CoBel-World: Embodied Multi-Agent LLM Framework

• CoBel-World is a framework for embodied multi-agent LLM systems that integrates environment perception with mental-state modeling via a symbolic belief language.
• It uses a zero-shot Bayesian-style belief collaboration procedure to update and predict states, effectively detecting miscoordination while reducing communication by 22–60%.
• The system improves task completion efficiency by 4–28% and scales to multi-agent scenarios by unifying environmental and teammate belief representations.

CoBel-World is a framework for embodied multi-agent LLM systems that equips agents with an explicit, intent-aware collaborative belief world: an internal, structured representation that jointly models the physical environment and teammates’ mental states. Its design has two pillars: a Symbolic Belief Language (SBL) for parsing open-world task knowledge into beliefs, and a zero-shot Bayesian-style belief collaboration procedure that updates and predicts beliefs from partial observations and communications in order to detect miscoordination and trigger adaptive communication only when needed. In the reported evaluations on TDW-MAT and C-WAH, CoBel-World reduces communication cost by 22–60% and improves task completion efficiency by 4–28% over strong LLM baselines (Wang et al., 26 Sep 2025).

1. Problem setting and formal model

CoBel-World is motivated by decentralized, partially observable embodied collaboration. Agents receive egocentric observations and sparse messages, must act without a central controller, and therefore must decide both what to do and when to communicate. The framework targets the failure modes identified for prior LLM collaboration systems: fixed communication protocols, lack of explicit modeling of collaborators’ internal states, inconsistent plans, and redundant communication.

The internal state is formalized as a collaborative belief world

$B_t = (E_t, M_t),$

where $E_t$ captures zero-order beliefs about the environment and $M_t$ captures first-order beliefs about what another agent believes. In the stated formulation, $E_t$ includes beliefs about world states, objects, rooms, containers, exploration progress, possession, and container contents, while $M_t$ includes beliefs about collaborators’ knowledge and plans (Wang et al., 26 Sep 2025).

The underlying decision problem is cast as a DEC-POMDP:

$$\mathcal{M} = \langle I, \mathcal{S}, \{\mathcal{A}_i\}, \{\mathcal{O}_i\}, T, \{O_i\}, R, h \rangle,$$

where $I$ is the agent set, $\mathcal{S}$ the environment state space, $\mathcal{A}_i$ actions, $\mathcal{O}_i$ observations, $E_t$0 transitions, $E_t$1 observation models, $E_t$2 a shared reward, and $E_t$3 the horizon. Each agent follows a local policy based on its observation-action history and maintains a belief $E_t$4 over $E_t$5. CoBel-World adheres to this belief-update paradigm conceptually, but replaces explicit numeric filtering with LLM-driven symbolic updating and prediction.

This formalization makes CoBel-World simultaneously a world model and a collaborator model. The environment and the inferred mental states of teammates are not maintained as separate auxiliary structures; they are co-resident within $E_t$6. That joint representation is central to the framework’s claim that proactive miscoordination detection requires both state estimation and intent inference.

2. Symbolic Belief Language and structured belief representation

The Symbolic Belief Language provides the representational substrate through which task instructions, observations, and dialogues are converted into machine-manipulable beliefs. Its atomic syntax is defined by expressions of the form:

• ?belief ::= ?entity PREDICATE ?entity | ?entity ATTRIBUTE ?state
• PREDICATE examples: IN, [HOLD](https://www.emergentmind.com/topics/higher-order-langevin-dynamics-hold), AT, CONTAIN, NEAR
• ATTRIBUTE examples: EXPLORED, CONTENTS

Entities include agents, objects, and locations, and are grounded as <name>(id), such as <kitchen>(2000) or <apple>(12123). Mental-state nesting is expressed by the BELIEVE operator. Zero-order beliefs take the form ?agent BELIEVE ?belief; first-order beliefs take the form ?agentA BELIEVE ?agentB BELIEVE ?belief (Wang et al., 26 Sep 2025).

The belief types represented in SBL include environmental beliefs, task beliefs, and intent or plan content maintained alongside beliefs. Examples given for zero-order rules include:

• ?agent BELIEVE ?object IN ?room
• ?agent BELIEVE ?agent HOLD ?object
• ?agent BELIEVE ?container CONTAIN ?object
• ?agent BELIEVE ?room EXPLORED ?exploration_state
• ?agent BELIEVE ?agent AT ?room

First-order rules mirror these with an outer agentA BELIEVE agentB BELIEVE ... structure.

The framework uses a propose-and-revise collaborative process to construct consensus belief rules from task descriptions, capabilities, and constraints. During execution, task instructions are parsed into goal-oriented rules and constraints; visual observation is parsed into zero-order beliefs; and dialogue is parsed into first-order beliefs. The examples in the paper make the semantics explicit. If Alice observes Bob holding a banana, the parsed belief is Alice BELIEVE Bob HOLD <banana>(12234). If Bob says, “I found an apple in the kitchen,” Alice parses Alice BELIEVE Bob BELIEVE <apple>(12123) IN <kitchen>(2000).

