acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

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. Code and data are available at https://anonymous.4open.science/r/substitution-F287.

Primary area:
foundation or frontier models, including LLMs
Keywords:
multi-agent systems, value of information, controlled experiments, message reliability

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_kVFTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how upstream messages affect downstream LLM’s answers, particularly when an incorrect message causes it to abandon an answer it could produce correctly on its own. The paper conducts controlled experiments across several benchmarks and shows the errors often follow the upstream agent’s specific wrong answer, even when the receiver has relevant evidence.

Critical strengths
  1. The paper uses designed controls to distinguish the influence of message content from alternative explanations, such as source labels and additional reasoning attempts.

  2. Experiments across five models and five benchmarks show that the observed failure is not limited to a single model or task.

Critical weaknesses
  1. The main concern is the LLM conformity, including abandoning correct judgments under incorrect peer influence has been investigated by previous work[1]. It would be better if the paper illustrate what new conclusion this design enables beyond the prior conformity studies.

  2. It is unclear how sensitive the findings are to the combination of upstream and downstream models, as the paper evaluates only selected combinations. Also, can the findings from controlled two-agent be extended to multi-agent settings?

  3. The paper focuses on demonstrating substitution, but do not have much investigation into the factors that could cause it. For example, does substitution depend on the stability or uncertainty of the receiver’s original answer? Does it occur more often when evaluating the evidence requires difficult inference than when the evidence directly contradicts the upstream message? Understanding these factors would provide more insights for future work.

[1] DO AS WE DO, NOT AS YOU THINK: THE CONFORMITY OF LARGE LANGUAGE MODELS (ICLR 2025)

Questions for authors
  1. Can you illustrate the novelty of the findings and the differences between this work and previous work?

  2. Would different upstream and downstream model combinations lead to different conclusions? Also, can this controlled two-agent conclusion been extended to multi-agent settings?

  3. Does substitution depend on the stability or uncertainty of the receiver’s original answer? Does it occur more often when evaluating the evidence requires difficult inference than when the evidence directly contradicts the upstream message?

8 comments
AuthorsOpenReview ↗

Response to Reviewer kVFT (1/5)

Dear Reviewer kVFT,

Thank you for your careful review and for recognizing the value of our controlled experiments and evaluation across models and benchmarks. Your suggestions have guided us in clarifying the contribution, extending the model and workflow comparisons, and examining factors associated with substitution.

1. What new conclusions does this design enable beyond prior conformity studies?

Weakness 1

The main concern is the LLM conformity, including abandoning correct judgments under incorrect peer influence has been investigated by previous work[1]. It would be better if the paper illustrate what new conclusion this design enables beyond the prior conformity studies.

[1] DO AS WE DO, NOT AS YOU THINK: THE CONFORMITY OF LARGE LANGUAGE MODELS (ICLR 2025)

Question 1

Can you illustrate the novelty of the findings and the differences between this work and previous work?

Thank you for encouraging us to clarify the findings beyond prior conformity studies. The original submission quantifies the risk that a receiver abandons an independently correct answer for the upstream model's specific wrong answer while its independent evidence remains available. Communication gains can coexist with these directed losses on independently solvable items, so overall accuracy alone does not reveal this handoff risk.

We examine this risk through item-level shown–hidden comparisons that hold the receiver's independent evidence fixed. Task-scored correct-to-wrong (c→w) changes occur in 11.1% of independently solvable model–question cases, with a maximum of 32% in one model–benchmark combination. A separate audit of c→w transitions finds that 94% match the upstream model's specific wrong answer. Reversing the upstream conclusion while keeping the evidence fixed also changes receiver answers, providing intervention evidence that receiver answers track the message's specific conclusion.

Weng et al. [1] examine interaction protocols, interaction time, and majority size, and evaluate persona and reflection strategies for mitigating conformity. Qu et al. [2] manipulate consensus structure and authority labels, compare harmful and beneficial revisions, and test chain-of-thought and reflect-then-revise interventions.

Our supplementary experiments further quantify how directed substitution risk varies across upstream–receiver pairings and complete multi-agent workflows while independent evidence remains available. Substitution occurs across all tested model pairings, with frequencies varying across pairings and tasks (Response 2.1), and specific upstream errors can reach final answers despite subsequent review, adjudication, or aggregation (Response 2.2). Lower original-answer stability and lower self-reported confidence are associated with higher substitution rates (Response 3.1); natural-question comparisons also link the inference needed to refute the actual upstream draft with higher substitution rates (Response 3.2).

AuthorsOpenReview ↗

Response to Reviewer kVFT (2/5)

2. How sensitive are the findings to model combinations, and do they extend to multi-agent settings?

Weakness 2

It is unclear how sensitive the findings are to the combination of upstream and downstream models, as the paper evaluates only selected combinations. Also, can the findings from controlled two-agent be extended to multi-agent settings?

Question 2

Would different upstream and downstream model combinations lead to different conclusions? Also, can this controlled two-agent conclusion been extended to multi-agent settings?

2.1. Do different upstream and downstream model combinations lead to different conclusions?

Thank you for asking about sensitivity to upstream and downstream model combinations. Substitution is observed in all 20 tested combinations on each of the five tasks.

