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

Yaxin Gong; Gangyi Zhang; Chongming Gao; Leyang Shen; Chenxiao Fan; Jiakai Wang; Dong Wang; Yang Liu

cs.AISep 29, 2026

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

Authors: Yaxin Gong, Gangyi Zhang, Chongming Gao, Leyang Shen, Chenxiao Fan, Jiakai Wang, Dong Wang, Yang Liu, +2 more

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

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.

Figures & tables

Figure 1: An erroneous peer message overrides an otherwise correct, evidence-supported answer.

Figure 2: Evidence–message interaction, conditional message value, and upstream model variation.

Figure 3: Override rates across 25 cells and per-benchmark transition breakdown.

Figure 4: Cross-model conclusion reversal. (a) Corrupting a correct upstream conclusion universally lowers accuracy (11/11 cells). (b) Correcting a wrong conclusion universally restores it (11/11 cells). Gray dots = original message; colored dots = after reversal.

Figure 5: Source-label control, delayed-receipt control, and CoT override rates.

Figure 6: Error-location interventions and gated recovery actions.

Component	BIRD ( $n{=}92$ )	LBM ( $n{=}70$ )	L2W ( $n{=}16$ )
A: Message removal	$+10.9$	$+6.7$	$+1.7$
B: Receiver replacement	$+10.9$	$+24.1$	$+22.3$
C: Total	$+21.7$	$+30.8$	$+24.0$

Table 1: Three-way decomposition on upstream-wrong items (paired bootstrap,

B{=}10{,}000

). On upstream-correct items, message removal is harmful (Appendix Table 9 ).

Appendix figures & tables29 assets

Supplementary material from the paper’s appendix.

Appendix

Benchmark	Task type	$n$	Upstream err.	Indep. evidence	Metric
BIRD ( Li et al., 2023 )	SQL generation	150	61.3%	schema desc.	exec. acc.
L2W ( Ho et al., 2020 )	Multi-hop QA	120	13.3%	retrieved pass.	token F1
LBM ( Trivedi et al., 2022 )	Multi-hop QA	160	43.8%	retrieved pass.	token F1
HotpotQA ( Yang et al., 2018 )	Multi-hop QA	200	50.0%	gold paragraphs	token F1
DROP ( Dua et al., 2019 )	Reading comp.	120	18.3%	passage + question	exact match

Appendix

Table 2: Benchmark overview. “Upstream err.” is the fraction of items where the upstream agent (gpt-4o-mini) produces an incorrect answer.

Receiver can solve without message?	$n$	Message value with evidence	Interpretation
Yes (score $>0.5$ )	18	$\mathbf{-22.8}$ pp	Substitution
No (score $\leq 0.5$ )	69	$+19.1$ pp	Reasoning helps
All upstream-wrong items	87	$+10.5$ pp	—

Appendix

Table 3: LBM items where kimi-k2.6 gives an incorrect upstream answer. Positive message value with independent evidence comes from items the receiver cannot solve alone; on items it can solve, the wrong conclusion causes harm.

Behavior	BIRD	LBM	HQA	DROP
Copies upstream error	84	66	38	86
Ignores error, correct	7	4	37	0
Can solve, still follows	9	6	23	9
Cannot solve, diverges	0	21	0	5
Diverges, incorrect	0	3	0	0
Diverges, correct	0	0	2	0

Appendix

Table 4: Behavioral classification for the gpt-4o-mini receiver (%). Percentages are rounded independently; “Follows upstream error” combines unconditional copying with cases where the receiver can solve alone but still follows. Column sample sizes: BIRD 92, LBM 70, HQA 100, DROP 22.

Condition	Shown (%)	Hidden (%)	$\tau$ pp
Without independent evidence	86.0	14.2	$+71.8$
With independent evidence	86.8	63.6	$+23.2$
$\Gamma$			$+48.6$	$p<0.0001$

Appendix

Table 5: DROP four-cell decomposition (gpt-4o-mini receiver,

n

=120).

\tau

= message value (shown

-

hidden accuracy).

