BaT: Towards Self-Evolving Medical Research Agent with Stage Rubrics

Junqi Liu; Yufan He; Yexiao He; Pengfei Guo; Dong Yang; Andriy Myronenko; Can Zhao; Hanrong Ye

cs.AIAug 17, 2026

BaT: Towards Self-Evolving Medical Research Agent with Stage Rubrics

Authors: Junqi Liu, Yufan He, Yexiao He, Pengfei Guo, Dong Yang, Andriy Myronenko, Can Zhao, Hanrong Ye,

Organizations: University of California, Santa Cruz · NVIDIA

Abstract

Long-horizon agents are beginning to automate complete workflows that produce code, reports, and research artifacts. Medical imaging workflows are multi-stage and data-sensitive, while expert trajectories remain scarce and difficult to share. Structured benchmarks can localize failures through stage-level rubrics, but standard post-training discards these diagnostics before the next training round. We present Benchmark-as-Teacher (BaT), a recursive self-improvement system for agent post-training. BaT contains two linked components: the asynchronous Stage Bank data pipeline and BiCuRL (Bilevel Curriculum Reinforcement Learning), its self-improving post-training method. Stage Bank synthesizes content-isolated training states outside the policy-update loop. BiCuRL uses a fixed held-out evaluation to select the next stage curriculum, verifies rollouts with task rubrics, updates the policy with GRPO, and returns the candidate checkpoint to evaluation. On AutoMedBench-Lite, BaT-4B and BaT-9B more than double the Overall scores of their Qwen Instruct baselines. BaT-9B Agent reaches 79.6 Overall, exceeding Claude Opus 4.6 with Claude Code at 77.5.

Figures & tables

Figure 1: Benchmark-as-Teacher turns an output-only benchmark into post-training infrastructure. (a) Top: conceptual overview. BaT turns the usual output-only benchmark into training infrastructure with the Stage Bank data-synthesis pipeline, Curriculum Teacher, and BiCuRL post-training method. (b) Lower left: BiCuRL loop. The system forms a closed loop with four operational phases. First, after an SFT cold start, benchmarking splits into stage-wise scores. Second, the aggregate scores feed the Curriculum Teacher, shown as the Domain-Shift Detector and Report for next round. Third, the environment sandbox pool selects targeted-stage (S-target) sandboxes together with mix-stage (S-mix) and end-to-end (E2E) sandboxes. Fourth, the agent executes rollouts in these sandboxes, an LLM verifier scores them with rubric-as-rewards ( Gunjal et al., 2026 ) , and GRPO ( Shao et al., 2024 ) updates the policy. The updated checkpoint loops back to benchmarking and restarts the cycle. (c) Lower right: training rounds. Dashed curves show each round’s evaluation, solid step lines track the best checkpoint so far, stars mark the auto-selected best checkpoints.

Figure 2: BaT-9B Agent leads AutoMedBench-Lite and trails the leaders by 9.3 points on ABRA and 14.8 points on MedXpertQA-Text. It scores 79.6 on AutoMedBench-Lite, 70.6 on ABRA, and 50.2 on MedXpertQA-Text. The strongest non-BaT systems score 77.5, 79.9, and 65.0, respectively. The legend shortens BaT-4B Agent and BaT-9B Agent to BaT-4B and BaT-9B. Each panel uses its benchmark’s own protocol, and each system keeps the execution setting shown in the legend ( Liu et al., 2026a ; Maksudov et al., 2026 ; Zuo et al., 2025 ; Qwen Team, 2026b ; Qwen Team, 2026c ; Google DeepMind, 2026 ; GLM-5 Team, 2026 ; OpenAI, 2026 ; Anthropic, 2026 ) .

Figure 3: Each BaT round combines all three sandboxes built from clean synthetic Stage Bank. The Data Factory fills the Stage Bank, and the report for the next round only picks the weak stage: E2E trains on the whole unsegmented trajectory, S-target applies the weak-stage rubric, and S-mix blends stage sandboxes from the other stages. No evaluation task content enters training rows.

	Qwen3.5-4B ( Qwen Team, 2026b )			Qwen3.5-9B ( Qwen Team, 2026c )
Training	Ovl.	Agt.	Task	Ovl.	Agt.	Task
Baseline	6.1 $\pm$ 1.7	12.1 $\pm$ 2.0	0.0 $\pm$ 1.1	19.9 $\pm$ 2.3	11.3 $\pm$ 2.0	28.4 $\pm$ 2.6
SFT ( 2022 )	18.1 $\pm$ 2.3	10.8 $\pm$ 1.9	25.4 $\pm$ 2.5	12.9 $\pm$ 2.1	21.6 $\pm$ 2.4	4.1 $\pm$ 1.5
GRPO ( 2024 )	11.4 $\pm$ 2.0	15.9 $\pm$ 2.2	6.9 $\pm$ 1.7	31.9 $\pm$ 2.7	39.2 $\pm$ 2.8	24.5 $\pm$ 2.5
BiCuRL	22.9 $\pm$ 2.4	28.8 $\pm$ 2.6	17.0 $\pm$ 2.2	53.4 $\pm$ 3.2	64.7 $\pm$ 3.1	42.1 $\pm$ 3.2

