ANTMAN: Adaptive Need Tracking for Multi-Agent Navigation in Large Information Spaces

Jerry Wang; Haibo Jin; Xiaopeng Yuan; Peng Kuang; Haohan Wang

cs.AISep 27, 2026

ANTMAN: Adaptive Need Tracking for Multi-Agent Navigation in Large Information Spaces

Authors: Jerry Wang, Haibo Jin, Xiaopeng Yuan, Peng Kuang, Haohan Wang

Organizations: University of Illinois Urbana-Champaign

Abstract

Information-seeking agents increasingly operate over information spaces that are too large to process exhaustively. Yet many multi-agent systems organize computation around static partitions of the available space, causing coordination to grow with how information is segmented rather than with what the query still requires. We introduce ANTMAN, an adaptive coordination framework that treats evolving unresolved information needs as the unit of runtime coordination. ANTMAN maintains a revisable Need Graph that tracks unresolved requirements, accumulated evidence, prior attempts, and search progress, and uses this state to control worker selection, routing, and task-local recovery as new evidence is discovered. By separating the coordination policy from substrate-specific search interfaces, the same need-conditioned mechanism can operate across different information spaces. Experiments across multi-document question answering, controlled long-context scaling, and realistic structured navigation show that ANTMAN remains effective across settings, including when execution is delegated to substantially smaller worker models. Under a 16x increase in searchable context, ANTMAN increases active coordination by only 1.23x, compared with more than 15x for partition-driven baselines, while preserving strong answer quality.

Figures & tables

Method	HotpotQA		2Wiki		MuSiQue		Avg.
Method	EM	F1	EM	F1	EM	F1	EM	F1
Retrieval and RAG Baselines
Iterative Sparse Retrieval (BM25)	.533	.684	.367	.515	.433	.582	.444	.594
Dense Retrieval (DPR) ( Karpukhin et al., 2020 ) EMNLP’20	.533	.696	.267	.379	.200	.292	.333	.456
ChainRAG ( Zhu et al., 2025 ) ACL’25	.567	.711	.733	.791	.500	.654	.600	.719
S2G-RAG ( Li et al., 2026 ) ACL’26	.667	.810	.800	.858	.633	.730	.700	.799

Table 2: Performance on standard multi-document QA. We report EM and token-level F1 over 30 questions per benchmark, with macro averages across the three benchmarks.

Required Evidence Units	Active Coordination	Calls / Query	Cost / Query	Complete Evidence	Answer Acc.
ANTMAN
1	5.0	35.5	$0.107	100%	100%
4	4.3	43.6	$0.111	100%	100%
16	8.0	86.2	$0.293	100%	100%

Table 3: Coordination under increasing information demand. Searchable space is fixed at 512K while required evidence increases.

Method	RepoProbe	SWE-QA-Pro	GAIA-Text-103
Method	RepoProbe	SWE-QA-Pro	L1	L2	L3	Overall
Retrieval and Iterative Retrieval Baselines
Direct	31.08	55.74	28.2	17.3	8.3	20.4
Iterative Sparse Retrieval	33.15	60.90	53.8	42.3	33.3	45.6
Dense Retrieval ( Karpukhin et al., 2020 )	31.51	58.30	43.6	26.9	25.0	33.0
S2G-RAG ( Li et al., 2026 ) ACL’26	22.87	60.40	25.6	19.2	0.0	19.4

Table 4: Performance in realistic structured information spaces. GAIA-Text-103 is broken down by difficulty level. Raw repository scores are reported in Appendix C .

Variant	512K F1	SWE-QA-Pro Score	GAIA Accuracy
Adaptive Coordination Ablations
Static ANTMAN	40.32 $\downarrow 52.02\%$	68.58 $\downarrow 15.78\%$	50.00 $\downarrow 16.67\%$
Graph-free Adaptive	64.01 $\downarrow 23.82\%$	78.88 $\downarrow 3.13\%$	50.00 $\downarrow 16.67\%$
w/o Need Revision	74.44 $\downarrow 11.41\%$	75.32 $\downarrow 7.50\%$	42.50 $\downarrow 29.17\%$
w/o Adaptive Rerouting	82.20 $\downarrow 2.18\%$	72.13 $\downarrow 11.42\%$	60.00 ( $\leftrightarrow$ 0.00%)
w/o Recovery	60.12 $\downarrow 28.45\%$	76.52 $\downarrow 6.03\%$	52.50 $\downarrow 12.50\%$

Table 5: Ablation of runtime adaptive coordination. Parentheses indicate relative quality drops from full ANTMAN.

