Social Circuits behind Multi-agent Echo Chambers

Chuiyang Meng; Wenlu Yu; Ming Tang; Cheng Li

cs.AISep 28, 2026

Social Circuits behind Multi-agent Echo Chambers

Authors: Chuiyang Meng, Wenlu Yu, Ming Tang, Cheng Li

Organizations: Simon Fraser University · University of Alberta · Southern University of Science and Technology

Abstract

Language-model agents exchange messages to combine evidence, but their communication can also create echo chambers that reinforce shared errors. However, overall task performance does not explain how a message changes the receiving agent's internal activations and affects its decision. In this work, we introduce Social Circuits, a framework for tracing message effects through receiver activations. We compare the receiver's answers before and after changing a message. Then, we restore selected activations recorded under the original message to determine how much of the message effect these activations reproduce. Based on Social Circuits, we propose Circuit-Guided Deliberation (CGD), which learns to select useful messages using receiver activation changes. We establish when activation replacement preserves receiver decisions and bound the gap between CGD's task performance and the best achievable through message selection. Experiments show that receiver activation changes explain the message effects and guide message selection that improves the task performance. Across three models and four datasets, CGD achieves the highest or joint-highest average accuracy in our main comparisons while generating fewer tokens than multi-agent baselines.

Figures & tables

Figure 1: Comparison of the error correction and an echo chamber. True evidence corrects the receiver’s answer, whereas a message presenting the agents’ shared error as supporting evidence reinforces the echo chamber.

Figure 2: Receiver accuracy on Gemma-3-4B-it as more messages support the correct (orange) or incorrect (blue) answer.

Figure 3: Illustration of Social Circuits and Circuit-Guided Deliberation (CGD). Matched reruns change the message, whereas activation replacement changes selected receiver activations while preserving the reference input. CGD uses sender summaries and receiver activation changes to select message combinations.

Panel of Figure 4 — Figure 4: (a) Fractions of messages that correct or introduce errors for Gemma-3-4B and Qwen3-4B. (b) Answer-score changes for Gemma-3-4B. (c) Answer-score changes for Qwen3-4B. (d) Message-effect recovery $\rho$ across three models. (e) Average task-performance changes for Gemma-3-4B. (f) Average task-performance changes for Qwen3-4B. Diagonal dashed lines in (b) and (c) mark equal answer-score changes. Figs. (a)–(c), (e), and (f) use the same candidate messages within each model. Task performance is the accuracy of the higher-scoring answer, and error bars show standard deviations across random seeds.

2Wiki (A/B answers)
		Single-agent		Multi-agent
Model	Metric	Single	SC6	Majority	MAD	MAD-M2	MADC	MOC	SafeSieve	CGD
Gemma-3-4B	Accuracy $\uparrow$	78.89 $\pm 0.48$	77.78 $\pm 1.46$	83.19 $\pm 0.64$	80.69 $\pm 1.05$	82.92 $\pm 1.50$	79.44 $\pm 1.27$	86.94 $\pm 0.96$	79.72 $\pm 3.07$	89.72 $\pm 1.46$
Gemma-3-4B	Tokens $\downarrow$	79	478	219	647	516	638	401	485	139
Qwen3-4B	Accuracy $\uparrow$	73.19 $\pm 1.73$	70.83 $\pm 2.32$	73.61 $\pm 1.68$	80.42 $\pm 2.08$	82.08 $\pm 1.44$	77.22 $\pm 1.27$	86.94 $\pm 0.64$	83.89 $\pm 1.05$	89.86 $\pm 1.68$
Qwen3-4B	Tokens $\downarrow$	171	1025	468	978	1028	961	864	874	299

Table 1: Accuracy (%, mean

\pm

standard deviation over three seeds) and mean tokens per task for generated messages and final answers under the original generation settings. On 2Wiki A/B, CGD sends passages and responses, and ollaboration baselines send responses alone.

