acceptodds
Under review as a conference paper at ICLR 2027

Social Circuits behind Multi-agent Echo Chambers

13 views

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.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-agent systems, Large language models, Mechanistic interpretability, Inter-agent communication, Message selection

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_p3yhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Social Circuits, a framework using message interventions and receiver activation replacement to trace communication effects in LLM-based multi-agent systems. It further introduces Circuit-Guided Deliberation(CGD), a learned message selector based on activation and message features. The paper provides theoretical analysis and evaluates the method on a synthetic task and four benchmarks, demonstrating strong answer-change recovery on 2Wiki and competitive performance with fewer generated tokens.

Critical strengths

S1. The paper investigates an important problem in multi-agent communication. Its intervention-based framework provides a useful approach to analyzing message effects, achieving approximately 97% answer-change recovery on 2Wiki, substantially exceeding other message controls.

S2. CGD demonstrates promising message-selection performance with limited training data. Controlled candidate-message comparisons and cross-benchmark evaluations provide evidence of its effectiveness.

Critical weaknesses

W1. Insufficiently controlled comparisons and efficiency evaluation. In the 2Wiki two-answer evaluation, CGD receives sender passages and responses, while collaboration baselines exchange only generated responses. Single-agent baselines also receive less evidence, making performance gains difficult to attribute to message selection. Moreover, generated-token comparisons exclude answer scoring and activation recording, while CGD requires receiver evaluations over exponentially many message subsets. Its actual efficiency advantage remains unclear.

W2. Limited validation of the proposed mechanism and feature construction. Activation replacement reproduces message effects but does not clearly isolate mechanisms specific to echo-chamber reinforcement. Answer recovery on Correlated-Trap is only 1.33%-16.39%, substantially below 2Wiki, although the protocols differ. Furthermore, CGD's activation-change features are algebraically related to those available to the Activation Products baseline, making the contribution of its specific parameterization unclear. Activation Products also slightly outperforms CGD in the Gemma free-generation setting.

Questions for authors

Q1. Can the authors provide evidence-matched comparisons in which all collaboration baselines receive the same sender passages and responses, including a single-agent baseline with access to all available evidence?

Q2. What are CGD's end-to-end latency and computational costs, including receiver processing, answer scoring, activation extraction, and selection? Does CGD retain its efficiency advantage under matched total compute, accounting for cache reuse and computations performed only during offline evaluation?

Q3. Does CGD consistently outperform Activation Products under matched input features, preprocessing, regularization, and multiple evaluation or training seeds?

Q4. What explains the substantial answer-recovery gap between 2Wiki and Correlated-Trap? Does it persist under matched replacement protocols, and are the terms in Theorem 2's bound empirically small enough to make it informative?

6 comments
AuthorsOpenReview ↗

Response to Reviewer p3yh Part 1

We thank the reviewer for the careful comments on the comparison conditions, efficiency, and feature construction. We describe the clarifications and additional comparisons in the revised manuscript below.

Weakness 1: W1. Insufficiently controlled comparisons and efficiency evaluation. In the 2Wiki two-answer evaluation, CGD receives sender passages and responses, while collaboration baselines exchange only generated responses. Single-agent baselines also receive less evidence, making performance gains difficult to attribute to message selection. Moreover, generated-token comparisons exclude answer scoring and activation recording, while CGD requires receiver evaluations over exponentially many message subsets. Its actual efficiency advantage remains unclear.

Author's Response: Thank you for highlighting the comparison conditions. With three senders using the same passages, FGD is more accurate than Single with all passages. With two senders using passages, FGD is more accurate than self-consistency under the time budget set by FGD. We have added these comparisons in Appendix E.5 and complete time measurements in Appendix E.6.

Table 1 compares the complete approaches, with FGD using passages and responses and the collaboration baselines using responses. Appendix D.1 specifies shared passage access in their initial free-generation messages. Table 2 compares matched generation budgets, and Table 10 fixes candidate messages and receiver settings.

Table 1. Accuracy (%, mean ± 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, FGD sends passages and responses, and collaboration baselines send responses alone.

Single-agent: Single and SC6. Multi-agent: Majority vote, MAD, MAD-M2, MADC, MOC, SafeSieve, and FGD.

2Wiki (A/B answers)

ModelMetricSingleSC6Majority voteMADMAD-M2MADCMOCSafeSieveFGD
Gemma-3-4BAccuracy ↑78.89±0.4877.78±1.4683.19±0.6480.69±1.0582.92±1.5079.44±1.2786.94±0.9679.72±3.0789.72±1.46
Tokens ↓79478219647516638401485139
Qwen3-4BAccuracy ↑73.19±1.7370.83±2.3273.61±1.6880.42±2.0882.08±1.4477.22±1.2786.94±0.6483.89±1.0589.86±1.68
Tokens ↓17110254689781028961864874299
Qwen3-8BAccuracy ↑75.69±0.4872.92±0.4280.97±2.5581.81±2.7178.47±2.7778.61±0.2485.83±0.4281.94±1.9787.36±0.64
Tokens ↓1569304399729009598811004286

GSM8K (free generation)

ModelMetricSingleSC6Majority voteMADMAD-M2MADCMOCSafeSieveFGD
Gemma-3-4BAccuracy ↑84.86±1.2088.19±1.0586.81±0.6486.53±1.7388.61±1.3486.53±1.7388.75±1.5087.64±0.2488.75±0.48
Tokens ↓224132967813861248138612111139656
Qwen3-4BAccuracy ↑89.86±0.8791.53±0.2489.72±0.6490.83±0.7291.11±1.6890.83±0.7291.11±0.6492.92±0.4293.33±1.10
Tokens ↓221132065612911226129111441268640
Qwen3-8BAccuracy ↑91.81±1.4692.92±0.8392.78±1.0593.89±1.7393.89±1.0593.89±1.7393.75±0.4292.64±0.4893.89±0.64
Tokens ↓229138368814271325142712121389677