\Gamma=\tau(\text{without evidence})-\tau(\text{with evidence})

. DROP shows the largest

\Gamma

among all five benchmarks. Unlike HotpotQA and LBM,

\tau(\text{with evidence})

remains positive (

+23.2

pp): the message is net-helpful even with evidence. Independent evidence sharply reduces the marginal value of the message.

Receiver	With evidence		No evidence		$\tau$ (w/o ev.)	$\tau$ (w/ ev.)	$\Gamma$ [95% CI]	c $\to$ w
	Shown	Hidden	Shown	Hidden
gpt-4o-mini	86.8	63.6	86.0	14.2	$+71.8$	$+23.2$	$+48.6$	5
deepseek-v3.2	90.8	70.8	90.0	14.2	$+75.8$	$+20.0$	$+55.8$ [45,66]	3
kimi-k2.6	90.8	83.3	90.8	21.7	$+69.2$	$+7.5$	$+61.7$ [52,71]	7
glm-5	92.5	90.8	90.8	20.0	$+70.8$	$+1.7$	$+69.2$ [61,78]	2
qwen3.6-plus	92.5	94.2	90.0	24.2	$+65.8$	$-1.7$	$+67.5$ [58,76]	6

Appendix

Table 6: DROP cross-receiver four-cell decomposition (

n

=120 per receiver). Accuracy is exact-match percentage.

\Gamma

is significantly positive for all five receivers (

p<0.0001

); 95% CIs from paired bootstrap (

B

=10,000).

Benchmark	Precision (%)	Recall (%)	Items flagged
BIRD ( $n$ =150)	91.8	60.9	61
LBM ( $n$ =160)	81.1	42.9	37
L2W ( $n$ =120)	35.0	43.8	20

Appendix

Table 7: Error-detector operating characteristics (gemini-2.5-pro). Precision = fraction of flagged items that are truly upstream-wrong; recall = fraction of upstream-wrong items that are flagged. L2W has low precision because most upstream answers are correct (86.7%), so even modest false-positive rates dominate.

Intervention	BIRD ( $n$ =150, fl.=61)		LBM ( $n$ =160, fl.=37)		L2W ( $n$ =120, fl.=20)
Intervention	Acc	95% CI	Acc	95% CI	Acc	95% CI
Default (no detection)	.400	[.320, .480]	.539	[.467, .609]	.742	[.671, .807]
No message (same receiver)	.460	[.380, .540]	.553	[.482, .624]	.730	[.656, .801]
Rerun (same receiver)	.447	[.367, .527]	.553	[.481, .623]	.755	[.686, .819]
CoVe	.407	[.327, .487]	.535	[.462, .606]	.746	[.675, .813]
LLM-judge repair	.407	[.327, .487]	.542	[.470, .614]	.746	[.676, .813]

Appendix

Table 8: Recovery experiment: gated accuracy under different interventions with gemini-2.5-pro as the shared detector. “No message (different family)” uses a receiver from another model family without the upstream message. “Answer, then review message” lets the receiver answer independently before reviewing the peer message.

Component	BIRD ( $n{=}150$ )	LBM ( $n{=}160$ )	L2W ( $n{=}120$ )
A: Message removal	$-1.3$	$-7.7$	$-21.8$
B: Receiver replacement	$+9.7$	$+17.4$	$+14.8$
C: Total	$+8.3$	$+9.7$	$-7.0$

Appendix

Table 9: Three-way decomposition on all with-evidence items (unconditional on upstream correctness; paired bootstrap,

B{=}10{,}000

). Component A is negative when most upstream answers are correct, because removing the message discards correct signals. On the upstream-wrong target population, Component A is non-negative on all three benchmarks (Table 1 ).

C=A+B

by construction.