	Single-agent		Multi-agent
Model	Single	SC6	Majority	MAD	MAD-M2	MADC	MOC	SafeSieve	CGD
Gemma-3-4B	24.72 $\pm 1.05$	25.42 $\pm 1.67$	26.39 $\pm 2.77$	7.22 $\pm 1.88$	20.97 $\pm 2.10$	7.22 $\pm 1.88$	20.28 $\pm 0.48$	25.42 $\pm 4.23$	42.92 $\pm 0.83$
Qwen3-4B	21.94 $\pm 0.24$	22.22 $\pm 0.48$	25.97 $\pm 1.97$	23.33 $\pm 2.53$	25.14 $\pm 3.13$	23.33 $\pm 2.53$	29.31 $\pm 2.44$	40.56 $\pm 2.37$	51.81 $\pm 0.64$
Qwen3-8B	23.61 $\pm 2.10$	23.89 $\pm 1.05$	30.00 $\pm 1.67$	14.17 $\pm 0.83$	30.83 $\pm 0.83$	14.17 $\pm 0.83$	28.75 $\pm 5.42$	29.86 $\pm 5.50$	61.53 $\pm 0.48$

Table 2: Free-generation accuracy (%) on 240 2Wiki tasks under matched total generation budgets.

Figure 5: Two-answer scoring accuracy on 2Wiki: 240 new tasks for the 4B models, and Table 1’s tasks for Qwen3-8B.

Panel of Figure 6 — Figure 6: 2Wiki learning curves. Lines and bands show mean accuracy and standard deviation.

Appendix figures & tables24 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Meaning
$\mathcal{A}$	Message combination, a subset of $\mathcal{C}$ .
$\widehat{\mathcal{A}},\mathcal{A}^{\star}$	Combinations with the highest CGD score and expected task utility, respectively.
$\mathcal{B}$	Another message combination in $\mathcal{S}$ .
$\mathcal{C}$	Candidate messages for one receiver.
$d$	Activation-summary dimension.
$e$	Incorrect answer favored by multiple agents before delivery.

Appendix

Table 3: List of key notations.

Comparison	Task	Questions
Collaboration approaches	Correlated-Trap	600
Message content and repetition	Correlated-Trap	240 per direction
Activation replacement	Correlated-Trap	240 per direction
Message-specific replacement	2Wiki	240, two candidates each
Two-option message selection	2Wiki	240 per test set
Two-answer scoring	MATH-500	200

Appendix

Table 4: Evaluation sets. Counts are per experiment.

Approach	Information
Single, SC6	The designated receiver’s assigned input, without messages from other agents.
Majority vote	Independent answers from three agents, each using its own assigned input.
MAD, MADC	Each agent’s assigned input and the responses exchanged according to the communication procedure.
MAD-M2	Each agent’s assigned input and the responses retained by memory masking.
MOC	The receiver’s assigned input, a summary of the two senders’ messages, and an updated sender response.
SafeSieve	Each agent’s assigned input and messages received through the learned communication graph.

Appendix

Table 5: Information used by each approach.

Panel of Figure 9 — Figure 9: Receiver accuracy when the answer without a message is (a) incorrect or (b) correct.

Panel of Figure 10 — Figure 10: Effects of additional messages: (a) correcting an incorrect answer; (b) switching from a correct answer to a shared incorrect answer.

Panel of Figure 11 — Figure 11: Gemma-3-4B with different agent counts and rounds. (a) Wrong answers corrected. (b) Correct answers changed to wrong answers. The first round uses independent answers, and the one-agent scenario uses self-review in later rounds.

Figure 12: Qwen3-8B receiver accuracy with different communication topologies and agent numbers.

Figure 13: Message-effect recovery across three models. Values are 100 times the mean replacement effect divided by the mean message effect, using 240 tasks per direction, averaged across seeds. Positive values follow the original message effect, while negative values reverse it.

Panel of Figure 14 — Figure 14: Receiver activation replacement on Gemma-3-4B-it: (a) current-task and other-task changes, with one layer replaced per run; (b) average changes from other tasks, applied across six layers on new tasks. Values use complete-answer scores. Circles in (b) show results of different seeds.

Figure 15: Message-specific activation replacement on 2Wiki. Mean absolute error between message and replacement effects on the correct answer’s score, evaluated on matched samples.

(a) 2Wiki
Model	Activation changes	Answer agreement (%)	Answer recovery (%)
Gemma-3-4B	Original message	$99.61\pm 0.27$	$\mathbf{97.81}\pm 0.23$
	Another message (same task)	$88.89\pm 1.88$	$41.07\pm 13.20$
Qwen3-4B	Original message	$99.78\pm 0.00$	$\mathbf{96.73}\pm 0.22$
	Another message (same task)	$94.23\pm 1.48$	$55.06\pm 11.67$
(b) Correlated-Trap

Appendix

Table 6: Receiver answers after activation replacement, reported as mean

\pm

sample standard deviation. Agreement uses the same matched message records for both replacements: 1,268 for Gemma and 1,334 for Qwen on 2Wiki, and 1,440 per model on Correlated-Trap. Recovery uses only records where the message changes the answer, including both corrections and changes to incorrect answers. Bold marks the higher mean recovery within each model and dataset.