MATH-500 (A/B answers)

ModelMetricSingleSC6Majority voteMADMAD-M2MADCMOCSafeSieveFGD
Gemma-3-4BAccuracy ↑42.67±0.7652.00±1.0046.33±1.0452.83±0.5853.83±2.0252.83±0.5866.33±0.5859.67±0.5868.83±1.76
Tokens ↓4192507126424292411242923532203897
Qwen3-4BAccuracy ↑55.83±0.7660.00±0.5057.50±1.3263.50±1.8062.67±2.3163.50±1.8069.50±2.1865.33±0.7670.83±0.76
Tokens ↓3852321116823092279230921432218818
Qwen3-8BAccuracy ↑59.17±0.7662.33±0.5859.00±1.3259.83±1.0462.00±1.3259.83±1.2664.67±1.2661.50±0.5065.83±1.61
Tokens ↓3872333117423632308236522272318912
AuthorsOpenReview ↗

Response to Reviewer p3yh Part 2

Weakness 1 (continued)

Table 1 (continued)

TruthfulQA (A/B answers)

ModelMetricSingleSC6Majority voteMADMAD-M2MADCMOCSafeSieveFGD
Gemma-3-4BAccuracy ↑62.67±1.7663.83±1.2668.00±1.3266.83±2.7570.17±0.2966.83±2.7568.67±1.0467.17±1.2670.67±1.04
Tokens ↓67405236583418583449617191
Qwen3-4BAccuracy ↑76.17±1.5375.83±0.7679.00±1.8079.83±1.4479.00±1.0079.83±1.4482.00±1.3280.00±1.5083.33±1.76
Tokens ↓105627353828691828677716208
Qwen3-8BAccuracy ↑82.67±0.5882.00±0.5083.67±1.0485.00±2.2982.50±1.5085.00±2.2986.33±2.0884.83±0.5887.00±1.50
Tokens ↓96576305746590746608698236

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

Single-agent: Single and SC6. Multi-agent: Majority vote, MAD, MAD-M2, MADC, MOC, SafeSieve, and FGD.

ModelSingleSC6Majority voteMADMAD-M2MADCMOCSafeSieveFGD
Gemma-3-4B24.72±1.0525.42±1.6726.39±2.777.22±1.8820.97±2.107.22±1.8820.28±0.4825.42±4.2342.92±0.83
Qwen3-4B21.94±0.2422.22±0.4825.97±1.9723.33±2.5325.14±3.1323.33±2.5329.31±2.4440.56±2.3751.81±0.64
Qwen3-8B23.61±2.1023.89±1.0530.00±1.6714.17±0.8330.83±0.8314.17±0.8328.75±5.4229.86±5.5061.53±0.48

Table 10. A/B accuracy (%) on 240 new 2Wiki tasks for the 4B models and the original 240 tasks for Qwen3-8B, reported as mean ± sample standard deviation. FGD achieves the highest mean accuracy across all three models.

StrategyGemma-3-4BQwen3-4BQwen3-8B
No message84.17±0.0078.75±0.0081.67±0.00
Sender 1 only86.11±1.2781.67±1.1084.58±2.17
Sender 2 only82.50±1.9179.58±1.4477.50±0.83
Both messages85.28±1.3483.89±1.3482.78±1.97
FGD89.72±1.2091.11±0.6487.36±0.64

Table 16 reports time per question, including prompt processing, activation extraction, scoring, selection, and decoding. FGD takes less time than MAD and SafeSieve on both models under their original protocols. FGD generates sender responses and the final answer, while those baselines add one communication round after their initial responses. Answers for unselected combinations are generated after timing.

Table 16. Time per question under the original protocols on A100 40GB. We show the timing columns from the manuscript. Entries show the mean and sample standard deviation across three seeds on 2Wiki.

ApproachGemma-3-4B: Time (s)Qwen3-4B: Time (s)
FGD10.65±2.3519.46±11.40
MAD39.25±9.0730.37±7.08
SafeSieve27.21±6.2324.69±0.14

Weakness 2: W2. Limited validation of the proposed mechanism and feature construction. Activation replacement reproduces message effects but does not clearly isolate mechanisms specific to echo-chamber reinforcement. Answer recovery on Correlated-Trap is only 1.33%-16.39%, substantially below 2Wiki, although the protocols differ. Furthermore, CGD's activation-change features are algebraically related to those available to the Activation Products baseline, making the contribution of its specific parameterization unclear. Activation Products also slightly outperforms CGD in the Gemma free-generation setting.

Author's Response: Thank you for pointing out this connection. FGD uses the same weight, with opposite signs, for the sender–receiver products before and after a message. Activation products learns a separate weight for each. Section 4.1 now explains this difference.

Table 11 shows higher mean A/B accuracy for FGD across all three models after matching candidate messages, training data, input scaling, and weight penalty. The advantage is modest, and the ordering varies across seeds. In free generation, Table 13 shows higher accuracy for FGD on Qwen3-8B and for Activation products on Gemma-3-4B. The Qwen3-8B FGD entry uses a separate run with the same tasks, seeds, and scoring.

AuthorsOpenReview ↗

Response to Reviewer p3yh Part 3

Weakness 2 (continued)

Table 11. A/B accuracy (%) on the original 240 2Wiki tasks for each model, before and after matching the scaling and penalty to FGD. Entries show the mean and sample standard deviation across three seeds.

ApproachGemma-3-4BQwen3-4BQwen3-8B
Activation products88.06±2.2987.92±0.8385.00±1.50
Activation products (adjusted)89.31±1.0588.75±1.5086.94±1.27
FGD89.72±1.4689.86±1.6887.36±0.64

Table 13. Free-generation accuracy (%) on 2Wiki, reported as mean ± sample standard deviation. The Qwen3-8B FGD entry uses a separate run with the same evaluation tasks, seeds, and scoring. Bold marks the highest mean in each column.

