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

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

ICLR 2026; HAMLET: hierarchical adaptive multi-agent framework for live interactive theater

Ask a follow-up

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

TL;DR¶

HAMLET is a hierarchical multi-agent framework for autonomous AI drama creation and live embodied performance, decoupling offline narrative blueprint generation (from a bare topic) from online improvisational enactment with physical environment interaction. A Perceive And Decide (PAD) module grounds actor agents in dual-process cognition (fast/slow/silence modes), enabling proactive rather than reactive behavior. A specialized 8B critic model, HAMLETJudge, enables cost-effective automated drama evaluation via pairwise 5-point Likert scoring.

Problem¶

Existing LLM-based drama systems are passive and reactive (agents await user turns rather than driving the narrative), lack embodied environment interaction (physical prop manipulation), and require heavy upfront input—full story outlines or elaborate guiding paragraphs—reducing flexibility and immersion. No established evaluation benchmark assesses complete live dramatic performance beyond textual coherence or single-character role-play fidelity.

Method¶

HAMLET operates in two decoupled stages:

Offline Planning — A four-agent collaborative workflow (actor designer, plot designer, reviewer, director) transforms a simple topic into a structured narrative blueprint. The director applies backward planning: generate the act's endpoint first, then retroactively construct logically coherent preceding narrative points. Each act specifies scenes (with enumerated interactive props and their absolute/relative positions) and ordered narrative points (each with a completion flag and expected result).

Online Performance — A hierarchical control layer (Planner, Transfer, Advancer) manages real-time execution. The Planner decomposes each narrative point's flag into a beat chain (sequences of actor actions). Transfer monitors flag satisfaction and advances the drama; Advancer intervenes if the plot stalls. A narrator agent adjudicates every physical environment interaction: it evaluates feasibility given current environment state and physics, confirms or rejects actions, updates global environment state, and broadcasts an objective description to all participants.

PAD Module — Each actor agent is backed by a Perceive And Decide module inspired by dual-process theory (Kahneman 2011 System I/II). PAD integrates an internal state (static: Persona, Subjective Relationships; dynamic: Goal, RAG-retrieved Memory) with external stimulus (Environment Description, Actor List, Dialogue History, Interactable Objects). It outputs a decision tuple \((r, o)\) where \(r \in \{\text{FAST}, \text{SLOW}, \text{SILENCE}\}\) and \(o\) is a structured action triplet (subject, verb, object) or \(\emptyset\). An optional internal monologue guides the actor's final concrete utterance/action.

HAMLETJudge — An 8B critic model fine-tuned for drama evaluation. It performs pairwise comparisons against a GPT-4o baseline and scores on a 5-point Likert scale across three dimensions: Character Performance (CP: Believability, Agency), Narrative Quality (NQ: Coherence, Resonance, Integrity), and Interaction Experience (IE: Responsiveness, Immersion, Fluency). Evaluation is holistic (full performance), not turn-by-turn, to accommodate literary devices like foreshadowing.

Key Contributions¶

HAMLET framework: two-stage (offline blueprint + online embodied performance) multi-agent architecture operable from a single topic input, no full story outline required.
PAD module: dual-process decision module producing FAST/SLOW/SILENCE strategies with structured action triplets; enables proactive, initiative-taking actor behavior in group chat scenarios.
Backward planning: narrative point construction starting from the act's endpoint to prevent plot deviation.
Narrator agent: physical environment adjudication loop ensuring credible prop interactions with global state broadcast.
RAG-based memory compression: prevents exponential context growth in long-form multi-act performance.
HAMLETJudge: 8B specialized critic achieving state-of-the-art evaluation reliability at low inference cost.
Evaluation dataset: 100 cases (25 classical Chinese + 25 English literary excerpts + 50 custom topics spanning 10 themes) with three-dimensional evaluation rubric.

Results¶

Claude-4-sonnet-Thinking achieved the highest overall performance across most metrics in both English and Chinese drama.
Reasoning models consistently outperform non-reasoning counterparts: Qwen3-32B-Thinking (average 73.85) significantly exceeds Llama-3.1-70B (45.75) and rivals Qwen3-235B-A22B (71.21), demonstrating reasoning efficiency over parameter scaling in multi-constraint drama tasks.
Partial English character performance scores from the leaderboard: Claude-4-sonnet 76.50, Higgs-Llama-3-70B 72.00, Qwen3-235B-A22B 69.50, Qwen3-32B 65.00, Claude-3.7-sonnet 65.80, Claude-3.5-sonnet 61.00, GPT-4.1 63.50, DeepSeek-V3-0324 54.00, Llama-3.1-70B 49.80, Gemini-2.5-flash 46.00, Llama-3.1-8B 35.00.
Both PAD and HAMLETJudge are 8B-parameter models achieving state-of-the-art at their respective tasks (decision-making and evaluation).
Ablation studies on PAD and control agents (Planner/Transfer/Advancer) confirm each component's contribution (full results truncated in provided text).

Limitations¶

Physical environment is fully simulated; no real robotic embodiment—"embodied" refers to virtual prop state changes only.
PAD requires training a separate 8B model; not plug-and-play with arbitrary base LLMs.
Evaluation depends on HAMLETJudge's pairwise win-rate methodology, which conflates absolute quality differences into relative rankings against a single GPT-4o baseline.
Dataset of 100 cases, though diverse, may not capture all dramatic genres or cultural contexts.
Human-actor interaction in the online performance loop was not systematically evaluated in the provided text; the immersive experience for real human participants remains to be validated at scale.
Long-form multi-act scenarios rely on RAG memory compression whose fidelity in retaining plot-critical information is not ablated in the provided text.

Relevance to Agentic AI / LLM Agents¶

HAMLET directly advances multi-agent coordination by shifting from passive, turn-reactive agents to proactive, goal-driven actors with a dual-process decision layer—a design transferable to any agentic system requiring initiative and strategic planning under social constraints. The PAD module's FAST/SLOW/SILENCE decision taxonomy, backed by internal monologue, is a concrete implementation of deliberate vs. automatic reasoning in agents that connects to broader work on chain-of-thought and tool-augmented agents. The narrator-mediated environment interaction pattern (agent proposes action → adjudicator validates → global state update → broadcast) is a generalizable architecture for grounded, world-consistent agentic behavior beyond drama. The finding that reasoning-enhanced models substantially outperform larger non-reasoning models in multi-constraint (goal + persona + context) tasks strengthens the case for inference-time reasoning as a key axis for agentic capability.