Receiver	Family	Solo acc. (%)	$\Delta$ vs gpt ( pp)	Recovery ( pp)	$p$
Same-family replacements (OpenAI):
gpt-4.1-mini	OpenAI	44.0	$+0.7$	$+17.4$	$<0.001$
gpt-4o	OpenAI	48.7	$+5.4$	$+21.7$	$<0.001$
Cross-family replacements (capability-matched, $\|\Delta\|\leq 5$ pp):
deepseek-v3.2	DeepSeek	40.7	$-2.7$	$+9.8$	$0.022$
glm-4.5-air	Zhipu	42.0	$-1.3$	$+17.4$	$<0.001$

Appendix

Table 10: Recovery by receiver identity on BIRD (

n

=92 upstream-wrong items, paired bootstrap). Solo accuracy is computed on 150 BIRD items without any peer message in an independent run. Recovery = paired accuracy gain over the default gpt-4o-mini receiver with the erroneous message; the default receiver’s same-model removal baseline is

+10.9

pp(Table 1 ). Same-family OpenAI replacements achieve recovery comparable to cross-family receivers at similar capability levels (

p=0.50

for same-family vs. cross-family mean difference), indicating that recovery does not require family diversity—any different model suffices.

Receiver	Bench	$n$	$\tau$ (w/o ev.) pp	$\tau$ (w/ ev.) pp	$\Gamma$ pp	95% CI
gpt-4o-mini	BIRD	150	$+26.7$	$+1.3$	$+25.3$	$[+17.3,+33.3]$
gpt-4o-mini	LBM	160	$+37.9$	$+7.7$	$+30.2$	$[+22.1,+38.6]$
gpt-4o-mini	L2W	120	$+46.6$	$+21.8$	$+24.8$	$[+15.2,+34.4]$
gpt-4o-mini	HQA	200	$+34.3$	$-5.7$	$+40.0$	$[+33.7,+46.6]$
deepseek-v3.2	BIRD	150	$+27.3$	$-4.0$	$+31.3$	$[+22.7,+40.0]$
deepseek-v3.2	LBM	160	$+38.3$	$+10.5$	$+27.8$	$[+20.7,+34.8]$

Appendix

Table 11: Full 20-cell

\Gamma

table.

\tau

(w/o ev.) = message value without independent evidence;

\tau

(w/ ev.) = message value with evidence;

\Gamma=\tau\text{(w/o\,ev.)}-\tau\text{(w/\,ev.)}

. All CIs from paired bootstrap (

B=10{,}000

, seed 42). All 20 cells show

\Gamma>0

(

p<0.0001

): independent evidence consistently reduces the marginal value of the upstream message.

Receiver	Benchmark	$n$	$\tau$ (w/ ev.) pp	95% CI	$p$
glm-5	BIRD	150	$-12.7$	$[-19.3,-6.7]$	${<}.001$	$\ast\ast\ast$
kimi-k2.6	BIRD	150	$-12.0$	$[-18.7,-5.3]$	${<}.001$	$\ast\ast\ast$
qwen3.6-plus	BIRD	150	$-4.7$	$[-11.3,+2.0]$	.191
deepseek-v3.2	BIRD	150	$-4.0$	$[-11.3,+3.3]$	.319
gpt-4o-mini	BIRD	150	$+1.3$	$[-5.3,+8.0]$	.769
kimi-k2.6	LBM	160	$-10.2$	$[-16.0,-4.4]$	${<}.001$	$\ast\ast\ast$

Appendix

Table 12: Paired bootstrap 95% CIs for

\tau

(with evidence) (

B=10{,}000

resamples, seed 42). Of 20 cells, 13 are significantly negative at

\alpha=0.05

; 6 survive Bonferroni correction (

\alpha=0.0025

for 20 tests).

\ast

p<.05

;

\ast\ast

p<.01

;

\ast\ast\ast

p<.001

Receiver	Bench	Bypass correct	$P(\text{c}{\to}\text{w})$	Bypass wrong	$P(\text{w}{\to}\text{c})$
qwen3.6-plus	BIRD	96	16.7%	54	16.7%
glm-5	BIRD	87	26.4%	63	6.3%
kimi-k2.6	BIRD	75	32.0%	75	8.0%
deepseek-v3.2	BIRD	70	25.7%	80	15.0%
gpt-4o-mini	BIRD	58	20.7%	92	15.2%
qwen3.6-plus	LBM	121	10.7%	39	12.8%

Appendix

Table 13: Per-item transitions under

\tau

(with evidence).

P(\text{c}{\to}\text{w})

: fraction of items correct without the message that become wrong with it.

