CommCP: Calibrated Multi-Agent Coordination

• CommCP is a decentralized communication framework for cooperative multi-agent tasks that integrates LLM natural language coordination with conformal prediction for message calibration.
• It employs semantic value maps and frontier-based exploration to intelligently route peers’ observations, reducing redundant paths and delays.
• Empirical studies show improved success rates and reduced normalized time cost, emphasizing the crucial role of calibrated message quality.

Searching arXiv for the specified paper and closely related context. to=arxiv_search 弘鼎 银航 code 申博太阳城 ＿老司机 玩北京赛车 վճարել 玩北京赛车 to=arxiv_search 天天彩票提现_code {"^^^^^^^^2query2 OR \2^^^^^^^^"CommCP: Efficient Multi-Agent Coordination via LLM-Based Communication with Conformal Prediction\"","max_results":5,"sort_by":"submittedDate"} Searching arXiv for the exact CommCP paper and acronym collisions. CommCP is a decentralized communication framework for fully cooperative multi-agent, multi-task embodied question answering (MM-EQA) that couples natural-language coordination by LLMs with conformal prediction–based calibration of message content. In the formulation introduced by "CommCP: Efficient Multi-Agent Coordination via LLM-Based Communication with Conformal Prediction" (&&&^^^^2query2^^^^&&&), multiple robots explore a shared 3D household environment, interpret assignments provided by humans in natural language, gather information relevant to embodied multiple-choice questions, and exchange peer-to-peer text messages whose content is filtered to suppress low-confidence or distracting claims. The framework is designed to maximize task success rate under a finite horizon while reducing redundant exploration and communication-induced delay.

^^^^2(Zhang et al., 5 Feb 2026) OR \2^^^^. MM-EQA formulation

MM-EQA is defined as a fully cooperative, multi-agent, multi-task embodied question answering problem in which PRESERVED_PLACEHOLDER_^^^^2query2^^^^ robots operate in a 3D scene PRESERVED_PLACEHOLDER_^^^^2(Zhang et al., 5 Feb 2026) OR \2^^^^ with dimensions $L \times W \times H$. The scene is discretized into a voxel map $M$ with voxel side length $l$. Each robot $i$ begins from an initial 2D pose $g^i_0$, and the joint set of poses at time $t$ is $G_t=\{g^i_t \mid i=1,\dots,N_a\}$. At every time step, robot $i$ receives RGB and depth observations, denoted PRESERVED_PLACEHOLDER_^^^^2(Zhang et al., 5 Feb 2026) OR \2query2^^^^ and PRESERVED_PLACEHOLDER_^^^^2(Zhang et al., 5 Feb 2026) OR \2(Zhang et al., 5 Feb 2026) OR \2^^^^, fuses depth via TSDF to update occupancy in PRESERVED_PLACEHOLDER_^^^^2(Zhang et al., 5 Feb 2026) OR \22^^^^, and maintains a 2D grid map aligned with its semantic value map (&&&^^^^2query2^^^^&&&).

The task layer assigns to each robot PRESERVED_PLACEHOLDER_^^^^2(Zhang et al., 5 Feb 2026) OR \2^^^^3 a set of PRESERVED_PLACEHOLDER_^^^^2(Zhang et al., 5 Feb 2026) OR \2^^^^4 embodied multiple-choice questions PRESERVED_PLACEHOLDER_^^^^2(Zhang et al., 5 Feb 2026) OR \2^^^^5 with answer choices PRESERVED_PLACEHOLDER_^^^^2(Zhang et al., 5 Feb 2026) OR \2^^^^6. The full question set is

PRESERVED_PLACEHOLDER_^^^^2(Zhang et al., 5 Feb 2026) OR \2^^^^7

and the ground-truth answers are

PRESERVED_PLACEHOLDER_^^^^2(Zhang et al., 5 Feb 2026) OR \2^^^^8

All robots know all questions, even when a question is assigned to another robot. This shared question context is central to CommCP’s communication design, because agents can report observations and candidate answers relevant to teammates’ assignments rather than only to their own.