Table 2: BiCuRL more than doubles the Qwen Instruct AutoMedBench-Lite Overall score at both model sizes. Each score is the mean over the seven track-level means, with

\pm

showing the source draft’s reported uncertainty half-width. Bold and underline mark the best and second-best result within each model size. Baseline denotes Qwen3.5 Instruct ( Qwen Team, 2026b ; Qwen Team, 2026c ) . SFT follows supervised instruction tuning ( Ouyang et al., 2022 ) , and GRPO follows group-relative policy optimization ( Shao et al., 2024 ) . Ovl. denotes Overall, and Agt. denotes Agentic.

Figure 4: BaT keeps improving across rounds, while GRPO with one final reward saturates early. Solid step lines track the best BaT checkpoint so far and stars mark the selected checkpoints; dashed curves show GRPO in the matching model color.

Figure 5: The full S-target, S-mix, and E2E sandbox mix leads every partial mix. The dot matrix on the left marks which pools enter training (columns 1–3: S-target, S-mix, End-to-End; dark included, pale omitted). Green bars show Agentic, red bars show Task, and gray circles show Overall. Rows from top to bottom: full BaT, E2E+S-mix, E2E+S-target, S-target+S-mix, E2E, S-mix, S-target, and the Qwen3.5-9B Instruct baseline. All scores use a 0–100 scale and one decimal.

Figure 6: BaT-9B ranks first, and BaT-4B ranks third among 10 tiny local LLMs under the same default runner. Here tiny local LLM is defined as parameter size no more than 12B. Bars report AutoMedBench-Lite Overall scores.

short-turn reasoning tasks
	Qwen3.5-4B ( Qwen Team, 2026b )			Qwen3.5-9B ( Qwen Team, 2026c )
Benchmark	Base	BaT	$\boldsymbol{\Delta}$	Base	BaT	$\boldsymbol{\Delta}$
AIME 2025 ( Mathematical Association of America, 2026 )	85.0	56.7	-28.3	88.3	83.6	-4.7
AIME 2026 ( Mathematical Association of America, 2026 )	89.7	53.3	-36.4	92.5	86.7	-5.8
GPQA-Diamond ( Rein et al., 2023 )	76.2	58.1	-18.1	81.7	78.3	-3.4
long-horizon tasks

Table 3: BaT-9B retains most reasoning accuracy and improves three of five long-horizon scores, while BaT-4B declines on all eight benchmarks. Base denotes each Qwen3.5 Instruct checkpoint ( Qwen Team, 2026b ; Qwen Team, 2026c ) . Delta is BaT minus Base in percentage points; green marks gains and red marks drops. Each pair uses the same benchmark protocol.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter	Qwen3.5-9B cold SFT
Parameter update	Full parameters, without adapters
Parallelism	8 GPUs, FSDP full_shard
Per-GPU microbatch	1
Gradient accumulation	8
Effective global batch	64
Learning rate	$2\times 10^{-5}$

Appendix

Table 4: Qwen3.5-9B cold-start SFT configuration. We use full-parameter training with PyTorch FSDP and AdamW ( Qwen Team, 2026c ; Zhao et al., 2023 ; Loshchilov and Hutter, 2019 ) .

Runtime setting	Qwen3.5-4B	Qwen3.5-9B
GPUs	8 $\times$ NVIDIA A100 80GB
Default engine	SGLang
Supported engine	vLLM
Context limit	256K tokens
Recorded throughput	320 tokens/s	220 tokens/s

Appendix

Table 5: Serving configuration and recorded throughput. The stack supports SGLang and vLLM, with SGLang as the default engine for agentic tasks ( Zheng et al., 2024 ; Kwon et al., 2023 ) . Each recorded rate uses an eight-GPU NVIDIA A100 80GB node ( NVIDIA, 2022 ) . The source run summary reports an unsplit token rate and omits the remaining serving controls.

Model	Overall	Agentic	Task
Qwen3.5-4B ( Qwen Team, 2026b )	6.1	12.1	0.0
Meissa-4B ( Chen et al., 2026 )	7.6	15.2	0.0
Gemma 4 E2B ( Gemma Team, 2026 )	9.0	18.0	0.0
Nemotron 3 Nano 4B ( NVIDIA, 2026a )	9.1	18.2	0.0
GPT-OSS-20B ( OpenAI, 2025 )	17.6	25.3	9.9
Nemotron 3 Nano 30B-A3B ( NVIDIA, 2025a )	18.0	22.5	13.5

Appendix

Table 6: Full local LLM comparison with the default runner. Scores use the AutoMedBench-Lite evaluation ( Liu et al., 2026a ) and a 0–100 scale, ranked by Overall score. The two shaded rows are BaT policies. Overall equals the arithmetic mean of Agentic and Task scores.