ApproachGemma-3-4BQwen3-8B
Text-only38.33±1.6756.67±1.10
Receiver-only40.69±1.5860.42±0.42
Text + receiver38.47±1.0559.17±0.72
Additive42.78±1.2753.61±1.05
Activation products43.33±1.9157.36±0.24
FGD42.92±0.8361.53±0.48

Table 3 in Section 5.2 shows high answer recovery on both tasks under the common answer prefix and A/B scoring, with higher recovery for the original message than for controls matched in norm. The recorded activations precede either answer token. Both replacement position and scoring change from the original Correlated-Trap protocol.

Table 3. Answer recovery (%) after replacement at the common answer prefix on Gemma-3-4B. Answer recovery uses records where the original and reference messages produce different answers.

TaskReferenceOriginal messageControl
2WikiMessage removal95.92±2.3544.13±17.80
Shuffled message97.81±2.1333.53±8.90
Correlated-TrapMessage removal99.85±0.260.00±0.00
Shuffled message99.84±0.280.00±0.00

Table 4 reports a further round of 2Wiki debate after the agents have given the same wrong answer. Among answers changed by shuffling the sender message, restoring the original activations recovers the original answer more often than both controls matched in norm in every seed. Recovered answers include the previously shared error. Section 5.2 presents the results, and Appendix E.3 describes the settings.

Table 4. Answer recovery (%) of further MAD debate on 2Wiki with Gemma-3-4B. We replace activations from the sender message through the final prompt token, using a shuffled message as the reference. Entries show the mean and sample standard deviation across three seeds.

Activation changesAnswer recovery (%)
Original message49.70±6.70
Different message37.60±5.46
Shuffled activation changes15.77±14.52

Table 8(b) reports the low recovery under the original Correlated-Trap protocol, which replaces tokens before the answer prefix under a shuffled reference. The original 2Wiki protocol replaces the last common answer-prefix token after message removal. Fig. 13 reports message-effect recovery, including negative recovery for correction on Qwen3-4B.

Table 8. Receiver answers after activation replacement, reported as mean ± sample standard deviation. Agreement uses the same matched message records for both replacements: 1,268 for Gemma-3-4B and 1,334 for Qwen3-4B on 2Wiki, and 1,440 per model on Correlated-Trap. Answer 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.

(a) 2Wiki

ModelActivation changesAnswer agreement (%)Answer recovery (%)
Gemma-3-4BOriginal message99.61±0.2797.81±0.23
Different message (same task)88.89±1.8841.07±13.20
Qwen3-4BOriginal message99.78±0.0096.73±0.22
Different message (same task)94.23±1.4855.06±11.67

(b) Correlated-Trap

ModelActivation changesAnswer agreement (%)Answer recovery (%)
Gemma-3-4BOriginal message23.68±2.6816.39±1.67
Swapped-role changes6.74±1.150.23±0.00
Qwen3-4BOriginal message18.47±1.071.33±0.92
Swapped-role changes14.51±1.050.27±0.27
Qwen3-8BOriginal message40.90±1.894.88±1.54
Swapped-role changes30.49±1.730.00±0.00
Reviewer_DF4LOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces the SOCIAL CIRCUITS framework to study the beneficial and detrimental effects of messages in multi-agent LLM systems, precisely quantifying the impact of individual messages through matched reruns and activation patching. Building on this framework, it proposes CGD, a novel algorithm for scoring message combinations. Experiments show that CGD achieves superior performance while substantially reducing token consumption, with notable accuracy gains on tasks such as 2Wiki under the same token budget.

Critical strengths
  1. Most research on multi-agent communication reports only aggregate accuracy, while shared errors and echo chambers remain important yet understudied issues. This paper models message effects as processes that can be traced through the receiving agent’s activations. This novel and compelling approach connects multi-agent debate with mechanistic interpretability.

  2. The experimental methodology is solid and rigorous, with extensive empirical evidence supporting the paper’s claims.

  3. The main results are strong, and the experimental comparisons are fair.

  4. The investigations into message content, interactions among multiple messages, and communication topologies provide valuable additional insights.

Critical weaknesses
  1. The paper’s title and central claims do not hold up because no “circuits” are actually identified. The interventions merely replace activations at six evenly spaced layers and fixed post-message token positions. The paper performs no sparsification, attribution, or search over layers and positions to identify minimal circuits. The name “SOCIAL CIRCUITS” therefore overstates what the paper demonstrates, which I find disappointing.

  2. The measurement of echo chambers relies on the synthetic Correlated-Trap task. On real benchmarks, “shared error” scenarios are only approximated through a constructed binary-choice scoring setup. For example, by adding or subtracting 1 from the correct answer to create distractors for MATH-500. Whether this activation-based tracing remains effective when error reinforcement arises naturally in free-form, multi-round debates remains unverified.

  3. All CGD experiments use only two senders. Although the paper is motivated by echo chambers among multiple agents (e.g., 3–5 agents), CGD is never evaluated with more than two candidate messages.

Questions for authors
  1. What are the wall-clock runtime and computational cost of CGD’s selection procedure, including activation extraction and scoring, compared with one round of MAD or one pass of SafeSieve? Please report these measurements explicitly for the 2Wiki setting.

  2. How does CGD perform with three or more senders, in terms of both accuracy and computational efficiency?

  3. Figure 4(a) reports message-level error correction rates of only 3.71%–7.18%. If roughly 92–96% of messages leave the answer unchanged, how stable and informative are the pairwise training labels when most message combinations have equal utility?

  4. I recommend reconsidering the “Social Circuits” title and the associated claims, or adding experiments that explicitly identify circuits to substantiate those claims.

5 comments
AuthorsOpenReview ↗

Response to Reviewer DF4L Part 1

We thank the reviewer for the constructive comments and for recognizing the connection between multi-agent communication and receiver activations. We describe the corresponding changes in the revised manuscript below.