Robot motion is governed by a collision-free planner PRESERVED_PLACEHOLDER_^^^^2(Zhang et al., 5 Feb 2026) OR \2^^^^9, which maps the current pose and a target position to the next feasible pose $L \times W \times H$^^^^2query2^^^^; a low-level controller executes the motion. Communication is peer-to-peer and natural-language based: robots exchange messages $L \times W \times H$^^^^2(Zhang et al., 5 Feb 2026) OR \2^^^^ containing relevant observations and potential answers. The formal MM-EQA tuple is

$L \times W \times H$2

Given the time horizon $L \times W \times H$3, the objective is to maximize task success rate (SR) and minimize normalized time cost (NTC), where NTC accounts for both robot movement and message sending latency. Success occurs when each assigned question is answered correctly before $L \times W \times H$4; failure occurs otherwise.

This formulation extends canonical embodied question answering by making communication and coordination first-class concerns. A plausible implication is that MM-EQA isolates a regime in which perceptual competence alone is insufficient: the principal difficulty lies in routing partial information across agents without creating redundancy or distraction.

2. Decentralized communication architecture

CommCP organizes multi-robot coordination as a decentralized pipeline. Robots first interpret human assignments and generate embodied questions. A visual-LLM then detects observed objects $L \times W \times H$5 from RGB input $L \times W \times H$6 using a structured prompt, producing object names and colors for downstream reasoning. Based on these detections and the globally shared question set, each robot decides whether some observation may be relevant to another robot’s target or to another robot’s answer options (&&&^^^^2query2^^^^&&&).

The communication stage is LLM-based and asynchronous. Messages are generated peer-to-peer without central coordination, and each agent independently decides what to send and when. The reasoning model is LLaMA3-8B-instruct, used at temperature $L \times W \times H$7, with zero-shot chain-of-thought prompting to reason about relevance and produce option-labeled outputs. Because CommCP requires token probabilities for calibration, the framework depends on a model that exposes them. The perception and question-answering component is Prismatic-VLM-^^^^2(Zhang et al., 5 Feb 2026) OR \2^^^^3B, which is used for object detection and for probabilities over answer options.

The pipeline can be summarized operationally. A robot observes its surroundings, detects objects, reasons about whether those objects are targets or semantically related to targets, and drafts text intended for a specific teammate. Conformal prediction then filters the message content before transmission. The receiving robot projects the calibrated message onto its semantic value map, which biases frontier-based exploration toward locations likely to contain a relevant object or direct answer. When a robot’s VLM becomes sufficiently confident, it can answer its own question or share the answer with the responsible robot if the question is assigned elsewhere.

CommCP’s decentralization is substantive rather than merely implementation-level. There is no central scheduler that aggregates beliefs or allocates subgoals. Instead, local perception, shared question context, and calibrated confidence jointly determine communication behavior. This suggests that the framework treats communication as a scarce planning resource rather than as a cost-free side channel.

3. Conformal prediction and message calibration

The defining mechanism in CommCP is the use of split conformal prediction to calibrate LLM-generated relevance judgments. For a request $L \times W \times H$8 associated with another robot’s question, the LLM classifies an observed object into one of four options:

• A: “Yes. Observed and Request are the same; you might directly find the Request.”
• B: “Yes. Highly relevant; Request should be close to Observed.”
• C: “No. Not strongly related.”
• D: “No. Observed is a common feature.”

Only options A and B are retained; C and D are discarded. For each observed object $L \times W \times H$9, the LLM outputs the probability $M$^^^^2query2^^^^ for the selected option letter. CommCP uses the token probability of the chosen option as the conformity score,

$M$^^^^2(Zhang et al., 5 Feb 2026) OR \2^^^^

where

$M$2

Separate calibration sets are constructed for A and B: $M$3

$M$4

To build these sets, the framework samples $M$5 pairs from 2^^^^2query2^^^^ HM3D scenarios, labels the highest-confidence option according to the LLM’s internal scoring, and records $M$6 and $M$7. Calibration and test pairs are assumed i.i.d., expressed as

$M$8

with $M$9 for all $l$^^^^2query2^^^^. The paper justifies this stronger-than-exchangeability assumption by random scenario sampling, independence of object-pair judgments, consistent LLM semantics across scenes, and frame-local computation (&&&^^^^2query2^^^^&&&).

For a test sample $l$^^^^2(Zhang et al., 5 Feb 2026) OR \2^^^^, the threshold $l$2 is the empirical $l$3-quantile of the calibration scores for the corresponding option: $l$4 The conformal prediction set is then

$l$5

Under standard CP marginal coverage for exchangeable data,

$l$6

where $l$7 corresponds to $l$8. In CommCP, $l$9 controls message strictness and thus the size of the prediction set.