We cross four upstream models with five receivers. Substitution means a correct hidden answer changes to the upstream model's specific wrong answer when the message is shown. Adoption increase is the shown-minus-hidden difference in upstream-error adoption, in percentage points (pp).

MeasureBIRDHotpotQALBMuSiQue2WikiMultihopQADROP
Questions15002001591200120
Pairings with substitution20/2020/2020/2020/2020/20
Adoption increase, pp [95% interval]+39.01 [30.16, 49.05]+14.73 [10.51, 19.40]+22.13 [18.22, 25.95]+36.61 [34.07, 39.10]+47.16 [38.77, 55.31]

Each cell below reports substitutions / hidden-correct answers on questions with an incorrect upstream answer (percentage).

Upstream modelReceiverBIRDHotpotQALBMuSiQue2WikiMultihopQADROP
GPT-4o-miniGPT-4o-mini53/99 (53.5%)3/62 (4.8%)4/13 (30.8%)7/38 (18.4%)2/2 (100.0%)
GPT-4o-miniDeepSeek-V3.287/219 (39.7%)6/61 (9.8%)9/17 (52.9%)9/55 (16.4%)3/4 (75.0%)
GPT-4o-miniKimi-K2.652/220 (23.6%)3/63 (4.8%)12/19 (63.2%)7/45 (15.6%)7/12 (58.3%)
GPT-4o-miniGLM-546/260 (17.7%)1/69 (1.4%)2/19 (10.5%)1/92 (1.1%)2/5 (40.0%)
GPT-4o-miniQwen3.6-Plus106/248 (42.7%)2/66 (3.0%)2/16 (12.5%)2/101 (2.0%)3/5 (60.0%)
GPT-5.4GPT-4o-mini24/89 (27.0%)5/5 (100.0%)4/9 (44.4%)11/11 (100.0%)3/3 (100.0%)
GPT-5.4DeepSeek-V3.268/117 (58.1%)5/5 (100.0%)8/8 (100.0%)96/107 (89.7%)5/6 (83.3%)
GPT-5.4Kimi-K2.675/129 (58.1%)3/3 (100.0%)6/9 (66.7%)41/41 (100.0%)11/12 (91.7%)
GPT-5.4GLM-5109/158 (69.0%)5/6 (83.3%)5/10 (50.0%)33/155 (21.3%)4/6 (66.7%)
GPT-5.4Qwen3.6-Plus76/149 (51.0%)6/6 (100.0%)6/9 (66.7%)99/130 (76.2%)3/5 (60.0%)
Kimi-K2.6GPT-4o-mini50/50 (100.0%)7/17 (41.2%)1/26 (3.8%)9/32 (28.1%)3/4 (75.0%)
Kimi-K2.6DeepSeek-V3.227/84 (32.1%)4/14 (28.6%)2/29 (6.9%)20/87 (23.0%)3/6 (50.0%)
Kimi-K2.6Kimi-K2.62/47 (4.3%)5/10 (50.0%)2/21 (9.5%)21/71 (29.6%)5/8 (62.5%)
Kimi-K2.6GLM-529/53 (54.7%)5/18 (27.8%)2/28 (7.1%)7/125 (5.6%)1/8 (12.5%)
Kimi-K2.6Qwen3.6-Plus24/68 (35.3%)6/16 (37.5%)1/25 (4.0%)5/94 (5.3%)1/8 (12.5%)
Qwen-PlusGPT-4o-mini13/25 (52.0%)6/21 (28.6%)2/31 (6.5%)45/45 (100.0%)1/1 (100.0%)
Qwen-PlusDeepSeek-V3.222/42 (52.4%)3/20 (15.0%)2/31 (6.5%)38/38 (100.0%)3/3 (100.0%)
Qwen-PlusKimi-K2.614/36 (38.9%)6/21 (28.6%)2/30 (6.7%)28/38 (73.7%)4/7 (57.1%)
Qwen-PlusGLM-523/61 (37.7%)4/23 (17.4%)4/37 (10.8%)38/50 (76.0%)2/4 (50.0%)
Qwen-PlusQwen3.6-Plus30/65 (46.2%)3/22 (13.6%)1/32 (3.1%)20/34 (58.8%)4/4 (100.0%)

The expanded matrix supports the same qualitative finding beyond the combinations selected in the original submission: substitution occurs across all tested pairings, while its frequency varies across combinations and tasks.

AuthorsOpenReview ↗

Response to Reviewer kVFT (3/5)

2.2. Does the controlled two-agent finding extend to multi-agent settings?

We appreciate your suggestion to examine multi-agent settings. Substitution persists in every tested multi-agent configuration across the five tasks. Within each task, workflows use the same questions and incorrect drafts, with five calls per question in each shown or hidden condition.

Homogeneous workflows use GPT-4o-mini; heterogeneous workflows use GPT-4o-mini, Kimi-K2.6, Qwen3.6-Plus, and GLM-5. “Roles swapped” exchanges Kimi-K2.6 and GLM-5 between review and adjudication. Substitution is reported over all questions (n/N).