Weakness 1: The paper’s title and central claims do not hold up because no “circuits” are actually identified. The interventions merely replace activations at six evenly spaced layers and fixed post-message token positions. The paper performs no sparsification, attribution, or search over layers and positions to identify minimal circuits. The name “SOCIAL CIRCUITS” therefore overstates what the paper demonstrates, which I find disappointing.

Author's Response: Thank you for the naming suggestion. Social Footprints describes the message-induced changes measured in the receiver's activations and answer. We have renamed the framework Social Footprints and the selection approach Footprint-Guided Deliberation (FGD). The revised title is Social Footprints in Multi-agent Echo Chambers, and the manuscript and figures use these names.

Section 3 now specifies the selected full decoder-block outputs and token positions. Table 8(a) shows higher recovery for the original message than for controls matched in norm. Fig. 14(a) shows how the message effect changes across layers.

Table 8. Receiver answers after activation replacement, reported as mean ± sample standard deviation. Agreement uses the same matched message records for both replacements: 1,268 for Gemma-3-4B and 1,334 for Qwen3-4B on 2Wiki, and 1,440 per model on Correlated-Trap. Answer 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.

(a) 2Wiki

ModelActivation changesAnswer agreement (%)Answer recovery (%)
Gemma-3-4BOriginal message99.61±0.2797.81±0.23
Different message (same task)88.89±1.8841.07±13.20
Qwen3-4BOriginal message99.78±0.0096.73±0.22
Different message (same task)94.23±1.4855.06±11.67

(b) Correlated-Trap

ModelActivation changesAnswer agreement (%)Answer recovery (%)
Gemma-3-4BOriginal message23.68±2.6816.39±1.67
Swapped-role changes6.74±1.150.23±0.00
Qwen3-4BOriginal message18.47±1.071.33±0.92
Swapped-role changes14.51±1.050.27±0.27
Qwen3-8BOriginal message40.90±1.894.88±1.54
Swapped-role changes30.49±1.730.00±0.00

Weakness 2: The measurement of echo chambers relies on the synthetic Correlated-Trap task. On real benchmarks, “shared error” scenarios are only approximated through a constructed binary-choice scoring setup. For example, by adding or subtracting 1 from the correct answer to create distractors for MATH-500. Whether this activation-based tracing remains effective when error reinforcement arises naturally in free-form, multi-round debates remains unverified.

Author's Response: Thank you for raising this question. Yes. Restoring the original activations recovers answers during a further round of debate, including errors the agents already shared.

We have added this comparison to Section 5.2 and described the settings in Appendix E.3. We continue saved MAD conversations in which the agents have already given the same wrong answer. We shuffle the sender message while preserving its length and tokens, then restore the original activations from the message through the final prompt token before free generation.

Table 4 shows higher recovery of the original answer than both controls matched in norm in every seed, among answers changed by shuffling. The controls use changes from a different message or the original changes with their dimensions shuffled, at the same replacement positions. Restoring the original activations can therefore bring back a shared wrong answer.

Table 4. Answer recovery (%) of further MAD debate on 2Wiki with Gemma-3-4B. We replace activations from the sender message through the final prompt token, using a shuffled message as the reference. Entries show the mean and sample standard deviation across three seeds.

Activation changesAnswer recovery (%)
Original message49.70±6.70
Different message37.60±5.46
Shuffled activation changes15.77±14.52

Table 9 uses message removal and replaces only the last prompt token. Under that setting, recovery is low for both the original activations and the control. Restoring all 75 common instruction tokens also recovers a smaller proportion of changed answers than the control. The comparison in Section 5.2 changes both the reference message and the replacement positions. Panel (a) shows that the original message preserves the shared wrong answer more often than removal.

AuthorsOpenReview ↗

Response to Reviewer DF4L Part 2

Weakness 2 (continued)

Table 9. Answers after a further MAD update on 2Wiki with Gemma-3-4B. Entries show means and sample standard deviations across three seeds, using 166 records with a shared wrong answer before the update, with 54, 57, and 55 records per seed. Message removal is the reference for replacement at the last prompt token. The control uses the same positions and norms. Answer recovery uses records whose normalized answer changes under message removal, with 16, 18, and 12 records per seed.

(a) Final answers

ConditionShared wrong answer (%)
Original message80.14±4.61
Message removal74.16±3.45
Activation replacement77.71±2.76
Control matched in norm72.86±3.99

(b) Recovery after replacing activations at the last prompt token

ConditionAnswer recovery (%)
Activation replacement13.43±13.91
Control matched in norm12.04±14.25

Fig. 24 provides a GSM8K example where either sender's message corrects a shared error, while sending both preserves it.

Weakness 3: All CGD experiments use only two senders. Although the paper is motivated by echo chambers among multiple agents (e.g., 3–5 agents), CGD is never evaluated with more than two candidate messages.

Author's Response: Thank you for suggesting comparisons with more senders. With three senders FGD using passages is more accurate than Single with all passages. We have added the three-sender comparison in Appendix E.5.

Table 14(b) shows the highest mean accuracy and F1 for FGD using passages among the three-sender approaches. With three senders using passages, FGD’s accuracy matches or exceeds every comparison approach in each evaluation seed. Selecting all passages reproduces Single's output tokens, so this comparison attributes the improvement to message selection. With generated responses, Additive and sending all three responses remain more accurate than FGD.

Table 14. Message selection on 2Wiki with Gemma-3-4B, reported as mean ± sample standard deviation across three seeds. Panel (a) compares parameters selected by validation accuracy from 2 or 70 choices using the same passages with two senders. Panel (b) compares separately collected three-sender results with parameters selected by validation accuracy. The last row uses all three responses in the Responses column and all sender passages in the Passages columns.

