acceptodds
Withdrawn from ICLR 2027

Do Agents Develop Tacit Understanding? From Repeated Collaboration to Partner-Specific Coordination

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In long-term agent collaboration, fixed partners gradually accumulate relational knowledge about each other's response patterns and coordination habits. Human long-term partners transform such knowledge into tacit understanding, thereby reducing repeated negotiation in similar situations. Prior work shows that LLM agents can coordinate implicitly, form conventions, and adapt to partner history, yet it remains unclear what behavioral evidence distinguishes genuine dyadic tacit understanding from coordination explained by current task structure, generic experience, partner predictability, or memorization. To address this gap, we formalize Dyadic Tacit Understanding (DTU) as a directional, history-induced shift from a history-neutral coordination-response law, and define Genuine Dyadic Tacit Understanding (GTU) by the joint requirements of Formation, Interaction Specificity, and Cross-Instance Generalization. We further propose a decentralized online relational learning framework that enables each agent to independently learn its partner's coordination-response patterns and transform them into private distributional guidance for its own decisions under limited explicit communication. We validate the proposed theory and framework on multiple collaborative benchmarks, showing that agents can develop reusable partner-specific coordination knowledge from shared experience while maintaining collaboration quality with reduced explicit communication. Code is available at https://anonymous.4open.science/r/tact-agents-F45C.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Dyadic Tacit Understanding, LLM Agents, Persistent Collaboration, Partner-Specific Coordination, Multi-Agent Communication

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_X9boOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper focuses on the problem of “tacit understanding”, the ability of a pair of agents, after repeated interaction, to predict the coordination patterns of one another in a way that generalises to unseen tasks and results in a reduced need for explicit communication.

They state that “tacit understanding” requires 3 properties: formation, specificity and generalisation. Next, they introduce a decentralised framework to enable tacit understanding for a pair of LLM agents and benchmark their method on the Overcooked, Hanabi and CoBlock environments.

Critical strengths

This paper raises an important question about understanding how deeply a pair of agents model each other after repeated interaction.

Also, the experiments indicate some promising results across three established coordination benchmarks.

Critical weaknesses

Poor writing; a lot of missing details

  1. The writing exposition is poor, making it difficult to follow the method and experiments.

    1. For example: Eq (3) introduces \rho, which is never used again in the main paper.

    2. Many symbols are introduced without defining what they represent, e.g., m_train, etc., in Eq 9.

  2. The paper misses key details, making it difficult to properly assess the contributions of the paper. This also makes reproducibility of the results challenging. Some examples are:

    1. “neutralised history” — how is this computed?

    2. “directional interventions” — how are they computed?

    3. “cell” — how is this computed? Who defines the mapping from states to cells? Is this scalable?

    4. Thresholds m_train, m_val, etc. — what are the exact values used?

    5. Eq(10): how are the predictions converted to text? Can you provide an example of this?

    6. Baselines and Ablations:

      1. The baselines miss key implementation details (e.g, how is memory retrieval implemented?), making it difficult to interpret the proposed contributions — the same applies to ablations.

Tacit understanding not well defined; not validated

  1. The distinction between “tacit” understanding and basic partner modelling is not sufficiently motivated.

  2. Why do formation, specificity and generalisation imply tacit understanding? Can you satisfy the GTU criteria but perform worse in coordination?

  3. How does tacit understanding differ from regular partner modelling? Under a standard prediction model, we would still assume that a) zeroing out the history and b) replacing it with that of another partner would change its predictions? Additionally, we would expect it to generalise to test scenarios.

  4. The proposed action prediction model is a fairly standard predictor of partner actions. Which aspects of the method make it better at tacit understanding?

  5. More importantly, the paper does not present any results comparing tacit understanding of the method against the baselines. p_ID,d is defined on Line 370, but not reported in the main paper. Does the proposed method lead to greater tacit understanding (as defined by the authors) compared to the baselines?

Statistical validity; missing key ablations

  1. Table 1 reports many mixed results, with the proposed method losing out to the baselines in some tasks. Also, the numbers often show small gains with overlapping error bars. How significant are the results?

  2. What is the motivation for the specialised architecture in Eq (6)? How does it compare against a standard MLP that takes context as input and outputs a probability distribution over usage labels?

Questions for authors

Please see the Weaknesses.

Reviewer_sMuROpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether LLM agents can develop partner-specific tacit understanding through repeated collaboration. It formalizes Genuine Dyadic Tacit Understanding (GTU) through three criteria: formation, interaction specificity, and cross-instance generalization. The authors propose TACT, a decentralized relational learning framework that learns partner coordination patterns and provides private guidance without modifying the underlying LLMs. Experiments on Collab-Overcooked, Hanabi, and CoBlock demonstrate improved coordination performance and reduced communication requirements across different LLM backbones.