ConfigurationWorkflowMeasureBIRDHotpotQALBMuSiQue2WikiMultihopQADROP
AllQuestions88354110120120
HomogeneousSequential reviewSubstitution, n/N (%)60/88 (68.2%)28/354 (7.9%)71/110 (64.5%)20/120 (16.7%)20/120 (16.7%)
Adoption increase, pp [95% interval]+70.45 [57.89, 83.75]+7.91 [5.08, 10.73]+64.55 [55.45, 73.64]+16.67 [10.00, 23.33]+16.67 [10.00, 23.33]
Final correct, hidden → shown81/88 → 20/88346/354 → 232/354110/110 → 27/110120/120 → 100/120120/120 → 100/120
Mutual review and adjudicationSubstitution, n/N (%)45/88 (51.1%)18/354 (5.1%)41/110 (37.3%)21/120 (17.5%)13/120 (10.8%)
Adoption increase, pp [95% interval]+51.14 [40.23, 62.50]+5.08 [2.82, 7.34]+37.27 [28.18, 46.36]+17.50 [10.83, 24.17]+10.83 [5.83, 16.67]
Final correct, hidden → shown82/88 → 36/88354/354 → 288/354110/110 → 58/110120/120 → 98/120120/120 → 106/120
Independent aggregationSubstitution, n/N (%)33/88 (37.5%)11/354 (3.1%)40/110 (36.4%)14/120 (11.7%)12/120 (10.0%)
Adoption increase, pp [95% interval]+38.64 [28.17, 49.35]+3.11 [1.41, 5.08]+36.36 [27.27, 45.45]+11.67 [6.67, 17.50]+10.00 [5.00, 15.83]
Final correct, hidden → shown82/88 → 47/88354/354 → 320/354110/110 → 61/110120/120 → 106/120120/120 → 107/120
HeterogeneousSequential reviewSubstitution, n/N (%)12/88 (13.6%)4/354 (1.1%)3/110 (2.7%)1/120 (0.8%)3/120 (2.5%)
Adoption increase, pp [95% interval]+13.64 [6.25, 19.44]+1.13 [0.28, 2.26]+2.73 [0.00, 6.36]+0.83 [0.00, 2.50]+2.50 [0.00, 5.83]
Final correct, hidden → shown77/88 → 66/88354/354 → 307/35483/110 → 80/110120/120 → 117/120120/120 → 116/120
Mutual review and adjudicationSubstitution, n/N (%)7/88 (8.0%)6/354 (1.7%)14/110 (12.7%)3/120 (2.5%)4/120 (3.3%)
Adoption increase, pp [95% interval]+7.95 [3.30, 13.51]+1.69 [0.56, 3.11]+12.73 [7.27, 19.09]+2.50 [0.00, 5.83]+3.33 [0.83, 6.67]
Final correct, hidden → shown67/88 → 62/88349/354 → 321/35489/110 → 67/110120/120 → 116/120120/120 → 115/120
Roles swappedSubstitution, n/N (%)14/88 (15.9%)8/354 (2.3%)6/110 (5.5%)6/120 (5.0%)5/120 (4.2%)
Adoption increase, pp [95% interval]+14.77 [4.23, 23.46]+2.26 [0.85, 3.95]+5.45 [1.82, 10.00]+5.00 [1.67, 9.17]+4.17 [0.83, 8.33]
Final correct, hidden → shown69/88 → 55/88348/354 → 324/354110/110 → 85/110120/120 → 110/120120/120 → 115/120

Every tested configuration on every task shows fewer final correct answers in the shown condition, alongside observed substitutions to the upstream model's specific wrong answer. Subsequent review, adjudication, and aggregation do not eliminate substitution in these tested settings.

Reviewer_oc3JOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies answer substitution during handoffs between LLM agents: an incorrect upstream message can induce a recipient to abandon the correct answer it could have derived from independent evidence. Holding recipient evidence fixed, the authors manipulate message visibility and conclusions across five benchmarks and five recipient model families. Answer overriding occurs in 11.1% of initially independently solvable examples, reaching 32% for one model–benchmark combination; in one audited subset, 94% adopted the specific incorrect answer supplied upstream. Additional experiments examine source labels, reception timing, matched review calls, conclusion-only edits, and chain-of-thought prompting. Interventions using oracle information and detector-gated recipient replacement explore recovery through dropping messages or switching recipients. The main contribution is a controlled empirical characterization of communication-induced errors and their boundary conditions.

Critical strengths
  1. The instance-level experimental design holds independent evidence fixed while separating benefits from communication from harmful correct-to-incorrect transitions. This is substantially more informative than reporting only overall pipeline accuracy.
  2. The experimental controls are comprehensive: matched review calls, delayed reception, modified source labels, minimal conclusion edits, and repeated evidence-only runs rule out several potential confounders.
  3. The study covers five recipient model families and five benchmarks, reports paired uncertainty estimates, and explicitly documents cases in which communication remains beneficial. The detector-gated recovery analysis also acknowledges the cost of discarding correct messages.
Critical weaknesses
  1. The main experiments rely on oracle-quality evidence and a single two-node handoff with verifiable answers. The three-layer experiment contains only 30 examples, so the measured incidence and intervention trade-offs cannot be assumed to generalize to noisy retrieval, iterative debate, or longer workflows.
  2. Evidence for a conclusion-specific causal effect is narrower than the general message-substitution results suggest. Full reversals change substantial supporting content, whereas minimal edits yield significant effects in only four of nine experimental combinations and none on BIRD.
  3. Evidence for practical recovery remains limited to one shared error detector and a small subset of benchmarks. Although detector precision and recall are reported, the paper does not characterize its operating curve or include the inference cost of detection and replacement in deployment trade-offs, making it difficult to assess selective gating beyond the tested settings.
Questions for authors

See Weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer oc3J (1/2)

Dear Reviewer oc3J,

Thank you for your time and for recognizing the value of separating communication benefits from harmful correct-to-wrong transitions on individual questions. We address your concerns about generalization, conclusion-specific causal evidence, and practical recovery below.

