uparrow$	Skill $\uparrow$	Overall $\uparrow$
w/o Thinking	7.92 ±0.15	1.85 ±0.05	4.73 ±0.08	1.40 ±0.06	3.52 ±0.01	2.84 ±0.02	2.22 ±0.04	2.89 ±0.03	1.47 ±0.02
w/ Thinking	8.19 ±0.12	2.14 ±0.13	5.15 ±0.30	1.49 ±0.04	3.70 ±0.05	2.93 ±0.03	2.34 ±0.01	2.95 ±0.02	1.69 ±0.02
TMI	8.66 ±0.04	2.71 ±0.10	5.43 ±0.32	1.65 ±0.02	3.91 ±0.06	2.96 ±0.02	2.45 ±0.02	2.98 ±0.01	1.90 ±0.003
DeepSeek-v4-Flash

Table 1: Sotopia-Hard and ESC-Eval results (mean

\pm

SD across three benchmark repeats). Rel.: Relationship; Know.: Knowledge; Fin.: Financial/Material Benefits; Expr.: Expression; Info.: Information. Bold and underline mark the best and second-best means within each model, ranked before rounding.

Tile: opening offer below $14
Model	w/o Thinking	w/ Thinking	TMI
Qwen	0/15	2/15	5/15
DeepSeek	0/15	0/15	5/15
Trip: commitment inquiry after refusal
Qwen	3/15	2/15	9/15
DeepSeek	0/15	3/15	5/15

Table 3: Action frequencies in two selected Sotopia roles (15 episodes per model and condition). Definitions and annotation procedures appear in Appendix C.1 .

Figure 3: Sotopia interaction endings (420 episodes per model and condition): (a) agent-initiated exits; (b) turn-limit terminations. Events are non-exclusive.

Elicit: new information
Scenario	w/o Thinking	w/ Thinking	TMI
Job search	2/6	0/6	3/6
Family conflict	0/6	2/6	6/6
Breakup	0/6	0/6	4/6
Total	2/18	2/18	13/18
Use: information used in support

Table 4: Information elicitation and use in three selected ESC-Eval scenarios. Counts indicate qualifying dialogues out of six per scenario and condition.

Figure 4: Model-selected strategy transitions in three gain-selected ESC-Eval scenarios (Table 4 ; 18 dialogues per condition). Node heights and ribbon widths encode dialogue proportions; ribbon colors identify the target strategy.

Figure 5: CB differences in opening offers, subsequent price movement, and agreed prices (TMI minus baseline), in percentage points of listing price. Values are equal-model means, rounded independently.

Figure 6: Reasoning tokens per decision (a–b; means with 95% cluster-bootstrap confidence intervals) and cross-turn four-gram reuse (c–f; dialogue-level ECDFs).

Outcome	w/ Thinking	Preserved	TMI
Sotopia Goal	8.19 ±0.12	8.01 ±0.23	8.66 ±0.04
ESC Overall	1.69 ±0.02	1.61 ±0.003	1.90 ±0.003
P4G Donation	0.68 ±0.03	0.67 ±0.03	0.76 ±0.02
CB Margin	0.51 ±0.02	0.57 ±0.02	0.73 ±0.01

Table 5: Qwen reasoning-preservation ablation (mean

\pm

SD). CB reports margin; P4G reports dollars.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

P4G: donation ($)
Group	Qwen	DeepSeek
Analytical	$+0.108\pm 0.047$	$-0.015\pm 0.066$
Behavioral	$+0.012\pm 0.075$	$+0.017\pm 0.025$
Conceptual	$+0.130\pm 0.066$	$+0.041\pm 0.004$
Directive	$+0.083\pm 0.024$	$+0.055\pm 0.017$
CB: buyer margin

Appendix

Table 6: Subgroup gains over w/ Thinking (mean

\pm

SD across three repeats). SD describes variability across repeat means.

