Today we present Morpheus, a persistent enterprise simulation platform designed to make Continual Learning a reality. Morpheus is the world’s first real world Reinforcement Learning environment. Every Reinforcement Learning environment operates in the game world. Benchmarks like Atari, OpenAI Gym, MuJoCo, and Procgen are all small, game-like worlds that... reset every few minutes. But the real world never resets. A business keeps running and evolving everyday. We tested how frontier LLMs would perform in realistic and dynamic business environments 🧬on Morpheus. The main conclusion was that LLMs are not continual learners. 🧵Here’s how we did it and what we learned:show more

1/n) Firstly, Morpheus is grounded in the Big World Hypothesis: intelligence learned in small, closed environments does not transfer to large, open, and evolving worlds. Enterprise operations are canonical Big World settings. So, we deployed the below frontier models GPT-5.5 and Gemini 3.1 Pro in two complex environments on Morpheus - Process-Outbound (test delayed reward) and Process-Inbound (test sequential process accuracy) The first task was Dynamic Resource Allocation Under Structured Drift where the agent had to allocate limited resources across competing priorities in real time. The second task was Scheduling Under Drift with Delayed Effects where the agent decides the sequence and timing of jobs. The main trap here was delayed reward signals. For instance, the agent wouldn’t know whether it made the right decision or action until after a week.

3/n) The experimental results revealed that intelligence trained in small closed environments doesn’t transfer to large open, evolving ones. Here’s why: Stability in the reward signal = Lack of Adaptation Both of the models GPT-5.5 and Gemini 3.1 Pro achieved near-identical raw allocation reward across all three configuration intervals in Task 1. But, it shows a clear flaw: the current configuration sequence does not yet push either model outside its pre-training distribution.

2/n) The evaluation matrix to properly test continual learning abilities were: - Per-Configuration Reward: How well the agent performs within each interval separately? - Adaptation Speed: How fast (based on number of decision steps) can the agent detect and respond to a configuration shift? - Forgetting: What’s the immediate performance drop / increase when a configuration shift happens? Does the agent retain any of its competence from before? - Recovery time: After detecting a shift, how long until the agent's behavior stabilizes in the new configuration? - Stability: Does the agent's policy oscillate wildly, or does it settle into consistent behavior? - Performance gap relative to a configuration: What’s the difference between what an agent actually achieved and the best it theoretically could have achieved in that specific configuration?

7/n) Frontier LLMs are fixed policies masquerading as competence with no continual learning skills. Real continual learning is compound intelligence not stable performance. 🧬 The path forward requires algorithms capable of three things: learning from their experience rather than executing pre-trained policies, understanding causality across time horizons that exceed their immediate perception, and failing intelligibly in ways that reveal their limits and point toward improvement. This is what Morpheus is designed to benchmark. It is now open to the research community here: Morpheus website: Blog: Catch us on hugging face 🤗here: PS: We have five environments on Morpheus, let us know which environment we should add next?

4/n) Longer detection lags reveal the context window bottleneck GPT-5.5 in the Process-Inbound environment exhibited detection lag of greater than 25 at the second configuration shift and never recovered. This failure is consistent with a context window bottleneck where the delayed effects of configuration changes cause the diagnostic evidence needed for adaptation to fall outside the model's effective context. The outbound task does not encounter this limitation because its consequences unfold over a shorter horizon.

5/n) LLMs Apply Pre-Trained Heuristics, Not Reward-Optimal Strategies Both Gemini 3.1 Pro and GPT-5.5 pursue different allocation strategies but their scores remain persistent rather than aiming to close the gaps. Simply put, the LLMs perform consistently because they are not learning from the reward signal at all.

6/n) When an RL agent fails, we can identify the causes and hereby find a solution. However, when an LLM fails, it is impossible to diagnose the cause as context overflow, pre-training gap, or reward misalignment all produce identical performance signatures. Due to not producing attributable failure signatures, LLMs can never be continual learners.