1. How do incidence and intervention trade-offs change with noisy retrieval, iterative debate, and longer workflows?

Weakness 1

The main experiments rely on oracle-quality evidence and a single two-node handoff with verifiable answers. The three-layer experiment contains only 30 examples, so the measured incidence and intervention trade-offs cannot be assumed to generalize to noisy retrieval, iterative debate, or longer workflows.

Thank you for asking how incidence and intervention trade-offs extend to these broader settings. Substitution persists in all three added settings, and intervention gated by the paper's Gemini-2.5-Pro error detector yields positive net gains over passing all messages (pass-all) in every tested benchmark–workflow combination.

In new experiments, we compare matched GPT-4o-mini workflows with the original upstream message shown or hidden at workflow entry. For questions with a correct initial independent answer and an incorrect upstream answer, cells report substitution (%) / excess error (pp). Substitution requires a correct hidden final answer that changes to the upstream's specific wrong answer when shown. Excess error is the shown-minus-hidden error rate, including both transition directions.

Added settingBIRDLBML2WHotpotQADROP
Noisy retrieval20.41 / +16.3336.25 / +34.4438.89 / +36.9432.00 / +32.0026.95 / +25.60
Iterative debate: 3 debaters + judge40.91 / +42.9519.58 / +18.6021.00 / +19.9517.28 / +16.4214.55 / +13.83
Longer workflow: 7-stage chain68.18 / +68.1832.63 / +30.9935.00 / +33.2528.80 / +27.3624.26 / +23.04

To assess intervention trade-offs, we use the paper's Gemini-2.5-Pro error detector to trigger intervention when it flags an upstream answer as likely wrong. The table reports net task-score gains (pp) over pass-all on the full cohort, including both correct and incorrect upstream answers.

Added settingBIRDLBML2WHotpotQADROP
Noisy retrieval+7.13+9.98+4.27+9.18+8.31
Iterative debate: 3 debaters + judge+16.40+8.22+4.26+12.32+7.19
Longer workflow: 7-stage chain+23.82+9.65+4.46+14.10+8.16

These results extend the incidence and intervention comparisons beyond oracle-quality evidence and a single two-node handoff, within tasks with verifiable answers. Their magnitudes vary by benchmark and workflow.

2. What conclusion-specific causal effect do minimal edits support?

Weakness 2

Evidence for a conclusion-specific causal effect is narrower than the general message-substitution results suggest. Full reversals change substantial supporting content, whereas minimal edits yield significant effects in only four of nine experimental combinations and none on BIRD.

Thank you for distinguishing conclusion-specific causal evidence from the general message-substitution results. Under the same minimal-edit protocol, all nine expanded comparisons now show positive net error increases that pass Holm correction, including all three BIRD combinations. Strict final-answer controls clarify which effects can be attributed to the answer field alone.

The minimal edit changes the final answer and at most one conclusion sentence while leaving the remaining supporting content unchanged. The table reports both transition directions and their net error increase, with Holm correction across the nine comparisons.

DatasetReceiverPaired nCorrect→wrongWrong→correctNet error increase, pp [95% CI]Holm-adjusted p
HotpotQAGPT-4o-mini650311+4.62 [3.08, 6.31]<0.001
HotpotQADeepSeek-V3.2650844+12.31 [9.69, 15.08]<0.001
HotpotQAKimi-K2.6650398+4.77 [2.77, 6.77]<0.001
MuSiQueGPT-4o-mini200342+16.00 [10.50, 21.50]<0.001
MuSiQueDeepSeek-V3.2200512+24.50 [18.50, 31.00]<0.001
MuSiQueKimi-K2.6200142+6.00 [2.50, 10.00]0.004
BIRDGPT-4o-mini650311+4.62 [3.08, 6.31]<0.001
BIRDDeepSeek-V3.2650160+2.46 [1.38, 3.69]<0.001
BIRDKimi-K2.6650162+2.15 [0.92, 3.54]0.003
AuthorsOpenReview ↗

Response to Reviewer oc3J (2/2)

Because minimal edits may also change a conclusion sentence, we add strict GPT-4o-mini controls across five benchmarks. The first two contrasts change only the final answer while holding the original or rewritten supporting content verbatim fixed; the third rewrites supporting content while holding the same wrong answer fixed. Values are net error increases in pp [95% CI].