P(\text{w}{\to}\text{c})

: the reverse. On HQA, all five receivers show non-trivial correct-to-wrong rates (5.7–13.2%).

Receiver	Bench	$n_{s}$	c $\to$ w	$\tau_{s}$ pp	95% CI	$n_{f}$	w $\to$ c
gpt-4o-mini	BIRD	58	12	$-20.7$	$[-31.0,-10.3]$	92	14
gpt-4o-mini	LBM	77	9	$-11.7$	$[-19.5,-5.2]$	83	23
gpt-4o-mini	L2W	67	1	$-1.5$	$[-4.5,0.0]$	53	29
gpt-4o-mini	HQA	159	21	$-13.2$	$[-18.9,-8.2]$	41	8
deepseek-v3.2	BIRD	70	18	$-25.7$	$[-35.7,-15.7]$	80	12
deepseek-v3.2	LBM	76	11	$-14.5$	$[-22.4,-6.6]$	84	28

Appendix

Table 14: Conditional decomposition of

\tau

(with evidence).

n_{s}

= items solvable without message (

k{=}3

majority vote);

\tau_{s}

= message value on solvable items (always

\leq 0

). CIs from bootstrap (

B=10{,}000

). All c

\to

w transitions across the primary 20 cells: 274 out of 2,184 solvable items (12.5% overall). Including DROP, the total is 297/2,667 (11.1%). In 18 of 20 primary cells, the c

\to

w rate exceeds 5%.

Upstream	Benchmark	Ups. corr.	$\tau$ (w/ ev.) pp	95% CI	$\Gamma$ pp
gpt-4o-mini	BIRD	38.7%	$+1.3$	$[-5.3,+8.0]$	$+25.3$
gpt-4o-mini	LBM	56.2%	$+7.7$	$[+1.3,+14.2]$	$+30.2$
gpt-5.4	HQA	84.5%	$+2.3$	$[-1.1,+5.9]$	$+43.1$
gpt-5.4	LBM	66.9%	$+29.3$	$[+21.8,+36.9]$	$+21.4$
kimi-k2.6	HQA	76.0%	$+1.0$	$[-2.5,+4.5]$	$+42.6$
kimi-k2.6	LBM	45.6%	$+20.4$	$[+13.5,+27.3]$	$+22.5$

Appendix

Table 15: Upstream model diversity: message value with evidence and interaction.

\tau

(w/ ev.) = message value with evidence;

\Gamma=\tau

(w/o ev.)

\,-\,\tau

(w/ ev.). All

\Gamma

values significant (

p<0.0001

\tau

(w/ ev.) is non-negative in all cells.

Figure 7: Conclusion-reversal experiment on items with an incorrect upstream answer (three QA benchmarks). Points show the accuracy difference between messages with corrected and original peer conclusions, with 95% bootstrap CIs. All six contrasts are significant (BIRD and DROP omitted; see text). A compact version appears in the main text as Figure 5 (a).

	Upstream wrong		Upstream correct
Benchmark	$\Delta$	95% CI	$\Delta$	95% CI
LBM	$+42.2$	$[31.6,\;53.1]$	$-39.4$	$[-49.3,\;-29.6]$
HotpotQA	$+22.7$	$[15.1,\;30.4]$	$-35.9$	$[-44.8,\;-27.6]$
L2W	$+46.0$	$[25.0,\;67.9]$	$-72.1$	$[-81.4,\;-61.9]$

Appendix

Table 16: Effects of reversing the peer conclusion, with 95% bootstrap CIs (20K draws). All six contrasts on the three QA benchmarks are significant (BIRD and DROP rows omitted due to SQL-execution scoring and source-wrong label divergence; raw values retained in comments). “Upstream wrong”: corrected

-

original; “Upstream correct”: corrupted

-

original.