Qwen3.8-Flash
Setting	Goal $\uparrow$	Bel. $\uparrow$	Rel. $\uparrow$	Know. $\uparrow$	Secret $\uparrow$	Rules $\uparrow$	Fin. $\uparrow$	Overall $\uparrow$
w/o Thinking	7.92 ±0.15	9.05 ±0.03	1.85 ±0.05	4.73 ±0.08	−0.04 ±0.01	−0.25 ±0.04	1.40 ±0.06	3.52 ±0.01
w/ Thinking	8.19 ±0.12	9.03 ±0.02	2.14 ±0.13	5.15 ±0.30	0.00 ±0.00	−0.08 ±0.06	1.49 ±0.04	3.70 ±0.05
TMI	8.66 ±0.04	9.09 ±0.004	2.71 ±0.10	5.43 ±0.32	−0.01 ±0.02	−0.13 ±0.03	1.65 ±0.02	3.91 ±0.06
DeepSeek-v4-Flash
w/o Thinking	8.02 ±0.10	9.05 ±0.02	2.48 ±0.03	5.33 ±0.27	−0.06 ±0.02	−0.26 ±0.05	1.55 ±0.08	3.73 ±0.04

Appendix

Table 7: Sotopia-Hard results (mean

\pm

SD). Overall averages the seven raw dimensions.

Qwen3.8-Flash
Setting	Fluency $\uparrow$	Expression $\uparrow$	Empathy $\uparrow$	Information $\uparrow$	Humanoid $\uparrow$	Skill $\uparrow$	Overall $\uparrow$
w/o Thinking	3.00 ±0.00	2.84 ±0.02	3.00 ±0.00	2.22 ±0.04	2.64 ±0.02	2.89 ±0.03	1.47 ±0.02
w/ Thinking	3.00 ±0.00	2.93 ±0.03	3.00 ±0.00	2.34 ±0.01	2.70 ±0.03	2.95 ±0.02	1.69 ±0.02
TMI	3.00 ±0.00	2.96 ±0.02	3.00 ±0.00	2.45 ±0.02	2.63 ±0.02	2.98 ±0.01	1.90 ±0.003
DeepSeek-v4-Flash
w/o Thinking	3.00 ±0.00	2.88 ±0.01	3.00 ±0.00	2.14 ±0.02	2.73 ±0.01	2.91 ±0.02	1.50 ±0.02

Appendix

Table 8: ESC-Eval scores (0–4; mean

\pm

SD). Overall is rated separately.

Tile: low-opening episodes
Model	Condition	Low openings	Agreements	Mean Goal
Qwen	w/o Thinking	0/15	–	–
Qwen	w/ Thinking	2/15	0/2	3.0
Qwen	TMI	5/15	5/5	8.0
DeepSeek	w/o Thinking	0/15	–	–
DeepSeek	w/ Thinking	0/15	–	–

Appendix

Table 9: Statistics supporting the Sotopia analysis (15 episodes per model and condition). Tile agreement counts and mean Goal scores cover only low-opening episodes; – indicates no such episode.

Model	Condition	Elicit	Use
Qwen	w/o Thinking	0/9	0/9
Qwen	w/ Thinking	2/9	0/9
Qwen	TMI	6/9	5/9
DeepSeek	w/o Thinking	2/9	1/9
DeepSeek	w/ Thinking	0/9	0/9
DeepSeek	TMI	7/9	2/9

Appendix

Table 10: Model breakdown of Table 4 (nine dialogues per model and condition). Counts sum to the totals in that table.

Scope	Cards	$n$	$\Delta$ Reflect.	$\Delta$ Suggest.
Selected scenarios	3	18	$-13.0$	$+16.7$
All scenarios	331	1,986	$-8.1$	$+5.4$

Appendix

Table 11: Strategy shares, TMI minus w/ Thinking (percentage points).

n

: dialogues per condition. Reflection averages

S_{3}

–

S_{5}

; suggestions are measured at

S_{5}

TMI minus w/ Thinking
Model	$n$	Initial	Movement	Final
Qwen	30	$-5.62$	$+3.28$	$-2.35$
DeepSeek	101	$-1.85$	$+0.29$	$-1.56$
Equal-model mean	–	$-3.74$	$+1.78$	$-1.95$
TMI minus w/o Thinking
Qwen	7	$-9.33$	$+5.31$	$-4.01$

Appendix

Table 12: CB price-path differences in percentage points of listing price.

n

counts eligible assignments, each contributing three rollouts per condition. Values are rounded independently.

Model	Personas	Thinking	TMI	Difference
Low amount mentioned only under w/ Thinking
Qwen	88	0.57	0.78	$+0.21$
DeepSeek	49	0.53	0.69	$+0.16$
No low amount mentioned under either condition
Qwen	96	0.79	0.78	$-0.01$
DeepSeek	146	0.71	0.70	$-0.02$

Appendix

Table 13: P4G donation means for the groups in Section 5.3 (dollars). Differences are TMI minus w/ Thinking; values are rounded independently.

