acceptodds
Under review as a conference paper at ICLR 2027

Context0: Zero-Label Self-Evolution for Context-Aware Clinical Dialogue

Abstract

Clinical dialogue is inherently context-dependent: an appropriate response must account for patient-specific constraints and healthcare settings, recognize when consequential information is missing, and avoid unnecessary clarification when the available context is sufficient. Developing these behaviors requires training experiences that capture not only medical knowledge, but also when and how that knowledge should be applied. We introduce **Context0**, a zero-label self-evolution framework that develops context-aware clinical dialogue capabilities without a pre-existing clinical dialogue dataset or example-level human annotations. Starting from zero curated training examples, a Teacher constructs context-sensitive dialogue tasks, while a Doctor learns from case-specific rubric feedback. A medical knowledge graph provides knowledge anchors for task construction, and feedback from the evolving Doctor guides the Teacher toward new learning opportunities. Alternating Teacher and Doctor updates jointly evolve the training curriculum and the response policy. On the Global Health subset of HealthBench, Context0 achieves a context-awareness score of 62.3%, exceeding GPT-5 by 3.2 percentage points in our evaluation. It also achieves full-rubric scores of 57.9% and 62.7% on the unclear- and clear-context subsets, respectively. These results highlight the potential of zero-label self-evolution to develop context-aware medical assistants through self-generated curricula and feedback. We will release the code and curriculum at https://anonymous.4open.science/r/Context0-2878.

Primary area:
foundation or frontier models, including LLMs
Keywords:
AI agent, reinforcement learning, clinical reasoning, medical question answering

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_4wWKOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents Context0, a zero-labeled self-evolutionary framework for context-aware clinical conversations. This method employs a Teacher–Doctor collaborative evolution mechanism: The Teacher uses the biomedical knowledge graph to generate multiple rounds of clinical conversations and dynamic scoring criteria, while the Doctor optimizes learning through the Rubric-based GRPO to learn how to utilize the existing context, identify missing key information, and avoid unnecessary inquiries.

Critical strengths

The method focuses on a practical but often overlooked issue in clinical conversations: not only do models need medical knowledge, they must also judge when to use existing information and when it requires further clarification.Combining context-sensitive task generation, dynamic scoring, and Teacher–Doctor co-evolution provides a valuable technical path to reduce reliance on manual annotation.

Critical weaknesses
  1. Context0 mainly integrates existing Teacher Student self evolution, knowledge graph guided synthetic data generation, dynamic Rubric evaluation, and GRPO optimization. Compared with related works such as Agent0, R-Zero, Socratic Zero, and InfiMed ORBIT, the paper has not fully explained what irreplaceable new contributions its key mechanisms bring at the algorithm level. For ICLR, simply combining existing components into clinical dialogue tasks may not be sufficient to demonstrate significant technological innovation of the method.

  2. Teacher-Doctor co-evolution is the most important technical claim of this paper. However, the existing ablation studies mainly focus on multi-round context, security filtering, and Rubric densification. The paper does not directly compare the effects of fixed Teacher, adaptive Teacher, static synthetic data training, and the complete bidirectional evolution mechanism. It also lacks performance changes across different evolution rounds. Therefore, it is currently impossible to determine whether the performance improvement comes from the true collaborative evolution, or merely from the larger-scale synthetic data and Rubric optimization. This directly weakens the causal evidence of the core contribution.

  3. Although the paper compared Agent0, R-Zero and several general language models, it did not adequately report the synthetic sample size, training computational budget, and optimization settings of each training method for comparability. In particular, there was a lack of standard GRPO, static Rubric RL, and non-adaptive curriculum training baselines that used the same base model and the same synthetic data scale. Given that Context0 achieved a 37.5 percentage point improvement in context-awareness compared to Qwen3-4B-Instruct-2507, it is necessary to reasonably control training resources and data differences.

  4. The main results are only from the Global Health subset of HealthBench and are not sufficient to demonstrate that the method generalizes to different diseases, healthcare systems, and real rounds of clinical interactions.More importantly, the paper emphasizes that the model can judge when clarification is needed and when a direct answer should be answered, but it lacks independent quantitative results for behaviors such as the need for clarification, redundant questioning, and incorrect contextual assumptions.

  5. The training tasks, scoring criteria, and feedback are highly dependent on automatic generation and model judgment. Although the knowledge graph provides structured medical evidence, it cannot guarantee that the generated clinical scenarios, recommendations, and Rubrics are completely correct. The appendix only describes the disease distribution and scoring criteria composition of 77 generated cases, lacking the factual correctness, Rubric consistency, and proportion of potentially harmful recommendations independently evaluated by clinical experts. Since the model directly optimizes based on these automated signals, systematic errors in the synthetic data may be repeatedly reinforced, affecting the credibility and clinical safety of the training results.

