C-MIG: Multi-view Information Gain-based Retrieval-Augmented Generation for Clinical Diagnosis Reasoning

arXiv:2605.27860 · cs.AI · Submitted 2026-05-27 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "C-MIG: Multi-view Information Gain-based Retrieval-Augmented Generation for Clinical Diagnosis Reasoning".

Jane: The gist:

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So we've covered how this C-MIG framework uses multi-view information gain to guide both retrieval and refinement, and how that leads to better credit assignment compared to older methods. The paper focuses on getting those process rewards right without relying on brittle exact matching.

Jane: The authors of C-MIG are Yuwei Miao, Gen Li, Yunsheng Zeng, Xiandong Li, Yujin Wang, Siyu Chen, Luning Wang, Yunhao Qiao, Junfeng Wang and Jianwei Lv. They’ve shown this framework can outperform RAGRL baselines on both in-domain and out-of-domain medical sets.

Lu: What this means is that instead of just rewarding the final answer, you get a signal guiding the model to perform better retrieval planning and knowledge digestion simultaneously, which is what they call multi-view information gain.

Meng: From a deployment standpoint, if we can achieve this with only three point four percent training overhead and no inference-time cost, that makes it very appealing for production systems where resources are tight.

Lalam: The paper concludes that C-MIG provides continuous information gain rewards suited for the naming ambiguity inherent in clinical diagnosis by leveraging a frozen reference model to compute drift-resistant rewards and employing a multi-subquery strategy to improve knowledge recall coverage.

Tom: So, the main implication is that we can get more reliable signals for complex reasoning tasks in medicine by rewarding the incremental benefit of new information at both the retrieval and refinement stages.

Jane: It suggests that for clinical diagnosis, focusing on how well you plan your search and how well you refine what you find is as important as just having a good final answer.

Conclusion: Tom: So we’ve been looking at C-MIG, which is this new framework for diagnosis that uses multi-view information gain to guide both what you retrieve and how you refine your answer.

Jane: It sounds like they’ve tackled a big problem where old methods gave zero credit for important reasoning steps.

Lu: Exactly. They figured out that instead of just rewarding the final output, you need to reward the incremental benefit of each step—the retrieval and the refinement.

Meng: I’m interested in how they handled that reward signal without it becoming noisy or unstable when you actually run it on real traffic at scale.

Lalam: The core idea is using a frozen reference model to calculate information gain, which gives a stable reward without needing some teacher to check everything manually.

Tom: So the title itself, C-MIG stands for Multi-view Information Gain-based Retrieval Augmented Generation for Clinical Diagnosis Reasoning. It’s quite descriptive.

Jane: It sounds very technical, but what does that mean in plain language for someone listening who isn't a machine learning expert?

Lu: Think of it like this: they break the diagnosis process into two parts—what to search and how to summarize—and they measure how much better each step makes the final answer than the step before it.

Meng: And that means you get a reward signal for planning your search well, not just for getting lucky with a single document hit.

Tom: The results are pretty strong though. They show C-MIG beats other RAG methods on both in-domain and out-of-domain medical tests, even when the questions are totally new.

Jane: That’s significant because clinical diagnosis is so messy; you need models that generalize well beyond just the specific training data they saw.

Lalam: They even show a high correlation between how accurate the final diagnosis is and how often they find the right documents, which suggests their reward signal is actually aligned with what matters.

Tom: It seems like this moves RAG from just being a retrieval tool to being an active reasoning supervisor for complex medical tasks.

Jane: It certainly does; it’s shifting the focus from just getting the right document to ensuring the entire sequence of actions—search and summarize—is medically sound.

Lu: This approach of using those two complementary views, retrieval quality and refinement quality, seems like a solid way to supervise heterogeneous reasoning capabilities in medical contexts.

Yuwei Miao, Gen Li, Yunsheng Zeng, Xiandong Li, Yujin Wang, Siyu Chen, Luning Wang, Yunhao Qiao, Junfeng Wang

Baidu Inc.

cs.AI

Submitted: 2026-05-27

Updated: 2026-10-14

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 91/100

The gist: The gist: C-MIG is a Multi-view Information Gain-based retrieval-augmented generation framework for Clinical diagnosis that estimates information gain under two complementary views to jointly guide

Key concepts

Information Gain (IG): Information gain measures how much a new piece of information reduces the model's uncertainty about the correct answer. C-MIG calculates this by comparing the logit probability of the answer with and without new retrieved documents or refined summaries, providing a unified reward signal that indicates diagnostic benefit.
Multi-View Reward Views: C-MIG uses two distinct views to supervise reasoning: one view (C-MIG-D) assesses how much the retrieval quality improves between successive steps. The second view (C-MIG-R) assesses how much the model's summary refines its understanding of the query alone, ensuring both retrieval and refinement are rewarded.
Frozen Reference Model: A frozen reference model is a stable, pre-trained LLM used to score information gain. By keeping this scorer fixed, C-MIG generates reproducible rewards that are not dependent on the constantly changing policy model's performance, effectively mitigating reward drift and providing reliable feedback.
Multi-Subquery Strategy: This strategy decomposes complex patient questions into several smaller, focused subqueries separated by semicolons. The model retrieves the best document for each subquery and then refines its knowledge based on these focused retrievals, improving overall information coverage.

