| --- |
| license: apache-2.0 |
| pretty_name: Physim Reasoning Tasks |
| language: |
| - en |
| size_categories: |
| - n<1K |
| task_categories: |
| - other |
| tags: |
| - reinforcement-learning |
| - agent-evaluation |
| - harbor |
| - biology |
| - physiology |
| - simulation |
| configs: |
| - config_name: default |
| data_files: |
| - split: test |
| path: task_catalog.jsonl |
| --- |
| |
| # Physim Reasoning Tasks |
|
|
| Ten whole-body physiology tasks. Each task combines interventions across organ systems and asks an agent to predict downstream changes relative to a matched simulated control. 33 endpoints are scored across the ten tasks: the set was reduced to the endpoints whose simulated direction is defensible from physiology, and the answer keys and graders are unchanged from the released task files. |
|
|
| The task catalog, model comparison, and project overview are available at [symbolic.life/tasks](https://symbolic.life/tasks). |
|
|
| ## Contents |
|
|
| - `tasks/`: runnable Harbor tasks, each with an instruction, container environment, verifier, and reference solution. |
| - `task_catalog.jsonl`: browsable task descriptions and prompts. |
| - `download/`: the exact task archive served on symbolic.life. |
| - `benchmark-results.json`: strict whole-task results for every model with a complete run. |
|
|
| ## Run |
|
|
| Install Docker and [uv](https://docs.astral.sh/uv/), then run from this repository’s directory: |
|
|
| ```bash |
| uv run --with harbor==0.23.0 python run_verified.py run --path ./tasks --agent oracle |
| ``` |
|
|
| To evaluate an agent, configure its provider credentials in your own environment and run: |
|
|
| ```bash |
| uv run --with harbor==0.23.0 python run_verified.py run --path ./tasks --agent terminus-2 --model PROVIDER/MODEL --ak reasoning_effort=high |
| ``` |
|
|
| The supplied launcher uses the tested Harbor version and mounts each task’s verifier separately and read-only. Keep `tests/` and `solution/` outside the evaluated agent’s environment. |
|
|
| ## Scoring |
|
|
| The answer gives increase, decrease, or unchanged for each scored output. The task prompt specifies the comparison and the unchanged threshold. The published metric is strict whole-task pass: an attempt counts only when every scored direction in it is correct. |
|
|
| ## Results |
|
|
| Five attempts per task per model, single-shot with no tools, reasoning effort high. Only models that completed every attempt of every task are listed. |
|
|
| | Model | Attempts passed | Strict pass rate | |
| |---|---:|---:| |
| | Claude Opus 5 | 50/50 | 100.0% | |
| | Gemini 4 Argon | 43/50 | 86.0% | |
| | Grok 4.7 | 42/50 | 84.0% | |
| | Grok 4.6 | 41/50 | 82.0% | |
| | GPT-6 Astra | 40/50 | 80.0% | |
| | Thinking Machines Inkling | 35/50 | 70.0% | |
| | Nemotron 3 Ultra | 20/50 | 40.0% | |
|
|
| Not listed: Claude Opus 5.5 (4 of 10 scenarios; 31 attempts refused by the provider's content filter); Claude Fable 5.1 (3 of 10 scenarios; 35 attempts refused by the provider's content filter); Gemini 3.8 Flash (10 of 10 scenarios but 48 of 50 attempts; 2 refusals). |
|
|
|
|
