Locating and Closing Information Gaps in Multi-Agent Systems

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

cs.AIOct 29, 2026

Locating and Closing Information Gaps in Multi-Agent Systems

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

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

Abstract

Multi-step Web tasks often require several LLM agents to gather and combine evidence. The communication graph of a fixed workflow restricts which agents can read each other's outputs: evidence retrieved by one agent may never reach the agent that needs it, leaving an information gap and an incomplete or incorrect answer. Adding a connection can deliver the missing evidence, but can also lead the receiver to adopt an incorrect upstream conclusion that displaces a correct answer; the final output alone does not separate the two. We propose Selective Probing and Link Insertion for Closing information gaps in multi-agent Execution (SPLICE), which locates information gaps per query and decides what to adopt after adding an edge. After the original workflow runs, SPLICE uses the query and node outputs to locate a possible gap, adds one directed edge from a source to a receiver to deliver the relevant evidence, and re-executes the receiver and affected downstream nodes, producing a candidate answer. The adoption decision reads both the final answers of the two runs and the process changes the edit caused inside the workflow: how far the receiver moved, where its new content came from, and whether the change reached the final answer. A selector trained on these process features chooses between adopting the candidate answer and retaining the original. On five benchmarks, SPLICE raises the task score on Web tool use by 8.0% and the mean score on four reasoning and planning benchmarks by 5.7% relative to the static skeleton, at 1.25 to 1.44 $\times$ the skeleton's cost, and transfers zero-shot to four unseen backbones.

Figures & tables

Figure 1. Comparison between existing fixed workflows (left) and our proposed SPLICE (right). Two stacked panels. The upper panel contrasts a workflow in which key upstream evidence is omitted, producing a false claim, with one in which an added edge delivers that evidence and the claim is corrected. The lower panel contrasts selecting between candidate outputs by answer text alone with selecting using the process evidence of each run.

Figure 2. Overview of SPLICE. The four sub-decisions are made in one timeline: diagnosis selects which queries are worth a trial and which candidate edge to open, targeted generation writes the message that edge carries, and the trial is then executed. Only afterwards does verification read the resulting execution and decide whether to adopt the candidate answer or retain the original $\hat{y}_{q}$ . The top half of the figure shows the core mechanism behind each step. A left-to-right pipeline. A base run is diagnosed to locate a missing delivery, a message is generated for the selected candidate edge, the modified workflow is executed as a trial, and the process evidence of that trial drives an adopt-or-retain decision between the candidate answer and the original.

	ToolBench ( $n{=}230$ )				Four benchmarks ( $n{=}2{,}142$ )				Total
Method	Micro	Served	Net %	Cost	MuSiQue	DROP	2Wiki	PlanCraft	Mean	Net %	Cost
Static	0.6347	0.3957	–	1.00 $\times$	0.4257	0.6734	0.7614	0.6988	0.6654	–	1.00 $\times$
Always	0.6514 $\uparrow$ .0167	0.3435 $\downarrow$ .0522	$+2.6$	1.75 $\times$	0.4184 $\downarrow$ .0073	0.7015 $\uparrow$ .0281	0.7649 $\uparrow$ .0035	0.6842 $\downarrow$ .0146	0.6725 $\uparrow$ .0071	$+2.0$	1.71 $\times$
Random	0.6384 $\uparrow$ .0037	0.3870 $\downarrow$ .0087	$+0.9$	1.67 $\times$	0.4343 $\uparrow$ .0086	0.7004 $\uparrow$ .0270	0.7607 $\downarrow$ .0007	0.6491 $\downarrow$ .0497	0.6678 $\uparrow$ .0024	$+1.9$	1.71 $\times$
Self-cons.	0.6282 $\downarrow$ .0065	0.3609 $\downarrow$ .0348	$-1.7$	3.00 $\times$	0.4346 $\uparrow$ .0089	0.6874 $\uparrow$ .0140	0.7627 $\uparrow$ .0013	0.6813 $\downarrow$ .0175	0.6692 $\uparrow$ .0038	$-0.1$	3.00 $\times$
Judge	0.6624 $\uparrow$ .0277	0.4304 $\uparrow$ .0348	$+6.1$	2.67 $\times$	0.4418 $\uparrow$ .0161	0.6985 $\uparrow$ .0251	0.7717 $\uparrow$ .0103	0.6871 $\downarrow$ .0117	0.6781 $\uparrow$ .0127	$+3.1$	1.88 $\times$

Table 1. End-to-end results on five benchmarks. Net is the per-question fix

-

harm rate (%). Best values per column are bolded .

