TL;DR for operators

A conversational agent must do more than generate a suitable reply. Across multiple exchanges, it must decide when to ask for information, revise a proposal, respond to resistance, or end the interaction. How those exchanges are represented inside the model may affect its ability to pursue a goal.

Mingyu Chen and colleagues1 investigate Tool-Mediated Interaction (TMI), a training-free protocol that treats dialogue actions as native tool calls and counterpart replies as returned observations. Across 15,678 simulated episodes, TMI achieves higher mean primary-outcome scores than conventional conversation with or without thinking across four benchmarks and two language models.

In simulated buyer negotiations, Qwen’s normalized margin rises from 0.51 to 0.73, while its target-price success rate increases from 13.17% to 28.83%. A separate experiment finds that simply preserving prior reasoning in conventional chat does not reproduce TMI’s advantages.

Cognaptus interpretation: Customer-service, sales, and information-gathering agent developers should consider interaction protocol a measurable design variable, alongside model selection, prompting, and memory. The results justify controlled prototypes, not assumptions about real customer outcomes.

The same conversation can run through different interfaces

Consider an automated buyer negotiating a price. Its opening offer influences the seller’s response. A refusal may reveal a constraint, while a counteroffer creates an opportunity to revise the proposal. Success depends on decisions made across the entire exchange.

A conventional conversational assistant receives messages, generates responses, and continues with an updated chat history. Even when reasoning is enabled, earlier private reasoning traces are ordinarily omitted from subsequent requests in the study’s conventional-chat configuration.

TMI changes that organization.

Instead of sending an ordinary assistant message, the model issues its dialogue action through a native interact tool call. The counterpart’s reply comes back as a tool result. The next decision receives a continuous history containing reasoning, actions, and observations.

The resulting execution structure resembles the reasoning–action–observation loop already used by tool-calling agents. It does not require new model weights, fine-tuning, or a separate planning algorithm.

This distinction matters because the model’s objective and permitted dialogue actions can remain unchanged while the execution interface changes. A question, offer, or refusal response is still a conversational act; TMI changes how that act and its consequences are organized for subsequent decisions.

The researchers compare three configurations: conventional conversation without thinking, conventional conversation with thinking, and TMI with thinking. Model parameters, assignments, simulated counterpart configurations, and generation settings are held fixed within the comparisons.

Four simulated tasks show improvements, with different magnitudes

The main experiment evaluates Qwen3.8-Flash and DeepSeek-v4-Flash across four environments: Sotopia-Hard for social goals, ESC-Eval for emotional support, Persuasion for Good for donations, and CraigslistBargain for buyer negotiation.

Each environment measures a distinct outcome, so the numerical gains should be interpreted on their own benchmark scales.

Primary outcome Qwen: conventional thinking → TMI DeepSeek: conventional thinking → TMI
Sotopia social Goal score 8.19 → 8.66 8.09 → 8.41
ESC emotional-support Overall score 1.69 → 1.90 1.72 → 1.92
P4G mean donation $0.68 → $0.76 $0.67 → $0.69
CraigslistBargain buyer margin 0.51 → 0.73 0.38 → 0.51

Source: Section 4.3, Tables 1–2. Results are means across three complete-benchmark repeats.

TMI leads on all four designated primary outcomes for both models. The donation improvement is relatively small for DeepSeek: the authors report 3.7% using unrounded means, compared with 12.3% for Qwen.

The negotiation results deserve closer attention. Buyer margin measures the agreed price relative to buyer and seller target prices, with no-agreement episodes assigned zero. TMI increases the buyer’s normalized margin and more than doubles the rate of agreements at or below the buyer’s target price for both models.

Yet the overall agreement rate declines from 92.17% to 91.17% for Qwen and from 97.17% to 95.33% for DeepSeek.

The agent therefore becomes more effective at achieving the buyer’s price objective without uniformly improving negotiation outcomes. For a commercial system, the distinction between obtaining a favorable price and securing an acceptable transaction is consequential.

The experiment’s 15,678 episodes reflect matched simulated assignments, three rollouts per assignment and condition, and two evaluated models. Reported benchmark standard deviations describe variation across three complete-benchmark repeats, not uncertainty estimated from independent human participants.

Remembering earlier reasoning is not enough

TMI retains previous reasoning traces, unlike the study’s conventional-thinking baseline. That creates an alternative explanation: perhaps the improvement comes entirely from allowing the model to see its earlier reasoning.

The researchers address this through a reasoning-preservation ablation. They add a Qwen-only condition that keeps previous reasoning within ordinary assistant–user message exchanges without changing the dialogue into tool calls and returned observations.

The additional experiment comprises 2,613 episodes.

Its results are informative:

Qwen primary outcome Conventional with reasoning preserved TMI
Sotopia Goal 8.01 8.66
ESC Overall 1.61 1.90
P4G donation $0.67 $0.76
Buyer margin 0.57 0.73

Source: Section 6, Table 5. This is an ablation testing an alternative explanation, not an independent human-dialogue validation.

