Pith. sign in

REVIEW 5 major objections 4 minor 30 references

Cura 1T claims a healthcare-specialized model can beat frontier generalists on five of six clinical benchmark panels by making the training-data mixture the object of a self-evolution loop, without collapsing general reasoning.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 02:12 UTC pith:MQGNL4DF

load-bearing objection The training loop literally fits the evaluation benchmarks, so the headline ranking is likely an artifact; the paper is transparent but the claim is unproven. the 5 major comments →

arxiv 2607.15314 v1 pith:MQGNL4DF submitted 2026-07-15 cs.AI

Cura 1T: Specialized Model for Agentic Healthcare

actAVA AI: Haolin Chen , Leon Qi , Steve Brown , Deon Metelski , Tao Xia , Joonyul Lee , Qixuan Wang , Kevin Riley
show 2 more authors
Frank Wang Weiran Yao
This is my paper
classification cs.AI
keywords healthcare LLMself-evolution loopdata mixture curationclinical reasoninghealthcare agentic tasksretention anchorsself-distillation fine-tuninglow-rank adaptation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to show that healthcare specialization is best treated as a data-mixture problem rather than a hyperparameter-search problem. It builds Cura 1T, a healthcare-focused large language model, by iteratively turning evaluation failures into targeted training data through a human-gated self-evolution loop. On six healthcare benchmark panels, the model ranks first on five and second on the remaining one, matching or approaching the best frontier general-purpose models. The same model remains competitive on out-of-domain math, science, and agentic benchmarks, suggesting that the specialization does not obviously destroy general capability. The claim matters because it offers a repeatable recipe for building capable specialized models in data-sparse, high-stakes domains.

Core claim

Cura 1T is post-trained from a general-purpose base model through a self-evolution loop in which the training-data mixture is the optimized object. In each round, a training agent plans a target capability, trains low-rank adapters through an SFT-to-RL-to-SDFT stack, evaluates the model on healthcare benchmarks, reads the graded trajectories, and synthesizes data that targets the observed failure mode while retaining already-correct behavior. The consolidated model outperforms or matches frontier baselines on five of six healthcare panels and holds its own on out-of-domain reasoning and agentic benchmarks. The authors conclude that healthcare specialization is above all about curating the ri

What carries the argument

The central object is the human-gated self-evolution loop with the data mixture as the search target. It cycles through plan, train, evaluate, and refine stages, using SFT as a cheap mixture screen, RL for reward-driven improvement, and self-distillation fine-tuning (SDFT) to consolidate the final adapter while staying close to the base model's own generation behavior. The loop's data-synthesis skills — retention anchors, reasoning correction, knowledge injection, and behavior calibration — turn graded failures into targeted training rows before the next round begins.

Load-bearing premise

The central ranking holds only if the evaluation harnesses — including the modified AgentClinic protocol and the MedAgentBench 'harness bug fix' — measure real clinical capability rather than inflating Cura 1T's scores, and if the loop's practice of synthesizing training data from benchmark failures does not amount to fitting the test set.

What would settle it

Run the entire evolution protocol with all six healthcare benchmark corpora quarantined, so that no benchmark trajectory or synthesized derivative enters the training mixture, then evaluate the resulting model on the untouched versions of those benchmarks against the same baselines. If the gains over the base model largely vanish, the loop was fitting the evaluation rather than building general healthcare capability.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Healthcare specialization can be achieved without an obvious loss of general reasoning or agentic ability, as measured by math, science, and tool-use benchmarks.
  • Different failure modes require different data actions: tool-use trajectories for EHR workflows, behavior calibration for rubric-graded responses, knowledge injection plus retention for expert reasoning, and interactive trajectories for diagnosis.
  • The self-evolution loop is a repeatable protocol: plan, screen with SFT, improve with RL, consolidate with SDFT, then curate the next mixture from the observed failures.
  • Retention anchors are a practical mechanism to prevent targeted repairs from eroding already-correct behavior, a risk the paper documents with a reverted round that improved aggregate scores but degraded a subset.
  • The consolidated model's strength comes from the accumulated mixture across rounds, not from a single generic medical-data update.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same loop could transfer to other data-sparse, high-stakes domains where failure modes are heterogeneous and evaluation is rubric-based, such as legal advice, clinical documentation, or social services.
  • Because the loop trains on synthesized derivatives of benchmark trajectories, the boundary between training and evaluation is blurred; a strong extension would quarantine benchmark corpora and validate on fresh clinical tasks or live pilots.
  • The human gates in the loop suggest that full automation of data curation still faces reliability hurdles; the loop's immediate value may be in accelerating human experts rather than replacing them.
  • If the loop scales with base-model capability, specialized models could be maintained over time by feeding updated clinical guidelines as knowledge-injection examples, giving a pathway for continual alignment with evolving standards.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper presents Cura 1T, a healthcare-specialized LLM obtained by post-training Kimi-K2.6 with LoRA adapters through a human-gated self-evolution loop. In each round, an LLM training agent plans a capability target, runs SFT/RL/SDFT, evaluates on healthcare benchmarks, analyzes failed trajectories, and synthesizes new training data from those failures. The paper claims that Cura 1T ranks at or near the top on five of six healthcare benchmark panels (MedAgentBench, HealthBench Professional/Hard, MedXpertQA, AgentClinic) while remaining competitive on out-of-domain reasoning and agentic benchmarks.