		Tokens / decision		Reduction	Tokens / dialogue		Decisions / dialogue
Environment	Model	Thinking	TMI	(%)	Thinking	TMI	Thinking	TMI
Sotopia	Qwen	490.4	196.7	59.9	3351.5	1143.4	6.86	5.80
	DeepSeek	269.1	142.0	47.2	1606.9	887.4	6.35	6.14
ESC-Eval	Qwen	162.2	94.0	42.0	810.8	469.9	5.00	5.00
	DeepSeek	200.9	43.1	78.6	1004.6	215.3	5.00	5.00
P4G	Qwen	426.4	168.8	60.4	979.0	424.9	2.30	2.56

Appendix

Table 14: Reasoning usage, with assignments weighted equally. Each model and condition has 420 dialogues in Sotopia, 993 in ESC-Eval, and 600 in P4G and CB.

Environment	Model	Dialogues	Assignments	Thinking	TMI	TMI $-$ Thinking	95% CI
		Paired sample		Full-text mean (%)		After length matching (pp)
Sotopia	Qwen	413	138	4.62	1.38	$-8.49$	$[-10.12,-6.99]$
	DeepSeek	419	140	3.11	1.17	$-3.26$	$[-4.30,-2.26]$
ESC-Eval	Qwen	988	331	2.07	1.05	$-1.18$	$[-1.31,-1.05]$
	DeepSeek	933	331	2.01	1.17	$+0.04$	$[-0.14,+0.21]$
P4G	Qwen	551	197	2.42	0.88	$-1.47$	$[-1.80,-1.15]$

Appendix

Table 15: Four-gram reuse with equal assignment weights. Paired counts apply to each condition. Length-matched differences use recomputed reuse rates on truncated texts.

Qwen
Environment	Pairs	Thinking	TMI	Reduction (%)
Sotopia	406	459.0	217.7	52.6
ESC-Eval	993	167.5	96.8	42.2
P4G	132	430.7	159.8	62.9
CB	588	303.3	99.9	67.1
DeepSeek

Appendix

Table 16: Mean reasoning tokens per decision over the first three decisions of paired dialogues reaching three decisions in both conditions.

Figure 7: Historical phrase reuse for

n=2,3,4

, using the same paired prefixes as Figure 6 c–f. Models receive equal weights; display limits retain full ECDF denominators.

Figure 8: Qwen reasoning tokens per decision under Thinking, Preserved, and TMI, with 95% cluster-bootstrap confidence intervals.

Figure 9: Qwen historical n-gram reuse under Thinking, Preserved, and TMI, using common prefixes across all three conditions. Display limits retain full ECDF denominators.

Sotopia-Hard: shared task prompt
Imagine you are {agent} , your task is to act/speak as {agent} would, keeping in mind {agent} ’s social goal. You can find {agent} ’s goal (or background) in the ’Here is the context of the interaction’ field. Note that {agent} ’s goal is only visible to you. You should try your best to achieve {agent} ’s goal in a way that align with their character traits. Additionally, maintaining the conversation’s naturalness and realism is essential (e.g., do not repeat what other people has already said before). {initial_history} . Your available action types are none speak non-verbal communication action leave.
Condition	Condition-specific instruction excerpts
Baseline	Note: You can "leave" this conversation if 1. you have achieved your social goals, 2. this conversation makes you uncomfortable, 3. you find it uninteresting/you lose your patience, 4. or for other reasons you want to leave.
TMI	For each turn, respond with exactly one interact tool call. Use the tool schema to specify your action; do not reply with text, JSON, or an explanation. Note: You can call interact with leave if 1. you have achieved your social goals, 2. this conversation makes you uncomfortable, 3. you find it uninteresting/you lose your patience, 4. or for other reasons you want to leave.

Appendix

Table 17: Sotopia-Hard task prompt and condition-specific instructions.