Terminology

Information Gain — 3× in this paper · explained in 2 episodes
Benchmark — 1× in this paper · explained in 2 episodes
Credit assignment — 1× in this paper · explained in 3 episodes
Grounding — 1× in this paper · explained in 3 episodes
RAG — 1× in this paper · explained in 3 episodes
Reinforcement Learning — 1× in this paper · explained in 4 episodes

Summary

The gist: C-MIG is a Multi-view Information Gain-based retrieval-augmented generation framework for Clinical diagnosis that estimates information gain under two complementary views to jointly guide what to retrieve and how to refine, alleviating issues of valuable reward signal loss and credit assignment.

Problem Addressed

Existing methods in clinical diagnosis often rely on exact-match binary rewards, which lead to two critical issues: (i) semantically relevant but non-verbatim steps receive zero signal, discarding valuable learning signals; and (ii) uni-dimensional rewards cannot effectively supervise heterogeneous reasoning capabilities. Hard process rewards discard valuable medical signals because they rely on exact string matching, assigning zero signal to intermediate steps that do not contain the ground truth verbatim. Furthermore, uni-dimensional signals lack multi-capability supervision, resulting in suboptimal credit assignment across heterogeneous reasoning steps.

C-MIG Framework and Information Gain

C-MIG formulates information gain as a unified process reward that captures two complementary aspects: (1) the incremental benefit of newly retrieved documents over the previous turn’s retrieval, and (2) the contribution of the model’s refined summary relative to the raw query alone. This mechanism is instantiated using a frozen reference model to compute logit probability scoring functions, defining unified information gain as ∆IG = LogitP(a Iextra) − LogitP(a Ibase). A positive ∆IG indicates that the new information reduces diagnostic uncertainty, while a non-positive value indicates noise introduction.

Multi-View Reward Views

C-MIG employs two complementary views to address the credit assignment problem: (1) C-MIG-D evaluates cross-turn retrieval quality by measuring information gain between successive retrieval rounds, defined by ∆(Dk) = LogitP(a q, Dk)−LogitP(a q, Dk−1). The reward is mapped via a truncated tanh function to avoid introducing noise in the first round and to handle negative gains. (2) C-MIG-R evaluates intra-turn refinement quality by measuring the information gain of knowledge summarization, defined as ∆(srefine) = LogitP(a q, srefine) − LogitP(a q). This is rewarded using a batch-rank reward mapping strategy to dynamically adjust the threshold according to batch difficulty.

Multi-Subquery Strategy and Training Objective

The framework decomposes colloquial patient inquiries into semantically complementary, medically standardized subqueries, prompting the model to generate up to 3 focused sub-queries separated by ";", each retrieving the single most relevant document. Following AutoRefine, the model summarizes and refines these documents in and provides a final diagnosis in tags. The overall reward is structured as Rtotal = Rformat + Rdiag + Rdoc + Rrefine, employing a non-linear composition strategy where the total reward reduces to process rewards when the diagnosis is incorrect, providing process-level feedback to guide the model toward better retrieval and refinement behaviors.

Experimental Results and Performance

Comprehensive experiments on four medical benchmarks demonstrate that C-MIG achieves state-of-the-art performance among all RAGRL methods on both in-domain and out-ofdomain sets, surpassing the strongest sparse-reward variant AutoRefineHardSearch by 9.30 points on MedDDx-Plus. C-MIG consistently outperforms IGPO by 6.71 points in in-domain performance and achieves the best out-ofdomain generalization. Furthermore, C-MIG achieves a Pearson correlation coefficient of r = 0.994 between diagnostic accuracy and document hit rate, indicating that the process reward effectively aligns intermediate retrieval behavior with the ultimate diagnostic objective.

Training Stability and Scalability

Training stability analysis shows that C-MIG maintains a gradient norm mean of 1.012 with a moderate coefficient of variation of 0.181, which is lower than IGPO’s highest CV of 0.250, confirming that computing information gain with the continuously updated policy model introduces gradient noise due to reward drift. Sequence-level rewards consistently outperform turn-level ones, yielding overall average gains of +0.84 / +1.28 on Qwen2.5-3B and +0.16 / +23.34 on Qwen2.5-7B for IGPO and C-MIG-D respectively. Scalability experiments confirm that introducing information gain at the retrieval-document and document-refinement levels yields overall performance gains of 2.04 and 1.39, validating the scalability of the proposed multi-view information gain mechanism.

Conclusion

C-MIG is a reliable RAG-RL solution for clinical diagnosis that leverages a frozen reference model to compute drift-resistant rewards and employs a multi-subquery strategy to improve knowledge recall coverage, achieving state-of-the-art performance across diverse medical benchmarks. The method successfully mitigates reward drift by using the frozen reference model, and its multi-view information gain mechanism effectively supervises both retrieval planning and knowledge digestion. The addition of each component—w Doc, w Refine, and Subquery—improves overall performance across diverse models. The study concludes that C-MIG provides continuous information gain rewards suited to the naming ambiguity inherent in clinical diagnosis.