Significance. If the results held under a clean evaluation protocol, the contribution would be significant: a data-mixture search loop for healthcare post-training that addresses patient care, clinical reasoning, and agentic workflows with parameter-efficient updates. The explicit documentation of reverted rounds and trace-level case studies is a strength. However, the evaluation protocol is compromised by the loop's use of evaluation benchmarks as training signal and by post hoc harness modifications. These issues must be resolved before the empirical claims can be accepted.

major comments (5)
  1. [§3.2, Fig. 3, App. A.3] The self-evolution loop converts failed trajectories from the evaluation benchmarks into training data. The data-construction skills defined in §3.2 (Reasoning Correction, Knowledge Injection, Behavior Calibration) and the case studies in App. A.3 (e.g., A.3.4, MedXpertQA Text-197 with gold answer D) show that the same task and gold answer used in evaluation appear in the training mixture. Consequently, the reported scores on these benchmarks are partly a measure of fit to the test set, not a prediction. The abstract's claim that Cura 1T 'ranks at or near the top among frontier baselines' is therefore not independently supported. The authors must either evaluate on a held-out healthcare benchmark never used in the loop or demonstrate that all test items were excluded from training.
  2. [§4.2, Table 3] Round 3 is labeled 'Harness bug fix' and raises the MedAgentBench score from 0.967 to 0.973 by 'restoring the intended prompt.' This changes the evaluation protocol after observing results. The table does not state whether the frontier baselines were run under the original or corrected harness, nor whether the fix affects all models equally. If the baselines were not re-run under the corrected protocol, the comparison is invalid. Please provide this information, and apply the same scrutiny to the AgentClinic tool-native harness modification described in §4.5 and App. A.2.
  3. [§4.6] The out-of-domain evaluation (AIME, GPQA-Diamond, τ2-Bench) does not serve as a control for healthcare specialization. These benchmarks are unrelated to the healthcare capabilities being optimized. A meaningful held-out check would use a healthcare benchmark that was not touched by the evolution loop (e.g., a different medical exam or EHR environment). The absence of such a control makes it impossible to distinguish genuine capability gains from memorization of the evaluation sets.
  4. [Table 3] The development path reaches 0.973 on MedAgentBench (Round 3), while the final consolidated Cura 1T model is reported at 0.940. The text does not explain why the consolidated model is worse than the development checkpoint despite being trained on the consolidated mixture. If the consolidated model is the one released, the headline ranking should be based on it; please clarify the relationship and the reasons for the drop.
  5. [§4.1, Tables 3–6] All metrics are point estimates at T=1.0 with no variance or confidence intervals. The margins between Cura 1T and the best baseline are small (e.g., 0.940 vs 0.937 on MedAgentBench). Without repeated runs, the ranking claims are not statistically meaningful. Please report at least 3–5 seeds or justify why the differences are significant.
minor comments (4)
  1. [Fig. 3 caption] The caption does not explain the color coding (solid vs dashed red arrows) or the meaning of 'retained' vs 'reverted'. Please make the caption self-contained.
  2. [Table 4] No frontier-model baselines are shown for HealthBench. Without these, the claim of 'at or near the top' is not supported for this panel. Please add baseline columns.
  3. [App. A.3] The case studies are labeled 'training stage' but do not state whether the synthesized rows are exact copies of the test item or paraphrased. This distinction is crucial for assessing leakage; please clarify.
  4. [References] The reference list contains several 2026 preprints and some items that appear to be future-dated. Please verify the publication status and accuracy of all citations.

