HAMLET: A Hierarchical and Adaptive Multi-Agent Framework for Live Embodied Theatrics¶

🕒 Published (v1): 2025-07-21 11:36 UTC · Source: Arxiv · Venue: ICLR 2026 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

HAMLET is a two-stage hierarchical multi-agent framework that converts a bare topic into a live, embodied theatrical performance: an offline planner synthesizes a narrative blueprint, then an online control system of specialized orchestrator agents drives improvisational actors through it in real time. Each actor runs a PAD (Perceive And Decide) module inspired by dual-process theory to decide when and how to act. A trained 8B critic model, HAMLETJudge, automates holistic drama quality evaluation.

Problem¶

Existing LLM-based drama systems are passive (turn-based, awaiting user instructions), require verbose inputs (full outlines or leading paragraphs), lack embodied physical interaction with scene environments, and have no evaluation framework for complete live theatrical performance — only for textual coherence or single-turn role-play fidelity.

Method¶

Offline Planning — Four agents (actor designer, plot designer, reviewer, director) collaborate in a pipeline: character profiles are generated and reviewed, a plot draft is structured by the director into a hierarchical narrative blueprint via backward planning (endpoint first, then causally preceding narrative points). The blueprint encodes acts, scenes, props, and narrative points with completion flags.

Online Performance — A three-agent control layer executes the blueprint: - Planner: decomposes each narrative point's flag into a beat-chain trajectory. - Transfer: polls flag satisfaction; triggers point transitions and actor entry/exit. - Advancer: detects stalls and injects directive nudges.

Actor agents use the PAD module, a two-stage pipeline: (1) given external stimulus (environment, dialogue history, interactable objects) and internal state (persona, memory via RAG, goal), PAD emits a strategy \(r \in \{\text{FAST}, \text{SLOW}, \text{SILENCE}\}\) plus a structured action triplet (subject, verb, object); (2) the LLM generates concrete dialogue/action conditioned on that strategy. A narrator agent adjudicates physical actions against environment state and broadcasts updates globally.

HAMLETJudge — an 8B model fine-tuned to perform pairwise drama comparisons on three dimensions: Character Performance (CP: believability, agency), Narrative Quality (NQ: coherence, resonance, integrity), and Interaction Experience (IE: responsiveness, immersion, fluency).

Key Contributions¶

Decoupled offline planning / online performance architecture requiring only a simple topic as input
PAD module: dual-process-theoretic pre-response mechanism for actor agents combining RAG memory, persona, and dynamic goal
Narrator agent for embodied environment interaction adjudication with global state broadcast
Backward-planning director for causally coherent narrative point generation
HAMLETJudge: 8B critic model for automated holistic drama evaluation via win-rate pairwise scoring
Benchmark dataset of 100 cases (50 literary excerpts, 50 custom topics; English and Chinese)

Results¶

Claude-4-sonnet-Thinking achieves the highest overall scores across both English and Chinese drama performance among all tested models
Qwen3-32B-Thinking (73.85 avg) significantly outperforms Llama-3.1-70B (45.75) and rivals Qwen3-235B (71.21), demonstrating reasoning efficiency over scale
Reasoning models (DeepSeek-R1, Claude-4-sonnet-Thinking) consistently outperform non-reasoning counterparts on all three evaluation dimensions
PAD and HAMLETJudge are both 8B models yet achieve state-of-the-art performance in their respective tasks (exact ablation numbers truncated in provided text)

Limitations¶

Full ablation numbers and comparison tables are partially truncated in the provided text; exact deltas for PAD and control-agent ablations are not reproducible from what is given
Evaluation relies heavily on LLM-as-judge (GPT-4o baseline + HAMLETJudge); human ground-truth agreement is not quantified in the provided excerpt
Embodied interaction is simulated via narrator adjudication, not grounded in a real physics engine or robotic actuator
The framework's scalability to very large casts or very long multi-act performances is untested
Dependency on GPT-4o as the backbone for most agents limits cost and latency in production settings

Relevance to Harnesses / Meta-Harnesses¶

HAMLET is a textbook meta-harness: it wraps multiple specialized agents (actor designer, plot designer, reviewer, director, planner, transfer, advancer, narrator, HAMLETJudge) in a two-stage orchestration pipeline where the offline phase produces structured control artifacts (the blueprint) that govern online agent behavior. The PAD module is itself a sub-harness — a pre-response decision layer that decouples strategic reasoning from language generation, a pattern directly analogous to planner–executor splits in general agentic harnesses. The three-agent online control layer (planner/transfer/advancer) provides a concrete example of hierarchical harness decomposition: macro-level narrative management separated from micro-level beat execution. Researchers tracking harness design patterns will find the backward-planning blueprint generation, the narrator-as-environment-arbiter role, and the trained critic-as-evaluator particularly transferable outside the drama domain.