Track	Held-out task	Cases	Runs
Classification	Skin-lesion classification	100	10
Detection	Wrist radiograph detection	100	10
Enhancement	Low-dose CT denoising	20	10
Report	Chest X-ray reporting	100	10
Segmentation	Multi-organ CT segmentation	40	10
Synthesis	Pancreas CT synthesis	20	10

Appendix

Table 7: AutoMedBench-Lite spans seven held-out medical AI tracks and 70 runs per system. Cases report the public evaluation subsets, and each system runs every track ten times. Agentic weights Plan, Setup, Validate, Inference, and Submit by 0.25/0.15/0.35/0.15/0.10; Overall averages Task and Agentic ( Liu et al., 2026a ) .

Figure 7: Per-track Overall comparison of the BaT Agents and Claude Opus 4.6 with Claude Code on AutoMedBench-Lite. Bars show the mean Overall score per track on a 0–100 scale. We order tracks left to right by ascending average output tokens per run, from the lightest question-answering tracks to the heaviest imaging pipelines. We estimate the Opus enhancement and report values from its published aggregate scores and measure the remaining values.

Parameter	Recorded setting
Initializer	Qwen3.5-9B SFT checkpoint
Train minibatch	8 / 8
Rollouts per prompt	4
Actor learning rate	$1\times 10^{-6}$
Actor KL coefficient	$1\times 10^{-3}$
Rollout temperature	0.7

Appendix

Table 8: Recorded controls for the matched Qwen3.5-9B pool-ablation contract.

LLM Agent Self-Improvement

Large language model agent self-improvement studies how agents use experience and feedback to improve their behavior, tools, and workflows.

See all 273 papers

AutoMedBench: Towards Medical AutoResearch with Agentic AI Models
Jun 1, 2026Junqi Liu, Selena Song, Yuhan Wang +12AI Agent Evaluation AI Agent Benchmarks
University of California, Santa Cruz · 2NVIDIA
Cura 1T: Healthcare Foundation Model via Recursive Self-Improvement
Jul 15, 2026Haolin Chen, Leon Qi, Steve Brown +7Medical VLMs LLM Agent Training
actAVA AI Team
BioMedArena: An Open-source Toolkit for Building and Evaluating Biomedical Deep Research Agents
May 7, 2026Jinge Wu, Hongjian Zhou, Mingde Zeng +8AI Agents for Scientific Discovery Deep Research Agents
University of Oxford · University College London · Technical University of Munich +1

Jun 1, 2026J/K move · Enter open · S save

AutoMedBench: Towards Medical AutoResearch with Agentic AI Models

Junqi Liu, Selena Song, Yuhan Wang, Jiawei Mao+11

University of California, Santa Cruz · 2NVIDIA

AI Agent Evaluation AI Agent Benchmarks Scientific Workflow Automation Autonomous Research Agents Medical AI Agents

Open paper AskPDF arXiv

Autonomous agents are increasingly expected to support end-to-end medical-AI research workflows, moving beyond isolated prediction tasks or short-form clinical question answering. However, existing medical agent benchmarks primarily evaluate final outputs, providing limited visibility into agent behavior within the research process. To address this gap, we present AutoMedBench, a workflow-aware benchmark for autonomous medical-AI research across diverse medical imaging and multimodal inference tasks, organizing agent execution into a unified five-stage workflow (S1-S5): Plan, Setup, Validate, Inference, and Submit. It comprises long-horizon tasks with each run averaging 33 agent turns, spanning five research tracks: segmentation, image enhancement, visual question answering (VQA), report generation, and lesion detection. Each task is evaluated under two difficulty tiers, Lite and Standard, which use the same data and metrics but differ in the amount of task-brief scaffolding, and each run is scored using both final task performance and S1-S5 stage scores, enabling stage-level analysis from the initial task brief to the final submitted artifact. Across thousands of recorded runs, stage-level scoring reveals that Validate is the weakest workflow stage on average, whereas Setup is the strongest, suggesting that current agents are better at making pipelines executable than at verifying their reliability. Post-run error analysis further shows that verification and submission failures dominate tagged errors, accounting for 37.7% and 38.1% of fired codes respectively, whereas task-understanding errors are rare at 0.9%, and runs with one fired error code have a 48% lower overall score than runs with no error code on average.

Abstract

LLM Agent Self-Improvement

Explore similar work

AutoMedBench: Towards Medical AutoResearch with Agentic AI Models

Cura 1T: Healthcare Foundation Model via Recursive Self-Improvement

BioMedArena: An Open-source Toolkit for Building and Evaluating Biomedical Deep Research Agents