ActiveCTReasoner-7B-RL

ActiveCTReasoner-7B-RL

GRPO-trained agent, initialised from ActiveCTReasoner-7B-SFT.

• Paper: ACTOR: Agentic CT Reasoning with Radiologist-Aligned Actions (EMNLP 2026)
• Code: https://github.com/qianxingyucode/ACTOR
• Benchmark: Qianxingyu/ACTOR-Bench

This model needs a tool environment

It is not a single-turn VQA model. It emits <tool_call> and expects <tool_response> with a fresh rendered slice, looping until it emits <answer>. Serving the weights and showing one image will score far below the table above, because that is a different task. The environment — the five radiologist-aligned tools, the window presets, the bounding-box overlay and the prior-scan alignment — ships in the code repository as actor/env/simulator.py.

vllm serve Qianxingyu/ActiveCTReasoner-7B-RL --port 8100 \
    --limit-mm-per-prompt image=35 --max-model-len 32768 --trust-remote-code
python -m benchmark.eval.eval_benchmark

Keep the image cap at or above the evaluator's --max-context-images (default 30): a DualSource episode runs up to 18 tool rounds and every observation adds a slice, so a lower cap makes the server reject precisely the longest episodes. 35 is what we evaluated with.

Not for clinical use

Research artefact only, evaluated on CT-RATE validation scans. It has not been validated for diagnosis and must not inform patient care. It can miss findings and can state measurements confidently that are wrong.

Licence

CC BY-NC-SA 4.0, non-commercial. The weights derive from training on CT-RATE-derived observations, so the data licence propagates here even though the ACTOR code itself is Apache-2.0.

CT-RATE is gated, and its Terms and Conditions section 5 state:

Redistribution of the dataset or any portion thereof is not allowed.

That is why no CT volumes or rendered slices are published with these weights, and why the evaluation data lives in Qianxingyu/ACTOR-Bench as text plus a render manifest that you replay locally against your own CT-RATE download.