The SBL schema is intended to enforce consistent, compositional encoding. Exploration states are discretized as none, part, and all, and the belief store is maintained in structured sets or maps keyed by entity ids and predicate types. This symbolic regularization is what allows the same framework to operate across open-world task descriptions while retaining enough structure for conflict analysis and belief revision.

3. Zero-shot Bayesian-style belief collaboration

CoBel-World adopts the Bayesian filtering paradigm for belief update and prediction, but instantiates it through zero-shot LLM reasoning rather than through explicit state-factorized probability models. The belief update and prediction equations are stated as:

$E_t$7

$E_t$8

and

$E_t$9

In CoBel-World, these are realized symbolically. The LLM generates candidate beliefs consistent with observations, messages, and belief rules during the update step, and predicts plausible future beliefs and intents for self and collaborator during the prediction step. Priors and likelihoods are approximated by reasoning rather than by explicit numeric estimation: $M_t$0 is implicitly represented by the set of candidate beliefs inferred from history and rules, and $M_t$1 is approximated by consistency checks that retain explanatory beliefs and prune inconsistent ones. Normalization is implemented as selection or ranking over candidate beliefs; the paper notes that a softmax over LLM preference scores can be used in implementations requiring weights (Wang et al., 26 Sep 2025).

The update routine is organized into four stages. First, the latest visual observations and messages are parsed into candidate SBL facts. Second, zero-order beliefs are updated by adding new facts and revising conflicting ones, such as exploration states. Third, first-order beliefs are updated using Theory-of-Mind-style prompting to interpret the collaborator’s perspective and any explicitly stated plans. Fourth, the framework predicts likely hidden targets, containers, and short efficient plans for the agent itself, and generates top-$M_t$2 likely collaborator plans from the updated first-order belief state.

A worked example illustrates the mechanics. Alice may initially believe that <livingroom>(1000) is EXPLORED none and that a <wicker_basket>(5388017) is in the living room, while also maintaining Alice BELIEVE Bob BELIEVE <kitchen>(2000) EXPLORED part. If Alice then sees Bob holding a banana and receives Bob’s message “Found an apple in the kitchen; I’ll grasp it next,” CoBel-World updates both environmental and mental-state beliefs, and extracts Bob’s plan as “go grasp <apple>(12123).” This example is intended to show how observation parsing, message interpretation, and collaborator-plan extraction are unified within the same belief world.

4. Miscoordination detection, adaptive communication, and action selection

CoBel-World uses the belief world to detect miscoordination before overt conflict occurs. Two sources of miscoordination are defined: conflicting plans and belief misalignment. If $M_t$3 is the agent’s planned action sequence and $M_t$4 is a candidate collaborator plan, a conflict function is defined such that

$M_t$5

when actions overlap in room-time with incompatible resource usage, and $M_t$6 otherwise. Belief misalignment is captured by the set

$M_t$7

Heavy misalignment is defined when $M_t$8 or when any fact in $M_t$9 implies wasted action, such as exploring a fully explored room.

The communication rule is correspondingly selective: the agent communicates when heavy miscoordination is detected by the LLM analyzer, instantiated as a prompt-driven classifier that outputs “Yes” if heavy miscoordination is present and “No” otherwise. If communication is triggered, the message is constructed to include only misaligned information and the sender’s current subplan. The example given is concise and operational: “I’ve fully explored <kitchen>(2000) and moved <apple>(12123). I’ll grasp <banana>(12234) and transport to the bed.” This is not general chit-chat but task-relevant state correction plus plan declaration (Wang et al., 26 Sep 2025).

Planning uses the updated beliefs to synthesize short-horizon actions under uncertainty and constraints such as carrying capacity, transport costs, and room exploration status. The canonical objective is

$E_t$0

where $E_t$1 captures expected progress under belief uncertainty. The planner checks whether the current subplan has been completed; if not, it selects the next action from a constrained list that includes going to a room, exploring the current room, going to grasp an object, putting an object into a container, and transporting.

The operational agent loop is correspondingly modular: perceive, parse, update_beliefs, predict, detect_conflicts, decide_communication, plan, and act. The system implementation described in the paper instantiates these modules with perception parsing, a belief updater, a belief predictor, a planner, and a communicator. The dominant computational cost is LLM inference per step, while belief-store operations are linear in the number of facts.

5. Embodied benchmarks, empirical results, and implementation parameters

The framework is evaluated on two embodied multi-agent benchmarks. TDW-MAT comprises food-transportation and object-transportation tasks. Each episode contains 10 target objects, 2–5 containers, and rooms drawn from living room, office, kitchen, and bedroom. Agents can carry two items, containers hold up to 3 items and are lost upon delivery, observations are partially observable and include egocentric RGB-D, optional oracle perception, messages, positions, and held objects, and the action space includes move forward, turn, grasp, put-in, drop, and send message. The primary metrics are transport rate, defined as the fraction of subtasks completed within 3000 frames, and communication cost, measured as tokens per episode.