Panel of Figure 16 — Figure 16: Message selection on 2Wiki with the same candidate messages and two-option scoring. (a) Accuracy across baselines. (b) Cases where CGD is correct and the baseline is wrong, or vice versa.

Strategy	Gemma-3-4B	Qwen3-4B	Qwen3-8B
No message	84.17 $\pm 0.00$	78.75 $\pm 0.00$	81.67 $\pm 0.00$
Sender 1 only	86.11 $\pm 1.27$	81.67 $\pm 1.10$	84.58 $\pm 2.17$
Sender 2 only	82.50 $\pm 1.91$	79.58 $\pm 1.44$	77.50 $\pm 0.83$
Both messages	85.28 $\pm 1.34$	83.89 $\pm 1.34$	82.78 $\pm 1.97$
CGD	89.72 $\pm 1.20$	91.11 $\pm 0.64$	87.36 $\pm 0.64$

Appendix

Table 7: Two-answer accuracy (%) on 240 new 2Wiki tasks for the 4B models and the original 240 tasks for Qwen3-8B, reported as mean

\pm

sample standard deviation. CGD achieves the highest average accuracy across all three models.

Approach	Gemma-3-4B	Qwen3-4B	Qwen3-8B
Activation products	88.06 $\pm 2.29$	87.92 $\pm 0.83$	85.00 $\pm 1.50$
Activation products (adjusted)	89.31 $\pm 1.05$	88.75 $\pm 1.50$	86.94 $\pm 1.27$
CGD	89.72 $\pm 1.46$	89.86 $\pm 1.68$	87.36 $\pm 0.64$

Appendix

Table 8: Two-answer accuracy (%) on the original 240 2Wiki tasks for each model, before and after matching the scaling and penalty to CGD. Entries show the mean and sample standard deviation across three seeds.

Figure 17: Message selection on new 2Wiki tasks without retraining. Points and error bars show average accuracy and standard deviation.

Approach	Gemma-3-4B	Qwen3-8B
Text-only	$38.33\pm 1.67$	$56.67\pm 1.10$
Receiver-only	$40.69\pm 1.58$	$60.42\pm 0.42$
Text + receiver	$38.47\pm 1.05$	$59.17\pm 0.72$
Additive	$42.78\pm 1.27$	$53.61\pm 1.05$
Activation products	$\mathbf{43.33}\pm 1.91$	$57.36\pm 0.24$
CGD	$42.92\pm 0.83$	$\mathbf{61.53}\pm 0.48$

Appendix

Table 9: Free-generation accuracy (%) on 2Wiki, reported as mean

\pm

sample standard deviation. The Qwen3-8B CGD entry uses a separate run with the same evaluation tasks, seeds, and scoring. Bold marks the highest mean in each column.

Approaches	Gemma-3-4B	Qwen3-4B	Qwen3-8B
Single	78.75 $\pm 0.00$	75.83 $\pm 0.00$	75.42 $\pm 0.00$
MAD	81.11 $\pm 1.05$	81.39 $\pm 1.73$	81.94 $\pm 2.93$
MAD-M2	82.92 $\pm 1.50$	81.53 $\pm 1.68$	78.47 $\pm 2.77$
MADC	79.72 $\pm 1.46$	79.03 $\pm 1.34$	81.53 $\pm 1.27$
MOC	86.81 $\pm 1.20$	86.94 $\pm 0.87$	85.69 $\pm 0.24$
SafeSieve	80.00 $\pm 2.60$	84.17 $\pm 0.72$	81.94 $\pm 2.10$

Appendix

Table 10: 2Wiki accuracy (%) using A/B answer scores. We report the mean and sample standard deviation across three seeds.

Figure 20: Free-generation accuracy and generated tokens over 720 runs per model and dataset. Points show means and error bars show standard deviations across seeds. MAD and MADC have overlapping results.

Figure 21: A GSM8K example with Qwen3-8B in which joint message delivery corrects the receiver’s answer.

Figure 22: A GSM8K example with Qwen3-4B in which the selected message supplies the receiver’s missing calculation.

Figure 23: A 2Wiki example with Gemma-3-4B in which excluding a misleading message corrects the receiver’s answer.

Figure 24: A GSM8K example with Gemma-3-4B in which single-message delivery corrects an error shared by both senders.