  6. The paper lists knowledge graph constraints and zero-annotation training as important features, but Appendix A.5 clearly states that the No-KG experiment only removed the knowledge graph triplet in the training anchor point and failed to confirm that all KG-derived Rubric templates were also disabled.Therefore, this experiment cannot completely isolate the role of the knowledge graph.In addition, Appendix A.2 explains that Zero-Label does not exclude manually designed templates and rules. Although this definition is reasonable in itself, the paper does not quantitatively analyze the impact of these artificial priors on performance.

  7. There is a significant misrepresentation in Section 5.1 regarding the ablation results of Figure 4. For example, after removing multiple rounds of context, the clear context score decreased from 61.9% to 53.0%, an actual decrease of 8.9 percentage points, while the main text report decreased by 18.9 percentage points; Removing Rubric densification corresponds to a decrease of 18.9 percentage points. In addition, after removing multiple rounds of context, the unclear context score increased from 50.3% to 50.4%, rather than the decrease described in the main text.

Questions for authors
  1. I particularly hope to see a comparison between the fixed Teacher and no Teacher updates, static course training, and complete co-evolution. It would be ideal to match the training data size and computing budget as well. This will directly affect my judgment on the innovation and core contribution of the method.

  2. It is hoped to understand the error rate of medical facts in generating scenarios, the consistency between the Rubric judgment and the clinical expert evaluation, as well as the specific performance of the model in providing dangerous suggestions, omitting critical warnings, and unnecessary clarifications. This will influence my assessment of the credibility of the automated training signals.

  3. In particular, it is necessary to confirm whether No-KG retains the KG-derived templates, to state the training budget of the main baseline, and to correct the inconsistent values in Section 5.1 and Figure 4.

Reviewer_dB8bOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Context0, a zero-label self-evolution framework for context-aware clinical dialogue. Context0 employs a Teacher–Doctor co-evolution process in which the Teacher generates clinical dialogue tasks grounded in a biomedical knowledge graph, while the Doctor learns through case-specific rubric feedback. The framework combines learner-dependent curriculum selection, dynamic rubric generation, and GRPO to improve the use of available context and the identification of missing information. Experiments on the Global Health subset of HealthBench show that Context0 outperforms the compared baselines, including GPT-5, on context-awareness and context-specific response-quality metrics. The paper also provides component ablations and analyses of knowledge-source sensitivity.

Critical strengths

S1. The paper addresses an important problem in context-aware clinical dialogue. Context0 focuses on whether an agent appropriately uses available patient information, identifies consequential missing context, and avoids unnecessary clarification. This goes beyond conventional medical question answering and addresses a meaningful challenge in developing reliable clinical dialogue agents.

S2. The experimental results are encouraging. Context0 achieves the highest reported context-awareness score in Table 1, outperforming GPT-5 and the evaluated self-evolution baselines. The component ablations in Figure 4 and knowledge-source comparisons in Table 2 provide additional evidence for the effectiveness of several design choices.

Critical weaknesses

W1. The contribution of the core Teacher–Doctor co-evolution mechanism is not sufficiently examined. Context0 combines Teacher optimization, learner-dependent curriculum selection, dynamic rubric generation, and rubric-based GRPO. However, the ablations in Section 5.1 mainly investigate multi-turn context, selected safety gates, and rubric densification. They do not directly isolate the contributions of Teacher optimization, Doctor-feedback-driven curriculum selection, or alternating Teacher–Doctor updates. As a result, it is unclear whether the improvements primarily come from the proposed co-evolution mechanism or from training on synthetic clinical dialogues with rubric-based rewards. Comparisons with a fixed Teacher, a non-adaptive curriculum, or Doctor-only training would help clarify the contribution of the core framework.

W2. The comparison with existing self-evolution and rubric-based learning methods is limited. Section 4.1 includes Agent0 and R-Zero as self-evolution baselines, but the paper does not clearly explain how these methods were adapted to the clinical dialogue setting or whether they received comparable training data and computational budgets. Furthermore, InfiMed-ORBIT [1], which also employs dynamic rubrics and rubric-guided reinforcement learning for complex open-ended tasks, is discussed in Section 2.2 but not included in the experiments. A more closely matched rubric-based baseline, or a clearer explanation of why such a comparison is not feasible, would help assess the specific contribution of Context0. [1] Wang, Pengkai, et al. "InfiMed-ORBIT: Aligning LLMs on Open-Ended Complex Tasks via Rubric-Based Incremental Training." arXiv preprint arXiv:2510.15859 (2025).

W3. The evaluation of interactive context-aware dialogue remains limited. Section 4.1 evaluates Context0 on HealthBench by generating a single response for each provided conversation history. While this measures context-sensitive response quality, it does not directly assess whether the model can effectively seek clarification and adapt its subsequent responses to newly provided information. Since these behaviors are central motivations of Context0, additional multi-turn interaction experiments would provide stronger evidence for its effectiveness. Furthermore, the reported 3.2-percentage-point improvement over GPT-5 in Table 1 lacks uncertainty estimates or statistical significance tests, making the reliability of this relatively modest gain difficult to assess.