Circularity Check

4 steps flagged

The self-evolution loop converts failures on the evaluation benchmarks into training data, and the appendix shows exact benchmark test items with gold answers used in training rounds; the headline healthcare ranking is therefore partly a fit to the evaluation set, not an independent prediction.

specific steps
  1. fitted input called prediction [Section 3.2, Figure 2, Table 2]
    "The agent reads a graded run, identifies the missing capability, and changes one mixture decision at a time... Refine uses the failed trajectories as evidence for root-cause analysis, converts the resulting failure modes into targeted data, curates the next mixture, and validates the candidate rows before proposing next steps."

    The 'graded run' and 'failed trajectories' come from the same healthcare benchmarks on which Cura 1T is later ranked (MedAgentBench, HealthBench, MedXpertQA, AgentClinic). Feeding those failures back into the training mixture and then reporting scores on the same benchmarks means the reported performance is partly a measure of how well the model memorized or fit the evaluation items, not an independent prediction of capability.

  2. fitted input called prediction [Appendix A.3.4]
    "The question asks which ductal carcinoma in situ patient gains the greatest local-control benefit from radiotherapy. The gold answer is D: a 0.9 cm grade-3 DCIS after lumpectomy with a negative margin. ... Round 2 (Knowledge injection + retention) ... Option D ...D. Reward 1."

    This is a MedXpertQA test question with its gold answer. The paper explicitly shows the training round receiving Reward 1 on the same item, then Table 5 reports Cura 1T's MedXpertQA overall score as a headline result. The score on this benchmark is therefore partly the result of training on the benchmark's own test items.

  3. fitted input called prediction [Appendix A.3.5]
    "The gold diagnosis is desquamative interstitial pneumonia. ... Interactive trajectory + retention ... 'The diffuse ground-glass opacities ... and macrophages in the airspaces ... [are] classic for Desquamative Interstitial Pneumonia.' Reward 1."

    This is an AgentClinic NEJM-Ext evaluation case with its gold diagnosis. The training trajectory is shown achieving Reward 1 on the exact case, and Table 6 then reports AgentClinic pass@1 for Cura 1T. The evaluation score is contaminated by construction: the model was trained on the same task and gold answer it is later evaluated on.

  4. fitted input called prediction [Appendix A.3.1]
    "Task task8_1 asks for an orthopedic surgery referral ServiceRequest... The grader requires SNOMED 306181000000106... Round 1 (Tool-use): fhir_post(ServiceRequest, priority=stat, code=306181000000106) Reward 1."

    This is a MedAgentBench task with its exact grader criteria. The paper shows a training round obtaining Reward 1 on this task, then Table 3 reports MedAgentBench task success for the released model. The benchmark result is partly a direct fit to the evaluation task rather than evidence of general EHR tool-use capability.

full rationale

The central claim—'Across the healthcare evaluation suite, Cura 1T ranks at or near the top among frontier baselines'—is not an independent prediction because the self-evolution loop explicitly uses the evaluation benchmarks' failed trajectories as training data. Table 2 describes Evaluate as 'Run benchmark harnesses and collect graded trajectories and failure summaries' and Refine as 'Categorize failures, synthesize targeted data, curate the next mixture.' The appendix then provides concrete instances where the exact test item and gold answer appear in training: MedAgentBench task8_1, HealthBench EROSTIN 10MG, MedXpertQA Text-197, and AgentClinic NEJM-Ext 21. Each is shown with a training-stage response receiving Reward 1, and each benchmark later appears in the main performance tables. This is not a case of paraphrased generalization or a held-out evaluation; it is fitting the benchmark by construction. The out-of-domain results (AIME, GPQA-Diamond, tau2-Bench) are genuinely independent and show no obvious capability collapse, which is why the score is not higher. However, those benchmarks are not healthcare benchmarks and do not validate the healthcare-specific headline. The paper also modifies the MedAgentBench harness after observing results ('Harness bug fix'), which is an additional evaluation-protocol concern, though the primary circularity is the training-on-test-items loop. Score 7 reflects that the headline healthcare ranking is substantially forced by the loop's design, while some independent content remains in the out-of-domain and non-exact synthetic data.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The paper's central benchmark claims rest on the assumption that the evaluation suites measure the intended healthcare capabilities and that modifying harnesses is benign. The data-mixture proportions chosen by the training loop are undisclosed free parameters; without them the core recipe is not replicable. No new physical or conceptual entities are postulated.

