Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Cooper: Coordinating Specialized Agents towards a Complex Dialogue Goal
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 Cooper, a framework where specialized agents each track and promote one aspect of a complex dialogue goal, with a learned global coordinator ranking their topic candidates to guide the next utterance, evaluated on ESConv and P4G.
Problem
Complex dialogue goals (emotional support, persuasion) are hard to measure objectively and hard to optimize directly; LLMs lack strategic awareness of overall dialogue progression across multiple turns.
Approach
Decompose a complex dialogue goal into interdependent aspects, assign each to a specialized agent that tracks its state, estimates progression via clustered target-state embeddings, and proposes topic candidates; a learned ranker coordinates candidates globally to guide utterance generation.
Key results
- On ESConv, Cooper (PT-G) with gpt-3.5-turbo achieves higher BLEU-1 (17.62 vs 17.16) and Distinct-2 (29.45 vs 25.53) than GPT-3.5 in static evaluation.
- Cooper (FT-G) reaches BLEU-2 9.54 and Distinct-2 24.22 on ESConv, above MultiESC (9.19 and 16.70).
- In interactive human evaluation, Cooper (PT-G) beats GPT-3.5 on Naturalness (78.5% win vs 10.0% lose) and Empathy (67.7% win vs 19.2% lose).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use specialized agents and global coordination instead of GPT-3.5 alone | Cooper (PT-G) · ESConv | Dist-1: 4.50 -> 5.22 | Helped |
| Use specialized agents and global coordination instead of GPT-3.5 alone | Cooper (PT-G) · P4G | Dist-1: 4.50 -> 5.28 | Helped |
| Use a fine-tuned generator with the Cooper framework instead of MultiESC | Cooper (FT-G) · ESConv | BL-1: 21.30 -> 22.76 | Helped |
| Use a fine-tuned generator with the Cooper framework instead of ARDM | Cooper (FT-G) · P4G | BL-1: 21.17 -> 23.88 | Helped |
| Add global coordination to rank and filter topic candidates (ablation) | Cooper (FT-G) · ESConv | BL-1: 19.73 -> 22.76 | Helped |
| Add progression signals to topic candidate ranking (ablation) | Cooper (FT-G) · ESConv | BL-1: 21.11 -> 22.76 | Helped |
| Add topic candidate generation (ablation) | Cooper (FT-G) · ESConv | BL-1: 20.51 -> 22.76 | Helped |
| Add state summaries to the generator input (ablation) | Cooper (FT-G) · ESConv | BL-1: 20.07 -> 22.76 | Helped |
| Use the prompt-based generator rather than the fine-tuned generator | Cooper · ESConv interactive evaluation | Natural win rate: 75.4% | Helped |
Surprising findings
- Prompted chain-of-thought (GPT-3.5+CoT) and MixInit perform worse than plain GPT-3.5 on static metrics.
- GPT-3.5 is only marginally better than random at selecting topic candidates.
Limitations
- Persuasion evaluation limited to static metrics; interactive evaluation only on ESConv.
- Relies on pseudo-labeling of topic rankings from strategy annotations and text similarity.
- Interactive evaluation uses ChatGPT as simulated seeker and five annotators.
Experimental setup
Datasets: ESConv P4G
Benchmarks: ESConv P4G
Metrics: BLEU-1 BLEU-2 BLEU-4 ROUGE-L METEOR Distinct-1 Distinct-2 Distinct-3 Precision@n win/lose/tie human preference
Baselines: GPT-3.5 GPT-3.5+CoT MixInit MultiESC KEMI ARDM ProAware
Builds on: MPNet BART BERT chain-of-thought prompting MultiESC
Closest papers by meaning
- Evaluating Generalization Capabilities of LLM-Based Agents in Mixed-Motive Scenarios Using Concordia (NeurIPS 2025 · similarity 0.86)
- COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.86)
- More Capable, Less Cooperative? When LLMs Fail at Zero-Cost Collaboration (ICML 2026 · similarity 0.86)
- Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action (NeurIPS 2026 · similarity 0.85)
- LLawCo: Learning Laws of Cooperation for Modeling Embodied Multi-Agent Behavior (ICML 2026 · similarity 0.85)
- Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs (ICML 2024 · similarity 0.85)
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.