(a) Two senders using passages, accuracy (%)

Approach2 parameter choices70 parameter choices
FGD47.36±0.8746.53±0.64
Additive47.08±1.1045.14±0.64
Activation products47.08±1.1045.56±0.64
Single with all passages46.81±0.8746.81±0.87

(b) Three senders

ApproachResponses: Accuracy (%)Passages: Accuracy (%)Passages: F1 (%)
FGD43.89±0.8749.44±2.8458.05±2.80
Additive45.97±1.6848.47±1.9256.37±1.51
Activation products43.33±0.0048.61±2.1456.88±1.87
All three responses / Single with all passages46.53±1.4647.92±1.4456.69±1.04

FGD scores all eight message combinations using mean sender activation summaries and concatenated messages. FGD, Additive, and Activation products use the same parameter choices and select by validation accuracy. Parameters are fixed before generating new answers to the previously used evaluation questions.

Question 1: What are the wall-clock runtime and computational cost of CGD’s selection procedure, including activation extraction and scoring, compared with one round of MAD or one pass of SafeSieve? Please report these measurements explicitly for the 2Wiki setting.

Author's Response: Thank you for asking about the complete processing cost. FGD takes less time per question than MAD and SafeSieve on both models under their original protocols. Table 16 in Appendix E.6 includes prompt processing, activation extraction, scoring, selection, and decoding. MAD and SafeSieve perform one communication round after their initial responses.

AuthorsOpenReview ↗

Response to Reviewer DF4L Part 3

Question 1 (continued)

Table 16. Time per question under the original protocols on A100 40GB. We show the timing columns from the manuscript. Entries show the mean and sample standard deviation across three seeds on 2Wiki.

ApproachGemma-3-4B: Time (s)Qwen3-4B: Time (s)
FGD10.65±2.3519.46±11.40
MAD39.25±9.0730.37±7.08
SafeSieve27.21±6.2324.69±0.14

Table 17 reports the comparisons on the same H100 GPU with the same weights, decoding settings, and software. FGD using passages is more accurate than self-consistency under the time budget set by FGD. With generated responses, self-consistency is more accurate under the corresponding budget. Single with all passages is faster and slightly less accurate than FGD using passages.

Table 17. Accuracy and time per question with Gemma-3-4B using the same H100 GPU and software. Entries show the mean and sample standard deviation across three seeds on 2Wiki. SC uses all passages and the elapsed-time budget of the indicated FGD implementation. Three-sender response runs are collected separately from the message-selection comparisons in Appendix E.5.

(a) Two senders

ApproachAccuracy (%)Time (s)
FGD using responses41.94±2.376.39±0.09
SC, time matched to FGD using responses47.78±1.206.41±0.09
FGD using passages47.92±1.102.41±0.06
SC, time matched to FGD using passages46.53±0.962.43±0.06
Single with all passages, no time limit47.22±1.272.11±0.05

(b) Three senders

ApproachAccuracy (%)Time (s)
FGD using responses45.69±0.247.94±0.14
All three responses48.89±0.647.70±0.15
FGD using passages49.44±2.842.48±0.05
Single with all passages47.92±1.442.21±0.03

Each message combination starts by processing its full input prompt. We disable key and value caching during activation collection. Answers for unselected combinations are generated after timing. Appendix E.6 reports generation calls, time limits and overruns, and the separate cost of collecting training and validation labels. Additive and Activation products reuse those records.

Question 2: How does CGD perform with three or more senders, in terms of both accuracy and computational efficiency?

Author's Response: With three senders FGD using passages is more accurate than Single with all passages. Single is faster. Please see our response to Weakness 3 for accuracy and Question 1 for timing and the included processing steps.

FGD processes all eight receiver message combinations in addition to the sender inputs. FGD using passages then generates one final answer. FGD using responses also generates one response per sender.

Question 3: Figure 4(a) reports message-level error correction rates of only 3.71%–7.18%. If roughly 92–96% of messages leave the answer unchanged, how stable and informative are the pairwise training labels when most message combinations have equal utility?

Author's Response: Thank you for asking about these labels. The collected labels are informative for message selection. FGD's mean accuracy advantage persists across the recorded training sizes and subsets. Training uses both corrections and introduced errors between message combinations. Equal task scores have zero weight in Equation (9). The comparisons also include the empty message combination. The correction rate in Fig. 4(a) therefore understates the available training information.

Table 18(a) in Appendix E.6 reports the proportions of questions and pairs of message combinations contributing to training. Fig. 6 shows higher mean accuracy for FGD than Activation products across the recorded training sizes and subsets. These subsets share collected labels. The bands show variation across evaluation seeds.

Reviewer_RJqDOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper asks how a message from one LLM agent changes another agent's answer, including when agents reinforce a shared error. Social Circuits compares the receiver's answer scores without the message, with it, and with the message omitted or its tokens shuffled but some of the receiver's activations copied back from the run with the message. Circuit-Guided Deliberation (CGD) uses receiver activations to choose which messages to send. It runs the receiver with each combination of two candidate messages (neither, one or both) and picks the one that a linear scorer, trained on 2Wiki, rates highest. The scorer's inputs include the change in the receiver's activations. Proposition 1 and Theorems 1 and 2 state when this activation replacement preserves the answer and bound CGD's gap to the best combination.

Critical strengths

The paper claims that the receiver's activations explain how a message changes its answer and help choose which messages to send, a claim that would matter if it held. For the reasons in Critical Weaknesses, I cannot count either as a separable contribution.

Critical weaknesses

The evidence that receiver activations explain message effects comes from a test that can hardly fail, and the one test that could fail, on the echo-chamber task of the title, does fail.

