ANTMAN: Adaptive Need Tracking for Multi-Agent Navigation in Large Information Spaces
Jerry Wang, Haibo Jin, Xiaopeng Yuan, Peng Kuang, Haohan Wang
cs.AI
2026-09-27
ANTMAN dispatches workers from a revisable Need Graph. At 16× context, coordination grows 1.23× vs over 15× for partition baselines; 8B workers keep most of the full-model score.
Long-context multi-agent systems often hire workers the way they slice the input. LongAgent gives each fixed-size chunk a member agent. Chain of Agents (CoA) walks those chunks in a relay. Grow the corpus and you grow the partition count, so team size follows the space, not the question. The evidence a query actually needs can stay small while coordination scales with how the files were cut. Even after relevant evidence is found, stretching the context can still hurt accuracy, and answers remain sensitive to where that evidence sits.
ANTMAN, from the University of Illinois Urbana-Champaign, treats unresolved information needs as the unit of runtime coordination. A revisable Need Graph, not the static partition map, decides who works.
The information space is deterministically split into territories. Each territory gets a WorkerCard, a short scope-and-content summary. That map is a candidate pool. It is not a roster that must all clock in. Workers need not use different models or toolkits; specialization is mostly about scope, and they share one search-inspect-return interface.
The Need Graph holds the live control state. A query is broken into initial needs; each node stores resolution status, gathered evidence, attempt history, and progress. Each step picks an unresolved need, routes it to a territory via WorkerCards, and has that worker do bounded local search. The worker returns a structured report. The coordinator then updates the graph: a need can close, be rewritten, grow a dependency, or spawn a new node. If search stalls, recovery rewrites the need, reroutes it, or takes a fallback path. Synthesis uses evidence attached to resolved nodes.
Active workers are bounded by min{M, ρ H(q)}, where M is the number of territories, H(q) is how many needs actually appear, and ρ is the per-need routing cap. Enlarge the space and M can grow without forcing more dispatches, as long as demand stays bounded. Routing itself is two-stage: a non-LLM ranker (lexical, exact-term, dense similarity, fused by reciprocal rank) shortlists WorkerCards, then the orchestrator decides whether and where to send work. The paper is explicit that this decoupling is about worker participation, not about retrieval cost being independent of corpus size.
Everyone uses GPT-4.1 except ANTMAN-H, which keeps that orchestrator and swaps non-orchestrator roles to Qwen3-8B.
On HotpotQA, 2WikiMultiHopQA, and MuSiQue (30 questions each), ANTMAN's macro EM/F1 is .711/.814, ahead of S2G-RAG at .700/.799, CoA at .567/.743, and LongAgent at .433/.616. ANTMAN-H lands at .622/.763. The 8B gap is mostly MuSiQue, where EM falls from .667 to .467.
In a Needle-in-a-Haystack PLUS setup (10 questions, Early/Middle/Late placement, 32K to 512K):
| Method | Coord. growth | Call growth | Cost growth | 512K F1 |
| LongAgent | 15.29× | 12.81× | 14.22× | .754 |
| CoA | 15.33× | 11.30× | 16.11× | .746 |
| ANTMAN | 1.23× | 1.21× | 6.90× | .840 |
At 32K ANTMAN uses 2.57 workers, 27.47 calls, $0.077; at 512K, 3.17 workers, 33.10 calls, $0.533. Headcount barely moves. Dollars still jump because each local search is heavier.
Hold space at 512K and raise required evidence from 1 to 16 units: workers 5.0→8.0, calls 35.5→86.2, cost $0.107→$0.293, with 100% evidence recovery and accuracy. At one required unit, 82% of worker actions miss the target evidence.
RepoProbe-Python has 108 questions over eight large Python repos. SWE-QA-Pro uses a frozen 80-question subset. GAIA-Text-103 is 103 text-only tool questions with Tavily search.
| Method | RepoProbe | SWE-QA-Pro | GAIA-Text-103 |
| Iterative Sparse | 33.15 | 60.90 | 45.6 |
| ReAct+RepoGraph | 25.71 | 67.38 | - |
| SWE-QA-Pro Agent | - | 77.58 | - |
| OWL | - | - | 40.8 |
| ANTMAN-H | 34.10 | 74.60 | 57.3 |
| ANTMAN | 38.21 | 79.78 | 58.3 |
Gains over the strongest non-benchmark-specific baselines are 15.3%, 18.4%, and 27.9%. ANTMAN-H keeps 89.2%, 93.5%, and 98.3% of full ANTMAN. The GAIA drop is concentrated at Level 3 (41.7 vs 25.0). Direct full-context inference has a middle-position gap of -.181; ANTMAN's is +.015.
Matched-budget ablations: graph-free adaptive replanning falls from 84.03 to 64.01 F1 at 512K (-23.82%) and from 60.00 to 50.00 on a 40-question GAIA slice. Frozen Static ANTMAN scores 40.32 at 512K. Killing recovery drops 512K F1 to 60.12; killing need revision drops GAIA to 42.50. On RepoProbe, ANTMAN spends $0.299 and 38.13 LLM calls per query; S2G-RAG uses 37.12 calls and scores 22.87. Call count does not explain the gap.
A long-context or repo agent can keep a large candidate worker pool and only wake people for what is still missing. At 512K, LongAgent runs 290 members and CoA 78 workers; ANTMAN uses a little over three. Swap workers to 8B and HotpotQA matches the full system; GAIA L1/L2 barely move. The expensive model can stay in the orchestrator.
The same Need Graph plugs into document retrieval, code-structure navigation, and web-plus-Python tools. That is an architecture choice, not a new backbone. Connecting adaptive retrieval and adaptive teaming through one runtime state is the portable idea.
The bill does not vanish. 512K still costs $0.533 per query. RepoProbe is about 45× the cost of a direct answer, moving the score from 31.08 to 38.21. Decoupling saves worker headcount. Token spend does not fall in lockstep.
There is no standalone Limitations section. The conclusion asks for richer substrates, more transferable need representations, and better revision and routing policies.
The evals are thin. Multi-document QA uses 30 questions per set; the scaling study uses 10 base questions. Demand scaling plants independent random values, so 100% recovery mainly shows the loop can find planted needles. GAIA and SWE-QA-Pro ablations shrink to 40-question subsets, so those absolute scores do not match the main tables.
The orchestrator is GPT-4.1 in every variant, including ANTMAN-H. Small workers can execute; the paper never shows a small model can coordinate. Appendix A.1 already notes that WorkerCard retrieval still grows with the territory count. At low demand, 82% of worker actions are off-target. GAIA has a single territory, so the rerouting ablation is a no-op. Cost grows 6.90× while calls grow 1.21×, and that gap is not decomposed.