Within-question contrast (GPT-4o-mini)BIRDMuSiQue2WikiMultihopQAHotpotQADROP
Paired questions2753338630
Change the answer, original supporting content fixed+3.70 [0.00, 11.11]+5.66 [0.00, 13.21]+3.03 [0.00, 9.09]+3.49 [0.00, 8.14]+3.33 [0.00, 10.00]
Change the answer, rewritten supporting content fixed+3.70 [0.00, 11.11]+26.42 [15.09, 37.74]+12.12 [3.03, 24.24]+10.47 [3.49, 17.44]+6.67 [0.00, 16.67]
Rewrite supporting content, wrong answer fixed+3.70 [-7.41, 14.81]+47.17 [33.96, 60.38]+90.91 [78.79, 100.00]+34.88 [25.58, 45.35]+86.67 [73.33, 96.67]

Across the four QA benchmarks, rewriting the supporting content with the wrong answer fixed produces the largest observed net error increases among the tested contrasts. The final-answer field also contributes: with rewritten supporting content fixed, changing the final answer is significant on MuSiQue and HotpotQA after Holm correction across the five benchmarks.

3. How do detector operating curves and inference cost inform selective gating?

Weakness 3

Evidence for practical recovery remains limited to one shared error detector and a small subset of benchmarks. Although detector precision and recall are reported, the paper does not characterize its operating curve or include the inference cost of detection and replacement in deployment trade-offs, making it difficult to assess selective gating beyond the tested settings.

Thank you for highlighting the need to assess selective gating through detector operating curves and complete inference costs. Across all five benchmarks, selective gating achieves higher task scores than the default workflow, fixed replacement, and 50% random replacement.

Both tables use the same questions and receiver pairs. Each cell reports task score / relative total API cost; cost includes detection and replacement and is normalized to the default workflow in the same column. Fixed replacement uses the replacement receiver on every question, random replacement on a uniformly selected half, and selective gating on detector-flagged questions. The random row reports the expected result; the last row identifies the gating detector.

PolicyBIRD (n=150)LBM (n=160)L2W (n=120)HotpotQA (n=200)DROP (n=120)
Default59.33 / 1.0067.25 / 1.0073.93 / 1.0078.29 / 1.0087.74 / 1.00
Fixed replacement57.33 / 1.4271.94 / 1.5255.64 / 0.4681.94 / 0.4187.98 / 1.26
Random 50%58.33 / 1.2169.60 / 1.2664.78 / 0.7380.11 / 0.7187.86 / 1.13
Selective gating62.67 / 1.6472.53 / 1.4475.23 / 0.9982.56 / 1.2089.40 / 1.22
Gating detectorGPT-4o-miniGPT-4o-miniGPT-4o-miniGPT-4.1-miniGPT-4o-mini

Among the four policies, selective gating achieves the highest task score on all five benchmarks, at a total API cost of 0.99–1.64 times the default.

On L2W, gating improves the task score over the default workflow while slightly reducing total API cost. Compared with fixed replacement, gating scores higher at lower cost on LBM and DROP, and scores higher at higher cost on BIRD, L2W, and HotpotQA. We have presented the detector operating curves in Appendix B.

We also compare three detector configurations to assess recovery beyond one shared error detector. Each selective-gating entry above matches the named detector's result in the same column below, at the same operating point.

Detector configurationBIRD (n=150)LBM (n=160)L2W (n=120)HotpotQA (n=200)DROP (n=120)
Default59.33 / 1.0067.25 / 1.0073.93 / 1.0078.29 / 1.0087.74 / 1.00
Gemini-2.5-Pro64.67 / 12.6371.62 / 4.1277.08 / 5.1583.45 / 14.9690.29 / 22.14
GPT-4o-mini62.67 / 1.6472.53 / 1.4475.23 / 0.9979.04 / 1.0889.40 / 1.22
GPT-4.1-mini62.00 / 1.8871.34 / 1.8274.35 / 1.6582.56 / 1.2087.74 / 1.36

Among the three detector configurations, GPT-4o-mini has the lowest total API cost on all five benchmarks and the highest task score on LBM. Gemini-2.5-Pro achieves the highest task scores on the other four benchmarks, at higher total cost.

Thank you again for your constructive feedback. We hope the clarifications and additional comparisons above address your concerns, and we welcome any further questions.

Sincerely,
The Authors

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer oc3J,

Thank you again for your constructive comments. Our response includes experiments with noisy retrieval, iterative debate, and longer workflows (Response 1), expanded minimal-edit comparisons and strict final-answer controls (Response 2), and comparisons of selective gating across detector configurations, including detection and replacement costs (Response 3).

As the discussion period draws to a close, we would appreciate your feedback on whether these comparisons address your main concerns. Please let us know if any points need further clarification.

Sincerely,
The Authors

Reviewer_M8b5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates a two-node draft to review handoff. The receiver gets task evidence along with an upstream message, which is either hidden, shown, or conclusion-reversed. Across five benchmarks and five receivers, the authors find that the message helps when the receiver would otherwise fail, and hurts when the upstream is wrong. Up to 32% of items the receiver can solve on its own flip from correct to wrong. In 94% of the 82 audited flips, the receiver copies the upstream's specific wrong answer. Several controls address source labels, delayed receipt, a matched second call, and minimal-edit reversal. Oracle-gated message removal plus receiver replacement recovers part of the loss.