1. The central claim holds only where the replacement copies the token the answer starts from, and fails on the paper's own echo-chamber task. The paper supports its abstract's claim that "receiver activation changes explain the message effects" by copying some of the receiver's activations from the run with the message into a run without it, and checking whether the answer returns. On 2Wiki, the copied activations are those of the last token before the answer, at six layers, and the run they are copied into has no message at all (Appendix D.2). That token is then the only position that carries the message, so every answer token, whether the answer is one token or a whole passage title, receives the message through it. The near-complete recovery shows only that the token from which the model starts its answer carries the message's effect, not where in the model the effect is mediated, and copying that token at several layers adds little beyond the deepest copy, whose depth the paper does not give.

On Correlated-Trap, the echo-chamber task of the title, the paper copies the prompt tokens after the message but leaves out the answer position (Appendix D.2), and recovery is at most 16.39% (Table 6b). The claim thus fails on the one test that could fail, which is also the task the paper asks about, and the main text does not say that the two results come from different protocols.

The theory does not change this. Proposition 1 amounts to the fact that a model given the same internal state produces the same output, and neither protocol meets its condition: the same token positions, with the message tokens and every later token used to score the answer copied at one layer. Theorem 1 states, for any two vectors of answer scores, that changes smaller than the relevant margins keep the top answer and the sign of a score difference. Neither uses the model, the message or the activations, or bounds the error it assumes to be small, and Theorem 2 is a generic regret bound for a pairwise ranking loss whose covariance and loss terms the paper does not estimate.

2. An unequal comparison, not the activation feature, produces CGD's advantage. On 2Wiki A/B, CGD's messages include the sender's passages, while the collaboration baselines send responses only (Table 1 caption). CGD's scorer is trained on 2Wiki labels, and CGD runs the receiver on every candidate combination and uses its activations before choosing. Each of these gives CGD information that the collaboration baselines lack. Against Activation products, which uses the same candidates and differs only in the feature, CGD's advantage is "0.42%–1.11%" with matched input scaling and weight penalty (Table 8). So the main comparison does not isolate CGD's receiver-change feature, which adds little once the other differences are removed.

Questions for authors
  1. Where in each decoder layer are the replaced activations taken, the full layer output or a sublayer? If they are the full layer outputs at the position before the answer, recovery is expected by construction, and only a narrower replacement, such as one attention or MLP sublayer, could show which parts mediate the effect.
  2. Do the findings hold beyond the dense models tested? Results on larger and mixture-of-experts models would show how general the framework is. In a mixture-of-experts model, a message can also change the routing.
  3. How does CGD compare when every method gets the same passages, and how much compute does it need, including the receiver runs used to score candidates? An advantage that remains with equal information and compute would answer Weakness 2.
  4. How are the Correlated-Trap tasks built: where do the questions come from, and how are the shared misleading passage and the correct evidence written and checked? Appendix D.1 describes the task in one sentence, and it is the only test of the echo-chamber setting in the title.
6 comments
AuthorsOpenReview ↗

Response to Reviewer RJqD Part 1

We thank the reviewer for the detailed comments. We respond to the weaknesses and questions below.

Weakness 1: The central claim holds only where the replacement copies the token the answer starts from, and fails on the paper's own echo-chamber task.

Author's Response: We disagree that high answer recovery is inevitable under the replacement used here. We restore full decoder-block outputs only at selected layers and token positions, whereas Proposition 1 requires all message tokens and later scoring tokens to be restored at one layer.

The changed message and the activations at unreplaced positions remain in the receiver run. Recovery therefore depends on whether the selected activation changes reproduce the original message effect through the remaining layers.

Table 3 shows higher recovery for the original message than for controls matched in norm on both tasks, with the same replacement positions and A/B scoring. All activations are recorded before either answer token. This comparison traces the message effect in the selected decoder-block outputs.

Table 3. Answer recovery (%) after replacement at the common answer prefix on Gemma-3-4B. Answer recovery uses records where the original and reference messages produce different answers.

TaskReferenceOriginal messageControl
2WikiMessage removal95.92±2.3544.13±17.80
Shuffled message97.81±2.1333.53±8.90
Correlated-TrapMessage removal99.85±0.260.00±0.00
Shuffled message99.84±0.280.00±0.00

Appendix D.2 specifies six depths from 15% to 90%, including the deepest replaced layer. Table 8(a) and Fig. 15 show higher answer recovery and lower mean absolute error in answer-score changes for the original message than for the matched controls. Fig. 14(a) reports the effects at individual depths, including reversals in direction.

Table 8. Receiver answers after activation replacement, reported as mean ± sample standard deviation. Agreement uses the same matched message records for both replacements: 1,268 for Gemma-3-4B and 1,334 for Qwen3-4B on 2Wiki, and 1,440 per model on Correlated-Trap. Answer 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.

(a) 2Wiki

ModelActivation changesAnswer agreement (%)Answer recovery (%)
Gemma-3-4BOriginal message99.61±0.2797.81±0.23
Different message (same task)88.89±1.8841.07±13.20
Qwen3-4BOriginal message99.78±0.0096.73±0.22
Different message (same task)94.23±1.4855.06±11.67

(b) Correlated-Trap

ModelActivation changesAnswer agreement (%)Answer recovery (%)
Gemma-3-4BOriginal message23.68±2.6816.39±1.67
Swapped-role changes6.74±1.150.23±0.00
Qwen3-4BOriginal message18.47±1.071.33±0.92
Swapped-role changes14.51±1.050.27±0.27
Qwen3-8BOriginal message40.90±1.894.88±1.54
Swapped-role changes30.49±1.730.00±0.00

The joint replacement measures the combined effect of the selected outputs. Establishing how much each shallower layer adds requires comparing their joint replacement with replacement at the deepest layer alone.

We agree that individual answer recovery is low under the original Correlated-Trap protocol. We disagree that its contrast with 2Wiki establishes failure on the echo-chamber task, because the two protocols replace different positions and use different scoring. Correlated-Trap replaces tokens before the answer prefix under a shuffled reference, whereas 2Wiki replaces the last common answer-prefix token after message removal. The common-prefix comparison above uses the same position and A/B scoring on both tasks. Fig. 13 preserves the original message-effect results, including negative recovery for correction on Qwen3-4B.

