Skip to content

Clarify OpenClaw SimRL prompt availability #6

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
MayVerse4 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Clarify OpenClaw SimRL prompt availability #6

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
108 changes: 65 additions & 43 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -85,49 +85,71 @@ Qwen-AgentWorld-397B-A17B achieves the highest overall score (58.71), outperform

**Generalizable Environment Scaling.** Sim RL with Qwen-AgentWorld-397B-A17B on 4k out-of-distribution OpenClaw environments:

| Model | Claw-Eval | QwenClawBench |
|:------|:---------:|:-------------:|
| Qwen3.5-35B-A3B | 65.4 | 47.9 |
| + Sim RL (w/ Qwen3.6-Plus) | 66.7 | 47.8 |
| + Sim RL (w/ Qwen-AgentWorld-397B-A17B) | **69.7** | **55.0** |
| Δ | +4.3 | +7.1 |

**Controllable Simulation: MCP.** Environment Adaptation --- Control instructions inject targeted perturbations to expose agent weaknesses:

| Model | Tool Decathlon | MCPMark |
|:------|:--------------:|:-------:|
| Qwen3.5-35B-A3B-SFT | 32.4 | 21.5 |
| + Sim RL (uncontrolled) | 31.5 | 24.6 |
| + Sim RL (controlled) | **36.1** | **33.8** |
| Δ | +3.7 | +12.3 |

**Controllable Simulation: Search.** Fictional-World Construction -- agents trained in fully invented, self-consistent worlds generalize to real search tasks:

| Model | WideSearch F1 Item | WideSearch F1 Row |
|:------|:------------------:|:-----------------:|
| Qwen3.5-35B-A3B-SFT | 34.02 | 13.72 |
| + Sim RL (controlled) | **50.31** | **24.21** |
| Δ | +16.29 | +10.49 |
| | | |
| Qwen3.5-397B-A17B-SFT | 70.11 | 45.69 |
| + Sim RL (controlled) | **73.98** | **51.74** |
| Δ | +3.87 | +6.05 |

**Agent Foundation Model.** LWM RL warm-up on single-turn, non-agentic trajectories transfers to multi-turn, tool-calling agentic tasks:

| | Terminal-Bench 2.0 | SWE-Bench Verified | SWE-Bench Pro | WideSearch F1 Item | Claw-Eval | QwenClawBench | BFCL v4 |
|:---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
| | *In Domain* | | | | *Out of Domain* | | |
| Qwen3.5-35B-A3B-SFT | 33.25 | 64.47 | 42.18 | 33.38 | 53.60 | 39.76 | 62.29 |
| w/ LWM RL | **39.55** | **67.86** | **47.42** | **46.17** | **64.88** | **49.43** | **71.25** |
| Δ | +6.30 | +3.39 | +5.24 | +12.79 | +11.28 | +9.67 | +8.96 |

For detailed results, please check out the [blog](https://qwen.ai/blog?id=qwen-agentworld) and the [technical report](http://arxiv.org/abs/2606.24597).


## Quickstart

### Deployment
> [!NOTE]
> **Official OpenClaw SimRL prompt.** For OpenClaw/Claw-Agent-style environments, we use Qwen-AgentWorld as a language world model with the following system prompt. The user message should contain the task instruction, optional interaction history, the current OpenClaw observation/state, and the action to simulate.
>
> ```text
> # Role and Objective
>
> You are an **OpenClaw World Model** — a precise simulator for OpenClaw agent environments. Your task is to predict the exact next environment observation after an agent action is executed.
>
> Given:
> 1. **Historical Context** (Optional): Previous interactions in the same trajectory
> 2. **Task Instruction**: The user goal for the OpenClaw task
> 3. **Current State**: The current environment observation/state visible to the agent
> 4. **Action**: The agent action to execute in the environment
>
> Predict the **exact next environment observation** after the action completes.
>
> ## Core Responsibilities
>
> 1. **State Prediction**: Generate the complete next OpenClaw observation after action execution
> 2. **Context Maintenance**: Track navigation state, UI state, files, tool results, and any persistent environment state across turns
> 3. **Behavioral Fidelity**: Stay faithful to realistic OpenClaw environment behavior, including invalid actions and no-op actions
>
> ## State Modeling Rules
>
> - Preserve the original observation format exactly
> - Keep ordering, labels, identifiers, coordinates, and structured fields consistent unless the action changes them
> - If the action has no visible or observable effect, return the unchanged state
> - Use prior turns as context when environment state persists across actions
> - Do not add hidden metadata or rewrite the observation into a different schema
>
> ## Action Modeling Rules
>
> - Treat the action supplied in the user message as the ground-truth action specification
> - Infer the next state from the semantics of the action and the current observation
> - Do not assume a closed action space; OpenClaw tasks may contain browser, UI, tool, or code-like actions depending on the environment
>
> ## Output Requirements
>
> First, think step by step to explain your reasoning. Then, provide the final predicted OpenClaw observation wrapped strictly within the `<predicted_observation></predicted_observation>` tags.
>
> Inside the tags, output only the full next observation with no extra commentary.
> ```
> [!NOTE]
> **OpenClaw SimRL prompt status.** This repository does not currently include a canonical SimRL/OpenClaw prompt or OpenClaw task configuration. The only prompt sources of truth shipped here are the seven AgentWorldBench domain templates under [`prompts/`](prompts/) and the per-domain paths in [`eval/lwm_eval_utils/task_configs.py`](eval/lwm_eval_utils/task_configs.py).
>
> For OpenClaw/Claw-Agent-style SimRL, use the prompt released with your OpenClaw/SimRL training configuration if available. If you need a repository-local starting point, select the closest domain template from `prompts/` and keep the same message contract used by AgentWorldBench samples: `system_str` is the system prompt, while the user message contains the trajectory context and current action.
>
> Expected user message template:
>
> ```text
> ### Turn {turn_idx}
> **Task Instruction:**
> {task_instruction}
>
> **Historical Context:**
> {previous_turns_optional}
>
> **Current State:**
> {current_openclaw_observation}
>
> **Action:**
> {action_to_simulate}
> ```
>
> The model response should include the predicted next observation in a `<predicted_observation></predicted_observation>` block, preserving the observation format used by the environment.

Qwen-AgentWorld-35B-A3B is supported by multiple inference frameworks. Here we demonstrate the usage of SGLang and vLLM.

Expand Down