Preserving reasoning improves buyer margin relative to ordinary thinking-enabled conversation, from 0.51 to 0.57, but remains below TMI’s 0.73. On the other three primary outcomes, preservation alone does not even improve the conventional-thinking baseline.

This makes reasoning retention an insufficient explanation for TMI’s performance. The broader execution organization appears relevant.

However, the ablation does not isolate every component of that organization, nor does it establish the internal cognitive mechanism responsible. The preservation comparison is also available only for Qwen.

The behavioral changes appear in offers, questions, and exits

The researchers examine conversational trajectories to investigate how the protocol affects observable decisions.

In three emotional-support scenarios selected for favorable TMI gains, the protocol elicits previously unavailable counterpart information in 13 of 18 dialogues, compared with 2 of 18 under conventional thinking. It then immediately incorporates newly elicited information into concrete assistance in 7 of 18 dialogues, compared with none under conventional thinking.

These observations suggest a possible route from interaction organization to better contextual assistance: the agent asks for information that can affect its next action, rather than continuing solely from what it already knows.

Negotiation behavior shows another pattern. In a restricted CraigslistBargain price-path analysis, TMI’s opening offers average 3.74 percentage points of listing price below conventional thinking, while final agreed prices average 1.95 points lower. These are equal-model averages for eligible assignments with recognized opening offers and completed agreements across repeated runs, not results for every negotiation.

The difference between opening and final prices suggests that TMI changes both initial positioning and subsequent adjustment.

Selected Sotopia scenarios also show more inquiries into existing commitments after a counterpart refuses a weekend-trip invitation. Those inquiries sometimes produce concrete alternative arrangements even though the original invitation remains unsuccessful.

Across the full Sotopia evaluation, TMI also produces more explicit agent-initiated exits and fewer turn-limit endings. This indicates different termination behavior, not necessarily better completion in every case.

These are behavioral associations rather than proven causal mechanisms. Some case studies deliberately focus on scenarios showing substantial TMI gains, making their frequencies unsuitable as estimates of behavior across all conversations.

Less repeated reasoning, but no established total cost saving

An agent that repeatedly reasons about the same dialogue may regenerate similar explanations at successive turns. TMI changes this measurable pattern.

Across all eight model–environment combinations, TMI reduces API-reported reasoning tokens per decision by 42.0–81.5% relative to conventional thinking.

The researchers also measure how frequently content-word four-grams in a new reasoning trace have already appeared in earlier reasoning from the same dialogue. Average historical phrase reuse decreases across all eight combinations. After matching reasoning-text lengths, the reduction remains in seven; the DeepSeek ESC-Eval comparison has a confidence interval spanning zero.

This supports a narrower conclusion than claiming superior reasoning quality: TMI generates less repeated reasoning text under the measured conditions.

It also does not establish cheaper production execution. The token analysis excludes counterpart and evaluator costs and does not measure total inference expenditure or latency. Input-context size, provider billing, and workflow completion costs remain separate operational questions.

What a business pilot should compare

For customer-service agents, TMI suggests an architectural experiment: retain the same model, task instructions, customer scenarios, and permitted actions, while changing whether dialogue proceeds through conventional messages or native tool execution.

A meaningful pilot should evaluate completed resolutions, correct information elicitation, use of new facts, refusal handling, appropriate escalation, and termination—not merely response preference.

For sales and negotiation agents, favorable margins should be assessed alongside agreement rates, transaction constraints, and counterpart experience. The paper’s simulated bargaining outcomes do not establish that a more aggressive opening offer benefits actual buyers, sellers, or continuing commercial relationships.

For inference engineering, reasoning tokens per decision can be a useful diagnostic, but total tokens, latency, tool overhead, and cost per satisfactorily completed task are more relevant deployment measures.

These are Cognaptus recommendations derived from the experimental findings, not workflows that the paper directly validates.

The deployment boundary is human interaction

The study evaluates two model configurations interacting with simulated counterparts. It does not observe real customers, validate emotional-support outcomes with human participants, or conduct actual financial transactions.

The outcome measurements also prioritize the agent’s assigned objective. In persuasive interactions, stronger goal attainment may create risks to counterpart autonomy or welfare. A production evaluation therefore needs safeguards for consent, refusals, permissible influence, and customer benefit alongside success metrics.

TMI demonstrates that interaction organization can materially affect simulated multi-turn agent performance without additional training. It does not establish that one protocol is universally preferable or that benchmark success translates automatically into safe, economical human-facing operation.

For an agent developer, the next step is a controlled runtime comparison on representative conversations. The same model may make different decisions when each reply is treated as an action followed by an observation. Whether those decisions improve the complete service outcome remains an application-specific question.


Cognaptus: Automate the Present, Incubate the Future.


  1. Mingyu Chen and Jingkai Lin and Xiaofen Xing and Liankai Huang and Ruiqi Xiong and Xiangmin Xu (2026). Can Large Language Models Extend Their Agentic Capabilities to Goal Pursuit in Open-Ended Multi-Turn Interaction?. arXiv:2611.01552. https://arxiv.org/abs/2611.01552 ↩︎