free parameters (3)
  • Final data mixture proportions (per-capability synthetic/curated rows) = not reported
    The training agent adjusts these proportions based on benchmark failure patterns (Section 3.2, Figure 3), but the actual mixture is not disclosed, so the central result depends on unobservable hand/agent-chosen weights.
  • Evolution round count and stopping criteria = not reported
    Rounds continue until benchmarks saturate; no formal rule, so the number of attempts before the reported scores is a free choice affecting final performance.
  • Benchmark evaluation temperature choice = T=1.0 for final; T=0.6 for MedAgentBench development path
    Evaluation temperature is set per benchmark; the choice can affect scores, especially for exact-match tasks.
axioms (4)
  • domain assumption HealthBench, MedXpertQA, MedAgentBench, and AgentClinic are valid measures of the healthcare capabilities claimed.
    The central ranking claim rests on these benchmarks' validity; the paper itself says strong scores do not establish safety, so the scores only support narrow benchmark competence.
  • domain assumption The modified AgentClinic tool-native harness and MedAgentBench 'harness bug fix' preserve relative model ordering.
    Appendix A.2 and Table 3 introduce harness changes that can alter difficulty; if they favor Cura 1T, the comparison against baselines is invalid.
  • domain assumption SDFT (Eq. 1) prevents forgetting of base capabilities while consolidating healthcare skills.
    Section 3.1 relies on Shenfeld et al. for this property; the paper offers no ablation isolating SDFT's anti-forgetting effect.
  • domain assumption LoRA adapters suffice to specialize a 1T-parameter base without full fine-tuning.
    Section 3.1; no full-parameter comparison is provided, so the method's ceiling is unknown.

pith-pipeline@v1.3.0-alltime-deepseek · 12647 in / 14817 out tokens · 144299 ms · 2026-08-02T02:12:52.741107+00:00 · methodology

0 comments
read the original abstract

Healthcare spans high-stakes communication, expert reasoning, and workflow execution, yet specialized LLMs that cover these use cases together remain limited. A healthcare model must handle patient consultation, clinical reasoning over text and images, interactive diagnosis, and electronic health record (EHR) tool use. These capabilities fail in different ways, and a narrow update for one task can degrade another. We present Cura 1T, a healthcare-specialized LLM trained through a human-gated self-evolution loop. In each evolution round, a training agent plans a target capability, trains the model, evaluates benchmark trajectories, and refines the data mixture from observed failures. This data-centered loop improves the model through targeted synthetic and curated examples rather than a single generic medical-data update. Across the healthcare evaluation suite, Cura 1T ranks at or near the top among frontier baselines, while remaining competitive on out-of-domain reasoning and agentic benchmarks.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 27 linked inside Pith

  1. [1]

    Healthbench: Evaluatinglargelanguagemodelstowardsimprovedhumanhealth.arXiv preprint arXiv:2505.08775,

    RahulK.Aroraetal. Healthbench: Evaluatinglargelanguagemodelstowardsimprovedhumanhealth.arXiv preprint arXiv:2505.08775,

    Pith/arXiv arXiv

  2. [3]

    Victor Barres, Honghua Dong, Soham Ray, Xujie Si, and Karthik Narasimhan.τ 2-bench: Evaluating conversa- tional agents in a dual-control environment.arXiv preprint arXiv:2506.07982,

    Pith/arXiv arXiv

  3. [5]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    Pith/arXiv arXiv

  4. [6]

    ResearchGym: Evaluating language model agents on real-world ai research.arXiv preprint arXiv:2602.15112,

    Aniketh Garikaparthi, Manasi Patwardhan, and Arman Cohan. ResearchGym: Evaluating language model agents on real-world ai research.arXiv preprint arXiv:2602.15112,

    arXiv

  5. [9]

    What disease does thispatienthave? alarge-scaleopendomainquestionansweringdatasetfrommedicalexams.arXivpreprint arXiv:2009.13081,

    12 Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. What disease does thispatienthave? alarge-scaleopendomainquestionansweringdatasetfrommedicalexams.arXivpreprint arXiv:2009.13081,

    Pith/arXiv arXiv 2009

  6. [13]

    Self-refine: Iterative refinement with self-feedback.arXiv preprint arXiv:2303.17651,

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. Self-refine: Iterative refinement with self-feedback.arXiv preprint arXiv:2303.17651,

    Pith/arXiv arXiv

  7. [14]

    Alexander Novikov, Ngan Vu, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J. R. Ruiz, Abbas Mehrabian, M. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Pushmeet Kohli, and Matej Balog. AlphaEvolve: Acodingagentforscientificandalgorithmicd...