Dataset	Task	Metric	Train	Test
ToolBench	Web tool use	Micro (Served)	535	230
MuSiQue	Multi-hop QA	Token F1	1,680	360
DROP	Discrete reasoning	Token F1	3,360	720
2WikiMultihopQA	Multi-hop QA	Token F1	3,360	720
PlanCraft	Embodied planning	Plan success	1,596	342

Table 2. Datasets and evaluation. Micro grades each required API by call and content; Served is the fraction of fully-scored requests; Token F1 and Plan success score the other four benchmarks against reference answers.

	Q1 (low)	Q2	Q3	Q4	Q5 (high)	All
Predicted	0.392	0.548	0.651	0.736	0.847	0.635
Actual	0.388	0.542	0.648	0.740	0.856	0.635
Trialled	100%	100%	50%	0%	0%	50%

Table 3. Query-level diagnosis calibration on ToolBench (

n{=}230

). Bins are quintiles of predicted base score

\hat{r}_{q}

Configuration	Score	$\Delta$	Net %	Cost
Static skeleton	0.6347	–	–	1.00 $\times$
w/o diagnosis	0.6813	$-.0043^{\scriptscriptstyle+.0073}_{\scriptscriptstyle-.0162}$	+7.8	1.90 $\times$
$\hat{r}$ = constant	0.6718	$-.0138^{\scriptscriptstyle-.0054}_{\scriptscriptstyle-.0296}$	$+6.1$	1.44 $\times$
$\hat{r}$ = parse failure	0.6648	$-.0208^{\scriptscriptstyle-.0146}_{\scriptscriptstyle-.0401}$	$+3.5$	1.44 $\times$
$\hat{r}$ = length	0.6642	$-.0214^{\scriptscriptstyle-.0147}_{\scriptscriptstyle-.0410}$	$+3.5$	1.44 $\times$
edge $=$ random	0.6654	$-.0202^{\scriptscriptstyle-.0137}_{\scriptscriptstyle-.0392}$	$+4.3$	1.44 $\times$

Table 4. Component ablation on ToolBench with the four locating layers varied in turn.

\Delta

is relative to Full (

b{=}0.5

); the intervals are 95% paired bootstrap CIs over questions.

Edge selection rule	Precision@1	$95\%$ CI
Random reachable edge	$0.41$	$[.34,\,.48]$
Terminal-distance (nearest)	$0.49$	$[.42,\,.56]$
Content coverage	$0.55$	$[.48,\,.62]$
$\rho$ (SPLICE)	$\mathbf{0.67}$	$[.60,\,.74]$

Table 5. Gap-hit precision@1 on the held-out ToolBench test set, with two judges supplying the gap labels and

\rho

compared against three heuristic edge-selection rules.

Condition	DeepSeek-v4-pro	Claude-opus-4-6
Source holds the fact	$0.88$ [.82,.93]	$0.86$ [.79,.91]
Receiver lacked the fact	$0.91$ [.85,.95]	$0.89$ [.83,.94]
Message delivered the fact	$0.85$ [.78,.90]	$0.83$ [.76,.89]
Revision used the fact	$0.82$ [.75,.88]	$0.80$ [.72,.86]
All four hold	$\mathbf{0.74}$ [.66,.81]	$\mathbf{0.71}$ [.63,.79]

Table 6. Gap-localisation audit on

n{=}120

adopted repairs, independently scored by two judges across four conditions whose joint rate is reported in the bold row.

Measure	Static	SPLICE
Fact-support rate	$0.71$ [.65,.77]	$\mathbf{0.83}$ [.78,.88]
Partially-supported rate	$0.11$ [.07,.15]	$0.08$ [.05,.12]
Unsupported-claim rate	$0.18$ [.13,.23]	$\mathbf{0.09}$ [.06,.13]
Task completion rate	$0.63$ [.57,.70]	$\mathbf{0.74}$ [.68,.80]

Table 7. Evidence support on the held-out ToolBench test set (

n{=}230

); the first three rows sum to one, and the judging protocol is detailed in Appendix A.10 .

Checkpoint	Archived observation
Node 0 return	API payload: $u^{\ast}$ + captcha PNG.
Terminal input	Parents (nodes 1–2): no payload and no $u^{\ast}$ .
Repair	Open $0\!\to\!3$ : deliver payload to terminal.
Answer	Static: “no API results”; SPLICE: $u^{\ast}$ + image.

Table 8. A held-out ToolBench case where SPLICE closes an API-payload gap: the UUID returned at node 0 reaches the terminal only after the added edge delivers it. Full archived exchange in Appendix A.8 .