C-WAH contains five household tasks—Prepare tea, Wash dishes, Prepare meal, Put groceries, and Set up table—with subgoals defined by ON/IN(x, y) predicates. It is evaluated in both symbolic and visual observation settings. The symbolic modality provides access to objects in the same room with name, location, and relational attributes; the visual modality provides egocentric RGB-D plus auxiliary observations. The action space includes navigate, turn, grasp, open/close, put, and send message. The reported metrics are average steps to finish all subgoals and communication tokens (Wang et al., 26 Sep 2025).

Setting	CoBel-World	Comparison
TDW-MAT, GPT-4o, transport rate	86.67	Capo 82.50; CoELA 80.42
TDW-MAT, GPT-4o, communication cost	1736	CoELA 2224; Capo 6940
C-WAH, GPT-4o, symbolic average steps	48	CoELA 63; Capo 66; MHP 68
C-WAH, GPT-4o, visual average steps	71	MHP 90; CoELA 98; Capo 103
C-WAH, GPT-4o, symbolic communication cost	377	CoELA 938; Capo 7231
C-WAH, GPT-4o, visual communication cost	360	CoELA 889; Capo 4696

Across both environments, the paper reports that CoBel-World reduces tokens by 22–60% and improves task efficiency by 4–28%, depending on the setting. For Qwen3-32B agents on TDW-MAT, similar trends are reported with smaller gains: transport rate 67.92 for CoBel-World versus 67.50 for Capo, and token cost 2155 for CoBel-World versus 3793 for CoELA and 7709 for Capo. In the C-WAH symbolic ablation, full CoBel-World achieves 51 average steps, compared with 55 for the system without SBR and 68 for the system without BBC. When scaling the number of agents, the reported values are 51 for 2 agents, 47 for 3 agents, and 43 for 4 agents; the paper notes that task complexity limits marginal gains beyond 3 agents.

Implementation details are explicit. The reported LLMs are Qwen3-32B via Aliyun API and ChatGPT-4o via OpenAI API, with temperature 0.7, top-p 1, and max tokens 512. Prompt templates for belief construction, update, prediction, adaptive collaboration, communication, planning, and replanning are described in the appendix, and code is linked through an anonymous repository. The reproducibility package therefore centers on prompt structure, benchmark specification, and the SBL-mediated belief store rather than on a learned model checkpoint.

6. Relations to adjacent paradigms, limitations, and name disambiguation

CoBel-World is explicitly connected to Theory-of-Mind, BDI, and DEC-POMDP frameworks. It encodes first-order beliefs via BELIEVE and uses ToM-style prompting to interpret others’ messages; it aligns with the belief component of BDI while explicit plan extraction and prediction serve as an operational representation of intentions; and it follows decentralized belief-update principles under partial observability while replacing numeric filters with LLM-driven symbolic updates and predictions. Relative to prior LLM collaboration frameworks such as CoELA and CaPo, the distinguishing design choice is explicit mental-state modeling combined with adaptive communication triggered by miscoordination analysis (Wang et al., 26 Sep 2025).

The limitations stated in the paper are centered on reliability, scaling, and transfer. Hallucinated or inconsistent beliefs can arise from LLM errors, misparsed observations, or incorrect likelihood approximation. Communication control can over- or under-trigger when borderline cases are misclassified. As agent count grows, first-order belief maintenance and plan conflict checking scale combinatorially, suggesting the need for further abstractions or probabilistic Theory-of-Mind. Visual ambiguity and partial observability can degrade belief accuracy, and domain-specific id assignment and predicate coverage must be extended for novel environments.

A terminological ambiguity arises from the overlap between “CoBel-World” and the earlier “2CoBel: An Efficient Belief Function Extension for Two-dimensional Continuous Spaces” (Pellicanò et al., 2018). The latter is a belief-function framework for continuous 2D worlds that represents focal elements as polygons, uses polygon clipping for set operations, and supports decision making through generalized pignistic probability and maximal intersections in a geometric evidential space. CoBel-World, by contrast, is the 2025 embodied multi-agent LLM framework centered on symbolic belief language, zero-shot Bayesian-style updates, and intent-aware collaboration. The shared vocabulary of “belief” and “world” should therefore be read as a naming overlap rather than as an indication of a common implementation or formalism.

Within the scope reported so far, CoBel-World’s main contribution is not a new probabilistic calculus but a structured interface between open-world LLM reasoning and decentralized embodied coordination. This suggests that its significance lies in how symbolic beliefs, collaborator modeling, and adaptive communication are coupled into a single control loop for partially observable multi-agent tasks.