Method	32K			64K			128K			Middle Gap
Method	E	M	L	E	M	L	E	M	L	Middle Gap
Direct Long-Context Reference
Full Context	.911	.590	.813	.767	.667	.733	.733	.613	.870	$-.181$
Ours
ANTMAN	.741	.783	.800	.747	.826	.780	.743	.748	.813	+.015

Table 6: Robustness to evidence position as the information space grows. We report F1 when required evidence appears at Early (E), Middle (M), or Late (L) positions. The final column reports

M-\frac{1}{2}(E+L)

after averaging each position across context lengths; negative values indicate a middle-position penalty.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Coord. Growth	Calls Growth	Cost Growth	$\Delta$ F1
Multi-Agent Baselines
LongAgent ( Zhao et al., 2024 )	15.29 $\times$	12.81 $\times$	14.22 $\times$	$+.023$
CoA ( Zhang et al., 2024 )	15.33 $\times$	11.30 $\times$	16.11 $\times$	$-.004$
Ours
ANTMAN	1.23 $\times$	1.21 $\times$	6.90 $\times$	$+.065$

Appendix

Table 7: Endpoint scaling from 32K to 512K. Growth factors compare the smallest and largest information spaces in the controlled evaluation.

\Delta

F1 reports the corresponding absolute change in answer quality.

Method	Length	Active Coord.	Calls / Query	Cost / Query	EM	F1
Multi-Agent Baselines
LongAgent ( Zhao et al., 2024 )	32K	19.00	22.83	$0.0983	.533	.731
	64K	37.00	39.50	$0.1797	.533	.737
	128K	73.10	75.60	$0.3549	.500	.670
	512K	290.50	292.50	$1.3974	.567	.754
CoA ( Zhang et al., 2024 )	32K	5.10	7.10	$0.0809	.567	.750

Appendix

Table 8: Complete results for controlled information-space scaling. Active coordination denotes the average number of participating workers or members per query. Calls and cost are averaged per query. Lower values are better for coordination, calls, and cost; higher values are better for EM and F1.

Method	RepoProbe-Python / 10	SWE-QA-Pro / 50
Retrieval and Iterative Retrieval Baselines
Direct	3.108	27.87
Iterative Sparse Retrieval	3.315	30.45
Dense Retrieval ( Karpukhin et al., 2020 )	3.151	29.15
S2G-RAG ( Li et al., 2026 )	2.287	30.20
Agentic and Repository-Aware Baselines

Appendix

Table 9: Raw scores for realistic structured navigation. RepoProbe-Python scores are reported out of 10 and SWE-QA-Pro scores out of 50.

Variant	Explicit Need State	Need Revision	Adaptive Rerouting	Recovery
Static ANTMAN	✓	$\times$	$\times$	$\times$
Graph-free Adaptive	$\times$	–	✓	✓
w/o Need Revision	✓	$\times$	✓	✓
w/o Adaptive Rerouting	✓	✓	$\times$	✓
w/o Recovery	✓	✓	✓	$\times$
Full ANTMAN	✓	✓	✓	✓

Appendix

Table 10: Definition of state and adaptive-coordination ablations. Checkmarks indicate enabled mechanisms; “–” denotes a mechanism that is not applicable without an explicit Need Graph.

Variant	Answer Quality	Active Coordination	Cost ($/query)	LLM Calls (/query)
Adaptive Coordination Ablations
Static ANTMAN	40.32	1.97	0.156	28.7
Graph-free Adaptive	64.01	3.20	0.354	20.7
w/o Need Revision	74.44	3.10	0.418	21.0
w/o Adaptive Rerouting	82.20	2.00	0.451	28.2
w/o Recovery	60.12	2.07	0.353	25.9

Appendix

Table 11: 512K controlled scaling. Answer quality is F1 score; cost and calls are per query.

Variant	Answer Quality	Active Coordination	Cost ($/query)	LLM Calls (/query)
Adaptive Coordination Ablations
Static ANTMAN	68.58	1.35	0.273	36.5
Graph-free Adaptive	78.88	2.45	0.165	22.1
w/o Need Revision	75.32	3.50	0.298	32.4
w/o Adaptive Rerouting	72.13	2.20	0.407	48.6
w/o Recovery	76.52	2.15	0.522	56.5

Appendix

Table 12: SWE-QA-Pro ablation. Results are reported on a 40-question subset sampled with seed 42. Answer quality is the direct score; cost and calls are per query.