Benchmark	Shown	Hidden	$\tau$	Shown	Hidden	$\tau$
	Without evidence			With evidence
BIRD	31.3	4.7	$+26.7$	40.0	38.7	$+1.3$
LBM	52.5	14.6	$+37.9$	53.9	46.2	$+7.7$
L2W	72.3	25.7	$+46.6$	74.1	52.4	$+21.8$
HQA	67.4	33.1	$+34.3$	67.2	72.9	$-5.7$

Appendix

Table 17: Accuracy (%) and message value

\tau

for four benchmarks (gpt-4o-mini receiver). Within each evidence condition,

\tau

is the paired per-item difference between message-shown and message-hidden accuracy. Classification uses

k{=}3

majority vote (Appendix A.20 ). DROP is reported separately in Table 5 .

Benchmark	$n$	Original (chars)	Reversed (chars)	Ratio
BIRD	150	$1167\pm 400$	$1221\pm 447$	1.05
LBM	160	$649\pm 175$	$667\pm 185$	1.03
DROP	120	$487\pm 130$	$461\pm 129$	0.95
L2W	120	$844\pm 248$	$874\pm 252$	1.04
HotpotQA	200	$1241\pm 315$	$921\pm 410$	0.74

Appendix

Table 18: Character length of original messages and messages with reversed conclusions (mean

\pm

std).

Bench	Upstream	$n$	No msg	Orig	Rev	Rev $-$ Orig	Rev $-$ No msg
LBM	Wrong	70	11.0	8.0	50.2	$+42.2$	$+39.2$
LBM	Correct	90	68.2	90.3	50.9	$-39.4$	$-17.3$
HQA	Wrong	100	57.4	37.0	66.7	$+29.7$	$+9.3$
HQA	Correct	100	90.8	96.0	57.0	$-39.0$	$-33.8$
L2W	Wrong	16	17.1	18.5	64.5	$+46.0$	$+47.4$
L2W	Correct	104	54.0	82.7	10.6	$-72.1$	$-43.4$

Appendix

Table 19: Accuracy (%) under three message conditions. “Rev” = reversed conclusion (correcting wrong or corrupting correct). All three conditions are from the same run; “Rev

-

Orig” is computed within this run. For bootstrapped CIs on paired effect sizes from the main experiment, see Table 16 . DROP and L2W have small source-wrong samples (

n

=22 and

n

=16). L2W source-wrong values (

n

=16) are mean token-F1

\times

100, not binary accuracy.

Benchmark	$n$	Coherent	Evidence-grounded	No artifacts
BIRD	150	99.3%	100.0%	98.0%
LBM	160	90.0%	98.8%	100.0%
HQA	200	96.3%	100.0%	100.0%
DROP	120	93.3%	100.0%	100.0%
L2W	120	92.5%	100.0%	100.0%
All	750	94.4%	99.7%	99.6%

Appendix

Table 20: LLM-judge semantic audit of all 750 messages with reversed conclusions. Overall, 93.8% pass all three criteria. The 5.6% flagged as incoherent are conservative: an incoherent message is harder for the receiver to adopt, working against our finding that reversing conclusions shifts accuracy.

Benchmark	$n$	Token-F1	ROUGE-L	Citation preservation
BIRD	150	0.73	0.68	50%
LBM	475	0.62	0.50	71%
HotpotQA	457	0.60	0.44	63%

Appendix

Table 21: Token-level overlap between original and reversed messages (filtered to messages

\geq

20 tokens).

n

counts message pairs across all receiver models that generated flips on each benchmark, hence exceeding per-benchmark item counts. Citation preservation measures the fraction of named entities and numbers shared between versions.

	Natural-error override		Injected-conflict override
Receiver	No CoT	CoT	No CoT	CoT
gpt-4o-mini	100% (9/9)	100% (9/9)	89% (8/9)	67% (6/9)
kimi-k2.6	93% (14/15)	56% (10/18)	93% (14/15)	67% (12/18)
deepseek-v3.2	—	—	90%	73%

Appendix

Table 23: CoT override rates on LBM. Parentheses show override count / eligible items. For kimi’s natural-error condition, the paired rate on the 11 items qualifying under both CoT conditions is 91%

\to

45%. Deepseek has no natural-error override cases on LBM in this CoT experiment (its 11 c

\to

w items on LBM in the main design, Table 14 , do not fall in the natural-error CoT-eligible subset).