The implementation uses $i$^^^^2query2^^^^ values leading to quantiles $i$^^^^2(Zhang et al., 5 Feb 2026) OR \2^^^^ for A and $i$2 for B. Only objects labeled A or B and passing the relevant threshold are included in a message. If no objects pass conformal filtering, the robot sends no message for that request. The message template is

“I see {relevant object} that may be relevant to your target {true target}, and {possible target object} may be your target at {position}.”

This calibration layer is intended to suppress distractors and misrouting. The paper’s central claim is not simply that communication helps, but that uncalibrated communication can be counterproductive. In that sense, CommCP treats calibration as part of the communication protocol rather than as a post hoc confidence score.

4. Semantic value maps, planning, and answer resolution

CommCP integrates communication into planning by projecting received text messages onto semantic value maps. Exploration is based on Frontier-Based Exploration, which identifies frontier points $i$3 at the boundary between explored and unexplored regions; Gaussian smoothing then spreads semantic value locally to smooth navigation preferences (&&&^^^^2query2^^^^&&&).

For frontier point $i$4 and task $i$5, the communication-derived semantic value is

$i$6

with $i$7 and $i$8. The first weight balances indirect relevance cues, whereas the second emphasizes direct target evidence. The final task-specific value is

$i$9

and the aggregated frontier score is

$g^i_0$^^^^2query2^^^^

If $g^i_0$^^^^2(Zhang et al., 5 Feb 2026) OR \2^^^^, the robot chooses a random frontier point.

Answer resolution is also confidence-gated. Let $g^i_0$2 be the VLM’s question-image relevance score, and let $g^i_0$3 be the probability assigned to answer $g^i_0$4. A question is accepted as answered and removed if

$g^i_0$5

where $g^i_0$6 is a global threshold shared across robots. If the question belongs to another robot, the answer is shared.

The end-to-end case study in the paper uses the question “Where is the red bear cushion?” Robot 2 explores several rooms and detects objects such as “basketboard,” “dolls,” “black chair,” and later “red pillow on blue chair.” The LLM assigns options A, B, C, or D with probability $g^i_0$7 for each observation. CP filters out low-confidence C and D judgments and also filters A or B predictions below threshold. Early messages mentioning “basketboard, dolls, black chair” are suppressed if they fail CP; only confident items such as “dolls” or “red pillow on blue chair” pass. Robot ^^^^2(Zhang et al., 5 Feb 2026) OR \2^^^^ updates $g^i_0$8, navigates toward frontier points with high $g^i_0$9, eventually observes the cushion, satisfies the $t$^^^^2query2^^^^ criterion, answers the question, and stops. The paper interprets this as evidence that CP reduces detours into irrelevant rooms and prevents distraction.

The system explicitly targets information gathering rather than manipulation execution. Heterogeneous manipulation capabilities are handled at the assignment level: questions and assignments are aligned with each robot’s capabilities, while shared exploration and answer-sharing improve overall throughput without forcing task transfer.

5. Benchmark, implementation, and empirical behavior

The MM-EQA benchmark introduced with CommCP uses photo-realistic HM3D scenes loaded in the Habitat simulator. It contains 7^^^^2query2^^^^ test scenarios for 2 robots, each robot having 3 assigned questions, for a total of 42^^^^2query2^^^^ embodied question tasks, plus an additional 2^^^^2query2^^^^ scenarios for conformal calibration (&&&^^^^2query2^^^^&&&). Questions were initially generated by GPT-4V and refined with human oversight. The reported question types are location, identification, counting, existence, and state. The experiments use onboard RGB and depth cameras and focus on information gathering rather than fine-grained manipulation.

Inference-time protocols combine decentralized messaging with LLM CP filtering, frontier-based exploration with semantic-value integration and Gaussian smoothing, and the confidence-check gate for early stopping and answer-sharing. Communication frequency is varied across $t$^^^^2(Zhang et al., 5 Feb 2026) OR \2^^^^, $t$2, $t$3, $t$4, and $t$5 messages per second, with a default of $t$6 message per second. Movement speed is $t$7 m/s. The system runs on dual NVIDIA 6^^^^2query2query2query2^^^^ Ada GPUs, with the VLM and LLM placed on separate devices. There is no model training or fine-tuning; CP is built from calibration data and the LLM relies on prompt engineering.

The evaluation uses two metrics: success rate across assigned questions and normalized time cost combining motion and messaging latency. The baselines are MMFBE, MMEuC, Ours-No-CP, Ours-Com-Control, and Ours-No-Answer-Sharing.

