acceptodds
Under review as a conference paper at ICLR 2027

When to Shape a Teammate: Inducing Decision-Relevant Belief Changes for Better Cooperative Futures

Abstract

Theory of Mind studies how agents infer others' hidden states from behavior, while legibility studies how behavior can shape such inference. In cooperative tasks, however, interpretation itself is not the objective: what matters is how it changes a teammate's behavior and future team outcomes. This raises a central question: when should an agent intentionally sacrifice its own immediate optimality to shape another agent's future behavior? A large belief change may have little effect when the teammate's decision remains unchanged, whereas a small change near a decision boundary can redirect its behavior and lead to a different cooperative future. We develop a closed-loop planning approach that evaluates candidate actions through the teammate responses and future joint outcomes they induce. The agent predicts how its behavior changes the teammate's belief and subsequent response, compares the resulting joint future with task-direct behavior, executes one action, and replans from the realized joint state. Across 400 held-out GridWorld scenes and 87 PARTNR-derived household instances, our approach reaches 97.0% and 62.1% complementary-task accuracy, compared with 95.5%/96.25% and 36.8%/46.0% for open- and closed-loop belief-based cooperative planning baselines.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Multi-Agent Collaboration, Embodied AI, Cooperative Planning

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_Xkm2OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper deals with agents in a cooperative setting and tries to answer the question of when an agent should try to shape another agent's future behavior. At each step, the focal agent (Main) predicts whether a shaping action changes the teammate's (Helper's) belief and therefore Helper's behavior. It evaluates whether Helper picks the correct complementary task, and the team completion time, under the shaping action, and compares this to task-direct behavior. It then decides whether to shape. Based on Helper's realized response, Main replans the next step. This approach is shown to be more effective when compared to belief-based cooperative planning (BSCP) baselines. The WSP framework is evaluated on 400 GridWorld scenes and 87 household instances derived from PARTNR, both with scripted Bayesian teammates. WSP outperforms the baselines of open and closed loop BSCP: it reaches 97.0% complementary-task accuracy in GridWorld (vs. 95.5% for Open-Loop and 96.25% for Closed-Loop BSCP) and 62.1% in Household (vs. 36.8% and 46.0%).

Critical strengths
  • Just knowing or influencing a change in belief without influencing a change in action does not help us in completing the task. Therefore the premise of the paper, influencing the teammate's belief only when it leads to a change in action, makes it clearly motivated.
  • The 2x2 comparison of Open-Loop BSCP, Closed-Loop BSCP, Timed Open-Loop BSCP and WSP lays out transparently the separation of contribution of timing from contribution of replanning.
Critical weaknesses
  1. The paper uses a fixed, scripted Bayesian teammate (Helper). A Bayesian teammate needs a prior, and its change in behavior follows directly from its belief update by design. If an LLM-based agent (or a human) acted as the teammate, it would provide some interesting insights on whether the approach generalizes beyond scripted teammates
  2. Both the open and closed BSCP are designed by the authors as baselines. The paper does not compare against legibility-based planners or other prior work.
  3. In the Household task, most of WSP's gain comes from choosing when to shape: Timed Open-Loop, which does not replan, already reaches 59.8%. Replanning only adds 2.3 points (about 2 of 87 instances), and no confidence interval is reported for this difference. The evidence for closed loop replanning being beneficial is weak.
  4. The evaluation size in the PARTNR-derived Household setting is small (87 instances from 30 episodes), and some confidence intervals are wide (e.g., [7.79, 27.16] for the 16.09-point gain over Closed-Loop BSCP).
Questions for authors
  1. Have any experiments been run with a teammate other than the assumed model (e.g., an LLM agent or a human study)?
  2. Have any experiments been run with a legibility-maximizing baseline or a published planner as a baseline?
  3. Are results on more Household episodes available?
  4. What is the confidence interval for the accuracy difference between WSP and Timed Open-Loop?
Reviewer_aEsMOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors introduce When-to-Shape Planning (WSP), a cooperative planning method that allows an agent to decide when to sacrifice immediate task efficiency for better joint outcome by shaping a teammate's belief. The approach weighs the task cost of deviating from a task-direct action against the additional shaping gain this deviation produces. At each step, WSP predicts how each candidate action changes a teammate's belief and response, compares the resulting joint future with task-direct behavior, executes one action, observes the teammate's realized response, and replans. The claim is that it builds on previous work by making the choice of 'when to shape' a state-dependent planning decision. The authors also report that their method performs better than other baseline approaches in two collaborative tasks: a simple GridWorld construction task and a more complex PARTNR-derived household task.

Critical strengths

The main contribution is a state-dependent criterion for deciding when to shape a teammate's belief, which weighs task cost against shaping gain. This is potentially a strong theoretical contribution to joint planning and the social cognition literature.

Critical weaknesses

The key theoretical contribution of WSP is rooted in the inequality 6 (pg 5). However, I do not fully understand where this is used in the decision process. The authors state that the controller optimizes the lexicographic objective in Eq. 8, but do not explain the role of inequality 6 in this objective, or how they are related. Based on fig. 1, I would expect the focal agent's decisions to be guided by inequality 6. There appears to be a theoretical disconnect between sections 3.2 and 4.1, and it will be helpful if the authors can clarify how the theory presented in section 3.2 is folded into the models described in section 4.1.

I don't agree with the statement "The results show that choosing when to shape and replanning from the realized response provide complementary benefits in longer interactions" on line 357 of pg 7. The confidence intervals for WSP's accuracy in the task include the accuracy for Timed Open-Loop. The large gains of both Timed Open-Loop and WSP methods over Closed-Loop BSCP suggest that there are benefits to choosing when to provide the instructions. However, I am skeptical about the claimed benefit of replanning on top of the Timed Open-Loop approach without any information about the confidence intervals for the paired difference between the accuracy values of the two approaches. Relatedly, in section 4.5, what are the confidence intervals for the number of steps mentioned in lines 430 and 431? The numbers of initial, delayed, and no-shaping cases for the two methods do suggest that they produce different behaviors, but to me it is not clear that one is better than the other.