Variant	Answer Quality	Cost ($/query)	LLM Calls (/query)
Adaptive Coordination Ablations
Static ANTMAN	50.00	0.196	42.8
Graph-free Adaptive	50.00	0.179	54.0
w/o Need Revision	42.50	0.115	28.7
w/o Adaptive Rerouting	60.00	0.221	47.4
w/o Recovery	52.50	0.325	69.5

Appendix

Table 13: GAIA ablation (40 questions). Answer quality is accuracy; cost and calls are per query.

Setting	Information access
Controlled long context	Lexical and semantic retrieval with bounded reading of selected document regions.
Repository navigation	Retrieval and bounded reading augmented with structural navigation over symbols, imports, references, inheritance, and call relations.
GAIA-Text-103	Web search, webpage text extraction, and isolated Python execution for tool-assisted information seeking.

Appendix

Table 14: Information-access capabilities used across evaluation substrates.

Required Evidence	Worker Activity	Relevant Activity	Irrelevant Activity
ANTMAN
1 unit	5.0	0.9	4.1
4 units	4.3	3.2	1.1
16 units	8.2	8.0	0.2

Appendix

Table 15: Trajectory diagnostics under increasing information demand. Worker activity is decomposed according to whether it contributes to required evidence acquisition.

Method	LLM Calls	Tool Calls	Cost / Query	RepoProbe Score
Retrieval and Iterative Retrieval Baselines
Direct	1.00	0.00	$0.0067	31.08
Iterative Sparse Retrieval	3.19	2.16	$0.0217	33.15
Dense Retrieval ( Karpukhin et al., 2020 )	1.00	1.00	$0.0124	31.51
S2G-RAG ( Li et al., 2026 )	37.12	24.00	$0.1788	22.87
Agentic and Repository-Aware Baselines

Appendix

Table 16: Execution diagnostics on RepoProbe-Python. LLM calls, tool calls, and generation cost are averaged per query. RepoProbe-Python scores are reported on the same 0–100 scale as Table 4 .

Explore similar work

MANTA: Multi-Agent Network Topology Adaptation for Self-Evolving Multi-Agent Systems
Jul 30, 2026Mao-xun Huang, Jerry Wang, Yi-Cheng Lai +3Model-Based Multi-Agent Systems Topology
Department of Computer Science, Cornell University, United States · Department of Computer Science, University of Illinois Urbana-Champaign, United States · Institute of Information Science, Academia Sinica, Taiwan
Multi-Agent Coordination Adaptation via Structure-Guided Orchestration
May 25, 2026Haoran Li, Shulun Chen, Shaoyuan Sun +1Multi-Agent Coordination Coordination
Nanjing University · University of Technology Sydney, Sydney, Australia · University of New South Wales, Sydney, Australia
Focus Is All You Need: Adaptive Goal-aware Attention Orchestration for Multi-Agent Graph Systems
Jul 26, 2026Mingzhou Fan, Siyuan Xu, Mingxuan YuanReasoning Paths

Jul 30, 2026, cs.AIJ/K move · Enter open · S save

MANTA: Multi-Agent Network Topology Adaptation for Self-Evolving Multi-Agent Systems

Mao-xun Huang, Jerry Wang, Yi-Cheng Lai, Zhengxin Zhang+2

Department of Computer Science, Cornell University, United States · Department of Computer Science, University of Illinois Urbana-Champaign, United States · Institute of Information Science, Academia Sinica, Taiwan

Model-Based Multi-Agent Systems Topology

Open paper AskPDF arXiv

Large language model-based multi-agent systems improve complex problem solving through task decomposition, agent specialization, information exchange, and intermediate validation. However, existing systems typically treat communication topology as a fixed design choice or an offline optimization target. We introduce MANTA, a framework for Multi-Agent Network Topology Adaptation that enables communication structures to self-evolve at inference time. Before execution, MANTA initializes a task-conditioned topology from prior structural experience. During deployment, it monitors collaboration traces and applies bounded structural updates when the current organization becomes insufficient. These updates can modify agent roles, communication links, execution order, information visibility, and validation pathways while preserving the task interface and agent budget. We evaluate MANTA against representative single-agent and multi-agent baselines on five benchmarks spanning information seeking, tool use, planning, workflow execution, and mathematical reasoning. MANTA achieves the highest average score of 74.0, outperforming the strongest baseline by 5.8 percentage points and obtaining the best result on PlanCraft. These results show that inference-time self-improvement can extend to the architecture of collaboration itself.

Abstract

Explore similar work

MANTA: Multi-Agent Network Topology Adaptation for Self-Evolving Multi-Agent Systems

Multi-Agent Coordination Adaptation via Structure-Guided Orchestration

Focus Is All You Need: Adaptive Goal-aware Attention Orchestration for Multi-Agent Graph Systems