Critical strengths
  • The paired, item-level design is sound: the evidence is held fixed, the message is manipulated, and c→w is measured against the same receiver's evidence-only answer. The matched-hidden control (Table 27) and the minimal-edit control (Table 24) are useful.
  • Many alternative explanations are tested, and the authors are candid about where tests lack power (e.g., LBM source-label test).
  • The cross-receiver replication is broad.
Critical weaknesses

1. The prompt makes "adopt the draft" the default action, which confounds the central claim. The VoI framing assumes "free disposal": the receiver should be able to ignore the message. But the prompts do not frame the task as independent answering with optional advice. On BIRD the receiver is told "Another agent … produced a draft … Produce the final query. The draft may be wrong; correct it if needed." On HotpotQA it is told "Review your answer considering all available information. If you want to change it, output the revised answer. Otherwise output your original answer", even though in the shown condition the receiver has never produced "its answer". The teammate's answer is therefore the only candidate on the table. Under this framing, keeping the upstream answer is the instructed default, and "substitution" may measure, in large part, compliance with a draft-editing instruction rather than an inability to use evidence.

2. Many "harmful" flips may be label ambiguity, not errors. The gold labels in HotpotQA, MuSiQue, and 2Wiki are known to be noisy and often admit multiple defensible answers, and token-F1 ≥ 0.5 scoring penalizes them. Three of the six traces in Appendix A.17, chosen as illustrations, involve ambiguous questions:

  • Trace 2: "Canary Islands" vs. "Macaronesia". The evidence itself states that Tenerife is the most populous island of both.
  • Trace 1: "Which director is from the US, Wayne Wang or W. D. Richter?" Wang is described as an "American film director".
  • Trace 4: "spouse of the leading lady", which can mean the real-life or the in-film spouse.

3. Novelty relative to existing conformity/sycophancy/knowledge-conflict work is thin. The qualitative finding is well established: LLMs adopt confidently stated wrong answers from context, peers, or users even when they could answer correctly. Qu et al. (2026), "Easier to mislead than to correct", is very close. The paper's differentiator is the "fixed evidence + paired manipulation" design. But several headline results are expected by construction and take up much of the main text:

  • Γ > 0 is a "sanity check" that the authors themselves call unsurprising.
  • "Corrupting a correct conclusion lowers accuracy" follows almost directly from the conformity literature.
  • "Messages help when the receiver would fail" is close to a tautology.

The paper would be stronger if it clearly stated what new, non-obvious quantitative insight the design yields. Two examples: when free disposal holds versus fails, as a function of receiver strength relative to upstream; or a predictive model of c→w from observable features such as receiver confidence, evidence-message conflict, and answer type.

Questions for authors
  1. In the main design, is c→w computed against the k=3 evidence-only majority, and the shown condition against a single run? Please report the hidden-vs-hidden churn baseline for all 25 cells, not only the HotpotQA subset in Table 27, and subtract it.
  2. How were the 82 audited c→w items chosen? Why only five cells when per-item upstream answers exist for 21 cells?
  3. Does the receiver in the HotpotQA "shown" condition ever see its own earlier answer? If not, what does "Review your answer" refer to?
8 comments
AuthorsOpenReview ↗

Response to Reviewer M8b5 (1/5)

Dear Reviewer M8b5,

Thank you for your time and for recognizing the paired design, matched hidden controls, and cross-receiver validation. Your suggestions have guided the clarifications and additional analyses below on prompt design, semantic correctness, quantitative insights, baseline correction, and audit coverage.

1. Does the error increase persist under independent answering with optional advice?

Weakness 1

The prompt makes "adopt the draft" the default action, which confounds the central claim. The VoI framing assumes "free disposal": the receiver should be able to ignore the message. But the prompts do not frame the task as independent answering with optional advice. On BIRD the receiver is told "Another agent … produced a draft … Produce the final query. The draft may be wrong; correct it if needed." On HotpotQA it is told "Review your answer considering all available information. If you want to change it, output the revised answer. Otherwise output your original answer", even though in the shown condition the receiver has never produced "its answer". The teammate's answer is therefore the only candidate on the table. Under this framing, keeping the upstream answer is the instructed default, and "substitution" may measure, in large part, compliance with a draft-editing instruction rather than an inability to use evidence.

Question 3

Does the receiver in the HotpotQA "shown" condition ever see its own earlier answer? If not, what does "Review your answer" refer to?

Thank you for pointing out the problem in our prompt description. We apologize for the confusion caused by mixing prompt templates from different stages when preparing the manuscript. In the main HotpotQA shown condition, the receiver did not see its own earlier answer. “Review your answer” belongs to the matched two-stage control (Table 27), where both hidden and shown reviews receive the same actual initial answer. We have corrected Appendix A.26 to present the main-experiment and two-stage prompts separately.

In the two-stage setting, the review instruction allows the receiver to retain its own initial answer or revise it after considering the upstream message. Following your suggestion, we added a direct-answering control with optional advice and compared it with two-stage review on the same questions, evidence, and upstream messages. Cells report automatically scored errors/outputs for GPT-4o-mini.