Critical strengths
  1. The train–candidate–active design offers a practical way to adapt frozen LLM agents. Testing predictions on later tasks before deployment, then monitoring them during use, helps control when learned partner information enters the agent’s decisions.
  2. The experiments show useful gains across tasks and model backbones. TACT achieves the highest mean quality under the restricted communication budget in all six Overcooked/CoBlock settings in Table 2, and the highest mean Hanabi score on all three backbones. The comparisons with memory and partner-modeling baselines, together with the component ablations, support the method’s practical value.
Critical weaknesses
  1. The method currently relies on fixed pairs and task-specific context, response, and usage mappings. Moving to a new workflow therefore requires interface design, which limits how easily the method can be reused.
  2. The deployment timeline needs clarification. Section 4.2 and Table 14 indicate a first qualification boundary at task 24, while the table’s footnote gives tasks 10 and 14. Please clarify the schedule used for the learning curves.
Questions for authors
  1. Which parts of the context, response, and usage mappings need to be redesigned for a new workflow?
  2. When was guidance first activated in the runs shown in Figure 2? Please make the schedule consistent with Section 4.2 and Table 14.
Reviewer_2qTWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks whether repeated collaboration leads to reusable, partner-specific coordination between LLM agents. It defines Genuine Dyadic Tacit Understanding (GTU) through formation, interaction specificity, and transfer to held-out tasks. TACT gives each frozen agent a private predictor of its partner's coordination responses. Predictions are checked on later instances before they can be turned into prompt guidance. Experiments span three task families and backbones. Under a message cap, TACT has the highest mean quality across all six Overcooked and CoBlock settings and meets the stated non-inferiority margin relative to full-budget B0. GTU tests pass in all three families.

Critical strengths

S1. The paper takes identification seriously. A response shift could reflect task structure, generic experience, or memorization. GTU therefore requires formation, specificity against alternative histories, and persistence on held-out tasks. The paper also keeps behavioral identification separate from task utility. This is more informative than treating better scores after repeated play as evidence of tacit understanding.

S2. The constrained-communication result is useful. At half of B0's median message budget, TACT has the highest mean quality in all six Overcooked and CoBlock backbone settings and stays within the prespecified five-point margin of full-budget B0. The Generic Guidance ablation also helps: replacing learned guidance with task text lowers GPT-5.5 Overcooked quality from 54.38 to 38.58 (Table 4). The learned content matters beyond merely filling a prompt slot.

Critical weaknesses

W1. The GTU test has not been checked against likely false positives. The paper reports the full GTU test only for TACT. Some methods that use past interactions also change their responses, but the paper does not give them the same full test. A useful control would be a system that remembers past examples without learning a general pattern. The GTU test should reject that system. Until these checks are done, a positive result shows that TACT meets the proposed criteria, but gives less confidence that the criteria can separate tacit understanding from simpler uses of history.

W2. The timeline of learning needs to be clearer. Figure 2 shows changes in responses and improvements in task quality before the first learned state can be used. The appendix says guidance is empty during these early interactions, while its table places the main change in responses later than Figure 2 does. The figure and table may measure different things or use different runs, but the paper does not explain this. Because learning over time is part of the GTU test, please explain how Figure 2 fits the schedule for using learned states and what causes the early gains in quality.

W3. The test of partner-specific learning is limited. In all main pairs, both agents use the same base model. The state-swap test asks whether a state learned by one pair works for another pair. This is useful, but the pairs may differ mostly in what happened during their interactions. It does not yet show whether TACT adapts to partners with different strategies or habits. Testing pairs with different kinds of agents would strengthen the claim. The paper also mentions a task-quality comparison for the state swap without giving the numbers. Those numbers would help show whether the change in behavior affects teamwork.

W4. The quality gain is clearest when messages are limited. TACT is not the best method in every comparison with the full message budget (Table 2). Its clearest advantage is that task quality remains high when messages are limited, as the paper notes. The abstract’s broader claim about using less communication could be misleading because much of the drop in sent messages comes from the message limit itself. TACT also produces fewer messages that are blocked before being sent, which may reduce the cost of running the model. Reporting how many messages are generated and how many tokens are used under both message limits would show the size of that saving.

W5. The learned pattern depends on categories chosen by the researchers. Each benchmark uses its own way to group situations and possible responses. This makes the test possible, but limits what the result shows about tacit understanding beyond those groups. Testing another reasonable way to group responses would help show whether the result holds up. A worked example would also make the paper easier to follow: show one pair’s early interactions, the predictor the system accepts, the guidance it gives, and the final GTU test.

Questions for authors
  1. Do the full GTU tests reject history-using baselines and a memorization control?
  2. How do Figure 2's early gains fit the formation and deployment schedule? What differs between the figure and the formation results of the appendix?
  3. Do the partner-specific results hold with different partner strategies or backbones? What is the task-quality effect of the cross-dyad state swap?
  4. How do generated messages and token costs compare at full and capped budgets, including blocked messages?
  5. How sensitive are the GTU conclusions to the chosen response categories?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.