    Pith/arXiv arXiv

  8. [15]

    Healthbench professional: Evaluating large language models on real clinician chats.arXiv preprint arXiv:2604.27470,

    OpenAI. Healthbench professional: Evaluating large language models on real clinician chats.arXiv preprint arXiv:2604.27470,

    Pith/arXiv arXiv

  9. [16]

    MedMCQA:Alarge-scalemulti-subject multi-choice dataset for medical domain question answering.arXiv preprint arXiv:2203.14371,

    AnkitPal,LogeshKumarUmapathi,andMalaikannanSankarasubbu. MedMCQA:Alarge-scalemulti-subject multi-choice dataset for medical domain question answering.arXiv preprint arXiv:2203.14371,

    Pith/arXiv arXiv

  10. [18]

    Agent- clinic: A multimodal agent benchmark to evaluate ai in simulated clinical environments.arXiv preprint arXiv:2405.07960,

    Samuel Schmidgall, Rojin Ziaei, Carl Harris, Eduardo Reis, Jeffrey Jopling, and Michael Moor. Agent- clinic: A multimodal agent benchmark to evaluate ai in simulated clinical environments.arXiv preprint arXiv:2405.07960,

    Pith/arXiv arXiv

  11. [19]

    Self-distillationenablescontinuallearning

    IdanShenfeld,MehulDamani,JonasHübotter,andPulkitAgrawal. Self-distillationenablescontinuallearning. arXiv preprint arXiv:2601.19897,

    Pith/arXiv arXiv

  12. [20]

    Reflexion: Language agents with verbal reinforcement learning.arXiv preprint arXiv:2303.11366,

    13 Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning.arXiv preprint arXiv:2303.11366,

    Pith/arXiv arXiv

  13. [21]

    Yu, and Jianfeng Gao

    Guiyao Tie, Jiawen Shi, Dingjie Song, Yixiao Huang, Ziji Sheng, Xueyang Zhou, Daizong Liu, Pan Zhou, Yongchao Chen, Ran Xu, Lifang He, Qingsong Wen, Manling Li, Cong Lu, Shuai Li, Pengtao Xie, Yixuan Yuan, Rui Meng, Lei Xing, Lichao Sun, Caiming Xiong, Philip S. Yu, and Jianfeng Gao. AutoResearch AI: Towards ai-powered research automation for scientific d...

    Pith/arXiv arXiv

  14. [22]

    Llama 2: Open foundation and fine-tuned chat models

    Hugo Touvron, Louis Martin, Kevin Stone, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288,

    Pith/arXiv arXiv

  15. [23]

    Chain-of-thought prompting elicits reasoning in large language models.arXiv preprint arXiv:2201.11903,

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models.arXiv preprint arXiv:2201.11903,

    Pith/arXiv arXiv

  16. [24]

    The AI scientist-v2: Workshop-level automated scientific discovery via agentic tree search.arXiv preprint arXiv:2504.08066,

    Yutaro Yamada, Robert Tjarko Lange, Cong Lu, Shengran Hu, Chris Lu, Jakob Foerster, Jeff Clune, and David Ha. The AI scientist-v2: Workshop-level automated scientific discovery via agentic tree search.arXiv preprint arXiv:2504.08066,

    Pith/arXiv arXiv

  17. [25]

    Le, Denny Zhou, and Xinyun Chen

    Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V. Le, Denny Zhou, and Xinyun Chen. Large language models as optimizers.arXiv preprint arXiv:2309.03409,

    Pith/arXiv arXiv

  18. [26]

    Scaling relationship on learning mathematical reasoning with large language models.arXiv preprint arXiv:2308.01825,

    Zheng Yuan, Hongyi Yuan, Chengpeng Li, Guanting Dong, Keming Lu, Chuanqi Tan, Chang Zhou, and Jingren Zhou. Scaling relationship on learning mathematical reasoning with large language models.arXiv preprint arXiv:2308.01825,