Benchmark	Receiver	$n$	c $\to$ w	McNemar $p$	$\Delta$ F1
LBM	gpt-4o-mini	48	12	$<0.001$	$-22.3$ pp
LBM	deepseek-v3.2	51	13	$<0.001$	$-21.7$ pp
LBM	kimi-k2.6	49	3	$0.250$	$-3.0$ pp
HQA	gpt-4o-mini	81	9	$0.004$	$-8.9$ pp
HQA	deepseek-v3.2	80	10	$0.002$	$-11.5$ pp
HQA	kimi-k2.6	83	2	$0.500$	$-1.5$ pp

Appendix

Table 24: Minimal-edit conclusion reversal across three receivers and three benchmarks.

n

= number of items answered correctly in an independent evidence-only response before either message condition, with valid minimal-edit messages (

\geq 0.90

overlap). c

\to

w = items that flip from correct (with the original message) to wrong (with the minimal edit).

\Delta

F1 = mean F1 difference (minimal-flip

-

original message).

Benchmark	Receiver	$n$	$a$	$b$
LBM	gpt-4o-mini	48	36	12
LBM	deepseek-v3.2	51	38	13
LBM	kimi-k2.6	49	46	3
HQA	gpt-4o-mini	81	72	9
HQA	deepseek-v3.2	80	70	10
HQA	kimi-k2.6	83	81	2

Appendix

Table 25: Full

2{\times}2

paired contingency tables for the minimal-edit control.

a

= correct under both messages;

b

= correct with original, wrong with minimal edit;

c

= wrong with original, correct with minimal edit;

d

= wrong under both. Items are selected by independent evidence-only correctness before either message condition. Because such items nearly always remain correct under the original (source-correct) message,

c{=}0

and

d{=}0

in every cell are structural consequences of this selection, not design constraints. Binarization threshold: token-F1

\geq 0.5

Receiver	$\tau$ pp	95% CI	c $\to$ w / w $\to$ c	$\tau$ pp	c $\to$ w	w $\to$ c	McNemar $p$
	Full sample			Upstream-wrong items
gpt-4o-mini	$-4.0$	$[-6.3,\;-1.7]$	34 / 14	$-10.4$	34	5	$<0.001$
deepseek-v3.2	$-0.0$	$[-3.4,\;+3.3]$	9 / 5	$-5.5$	9	0	$0.004$
kimi-k2.6	$-1.2$	$[-3.4,\;+0.7]$	4 / 1	$-3.1$	4	0	$0.125$

Appendix

Table 26: Matched-review control on HotpotQA (

n=400

for gpt-4o-mini,

n=200

for deepseek and kimi). Both message-shown and message-hidden branches use the same two-step review prompt; the hidden branch receives a neutral placeholder instead of the peer message.

\tau

= paired per-item accuracy difference (shown

-

hidden). On upstream-wrong items, c

\to

w = items correct under hidden but wrong under shown; w

\to

c = reverse. The gpt-4o-mini result is significant on the full sample (McNemar

p=0.006

) and highly significant on upstream-wrong items (

p<0.001

). Deepseek is also significant (

p=0.004

). Kimi shows the same directional pattern but lacks power.

Receiver	Condition	$n$	c $\to$ w	Rate
gpt-4o-mini	Matched hidden	64	1	1.6%
	Teammate	64	17	26.6%
	Unverified tool	64	16	25.0%
	Unlabeled	64	18	28.1%
	Evidence-priority	64	16	25.0%
deepseek-v3.2	Matched hidden	64	1	1.6%

Appendix

Table 27: Source-label attribution control on HotpotQA (items where the receiver is independently correct and the upstream answer is wrong). Every labeled condition produces more correct-to-wrong transitions than the hidden baseline; no pairwise comparison between labeled conditions is significant. Pooling across receivers, all labeled-vs-hidden comparisons are significant (Fisher exact

p<0.001

). At the individual-receiver level, all comparisons are significant for gpt-4o-mini and deepseek (

p<0.05

); kimi’s low base rate (3–6 events out of 70) leaves most comparisons underpowered (teammate

p=0.12

, unlabeled

p=0.014

Dimension	Qu et al.	Cho et al.	Xie et al.	Ours
Setting	Multi-agent discussion	Simulated herd	Single-model context	Pipeline handoff
Evidence control	None	None	Parametric vs. context	Fixed gold evidence
Message manip.	Observe only	Majority injection	Context injection	Show/hide/reverse
Upstream errors	Natural	Simulated majority	Constructed	Natural
Causal granularity	Aggregate	Aggregate	Aggregate	Per-item paired
Trace analysis	No	No	No	60 annotated CoT

