acceptodds
Withdrawn from ICLR 2027

SEGUE: Bridging Symbolic Coordination and Embodied Execution for Sequential Social Dilemmas

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Long-horizon coordination in embodied shared-resource environments requires agents to follow collective strategies while adapting to local observations, execution constraints, and failures. Symbolic solvers such as Value-Set Iteration (VSI) can compute effective high-level coordination policies, but their abstract models do not specify how recommendations should be grounded under changing embodied conditions. We propose SEGUE (Symbolic-to-Embodied Guidance for Uncertain Execution), which treats private VSI recommendations as revisable guidance and grounds action type and quantity in each agent's current context. We further introduce ESD-Bench, a closed-loop benchmark that evaluates the full symbolic-to-embodied process from recommendation and program generation to execution and resource transition. In free-action evaluation, SEGUE achieves welfare of versus for the strongest baseline in the Balanced setting, achieves the best observed welfare of under Scarcity, and executes of requested actions and quantities in both settings. It also matches symbolic VSI welfare under aligned dynamics and maintains joint feasibility under action-set perturbations. Ablations show that the gains stem from current coordination guidance rather than additional context or a particular backbone.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
sequential social dilemmas, long-horizon planning, multi-agent systems

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_HzZ8OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces SEGUE, a framework that combines long-term coordination guidance from the symbolic solver VSI with decentralized LLM-based action selection and program execution. Agents can revise recommendations based on local constraints and use realized execution outcomes to update subsequent plans. The authors also introduce ESD-Bench, a Minecraft Cleanup benchmark, and report improved social welfare and execution reliability in selected settings. However, evaluation is limited to a single task with two agents, and the methodological contribution primarily involves integrating existing components.

Critical strengths
  1. The paper addresses failures that can arise when symbolic coordination policies are executed in an environment. Agents can adjust recommendations to local constraints and update subsequent plans using actual execution outcomes.
  2. The evaluation tracks action choices, requested and completed quantities, and resource transitions, helping distinguish poor coordination from execution failures. Controlled experiments and ablations also help assess the contribution of VSI guidance.
Critical weaknesses
  1. The paper motivates SEGUE through the symbolic-to-embodied deployment gap, but the evaluation does not sufficiently capture embodied complexity. It considers only a single task with a primarily text-based planning interface, providing limited evidence on embodied challenges such as multimodal perception, spatial reasoning, and recovery from physical execution failures.
  2. The evaluation is limited to a single Cleanup task with two agents. Variations in initial resource levels and action availability do not introduce substantially different coordination problems. This narrow scope makes it difficult to assess whether the method’s benefits extend to other social dilemmas, task structures, or larger teams.
Questions for authors
  1. What information is included in each agent’s local observation? Are spatial information and reachability provided by the environment interface, or must agents infer them?
  2. Is there evidence that SEGUE’s benefits extend beyond the two-agent Cleanup setting, particularly to tasks with different coordination structures or larger teams?
Reviewer_cVwpOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper concerns a multi-agent system with a central coordinator giving recommendations to each agent. Each agent can deviate from the recommendation. The central coordinator aims to maximize its own reward while ensuring that individual agents cannot improve by deviating.

The planner encodes the environment into a finite system and computes recommendations within the simpler system. However, the symbolic model may leave out details that may render recommendations impossible.

The paper proposes SEGUE, a method where each agent is equipped with an LLM that adjusts the recommendation. Furthermore a Minecraft benchmark is introduced for evaluation.

Critical strengths
  1. The problem seems to be well-motivated, and the overall approach appears to be a promising direction.
Critical weaknesses
  1. The presentation of Section 3.3 is hard to follow.

The section introduces local observations, interaction history, judge feedback, reflection memory, and an agent planner without clearly explaining what these components contain or do. What information does each agent observe? What are the judge, memory, and feedback? Eq. (16) gives the planner’s inputs and outputs, but does not explain how it selects an action and quantity. What exactly does the LLM do, and how does the VSI recommendation influence its decision? And so on.

Appendix D clarifies the execution order, but largely repeats the same abstract operations. Unfortunately, even after careful reading, I cannot reconstruct how the method actually works. This substantially limits my ability to assess the contribution.

  1. The connection between the formal description and the benchmark is unclear.

The formal model introduces a coordinator with its own reward r_0. What is this reward in the benchmark? Is it the sum of the agents’ rewards reported as social welfare, or something else? What are the concrete reward and transition functions used by VSI? The method also introduces quantities and executable programs, but does not explain how these correspond to actions in the formal game, particularly when execution is partial or unsuccessful. I therefore find it difficult to determine precisely which formal problem the implementation solves and what the experiments establish about the theoretical framework.

Questions for authors
  1. Please specify the benchmark as an instance of the formal game.
  2. Please provide a precise description of 3.3. In particular, the role of the LLM.
Reviewer_neeSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the gap between symbolic coordination and physical execution in embodied multi-agent environments. It proposes SEGUE, a framework that combines the long-horizon coordination capabilities of Value-Set Iteration (VSI) with LLM agents that adapt action types and quantities to local conditions and generate executable programs. The paper also introduces ESD-Bench, a Minecraft-based benchmark for evaluating the complete pipeline from symbolic recommendations to realized outcomes. Experiments across different resource regimes and execution conditions show that SEGUE achieves competitive or improved social welfare while maintaining reliable execution in the tested settings.

Critical strengths
  1. The paper addresses an important gap between symbolic coordination and executable behavior in embodied multi-agent systems. Combining long-horizon VSI guidance with context-dependent action adaptation offers a useful approach to this problem.
  2. ESD-Bench records symbolic recommendations, intended actions and quantities, generated programs, and actual execution outcomes. These traces and the separate welfare and execution metrics help distinguish coordination failures from execution failures.
Critical weaknesses
  1. The manuscript describes the overall loop, but several details needed to reconstruct ESD-Bench and the planner are not specified at an implementation level: the map/layout and exact reward, pollution, and regeneration rules; the symbolic reward/transition model; the action-quantity scoring rule and admissible candidates; and key prompts. Sections 5.6 and A.6 identify which planning components are removed in the 100-turn ablations, but do not specify their computation or removal precisely. These gaps limit both reproducibility and interpretation of the reported ablations.
  2. The action-set perturbations expose current availability to embodied agents, while Symbolic VSI follows its original recommendations and can request disabled actions. This establishes adaptation relative to an unrepaired symbolic policy, but does not isolate what the proposed planner adds beyond simple action repair. A VSI baseline with the same availability information and execution interface, plus a feasibility filter and fixed fallback, is needed to assess that contribution.
Questions for authors
  1. Regarding Weakness 1, could the authors provide the exact environment and symbolic-model configuration, planner scoring/candidate rules, and key prompts, including the implementation of the two 100-turn ablations?
  2. Regarding Weakness 2, how does SEGUE compare with VSI given the same availability information and executor, followed by a feasibility filter and fixed fallback?

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

Related papers

Discussion (0)

Sign in to comment.