DatasetTwo-stage: hiddenTwo-stage: shownDirect: hiddenDirect: shown
HotpotQA9/20727/20711/20735/207
BIRD1/10531/1051/10558/105
LBMuSiQue12/5115/5113/5121/51
2Wiki1/485/481/486/48
DROP1/727/722/7210/72

Observed error counts increase in both procedures on all five benchmarks. The direct-answering control removes the draft-editing instruction: the receiver answers from the evidence, and the upstream answer is presented only as optional advice. Thus, the same directional pattern appears when keeping the upstream answer is no longer the instructed default.

2. Do harmful flips reflect label ambiguity or genuine errors?

Weakness 2

Many "harmful" flips may be label ambiguity, not errors. The gold labels in HotpotQA, MuSiQue, and 2Wiki are known to be noisy and often admit multiple defensible answers, and token-F1 ≥ 0.5 scoring penalizes them. Three of the six traces in Appendix A.17, chosen as illustrations, involve ambiguous questions:

  • Trace 2: "Canary Islands" vs. "Macaronesia". The evidence itself states that Tenerife is the most populous island of both.
  • Trace 1: "Which director is from the US, Wayne Wang or W. D. Richter?" Wang is described as an "American film director".
  • Trace 4: "spouse of the leading lady", which can mean the real-life or the in-film spouse.

We appreciate your close reading of the illustrative traces and your concern about label ambiguity. Most audited harmful flips remain correct-to-wrong (c→w) transitions after reassessment against the full evidence. The three traces you identify are ambiguous; we have replaced these examples in Appendix A.17.

One human annotator blindly reassessed all 82 original records against the question and full evidence, accepting defensible answers supported by that evidence. GPT-6-Astra separately applied the same criteria to 152 additional records from GPT-4o-mini across three datasets.

AuthorsOpenReview ↗

Response to Reviewer M8b5 (2/5)

The table reclassifies the automatically scored c→w records by semantic correctness. “Both correct” means both answers are supported by the evidence; “Unresolved” means at least one cannot be judged because of ambiguity or insufficient evidence.

RecordsReviewTotalBoth correctCorrect→wrongWrong→correctUnresolved
Original archiveSingle human annotator8237504
HotpotQAGPT-6-Astra8347414
LBMuSiQueGPT-6-Astra3432713
2WikiGPT-6-Astra3542713
Three-dataset totalGPT-6-Astra15211128310

Human review confirms c→w in 75/82 audited records (91.5%); model review confirms it in 128/152 additional audited records (84.2%). Label ambiguity and valid answer variations therefore account for some flips, but not most of the reviewed records. Of the 75 human-confirmed transitions, 70 (93.3%) adopt the upstream model’s specific wrong answer, which itself contradicts the evidence.

The following CoT records show how the receiver’s answer changes when the message is shown:

Denying an award recorded in the evidence. The question asks what the singer of Moon River in Breakfast at Tiffany’s won a Tony Award for. The evidence identifies Audrey Hepburn and her 1954 award for Ondine. DeepSeek-V3.2 answers Ondine with the message hidden. With it shown, the model still identifies Hepburn but claims that no Tony Award is recorded and repeats the upstream answer: “Audrey Hepburn did not win a Tony for ‘Moon River’ or any other performance.”

Switching to the other co-creator’s birth year. Asked for the birth year of the news anchor who co-created an award with Sagarika Ghose, Kimi-K2.6 answers 1965 for Rajdeep Sardesai both initially and in the hidden review. With the message shown, its reasoning still gives Sardesai’s birth date as 24 May 1965, but then treats Ghose as the requested person and adopts the upstream answer, 1964.

3. What new, non-obvious quantitative insight does the design yield?

Weakness 3

Novelty relative to existing conformity/sycophancy/knowledge-conflict work is thin. The qualitative finding is well established: LLMs adopt confidently stated wrong answers from context, peers, or users even when they could answer correctly. Qu et al. (2026), "Easier to mislead than to correct", is very close. The paper's differentiator is the "fixed evidence + paired manipulation" design. But several headline results are expected by construction and take up much of the main text:

  • Γ > 0 is a "sanity check" that the authors themselves call unsurprising.
  • "Corrupting a correct conclusion lowers accuracy" follows almost directly from the conformity literature.
  • "Messages help when the receiver would fail" is close to a tautology.

The paper would be stronger if it clearly stated what new, non-obvious quantitative insight the design yields. Two examples: when free disposal holds versus fails, as a function of receiver strength relative to upstream; or a predictive model of c → w from observable features such as receiver confidence, evidence-message conflict, and answer type.

Thank you for encouraging us to clarify the quantitative insight yielded by the design. The original submission quantifies the risk of substitution at a handoff: receivers that can answer correctly on their own instead adopt the upstream model’s specific wrong answer while their independent evidence remains available. Paired item-level comparisons with fixed evidence, together with interventions on the message conclusion, link these answer changes to message content. It also measures how much accuracy can be recovered through message removal and receiver replacement when upstream errors are known.

Qu et al. [1] examine how peer consensus and authority cues affect harmful and beneficial revision. Weng et al. [2] examine conformity across interaction protocols, different numbers of prior discussion rounds, and majority sizes.

