Skyfall AI Releases MORPHEUS: A Continuous Business Simulation Benchmark That Enables Continuous Reinforcement Learning Under Systematic Discontinuity

Most reinforcement learning benchmarks reset the world after each episode. Real jobs are never reset. Skyfall AI’s MORPHEUS addresses that gap. It is a business simulation platform for continuous reinforcement learning (CRL).

What is MORPHEUS?

MORPHEUS is based on the Big World Hypothesis (Javed & Sutton, 2024). It says that the complexity of the world exceeds the representational capacity of agents. As a result, the environment looks static even under constant dynamics.

To enforce continuous learning, MORPHEUS requires three characteristics: persistence, persistence, and operational complexity. Persistence refers to past decisions that involve future changes. Non-stationarity means that any consistent policy eventually becomes superior. The complexity of the operation means that no consistent policy exists.

Each domain is a self-contained global TypeScript plugin. Exports Operational Descriptors (ODs), simulation editor, seed data, and scripts. OD describes a step-by-step implementation plan for a skill. Agents work with the power API, and each call triggers an OD execution.

How the Platform Works?

Building on that diagram, non-stop comes from two engines. First, a failed injection engine introduces a typed interruption between the OD steps. It draws on eleven types of failure, including missing_data, dependency_failureagain rate_limit. It works with four preset values: light (5%), realistic (8%), medium (15%), and aggressive (30%).

Second, the asynchronous configuration shift control switches the failover reset and demand to constant timestamps. It works independently of the training loop, so shifts are never aligned with gradient updates. This prevents the agent from using the update periodicity as the proxy clock.

In addition to these engines, the reward comes from three performance indicators that are natively installed by the platform. These are event signals of failure, balance sheet status, and resource outflows. A compound reward brings them together. Default weights are w_f = 0.5 again w_l = w_p = 0.25.

# Composite reward — MORPHEUS, Appendix C (default weights).
def clip(x, lo, hi):
    return max(lo, min(hi, x))

def composite_reward(tickets, actual_cost, planned_cost, units, capacity,
                     w_f=0.5, w_l=0.25, w_p=0.25):
    r_f = -sum(t["severity"] for t in tickets)          # failure event signal
    r_l = clip(1 - actual_cost / planned_cost, -1, 1)   # financial ledger
    r_p = clip(units / capacity, 0, 1)                  # resource throughput
    return w_f * r_f + w_l * r_l + w_p * r_p

Under optimal assumptions (zero failure, low cost, full output), the bond in each configuration is equal to 0.50.

Policy Implementation

Because the action space is large, pure RL from scratch is ineffective. MORPHEUS therefore uses a two-stage pipeline. The frontier model (Gemini 3.1 pro) collects trajectories using the ReAct framework. These tracks are then fine-tuned by Qwen3-14B with supervised fine-tuning (SFT).

As a result, all RL runs start at this assigned SFT checkpoint. This separates progressive learning behaviors from basic functional skills. All bases then use PPO as the basic foundation for post-internet training.

Six-Metric Evaluation Protocol

In terms of training, cumulative reward alone is not enough. The scalar sum hides the operation at infinity. So the research team proposes six metrics instead. This is the reward per stop, adaptation speed, forgetfulness, recovery time, stability, and performance gap.

Among these, the speed of adaptation is a topic metric. It counts the steps until the running average reward reaches half of the upper limit. Two additional diagnoses also track relative adaptation gain (RAA) and plasticity with functional positions.

Primary Results

Using this protocol, the research team evaluates four families of algorithms from the SFT shared testbed. Two functions are defined. Task 1 is dynamic resource allocation under systematic drift. Activity 2 is organized under drift with delayed effects.

In all these results, no single family dominates. Opened process-outbound Task 1, EWC leads the reward and LCM adapts quickly. In Task 2, the HER leads the reward while the LCM loses the edge under the delayed reward. Meanwhile, the performance gaps mean they stay close to 1.0 all the way. That shows a big country deficit, not a small tuning gap.

Notably, PPO and HER generally adapt only to the initial configuration. Then they fail to adapt to the latest trends, even without the signals on the label.

Use Cases with examples

In practice, MORPHEUS fits several student roles. For AI developers, it tests whether the agent detects state changes without labels. For example, the need to change from low to high, and the policy should agree without a signal.

For data scientists, it emphasizes a delayed share of credit. For example, On-Time In-Full (OTIF) delivery is seen only days after the shipping decision. For software developers, the TypeScript plugin format allows swapping rewards or changing visibility without changing functionality.

Strengths and Weaknesses

Power:

• Persistence worlds with no resets, which match business systems used.
• Parameterisable, reproducible state shifts for convenient cross-algorithm comparisons.
• Rewards from native performance guarantors, requiring no external annotation.
• Open source test code (Skyfall-Research/morpheus-evals).

Weaknesses:

• Only two of the five locations are being tested so far.
• More responsibility assumes zero failure, so we’re always optimistic.
• Change is external, not driven by collective decisions.
• Reward weights are research variables, not guaranteed industry objectives.

Key Takeaways

• MORPHEUS runs persistent business worlds that never reset, unlike RL episodic benchmarks.
• It serves five locations; two are examined here: process-outbound again process-inbound.
• The protocol scores six metrics for each configuration, adaptation, forgetting, recovery, stability, and gap-to-upper-bound.
• The four bases (PPO, HER, EWC, LCM) all remain well below the theoretical upper limit.
• There is no single winning algorithm; reward and speed adaptation to select different winners.

Check it out Paper again Project Page. Also, feel free to follow us Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to Our newspaper. Wait! are you on telegram? now you can join us on telegram too.

Need to work with us on developing your GitHub Repo OR Hug Face Page OR Product Release OR Webinar etc.? contact us