Questions for authors

Q1. How much of Context0's improvement comes specifically from Teacher–Doctor co-evolution? Could the authors provide ablations with a fixed Teacher, a non-adaptive curriculum, or Doctor-only rubric-based training?

Q2. How were Agent0 and R-Zero adapted to the clinical dialogue setting, and were the baselines trained with comparable data and computational budgets? How does Context0 compare with a closely related rubric-based approach such as InfiMed-ORBIT?

Q3. Can the authors provide additional evaluations of interactive clarification behavior, particularly whether the Doctor asks appropriate follow-up questions and incorporates newly provided information?

Reviewer_WU6cOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Goal: To build a context-aware medical dialogue system without needing curated data or human annotation/intervention

Details:
The paper argues the importance of context in the medical dialogue setting. Specifically, during user/patient interaction, an agent/doctor must employ the clinical knowledge in context-aware fashion, i.e., it should ground on the available context and avoid requesting repeating clarifications, identify consequentially missing context and determine clarifications that would not change the answer.

To this end, the paper proposes Context0, a data-free co-evolution approach for building a medical dialogue system. On a high-level, a teacher agent proposes tasks and case-specific rubrics anchored on a given medical knowledge graph. A doctor agent then conducts a simulated dialogue with the user based on the teacher’s proposals. Then, a judge evaluates the dialogue based on the proposed rubrics. This feedback is then used to adapt both the doctor agent and the teacher agent using RL based training.

The paper benchmarks Context0 on the Global Health subset of HealthBench dataset. The results indicate that Context0 agent improves performance on the benchmark, surpassing proprietary models such as GPT-5.

Critical strengths

I like how Context0 adapts a self-improving loop to a medical dialogue setting targeting the common failure modes such as repeated questioning and asking inconsequential clarifications. Surprisingly, Context0 improves greatly even with the model as small as Qwen3 4B Instruct, which is used as backbone and judge. Further, I appreciate using case specific rubrics making training more focused.

Critical weaknesses
  1. The scope of the work is narrow and so is the evaluation. Context0 improves upon untrained baselines for the context aware response generation task. But it is not clear whether that translates into overall improvement in the medical dialogue abilities such asking the correct clarification over turns or reaching the correct diagnostic in the end. The approach is educational but how can this be extended beyond context-awareness into a more realistic medical dialogue setting is unclear. Can authors test this on any other dataset?
  2. It is concerning that the baselines such as R-Zero and Agent0 do not improve much over the instruct baseline in table 1. Authors need to discuss how these baselines are trained and why there is no gain. Is there an adaptation issue? Providing this information would strengthen Context0 position over baselines.
  3. Some ideas in the paper are explored in literature. For instance, AMIE [1] has outer and inner loops for data generation and model improvement, with both loops iteratively improving one another. Context0 has the benefit of zero-data training, but beyond that approach is similar to AMIE. I request authors to clarify their novelty against this approach.
  4. Paper does not provide full hyperparameters and training details. How many steps of teacher and student updates are performed. What is the training setup and hyperparameters? The anonymous github link provides some but not all details.
  5. As suggested in the paper title and the introduction, Context0 system is aimed at context-aware dialogues. But section 3 develops the Context0 framework around (history, response) pairs. It is not clear whether complete dialogue, from beginning to end, is ever generated during the training. Thus, I am skeptical whether resulting system can hold mult-turn conversation reliably. For my benefit, I request the authors to provide one-two examples where the learned doctor agent conducts an entire session with the simulated patient.
  6. Section 3 is quite dense and hard to follow. I suggest using a running example in figure 2 to better explain Context0. For instance, self-pool RAG and densify operation in figure 3 feel intuitive in spirit but are ill-motivated in the text. Also, a concrete example for what does a strata hold and how are the three contextual targets incorporated in it could help understand the approach better. Same goes for the rubric. Please provide examples.
  7. I could not find any qualitative or human evaluation studies to verify Context0 gains.

[1] Tu, T., Schaekermann, M., Palepu, A. et al. Towards conversational diagnostic artificial intelligence. Nature 642, 442–450 (2025). https://doi.org/10.1038/s41586-025-08866-7

Questions for authors

I re-iterate the need for improving section 3 of the paper with examples.

  1. What does context sensitive mean? What are patient constraints? The paper states in the related work section (121-123, 141-143) that the learner may not use context correctly, assume missing information or ask repeated questions. Can the authors provide some examples to clearly substantiate these behaviors?
  2. Terms in equation 2 are very unclear. How are you computing a reward based on the band shaping? Assume that the doctor agent is fixed and the teacher agent is being optimized. Can authors elaborate what exactly is the teacher striving for during this phase? Is the goal of the teacher to produce data in the acceptable band? If so, can you motivate this? Why would you choose a band of 0-0.75?
  3. Why are pre-trained only models Qwen3 4B Base and Qwen3 8B Base part of the baseline? I do not expect them to even follow the instructions in the prompt.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.