ESC-Eval supporter: shared task prompt
You are a supportive and empathetic assistant for an emotional support context. Strategies for emotional support: - question: Ask for details to better understand the seeker’s situation and emotions. - restatement: Paraphrase the situation or experience described by the seeker. - reflection: Reflect the seeker’s feelings about their situation. - self_disclosure: Share a relevant personal experience to show empathy only when grounded in your assigned identity or background. Do not invent lived experiences. - affirmation: Affirm the seeker’s strengths, motivation, and capabilities; offer reassurance and encouragement. - suggestions: Offer practical advice or actionable steps relevant to the seeker’s situation. - information: Provide factual information relevant to the seeker’s situation. - others: Exchange pleasantries or use support strategies outside the categories above. Respond in a calm and non-judgmental manner, acknowledging and validating the seeker’s emotions. Ask gentle clarifying questions only when helpful, and offer practical coping suggestions when appropriate. Do not diagnose, prescribe, or claim professional authority. Keep responses concise and natural (2-4 sentences).
Condition	Condition-specific instruction excerpts
Baseline	only Response is shown to the seeker. Do not include explanations or Markdown fences.
TMI	Make exactly one interact tool call following the tool schema; only Response is shown to the seeker. Do not send a separate text response or explanation. The tool returns the seeker’s next reply, or a notice that the interaction has ended.

Appendix

Table 19: ESC-Eval supporter task prompt and condition-specific instructions.

Tool-Using Agents

Tool-using agents plan and act through external tools, APIs, software interfaces, and environments to complete tasks beyond language generation.

See all 300 papers

Think Thrice Before You Speak: Dual knowledge-enhanced Theory-of-Mind Reasoning for Persuasive Agents
May 21, 2026Minghui Ma, Bin Guo, Runze Yang +7Theory of Mind Conversational Agents
Northwestern Polytechnical University, Xi’an, China · Peking University, Beijing, China · Harbin Engineering University, Harbin, China
TAC-CoT: Tool-Empowered Implicit Chain-of-Thought via Interleaving Explicit Tokens
Nov 1, 2026Juseon Do, Sangmin Bae, Jeesu Jung +1Implicit Reasoning in Language Models Adaptive Reasoning in Language Models
Data Intelligence System Lab, KAIST · Optimization and Statistical Inference Lab, KAIST AI · Cluvion
Unlocking Proactivity in Task-Oriented Dialogue
May 21, 2026Azure Zhang, Ning Gao, Yuqin Dai +7LLM Post-Training Conversational Agents
1Keeta AI, Meituan · 2Independent Researcher · 3CUHK +1

May 21, 2026J/K move · Enter open · S save

Think Thrice Before You Speak: Dual knowledge-enhanced Theory-of-Mind Reasoning for Persuasive Agents

Minghui Ma, Bin Guo, Runze Yang, Mengqi Chen+6

Northwestern Polytechnical University, Xi’an, China · Peking University, Beijing, China · Harbin Engineering University, Harbin, China

Theory of Mind Conversational Agents LLM Reasoning

Open paper AskPDF arXiv

Persuasive dialogue requires reasoning about others' latent mental states, a capability known as Theory of Mind (ToM). However, due to reliance on simple prompting strategies and insufficient ToM knowledge, existing LLMs often fail to capture the intrinsic dependencies among mental states, leading to fragmented representations and unstable reasoning. To address these challenges, we introduce the ToM-based Persuasive Dialogue (ToM-PD) task, grounded in the Belief-Desire-Intention (BDI) framework, which explicitly models the sequential dependencies among mental states in multi-turn dialogues. To facilitate research on this task, we construct a large-scale annotated dataset, ToM-based Broad Persuasive Dialogues (ToM-BPD), capturing fine-grained mental states and corresponding persuasive strategies. We further propose Think Thrice Before You Speak (TTBYS), a knowledge-enhanced stepwise reasoning framework that leverages both explicit and implicit prior experiences to improve LLMs' inference of desires, beliefs, and persuasive strategies. Experimental results demonstrate that Qwen3-8B equipped with TTBYS outperforms GPT-5 by 1.20%, 22.80%, and 16.97% in predicting desires, beliefs, and persuasive strategies, respectively. Case studies further show that our approach enhances interpretability and consistency in reasoning.

Abstract

Tool-Using Agents

Explore similar work

Think Thrice Before You Speak: Dual knowledge-enhanced Theory-of-Mind Reasoning for Persuasive Agents

TAC-CoT: Tool-Empowered Implicit Chain-of-Thought via Interleaving Explicit Tokens

Unlocking Proactivity in Task-Oriented Dialogue