Explore similar work

Oct 1, 2026cs.AI

Beyond Final Accuracy: Auditing Communication in LLM Multi-Agent Systems

Multi-agent communication aims to help agents benefit from one another's information. Yet improvements in system performance leave a fundamental ambiguity: do they reflect effective communication, a favorable agent architecture, or simply additional reasoning? Because communication methods are commonly evaluated within the systems they were designed for, these factors are difficult to disentangle. Final accuracy further merges corrected errors and corrupted answers into a single outcome, obscuring how communication changes decisions. We introduce Independent--Communicate--Revise (ICR), a controlled framework that evaluates communication as answer revision following independent reasoning. ICR fixes initial reasoning trajectories, measures correction and preservation conditional on both agents' initial correctness, and uses a no-message revision control to quantify gains beyond additional reasoning. Across four reasoning benchmarks, our audit of textual and latent communication reveals that similar aggregate accuracy can conceal substantially different revision behaviors. Compared with transmitting answers alone, full reasoning increases correction while reducing preservation on all four benchmarks, so richer messages amplify beneficial and harmful influence alike. Receiver-policy comparisons on MedQA and GPQA-D further show that a structured verification policy shifts every channel toward greater preservation and lower correction, while its effect on selectivity varies across channels and tasks. These findings challenge treating communication quality as an intrinsic property of a channel. ICR therefore recenters evaluation on selective revision, providing a unified framework for examining how message content and receiver policies jointly produce benefits and harms.

Shixuan Li, Wei Yang, Peiyu Zhang +3

Ming Hsieh Department of Electrical and Computer Engineering University of Southern California, Los Angeles, CA 90089, USA · Thomas Lord Department of Computer Science University of Southern California, Los Angeles, CA 90089, USA

Multi-Agent Communication Model Auditing Answer Reasoning Skills

Ask PDF

Sep 29, 2026cs.AI

When Upstream Messages Override Correct Answers: A Controlled Study of Multi-Agent LLM Collaboration

Multi-agent LLM systems rely on message passing among specialized agents to accomplish complex tasks. However, an upstream agent may provide useful information or an incorrect answer that causes a downstream agent to override a correct answer supported by its own evidence. Prior work has not clearly separated the benefits of communication from the damage caused by incorrect messages. We study this problem with controlled experiments across five benchmarks and five receivers, keeping the downstream task and evidence fixed while comparing answers under three conditions: no message, the upstream agent's original message, or a message with the opposite conclusion. Our experiments reveal three key findings. First, messages often help when the downstream agent would otherwise answer incorrectly. Second, messages can also hurt: when the downstream agent would answer correctly without a message, an incorrect upstream message changes the answer in up to 32% of cases. Third, in 94% of audited harmful cases, the downstream agent copies the upstream's specific wrong answer--a pattern we term answer substitution. Removing unreliable messages recovers part of the lost accuracy, suggesting that communication should be selective based on upstream reliability and the evidence already available to the downstream agent.

Yaxin Gong, Gangyi Zhang, Chongming Gao +7

University of Science and Technology of China · Qwen Business Unit of Alibaba · National University of Singapore

Multi-Agent Large Language Model Systems Downstream Reasoning Collaboration Answer Substitute Controlled Study

Ask PDF

May 24, 2026cs.AI

DarkForest: Less Talk, Higher Accuracy for Multi-Agent LLMs

Multi-agent LLM systems improve reasoning by combining outputs from multiple agents, but interaction-heavy methods can introduce error propagation and high communication overhead. When agents exchange raw responses or reasoning traces, incorrect intermediate reasoning may be adopted and amplified, leading to confident but wrong consensus; multi-round communication also increases token consumption, latency, and inference cost. In this paper, we propose a controlled-communication coordination framework named DarkForest. DarkForest first keeps agents independent, so each agent produces an answer without seeing the others' outputs. It then parses the raw responses into structured candidate records, groups semantically equivalent candidates into clusters, and estimates a calibrated belief distribution over these clusters using agent reliability, confidence, parse quality, support-pattern reliability, and independence corrections. A coordinator receives only policy-permitted evidence from this belief state with controlled communication. Experiments on six reasoning benchmarks show that DarkForest achieves leading overall quality, improves the strongest baseline by up to 30.7% on benchmark metrics, and reduces token consumption by up to

6.5\times

compared with communication-heavy baselines.

Yi Li, Songtao Wei, Dongming Jiang +3

University of Texas at Dallas · Independent Researcher · University of California, Davis

Multi-Agent Large Language Model Systems Reasoning Benchmark Talk

Ask PDF