Research map / Multi-agent systems and games / LLM agents
Autonomous Agents for Collaborative Task under Information Asymmetry
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes iAgents, a multi-agent paradigm where agents mirror human social networks and proactively exchange private human information via the InfoNav planning mechanism and a mixed (clear + fuzzy) memory, evaluated on the new InformativeBench.
Problem
Multi-person collaborative tasks where each LLM agent only sees its own human user's information, creating information asymmetry that existing shared-context LLM multi-agent systems cannot handle.
Approach
Mirror the human social network in an agent network where agents autonomously retrieve and exchange private human information; InfoNav keeps an explicit natural-language plan of unknown/known rationales that navigates communication, and a mixed memory (SQL-style clear memory plus embedding-based fuzzy memory) supplies the information; agents can also spawn recursive sub-conversations with neighbor
Key results
- GPT-4 backend reaches 56.67% on Schedule Easy, 64.00% on NP, and 57.94% on FriendsTV, with an average of about 50.48% across tasks per the paper; hardest Schedule Hard reaches 22.8%.
- Removing InfoNav drops Schedule Easy from 36.67% to 10.00% and Medium from 18.00% to 3.56% (GPT-3.5 backend), and NP from 51% to 39%.
- Removing recursive communication lowers FriendsTV accuracy from 35.71% to 23.02%, the largest ablation gain on that dataset.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add InfoNav (ablation) | iAgents with GPT-3.5 · Schedule Easy | precision: 10.00% -> 36.67% | Helped |
| add InfoNav (ablation) | iAgents with GPT-3.5 · Schedule Medium | F1: 3.56% -> 18.00% | Helped |
| add InfoNav (ablation) | iAgents with GPT-3.5 · Schedule Hard | IoU: 7.34% -> 12.25% | Helped |
| add InfoNav (ablation) | iAgents with GPT-3.5 · Needle in the Persona (NP) | precision: 39.00% -> 51.00% | Helped |
| add InfoNav (ablation) | iAgents with GPT-3.5 · FriendsTV | precision: 34.92% -> 35.71% | Helped |
| add recursive communication (ablation) | iAgents with GPT-3.5 · Needle in the Persona (NP) | precision: 48.00% -> 51.00% | Helped |
| add recursive communication (ablation) | iAgents with GPT-3.5 · FriendsTV | precision: 23.02% -> 35.71% | Helped |
| add Fuzzy Memory (ablation) | iAgents with GPT-3.5 · FriendsTV | precision: 29.37% -> 35.71% | Helped |
| add Clear Memory (ablation) | iAgents with GPT-3.5 · FriendsTV | precision: 33.33% -> 35.71% | Helped |
| anonymize primary character names | iAgents with GPT-3.5 · FriendsTV | precision: 35.71% -> 32.54% | Hurt |
| emphasize privacy preservation in the system prompt | iAgents with GPT-3.5 · FriendsTV | precision: 35.71% -> 30.95% | Hurt |
Surprising findings
- Agents without InfoNav fail to exceed 10% accuracy on all Schedule difficulty levels.
- Recursive communication contributed more than mixed memory on FriendsTV, contrary to the expected primacy of retrieval; the paper attributes the ablation drop mainly to recursive communication (12.7 points).
Limitations
- Privacy: output-side privacy prompting lowers accuracy (35.71% to 30.95%) and full privacy is impractical; edge small models struggle at high privacy levels.
- High input token cost (~30,000 tokens per task).
- Network modeling is simplistic, ignoring collaboration history and nuanced topology; human verification still needed.
Experimental setup
Datasets: Needle in the Persona (built from SPC dataset) FriendsTV (built from Friends season 1 script and FriendsQA) Schedule (synthetic)
Benchmarks: InformativeBench (Needle in the Persona, Schedule Easy/Medium/Hard, FriendsTV)
Metrics: accuracy F1 IoU
Baselines: w/o InfoNav w/o Recursive Comm w/o Fuzzy Memory w/o Clear Memory
Builds on: ChatDev MetaGPT Dialogue Status Tracking
Closest papers by meaning
- Systematic Failures in Collective Reasoning under Distributed Information in Multi-Agent LLMs (ICML 2026 · similarity 0.90)
- Emergent Coordination in Multi-Agent Language Models (ICLR 2026 · similarity 0.89)
- More Capable, Less Cooperative? When LLMs Fail at Zero-Cost Collaboration (ICML 2026 · similarity 0.88)
- AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors (ICLR 2024 · similarity 0.88)
- Latent Collaboration in Multi-Agent Systems (ICML 2026 · similarity 0.88)
- COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.