Method	Brief characterization
MMFBE	Multi-agent, multi-task frontier-based exploration; no semantic SV mapping for exploration, no communication
MMEuC	Multi-robot extension of Explore-Until-Confident; independent operation, no communication
Ours-No-CP	Communication without CP calibration
Ours-Com-Control	Randomly samples communicated objects to match CP count
Ours-No-Answer-Sharing	Shares observations and calibrated predictions but not final answers

The reported results emphasize that message quality matters more than message quantity. At $t$8, CommCP achieves $t$9, whereas MMFBE reaches $G_t=\{g^i_t \mid i=1,\dots,N_a\}$^^^^2query2^^^^ only at $G_t=\{g^i_t \mid i=1,\dots,N_a\}$^^^^2(Zhang et al., 5 Feb 2026) OR \2^^^^. CommCP completes on average in $G_t=\{g^i_t \mid i=1,\dots,N_a\}$2 s, compared with $G_t=\{g^i_t \mid i=1,\dots,N_a\}$3 s for MMFBE. MMEuC underperforms MMFBE because it lacks communication. Ours-No-CP performs similarly to MMEuC, which the paper interprets as evidence that uncalibrated messages are distracting. Ours-Com-Control fails to match CommCP despite communicating at matched volume, highlighting calibration quality rather than raw message count as the critical variable.

Answer-sharing also contributes materially. Ours-No-Answer-Sharing yields higher NTC and lower SR than the full system, indicating that sharing final answers accelerates convergence. In larger HM3D scenes labeled Size 3, CommCP improves NTC by approximately $G_t=\{g^i_t \mid i=1,\dots,N_a\}$4 over MMFBE, suggesting better scalability as scene size grows. Increasing message speeds improves the early rise in SR, although final SR converges similarly across rates; CommCP consistently outperforms MMFBE at all tested speeds. With 3 robots, the framework scales with minimal overhead and maintains faster SR growth than the baselines, while Ours-No-CP degrades early-stage SR because of distractors.

6. Assumptions, limitations, and nomenclature

CommCP’s theoretical support is inherited from split conformal prediction with conformity score $G_t=\{g^i_t \mid i=1,\dots,N_a\}$5 and per-option calibration sets. Under exchangeability, the conformal set provides marginal coverage, and the paper argues that the stronger i.i.d. assumption is reasonable in this domain because of random HM3D sampling, local per-frame judgments, and consistent semantics across scenes (&&&^^^^2query2^^^^&&&). The framework’s empirical benefits therefore depend on a calibration regime that remains compatible with deployment-time object-target relationships.

The paper identifies several limitations. Calibration is built on HM3D household patterns, so significant domain shift, such as atypical layouts or different cultures of object organization, may reduce coverage quality; the suggested mitigation is further calibration or adaptive CP. Reliance on open-source LLM probabilities limits model choice, since GPT-4V does not provide the token probability API used in these experiments. LLM and VLM hallucinations remain possible; CP reduces but does not eliminate errors. Communication cost and latency are simplified to message rate rather than deep network modeling. Real-robot deployment would require integration with physical controllers and robust perception, and mapping errors could degrade semantic value maps. Heterogeneity is currently handled only at the assignment level. Future directions proposed in the paper include larger teams and environments, richer modalities such as structured semantic maps, adaptive CP with online threshold updates, integration of manipulation beyond information gathering, and ranked or clustered message aggregation with multi-hop routing.

The acronym “CommCP” is not unique in the arXiv literature. "Communicating Concurrent Processes" uses CCP, also referred to as CommCP, for a modification of Hoare’s CSP in which traces are sequences of simultaneous event-sets rather than single visible events (Wang, 2018). "Comprehensive Monitor-Oriented Compensation Programming" uses CommCP for a monitor-driven architecture that separates system building from compensation modeling in service-oriented transactions (Colombo et al., 2014). These usages are unrelated to the MM-EQA framework of (&&&^^^^2query2^^^^&&&), but they illustrate that the abbreviation has appeared in multiple technical domains.

Within embodied multi-agent information gathering, however, CommCP denotes a specific synthesis: decentralized LLM-based natural-language communication, calibrated by conformal prediction, integrated with semantic-value-guided exploration and confidence-gated answer-sharing. A plausible implication is that the work’s main contribution is methodological rather than purely benchmark-driven: it frames communication reliability as a statistical calibration problem embedded inside embodied coordination.