Host	Bench	Static	Always	SPLICE	$p$
Chain	ToolBench	0.4222	0.4604 $\uparrow$ .0382	0.4832 $\uparrow$ .0610	0.0003
AgentPrune ( Zhang et al., 2025b )	ToolBench	0.4180	0.4428 $\uparrow$ .0248	0.4713 $\uparrow$ .0533	0.0003
G-Designer ( Zhang et al., 2025c )	ToolBench	0.3995	0.4212 $\uparrow$ .0217	0.4433 $\uparrow$ .0438	0.0008
Chain	2Wiki	0.6009	0.5942 $\downarrow$ .0067	0.6242 $\uparrow$ .0233	0.0828
AgentPrune	2Wiki	0.5616	0.5728 $\uparrow$ .0112	0.6076 $\uparrow$ .0460	0.0276
G-Designer	2Wiki	0.6625	0.6689 $\uparrow$ .0064	0.6833 $\uparrow$ .0208	0.0100

Table 9. Composability on three host workflows;

p

compares SPLICE against each host’s own static output.

Backbone	Static	Always	SPLICE	$95\%$ CI	$p$
Kimi-K2.6	0.7150	0.7436	0.7652 $\uparrow$ .0502	[.026, .076]	$1.3\!\times\!10^{-4}$
DeepSeek-V3.2 ( DeepSeek-AI et al., 2025 )	0.7245	0.7167	0.7758 $\uparrow$ .0513	[.030, .076]	$1.2\!\times\!10^{-5}$
Qwen3.7-Max	0.8355	0.8406	0.8673 $\uparrow$ .0317	[.014, .052]	$1.2\!\times\!10^{-3}$
Qwen3.6-Plus	0.7636	0.7714	0.7921 $\uparrow$ .0284	[.012, .047]	$1.7\!\times\!10^{-3}$

Table 10. Backbone transfer on DROP, 2Wiki and PlanCraft, with the selector fitted once on gpt-4o-mini and reused on four unseen backbones with no retraining.

Figure 3. Skeleton scan on DROP across three topology families (Chain, Tree, Dense) and three agent scales (4, 8, 12). Two rows of panels: top row shows task score for Chain, Tree and Dense topologies at 4, 8 and 12 agents, with SPLICE consistently above Static; bottom row shows the corresponding cost multiples.

Figure 4. Score and cost against the trial-budget fraction

b

on ToolBench and the four reasoning and planning benchmarks.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Skeleton	Adjacency list	$\|\mathcal{C}\|$
Main experiments
MuSiQue (most frequent)	1:{0}; 2:{1}; 3:{2}	3
DROP, 2Wiki	1:{0}; 3:{1,2}	3
PlanCraft, ToolBench	1:{0}; 2:{0}; 3:{1,2}	2
Skeleton scan (Figure 3 )
Chain-4	1:{0}; 2:{0}; 3:{1,2}	2

Appendix

Table 11. Skeletons as adjacency lists.

Configuration	Tokens / question	Multiple
Static skeleton	$4{,}509$	$1.00\times$
Diag. $+$ Sel., $b{=}0.3$	$5{,}059$	$1.12\times$
Diag. $+$ Sel., $b{=}0.5$	$5{,}426$	$1.20\times$
Nexa	$6{,}718$	$1.49\times$
Ours, $b{=}0.5$	$5{,}652$	$1.25\times$
Always / random / no selector	$7{,}998$	$1.77\times$

Appendix

Table 12. Token-level cost decomposition on the four reasoning and planning benchmarks (per-question token counts and multiples relative to the static skeleton).

Field	Definition	Range
Paired text
Message	content delivered along edge	text
Receiver before	receiver answer in $\tau_{q}$	text
Receiver after	receiver answer in $\tau_{q}(e,m)$	text
How far the receiver moved
Revised	receiver’s answer changed	{0, 1}