Limitations

This work is subject to several limitations. Our method relies on a frozen reference model, the evaluation is restricted to text-only medical question answering and diagnostic dialogues without incorporating multi-modal evidence such as medical images, and the benchmarks used may not fully reflect the diversity of patient populations across different regions, languages, and healthcare systems. Accordingly, the experimental results presented in this paper should be interpreted as methodological evidence for the effectiveness of retrieval-augmented diagnostic reasoning rather than as evidence of readiness for clinical deployment. Future work will focus on scaling C-MIG to larger and multi-modal foundation models, conducting broader cross-institutional and cross-lingual validation, and incorporating evaluations by professional clinicians together with safer evaluation protocols.

References

Baichuan. 2023. Baichuan 2: Open large-scale language models. arXiv preprint arXiv:2309.10305.

Junying Chen, Xidong Wang, Anningzhe Gao, Feng Jiang, Shunian Chen, Hongbo Zhang, Dingjie Song, Wenye Xie, Chuyi Kong, Jianquan Li, Xiang Wan, Haizhou Li and Benyou Wang. 2023. Huatuogptii: one-stage training for medical adaption of llms. Preprint arXiv:2311.09774.

Mingyang Chen and 1 others. 2025a. Research: Learning to reason with search for llms via reinforcement learning. In Advances in Neural Information Processing Systems (NeurIPS).

Qiang Chen and 1 others. 2025b. Lets: Learning to think-and-search via process-and-outcome reward hybridization. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP).

Yiqun Chen and 1 others. 2025c. Improving retrievalaugmented generation through multi-agent reinforcement learning. In Advances in Neural Information Processing Systems (NeurIPS).

Clément Christophe, Praveen K Kanithi, Tathagata Raha, Shadab Khan, and Marco AF Pimentel. 2024. Med42-v2: A suite of clinical llms. arXiv preprint arXiv:2408.06142.

DeepSeek-AI. 2026. Deepseek-v4: Towards highly efficient million-token context intelligence.

Chengfeng Dou, Chong Liu, Fan Yang, Fei Li, Jiyuan Jia, Mingyang Chen, Qiang Ju, Shuai Wang, Shunya Dang, Tianpeng Li and 1 others. 2025. Baichuanm2: Scaling medical capability with large verifier system. arXiv preprint arXiv:2509.02208.

Arsene Fansi Tchango, Rishab Goel, Zhi Wen, Julien Martel, and Joumana Ghosn. 2022. Ddxplus: A new dataset for automatic medical diagnosis. Advances in neural information processing systems, 35:31306–31318.

Yifan Feng, Hao Hu, Shihui Ying, Xingliang Hou, Shiquan Liu, Mingyuan Yang, Junchang Li, Shaoyi Du, Nanning Zheng and 1 others. 2026. Hyperrag: Combating llm hallucinations using hypergraphdriven retrieval-augmented generation. Nature Communications.

Improvements for AI systems

Bold header: Multi-view Information Gain Reward Mechanism

The C-MIG framework "formulates information gain as a unified process reward that captures two complementary aspects: (1) the incremental benefit of newly retrieved documents over the previous turn’s retrieval, and (2) the contribution of the model’s refined summary relative to the raw query alone. This jointly supervises “what to retrieve” and “how to refine, alleviating the issues of valuable reward signal loss and credit assignment.

Bold header: Frozen Reference Model for Reward Stability

The system employs a frozen reference model for all reward computations, fundamentally eliminating reward drift by computing information gain via the log-probability scoring function: "LogitP(a I) = 1/L Σ l=1 log πref(al I, a<l). This provides a time-invariant anchor, guaranteeing reward consistency throughout training."

Bold header: Multi-Subquery Retrieval Strategy

The framework introduces a strategy to improve knowledge recall under ambiguous inquiries by prompting the model to generate k semantically complementary subqueries, each retrieving the single most relevant document. This decomposes colloquial queries into medically standardized subqueries to ensure multi-facet evidence coverage.

Bold header: Non-linear Outcome Reward Composition

The system adopts a strategy where reward composition is conditional: when the diagnosis is correct, the total reward is computed as Rtotal = Rformat+Rdiag+Rdoc, bypassing other process rewards to prevent over-crediting intermediate steps when the final outcome is already successful. This design prioritizes final answer correctness while still fostering robust retrieval and refinement capabilities.

Bold header: Dynamic Threshold for Refinement Reward

The C-MIG-R reward uses a batch-rank reward mapping strategy where it rewards only the upper half of positive-gain samples, defined by the median gain: "reward only the upper half: R(i)refine = wr, if ∆(srefine)(i) ≥ ∆˜ + > 0, otherwise 0, which dynamically adjusts the threshold according to batch difficulty."

Bold header: Sequence-Level Process Reward Aggregation

The paper demonstrates that Sequence-level approaches consistently outperform turn-level ones in stability and performance, as aggregated signals provide a more stable optimization target than dense turn-level rewards, especially for handling high-variance medical knowledge distributions.