Building on the original findings, we followed your suggestions to examine receiver strength relative to the upstream model and item-level prediction of c→w. With the sender model, questions, evidence, and erroneous messages held fixed across receivers, higher independent accuracy accompanies lower excess error on BIRD, but the relationship is not monotonic on HotpotQA. Independent accuracy alone therefore does not fully characterize the risk of excess error in the tested configurations. Precommunication features selected on development questions and frozen before evaluation also improve c→w prediction over the base model on new questions across all five benchmarks. These signals help distinguish which initially correct answers are more likely to become wrong after the message is shown. We report these two comparisons below.

AuthorsOpenReview ↗

Response to Reviewer M8b5 (3/5)

3.1. Receiver strength relative to upstream

We appreciate your suggestion to test free disposal as a function of receiver strength relative to upstream. Higher independent accuracy accompanies lower excess error on BIRD, but the relationship is not monotonic on HotpotQA.

We fix GPT-4o-mini as the sender and compare receivers on the same evidence and erroneous upstream messages, with permission to ignore advice. Excess error is the shown-minus-hidden error rate, compared on the common qualified questions with complete paired outputs.

Receiver: independent accuracyBIRDHotpotQA
GPT-4o-mini53/99 (53.5%)72/85 (84.7%)
Kimi-K2.656/99 (56.6%)70/85 (82.4%)
Qwen3.6-Plus59/99 (59.6%)76/85 (89.4%)
GLM-5.261/99 (61.6%)79/85 (92.9%)
Receiver: excess error, pp [95% CI]BIRDHotpotQA
GPT-4o-mini+52.6 [31.6, 73.7]+12.7 [5.5, 21.8]
Kimi-K2.6+31.6 [5.3, 57.9]+9.1 [0.0, 18.2]
Qwen3.6-Plus+10.5 [0.0, 26.3]+18.2 [9.1, 29.1]
GLM-5.2+5.3 [0.0, 15.8]+3.6 [0.0, 9.1]
Common eligible paired questions1955

On BIRD, higher independent accuracy accompanies lower excess error. GLM-5.2 has 47.4 percentage points [21.1, 73.7] less excess error than GPT-4o-mini; the HotpotQA difference is 9.1 percentage points [0.0, 18.2]. On HotpotQA, however, excess error is not monotonic in independent accuracy: Qwen3.6-Plus is more accurate independently than GPT-4o-mini but incurs greater excess error. Positive excess error corresponds to a failure of free disposal; every receiver’s point estimate is positive on both benchmarks. Independent accuracy alone therefore does not fully characterize the risk of excess error under an erroneous upstream message.

A capability advantage also does not eliminate answer substitution. On GLM-5.2’s own qualification set, distinct from the common set above, six of 53 BIRD questions with complete paired outputs change from correct hidden execution to a shown result matching the upstream error.

3.2. A predictive model of c→w from observable features

Thank you for suggesting a predictive model of c→w from observable features. It helped us sharpen the quantitative contribution by testing whether information available before the message is shown can predict a correct-to-wrong change. We separate feature selection on development questions from evaluation on new questions.

We first compare four features on development questions: the three you suggest (receiver confidence, evidence-message conflict, and answer type), and agreement among three independent receiver answers. We select the two features with the largest ΔAUROC for each benchmark. All features are available before the message is shown; each ΔAUROC is measured relative to the same base covariates: receiver, message source, and text length.

Feature comparison and selection on development questions.

Added feature(s): ΔAUROC [95% CI]BIRDHotpotQAMuSiQue2WikiDROP
Development questions200807980384
Confidence+0.0396 [-0.0094, 0.0824]+0.0757 [0.0117, 0.1634]+0.1085 [0.0474, 0.1754]+0.0630 [0.0102, 0.1332]+0.0250 [0.0083, 0.0480]
Evidence–message conflict+0.0292 [-0.0261, 0.0698]+0.0872 [-0.0501, 0.2160]+0.1244 [0.0016, 0.2479]-0.0108 [-0.0831, 0.0617]-0.0009 [-0.0513, 0.0561]
Answer type-0.0109 [-0.0657, 0.0309]+0.0181 [-0.0126, 0.0535]+0.0045 [-0.0779, 0.0809]-0.0336 [-0.0833, 0.0125]-0.0521 [-0.1099, -0.0038]
Agreement+0.0350 [0.0046, 0.0548]+0.0485 [0.0000, 0.1266]+0.0213 [-0.0319, 0.0724]+0.0017 [-0.0320, 0.0363]+0.0029 [-0.0004, 0.0073]
Selected feature pairAgreement + ConfidenceEvidence–message conflict + ConfidenceEvidence–message conflict + ConfidenceAgreement + ConfidenceAgreement + Confidence
Selected joint ΔAUROC vs base+0.0605 [0.0083, 0.1106]+0.1111 [-0.0267, 0.2477]+0.1709 [0.0434, 0.3001]+0.0643 [-0.0007, 0.1395]+0.0278 [0.0097, 0.0511]

Receiver confidence gives the largest single-feature ΔAUROC on BIRD, 2Wiki, and DROP; evidence-message conflict gives the largest on HotpotQA and MuSiQue. The two highest-ranked features yield agreement + confidence for BIRD, 2Wiki, and DROP, and evidence-message conflict + confidence for HotpotQA and MuSiQue.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.