Appendix

Table 13. The process-evidence block

\phi_{q,e}

	ToolBench ( $n{=}230$ )		Four bench. ( $n{=}2{,}142$ )
vs.	$b{=}0.5$	Full	$b{=}0.5$	Full
Static	$+.0509^{\scriptscriptstyle+.0758}_{\scriptscriptstyle+.0280}$	$+.0466^{\scriptscriptstyle+.0725}_{\scriptscriptstyle+.0227}$	$+.0377^{\scriptscriptstyle+.0452}_{\scriptscriptstyle+.0304}$	$+.0447^{\scriptscriptstyle+.0524}_{\scriptscriptstyle+.0369}$
Always	$+.0342^{\scriptscriptstyle+.0554}_{\scriptscriptstyle+.0135}$	$+.0299^{\scriptscriptstyle+.0499}_{\scriptscriptstyle+.0101}$	$+.0307^{\scriptscriptstyle+.0377}_{\scriptscriptstyle+.0238}$	$+.0376^{\scriptscriptstyle+.0457}_{\scriptscriptstyle+.0297}$
Random	$+.0472^{\scriptscriptstyle+.0719}_{\scriptscriptstyle+.0215}$	$+.0429^{\scriptscriptstyle+.0678}_{\scriptscriptstyle+.0174}$	$+.0354^{\scriptscriptstyle+.0422}_{\scriptscriptstyle+.0290}$	$+.0423^{\scriptscriptstyle+.0489}_{\scriptscriptstyle+.0354}$
Self-cons.	$+.0574^{\scriptscriptstyle+.0849}_{\scriptscriptstyle+.0308}$	$+.0531^{\scriptscriptstyle+.0810}_{\scriptscriptstyle+.0265}$	$+.0339^{\scriptscriptstyle+.0414}_{\scriptscriptstyle+.0266}$	$+.0408^{\scriptscriptstyle+.0485}_{\scriptscriptstyle+.0332}$
Judge	$+.0232^{\scriptscriptstyle+.0394}_{\scriptscriptstyle+.0074}$	$+.0189^{\scriptscriptstyle+.0346}_{\scriptscriptstyle+.0062}$	$+.0250^{\scriptscriptstyle+.0308}_{\scriptscriptstyle+.0189}$	$+.0319^{\scriptscriptstyle+.0384}_{\scriptscriptstyle+.0253}$

Appendix

Table 14. Paired differences against every competitor.

LLM Agent Workflow Optimization

LLM agent workflow optimization studies how to design and run multi-step agents for better task success, reliability, latency, and cost.

See all 262 papers

Right Answers, Wrong States: Hidden Information Failures in Multi-Agent Collaboration
Oct 1, 2026Herun Wan, Jiaying Wu, Minnan Luo +4AI Agent Reliability Multi-Agent Collaboration
Xi’an Jiaotong University · National University of Singapore · Yunnan University +1
InfoSeeker: A Scalable Hierarchical Parallel Agent Framework for Web Information Seeking
Apr 3, 2026Ka Yiu Lee, Yuxuan Huang, Zhiyuan He +5Multi-Agent LLM Systems Web Search Agents
Huawei Noah’s Ark Lab · University of Liverpool · University College London
When Upstream Messages Override Correct Answers: A Controlled Study of Multi-Agent LLM Collaboration
Sep 29, 2026Yaxin Gong, Gangyi Zhang, Chongming Gao +7Multi-Agent LLM Systems Multi-Agent Collaboration
University of Science and Technology of China · Qwen Business Unit of Alibaba · National University of Singapore

Oct 1, 2026J/K move · Enter open · S save

Right Answers, Wrong States: Hidden Information Failures in Multi-Agent Collaboration

Herun Wan, Jiaying Wu, Minnan Luo, Zihan Ma+3

Xi’an Jiaotong University · National University of Singapore · Yunnan University+1

AI Agent Reliability Multi-Agent Collaboration

Open paper AskPDF arXiv

Multi-agent systems are often judged by whether they reach the correct answer. This can miss a distinct failure: collaboration may leave behind a corrupted information state even when the immediate decision is correct. We call this an off-query failure. To study this failure in collaborative decision support, we introduce OffQuery, which separately evaluates evidence verification (T1), shared-state reconstruction (T2), and task resolution (T3) in two representative high-stakes settings: healthcare and disaster response. Across GPT, Gemini, and Qwen models, standard collaboration shows much stronger task performance than state reliability. Averaged over 21 model--setting combinations, task resolution reaches 64.7%, while evidence verification and state reconstruction reach only 14.3% and 43.1%. We trace this gap to selective information use: current queries often bypass corrupted facts, which become consequential when later tasks require them. We further introduce ReGround, which resolves conflicting evidence, verifies shared facts, reconstructs a trusted state, and reasons over that state. Across seven models from three families, ReGround improves all three capabilities in every evaluated setting, with average relative gains of 309.0%, 82.9%, and 17.6% on T1, T2, and T3. Reliable collaboration therefore requires both a correct decision and a reliable shared state for future reasoning.

Abstract

LLM Agent Workflow Optimization

Explore similar work

Right Answers, Wrong States: Hidden Information Failures in Multi-Agent Collaboration

InfoSeeker: A Scalable Hierarchical Parallel Agent Framework for Web Information Seeking

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