Questions for authors

See weaknesses

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

Research on theory of mind often asks how an agent reads another’s hidden goals from behavior and work on legibility explores how to act so one’s goals are...err...legible. The authors argue cooperation needs a further question. Changing what a teammate believes is pointless if the teammate would do the same thing anyway, and a tiny belief change can matter substantially if it tips the teammate into a different choice. Shaping also costs something, because the shaping action usually delays one’s own task. So their real decision is when, if ever, it pays to deviate from your own job in order to change what your partner will do next. The aspiration is a planning rule that answers this by looking through the belief change to the teammate’s future behavior and the joint outcome. The authors formalized "shaping value" as the difference between the team’s future with and without the belief change, set it against the task cost of the deviating action, and made the agent re-evaluate after every step as it sees how the teammate actually responds. In a small grid(world) task and in a longer household task built on an embodied benchmark, this planner coordinated more reliably than belief-aware planners that either commit to a plan or replan without considering timing. Control experiments show the benefit flows through the teammate’s changed behavior rather than the action’s physical effect, vanishes when the teammate stops reading behavior at all, and depends heavily on timing. The best moment to shape is often later than the moment with the strongest signal. The idea is elegant and the controls are convincing within the setup (with a fully known, hand-built teammate.)

Critical strengths

The question seems contemporary and important. The framing (decision relevance, not belief magnitude) is clear and the controls are the best of the papers I have evaluated in this back: 1) timing-aware open-loop isolates onset selection from replanning, 2) the matched counterfactual isolates the belief channel, 3) the q sweep shows the mechanism collapse, and 3) the GridWorld saturation is reported rather than hidden. Nicely done.

Critical weaknesses
  • Propositions 1 and 2 are identities. Eq. 6 reduces to Q(a) > Q(a_task). With a known teammate model, this is standard belief-space planning, and the paper should probably say so. What distinguishes WSP from Closed-Loop BSCP is onset selection over candidate states, and exhaustive onset search (Table 8) beats WSP’s E(s)-peak heuristic. With horizons of 20–30 steps, why not just search exhaustively and drop the heuristic?

  • The teammate is a hand-built Bayesian receiver with known q, β, η, and a one-step-lag follow rule that makes shaping valuable. There is no learned, human, or LLM teammate. The mismatch sweep perturbs one parameter at a time over narrow ranges.

  • There are some missing baselines: a legibility planner (Dragan-style) with an explicit cost trade, and on GridWorld a direct solver of the joint problem to show how far WSP sits from optimal.

  • Scale. 87 household instances from 30 episodes; CIs are wide ([7.79, 27.16] on the headline gain). Perfect information sits at 100% vs 62.1%.

  • The lexicographic objective subordinates time to correctness, so the shaping-cost trade the paper motivates is never exercised as a trade.

  • E(s) is used in 4.1 and defined only in Appendix B.6--let's bring the defeinition back!

Questions for authors
  • What is WSP computationally beyond Closed-Loop BSCP plus onset search?

  • How do the results stack up against an exact joint solver on GridWorld?

  • Have you explored the robustness to a teammate whose response rule, not just parameters, differ from the model?

Reviewer_ueHgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

The paper claims that an agent should shape a teammate’s behavior only when it actually changes the teammate’s decision and the better team outcome outweighs the task cost.

  • Problem: Previous Theory-of-Mind and legibility works only interpret how an action shifts teammate's belief, but not decide whether worth it or when.
  • Approach: The author develop a closed-loop planning approach that evaluates candidate actions through the teammate responses and future joint outcomes they induce.
  • Main evidence: WSP improves coordination (table 1), where timing matters most.
Critical strengths
  1. A reframing of shaping as decision-relevant belief change, made by the matched-counterfactual decomposition. It explains why an informative action can be useless and a small one decisive. That gives the field a vocabulary for separating “signal sent” from “value created.”
  2. A clear controlled 2 by 2 design with cluster-bootstrap. Within this setting, timing matters is clearly supported statistically.
Critical weaknesses
  1. The paper never explains why prior methods are excluded. With every reference point author-built, the paper’s improvement over prior work cannot be measured against an outside reference.
  2. The claim that WSP adds a new “when to shape” capability beyond “standard belief-space cooperative planning” is tested only against author-built, restricted baselines.
  3. The central contrast with legibility, that decision relevance rather than belief-change magnitude determines shaping value (contribution 2), is supported only formally but not emprically.
  4. The main storyline is hard to follow because the pieces linking the idea to the experiments are left undefined in the main text. For example, “BSCP” is never defined as an acronym and has no citation, yet it is the main baseline family. The candidate-selection score is used in §4.1 but never defined in main body, only in the appendix.
Questions for authors

Q1 (for W1). You cite legibility and observer-aware planning (Dragan & Srinivasa, 2013; Miura & Zilberstein, 2021) and influence-aware planning (Sadigh et al., 2016; Nikolaidis et al., 2017) as the closest prior work. Why is none of them included as a baseline, and what prevents applying them to your tasks?

Q2 (for W2). Sadigh et al. (2016) replans at every step while modeling how its actions change the other agent’s behavior. How does WSP compare with such a planner when it searches all shaping onsets with your Helper model?

Q3 (for W3). Can you run a legibility or observer-aware controller (e.g., L-MDP or PoLMDP) end-to-end on Household? Can you also show how belief-shift size relates to whether the Helper’s decision changes?

I am happy to raise my score if you answer these questions that clearly support your arguments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.