Appendix

Table 28: Experimental-design comparison with closest prior work. Qu et al. study conformity in multi-agent discussion without controlling receiver evidence; Cho et al. inject simulated majorities without per-item pairing; Xie et al. study parametric-vs-contextual conflicts, not conflicts between two external inputs. Our design fixes downstream evidence and manipulates the upstream message item by item, enabling causal attribution.

Paper name	Role	Provider	API identifier	Max tok.
gpt-4o-mini	Upstream/receiver	OpenAI	gpt-4o-mini	4096
gpt-5.4	Upstream/receiver	OpenAI	gpt-5.4-0305-global	4096
deepseek-v3.2	Receiver	DeepSeek	deepseek-v3.2	4096
kimi-k2.6	Upstream/receiver	Moonshot	kimi-k2.6	4096
glm-5	Receiver	ZhiPu	glm-5	4096
qwen3.6-plus	Receiver	Alibaba	qwen3.6-plus	4096

Appendix

Table 29: Model manifest. The conclusion editor generates messages with reversed peer conclusions; the detector identifies candidate errors in recovery experiments. All models were accessed in September 2026.

Explore similar work

May 19, 2026cs.MA

What Do Agents Communicate? Characterizing Information Exchange in Multi-Agent Systems

Large Language Models (LLMs) have enabled collaborative Multi-Agent (MA) systems, where interacting agents improve performance through diverse reasoning and iterative refinement. However, these systems remain vulnerable to error propagation, where early-stage information degrades downstream reasoning. To address this, we conduct a systematic analysis of inter-agent communication to identify which information drives MA performance. We find that the absence of reasoning and verification in inter-agent communication significantly degrades performance. Based on these insights, we propose Category-Aware Recovery Augmentation (technique), which enforces the presence of critical information during communication. recovers up to 86.2% of failed cases. Our results highlight the key role of information quality in effective MA collaboration. Our code is available at https://anonymous.4open.science/r/cara_mas

Yong Jin Chun, Iftekhar Ahmed

Department of Informatics University of California, Irvine Irvine, CA, 92618

Downstream Reasoning

Ask PDF

Jun 1, 2026cs.CL

Easier to Mislead Than to Correct: Harmful and Beneficial Revision in LLM Conformity

Large language models are increasingly used in multi-agent systems, where they see and respond to other agents' answers. A key risk is conformity: a model may abandon its own answer simply because others agree on a different one. Prior studies show that LLMs often revise toward a majority answer, but it remains unclear whether these revisions help correct mistakes as often as they introduce new errors. In this paper, we conduct a controlled study in which an LLM first answers a question, then sees simulated peer responses before making a final decision. We manipulate two social cues: consensus structure and authority labels assigned to peers, and measure how they influence beneficial and harmful revisions. Across four open-weight LLMs and seven QA datasets, we find that peer agreement makes it much easier to mislead initially correct models than to correct initially wrong ones. Authority labels make models more likely to choose the endorsed answer, regardless of whether it is correct. More concerningly, generic reasoning interventions such as chain-of-thought and reflection do not reliably reduce harmful revision while preserving beneficial revision. These findings suggest that multi-agent LLM systems should verify peer answers rather than simply aggregate them.

Jiaming Qu, Lucheng Fu, Yibo Hu

Amazon · Georgia Institute of Technology · Illinois Institute of Technology

Multi-Agent Large Language Model Systems Conformity

Ask PDF

Oct 1, 2026cs.AI

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

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

Shixuan Li, Wei Yang, Peiyu Zhang +3

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

Multi-Agent Communication Model Auditing Answer Reasoning Skills

Ask PDF