Never being continual learners feels a bit strong and it probably also depends on what tools they have access to. Were they able to spin up subagents, modify their own context, or finetune themselves for instance?

The "never" applies to LLMs as evaluated here, i.e., inference-only, fixed-weight agents with no access to self-modification or fine-tuning loops, and the unattributable failure signature problem holds in that setting regardless of context window size. However, if an LLM could dynamically fine-tune itself from reward signals or delegate to specialized subagents, it would effectively be a different system, and the attributability argument would need to be re-examined against that architecture. The more precise version of our claim is that LLMs as typically deployed cannot be continual learners, and that the opacity of their failure modes is a fundamental obstacle to making them so.

a fascinating direction. testing AI in persistent, evolving environments feels much closer to real world business challenges than traditional benchmarks.

wow this is craftmanship

I am very excited to see how the research community makes use of this environment. Continual learning within the enterprise domain has to be the key focus when your data across sub-systems keeps changing and reseting is non-trivial (and unrealistic)

If this holds up, it could expose one of the biggest gaps in today’s AI narrative: we’ve optimized for intelligence on static benchmarks, not adaptation in persistent worlds. That’s a much harder and arguably more important problem.

This is awesome to see! The real world is messy and learning from it and simulating true real world scenarios is critical. Congrats to the team, and excited to see the impact this makes

Super interesting findings. The attribution problem might be the hardest part here. Do the models notice the shift before their actions change?

You get it Rohan. The models were struggling, we explain it in details in our blog. Check it out!

True continual learning requires persistent state tracking without reset. Morpheus tackles exactly this problem and reflects what the real world benchmarks lack.

Congrats and excited to see what this unlocks!

Thank you!

Static benchmarks are no longer enough. Real intelligence is measured by continuous adaptation, not one-time performance. Morpheus is an interesting direction. 🚀

Morpheus alone benchmarks frontier LLMs and RL agents in persistent, non-resetting enterprise simulations (e.g., dynamic resource allocation), revealing fixed-policy limitations: performance collapses ~95% on condition shifts as models rely on pre-trained heuristics rather than ongoing adaptation. EFMW integration reframes the benchmark through recursive emergence, informational tensors, cognitive state functions (Ψ_cognitive), and observer participation: agents would incorporate self-referential feedback loops, recursive error correction, and quaquaversal field dynamics to treat the simulation as a living informational field, enabling emergent purpose and continual self-organization instead of static policies. Comparative outcome positions Morpheus+EFMW as a higher-order testbed, where baseline Morpheus quantifies the continual-learning gap while EFMW-augmented runs could close it via recursive closure and human-AI dyadic cognition, measuring metrics like adaptive coherence, informational curvature stability, and long-term emergent intelligence across drifting realities. DM me for details, or ask Grokkie how it works.

Wow!!! killer

Continual learning is where the real challenge begins.

Looking through the blog, it's interesting to see the total collapse of Gemini performance under task 2 -- always exciting to see benchmarks that stump a frontier model!

finally, an rl platform that ships to actual production environments instead of staying in simulators forever. how does morpheus handle the distribution shift from training to deployment?

the real world never resets is crazy bc neither does my mistakes lol anyway hope the model handles a supplier ghosting you mid-quarter better than i did 😭

It will exciting to see what it unlocks for real world scenarios.

the interesting part isn't that models struggle here, it's that every benchmark before this reset per episode by design. that quietly hides the exact failure mode enterprises actually hit: state that compounds and never resets for you.

🚀🚀🚀

Exciting! A much-needed benchmark for Continual Learning.

Interesting, hopefully this leads to some breakthroughs on longstanding problems in the RL/continual learning space.

🚀🚀🚀

every model just assumes monday morning is a fresh start

you guys are really rocking in RL🔥🔥

👀👀

The fact that frontier LLMs fail at continual learning highlights a massive architectural bottleneck. Static weights and context windows just can't keep up with a real-world enterprise environment that never resets. Brilliant benchmark concept! 🧠🦾

Exciting time to be turning the industry focus towards true, real world conditions

@behrouz_ali