    Pith/arXiv arXiv

  19. [27]

    differentiation

    Mert Yuksekgonul, Federico Bianchi, Joseph Boen, Sheng Liu, Zhi Huang, Carlos Guestrin, and James Zou. TextGrad: Automatic “differentiation” via text.arXiv preprint arXiv:2406.07496,

    Pith/arXiv arXiv

  20. [28]

    Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah D. Goodman. Star: Bootstrapping reasoning with reasoning. arXiv preprint arXiv:2203.14465,

    Pith/arXiv arXiv

  21. [29]

    Medxpertqa: Benchmarking expert-level medical reasoning and understanding.arXiv preprint arXiv:2501.18362,

    YuxinZuo,ShangQu,YifeiLi,ZhangrenChen,XuekaiZhu,ErmoHua,KaiyanZhang,NingDing,andBowen Zhou. Medxpertqa: Benchmarking expert-level medical reasoning and understanding.arXiv preprint arXiv:2501.18362,

    Pith/arXiv arXiv

  22. [30]

    what medication is EROSTIN 10MG

    14 A Appendix A.1 Training Hyperparameters Table 7 reports the final training hyperparameters of Cura 1T. Hyperparameters for these screening runs vary by capability; the table reports the final consolidation rather than capability-specific screening runs or mixture proportions. Table 7Training hyperparameters for Cura 1T. Configuration Cura 1T Algorithm ...

    2023

  23. [2019]

    Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685,

    EdwardJ.Hu,YelongShen,PhillipWallis,ZeyuanAllen-Zhu,YuanzhiLi,SheanWang,LuWang,andWeizhu Chen. Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685,

    Pith/arXiv arXiv

  24. [2020]

    Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts

    Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T. Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts. DSPy: Compiling declarative language model calls into self-improving pipelines.arXiv preprint arXiv:2310.03714,

    Pith/arXiv arXiv

  25. [2021]

    Black, Gloria Geng, Danny Park, James Zou, Andrew Y

    Yixing Jiang, Kameron C. Black, Gloria Geng, Danny Park, James Zou, Andrew Y. Ng, and Jonathan H. Chen. MedAgentBench: A realistic virtual EHR environment to benchmark medical LLM agents.arXiv preprint arXiv:2501.14654,

    Pith/arXiv arXiv

  26. [2022]

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R. Bowman. Gpqa: A graduate-level google-proof q&a benchmark.arXiv preprint arXiv:2311.12022,

    Pith/arXiv arXiv

  27. [2023]

    HealthAgentBench: A unified benchmark suite of realistic agentic healthcare environments for challenging frontier AI agents

    Qianchu Liu, Sheng Zhang, Guanghui Qin, Jeya Maria Jose Valanarasu, Maximilian Rokuss, Mingyu Lu, Timothy Ossowski, Juan Manuel Zambrano Chaves, Cliff Wong, Peniel Argaw, et al. HealthAgentBench: A unified benchmark suite of realistic agentic healthcare environments for challenging frontier AI agents. arXiv preprint arXiv:2606.31179, 2026a. Ruoqi Liu, Imr...

    Pith/arXiv arXiv

  28. [2024]

    Clinalign: Scaling healthcare alignment from clinician preference.arXiv preprint arXiv:2602.09653,

    Shiwei Lyu, Xidong Wang, Lei Liu, Hao Zhu, Chaohe Zhang, Jian Wang, Jinjie Gu, Benyou Wang, and Yue Shen. Clinalign: Scaling healthcare alignment from clinician preference.arXiv preprint arXiv:2602.09653,

    arXiv

  29. [2025]

    HaolinChen,DeonMetelski,LeonQi,TaoXia,JoonyulLee,SteveBrown,KevinRiley,FrankWang,T.Y.Alvin Liu, Hank Capps, et al.χ-Bench: Can AI agents automate end-to-end, long-horizon, policy-rich healthcare workflows?arXiv preprint arXiv:2605.16679,

    Pith/arXiv arXiv

  30. [2026]

    Baichuan-m4: Aclinical-grademedicalagentsystem for continuous care.arXiv preprint arXiv:2606.08982,

    BaichuanAIandTHUBPMGroup,TsinghuaUniversity. Baichuan-m4: Aclinical-grademedicalagentsystem for continuous care.arXiv preprint arXiv:2606.08982,

    Pith/arXiv arXiv