Table 4 reports freely generated answers during a further 2Wiki debate. Among answers changed by shuffling the sender message, restoring the original activations recovers the original answer more often than both controls matched in norm in every seed. The recovered answers include errors the agents already shared. Section 5.2 presents this comparison, and Appendix E.3 describes its settings.

Table 4. Answer recovery (%) of further MAD debate on 2Wiki with Gemma-3-4B. We replace activations from the sender message through the final prompt token, using a shuffled message as the reference. Entries show the mean and sample standard deviation across three seeds.

Activation changesAnswer recovery (%)
Original message49.70±6.70
Different message37.60±5.46
Shuffled activation changes15.77±14.52
AuthorsOpenReview ↗

Response to Reviewer RJqD Part 2

Weakness 1 (continued)

Our theoretical contribution is to connect activation recovery with the task performance of message selection. Proposition 1 specifies sufficient activation replacements for exact recovery. Theorem 1 specifies when imperfect recovery still preserves the receiver's answer and the direction of the message effect. Theorem 2 then bounds FGD's expected task-utility difference from the optimal message combination. The bound accounts for the training loss and covariance between task utilities. It also includes answer disagreement after activation replacement. Reducing the training loss toward its minimum tightens the bound when recovery preserves answers and the covariance term vanishes.

We have also calculated the covariance and loss terms requested by the reviewer. Table 12 reports these terms after training and validation on separate questions. The training-loss term contributes most to the bound. The two sampling runs use their own answer probabilities and the same uniform random draw.

Table 12. Empirical terms in Theorem 2 over 720 runs. The receiver either selects the higher-scoring answer or samples A or B from the token probabilities at temperature 1. Sampling uses the same uniform random draw for original-message and activation-replacement answers. The last row reports the proportion of all runs with a bound below one.

QuantityHigher-scoring answerSampled answer
Task-utility difference0.045830.04790
Covariance term0.000000.00088
Training-loss term0.308360.37777
Answer-disagreement term0.050000.04525
Mean bound0.358360.42390
Runs with bound below one (%)88.0684.72

Weakness 2: An unequal comparison, not the activation feature, produces CGD's advantage.

Author's Response: We agree that the main comparison combines differences in message content, training, and receiver processing. We disagree that unequal information alone explains all of FGD's gains, because the separate feature comparison fixes candidate messages, training data, input scaling, and weight penalty. Table 1 compares FGD using sender passages and responses with collaboration baselines using responses. Appendix D.1 specifies shared passage access in their initial free-generation messages. Table 2 compares equal generation budgets, and Table 10 fixes messages and receiver settings.

Table 1. Accuracy (%, mean ± 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, FGD sends passages and responses, and collaboration baselines send responses alone.

Single-agent: Single and SC6. Multi-agent: Majority vote, MAD, MAD-M2, MADC, MOC, SafeSieve, and FGD.

2Wiki (A/B answers)

ModelMetricSingleSC6Majority voteMADMAD-M2MADCMOCSafeSieveFGD
Gemma-3-4BAccuracy ↑78.89±0.4877.78±1.4683.19±0.6480.69±1.0582.92±1.5079.44±1.2786.94±0.9679.72±3.0789.72±1.46
Tokens ↓79478219647516638401485139
Qwen3-4BAccuracy ↑73.19±1.7370.83±2.3273.61±1.6880.42±2.0882.08±1.4477.22±1.2786.94±0.6483.89±1.0589.86±1.68
Tokens ↓17110254689781028961864874299
Qwen3-8BAccuracy ↑75.69±0.4872.92±0.4280.97±2.5581.81±2.7178.47±2.7778.61±0.2485.83±0.4281.94±1.9787.36±0.64
Tokens ↓1569304399729009598811004286

GSM8K (free generation)

ModelMetricSingleSC6Majority voteMADMAD-M2MADCMOCSafeSieveFGD
Gemma-3-4BAccuracy ↑84.86±1.2088.19±1.0586.81±0.6486.53±1.7388.61±1.3486.53±1.7388.75±1.5087.64±0.2488.75±0.48
Tokens ↓224132967813861248138612111139656
Qwen3-4BAccuracy ↑89.86±0.8791.53±0.2489.72±0.6490.83±0.7291.11±1.6890.83±0.7291.11±0.6492.92±0.4293.33±1.10
Tokens ↓221132065612911226129111441268640
Qwen3-8BAccuracy ↑91.81±1.4692.92±0.8392.78±1.0593.89±1.7393.89±1.0593.89±1.7393.75±0.4292.64±0.4893.89±0.64
Tokens ↓229138368814271325142712121389677
AuthorsOpenReview ↗

Response to Reviewer RJqD Part 3

Weakness 2 (continued)

Table 1 (continued)

MATH-500 (A/B answers)

ModelMetricSingleSC6Majority voteMADMAD-M2MADCMOCSafeSieveFGD
Gemma-3-4BAccuracy ↑42.67±0.7652.00±1.0046.33±1.0452.83±0.5853.83±2.0252.83±0.5866.33±0.5859.67±0.5868.83±1.76
Tokens ↓4192507126424292411242923532203897
Qwen3-4BAccuracy ↑55.83±0.7660.00±0.5057.50±1.3263.50±1.8062.67±2.3163.50±1.8069.50±2.1865.33±0.7670.83±0.76
Tokens ↓3852321116823092279230921432218818
Qwen3-8BAccuracy ↑59.17±0.7662.33±0.5859.00±1.3259.83±1.0462.00±1.3259.83±1.2664.67±1.2661.50±0.5065.83±1.61
Tokens ↓3872333117423632308236522272318912

TruthfulQA (A/B answers)

ModelMetricSingleSC6Majority voteMADMAD-M2MADCMOCSafeSieveFGD
Gemma-3-4BAccuracy ↑62.67±1.7663.83±1.2668.00±1.3266.83±2.7570.17±0.2966.83±2.7568.67±1.0467.17±1.2670.67±1.04
Tokens ↓67405236583418583449617191
Qwen3-4BAccuracy ↑76.17±1.5375.83±0.7679.00±1.8079.83±1.4479.00±1.0079.83±1.4482.00±1.3280.00±1.5083.33±1.76
Tokens ↓105627353828691828677716208
Qwen3-8BAccuracy ↑82.67±0.5882.00±0.5083.67±1.0485.00±2.2982.50±1.5085.00±2.2986.33±2.0884.83±0.5887.00±1.50
Tokens ↓96576305746590746608698236

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

Single-agent: Single and SC6. Multi-agent: Majority vote, MAD, MAD-M2, MADC, MOC, SafeSieve, and FGD.

ModelSingleSC6Majority voteMADMAD-M2MADCMOCSafeSieveFGD
Gemma-3-4B24.72±1.0525.42±1.6726.39±2.777.22±1.8820.97±2.107.22±1.8820.28±0.4825.42±4.2342.92±0.83
Qwen3-4B21.94±0.2422.22±0.4825.97±1.9723.33±2.5325.14±3.1323.33±2.5329.31±2.4440.56±2.3751.81±0.64
Qwen3-8B23.61±2.1023.89±1.0530.00±1.6714.17±0.8330.83±0.8314.17±0.8328.75±5.4229.86±5.5061.53±0.48

Table 10. A/B accuracy (%) on 240 new 2Wiki tasks for the 4B models and the original 240 tasks for Qwen3-8B, reported as mean ± sample standard deviation. FGD achieves the highest mean accuracy across all three models.

StrategyGemma-3-4BQwen3-4BQwen3-8B
No message84.17±0.0078.75±0.0081.67±0.00
Sender 1 only86.11±1.2781.67±1.1084.58±2.17
Sender 2 only82.50±1.9179.58±1.4477.50±0.83
Both messages85.28±1.3483.89±1.3482.78±1.97
FGD89.72±1.2091.11±0.6487.36±0.64

Table 11 shows a modest mean gain over Activation products under these matched conditions. Table 13 shows higher free-generation accuracy for FGD on Qwen3-8B and for Activation products on Gemma-3-4B. The Qwen3-8B FGD entry uses a separate run with the same tasks, seeds, and scoring.

Table 11. A/B accuracy (%) on the original 240 2Wiki tasks for each model, before and after matching the scaling and penalty to FGD. Entries show the mean and sample standard deviation across three seeds.

ApproachGemma-3-4BQwen3-4BQwen3-8B
Activation products88.06±2.2987.92±0.8385.00±1.50
Activation products (adjusted)89.31±1.0588.75±1.5086.94±1.27
FGD89.72±1.4689.86±1.6887.36±0.64

Table 13. Free-generation accuracy (%) on 2Wiki, reported as mean ± sample standard deviation. The Qwen3-8B FGD entry uses a separate run with the same evaluation tasks, seeds, and scoring. Bold marks the highest mean in each column.

ApproachGemma-3-4BQwen3-8B
Text-only38.33±1.6756.67±1.10
Receiver-only40.69±1.5860.42±0.42
Text + receiver38.47±1.0559.17±0.72
Additive42.78±1.2753.61±1.05
Activation products43.33±1.9157.36±0.24
FGD42.92±0.8361.53±0.48

With three senders, Table 14(b) shows higher mean accuracy and F1 for FGD using passages than for Additive, Activation products, and Single with all passages. Additive and Activation products share FGD’s activation summaries, message combinations, prompt template, and decoding settings. FGD’s accuracy matches or exceeds every comparison approach in each evaluation seed.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to the Area Chair

We thank the area chair and reviewers for their careful reading and feedback. The revised manuscript is titled Social Footprints in Multi-agent Echo Chambers. The selection approach is now called Footprint-Guided Deliberation (FGD). We summarize the main concerns and corresponding changes below.

To Reviewer RJqD:

  • Answer recovery: Table 3 in Section 5.2 shows higher answer recovery for the original message than for controls matched in norm on both 2Wiki and Correlated-Trap. Both tasks use activation replacement at the common answer prefix with A/B scoring. These controls show that recovery depends on the message that produced the activations.
  • Theoretical contribution: Our theory connects activation recovery with the task performance of message selection. Sections 3.3 and 4.3 establish recovery conditions and bound FGD's expected task-utility difference from the optimal message combination. Table 12 in Appendix E.4 reports the numerical terms and shows that the training-loss term contributes most to the bound.

To Reviewer DF4L:

  • Naming and shared errors: Social Footprints describes the message effects traced through receiver activations. Section 5.2 now includes further debate after the agents have given the same wrong answer. Table 4 shows higher recovery for the original activations than for both controls matched in norm. The recovered answers include errors already shared by the agents. Appendix E.3 describes the settings.
  • More senders: Table 14(b) shows the highest mean accuracy and F1 for FGD using passages with three senders. Appendix E.5 holds the available passages and receiver settings fixed across the comparisons. This supports the benefit of message selection.

To Reviewer p3yh:

  • Feature contribution: Section 4.1 explains how FGD shares weights with opposite signs for the sender–receiver products before and after a message. Table 11 shows higher mean A/B accuracy across all three models under matched training and input settings. Table 13 shows higher free-generation accuracy for FGD on Qwen3-8B and for Activation products on Gemma-3-4B.
  • Equal information and elapsed time: Table 17 shows higher accuracy for FGD using passages than for self-consistency under the time budget set by FGD. Both receive the same available passages. Self-consistency is more accurate when FGD uses generated responses. Single with all passages is faster than FGD using passages. Appendix E.6 includes all online processing steps and separately reports training and validation data collection costs.

Finally, we thank the area chair for carefully reading this response and the reviewers for their constructive comments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.