Selected work

Representative Papers

Routing Divergence Is Not Evidence of Behavioral Influence in Same-Weight MoE Self-Distillation

Aug 16, 2026

This study addresses the controversy regarding the correlation between routing divergence and behavioral impact in equal-weight MoE self-distillation. We propose a block-level decomposition method to disentangle routing from content components, complemented by residual stream exposure analysis and noise-controlled experiments. Results demonstrate that routing shifts do not equate to behavioral changes; their effects are significantly weaker than those induced by natural contexts and can be replicated by matched noise. Consequently, this work reveals the non-behavioral nature of routing divergence and introduces an innovative "exposure-first" evaluation mechanism. By advocating for exposure assessment prior to intervention, we establish a novel paradigm for interpreting MoE internal mechanisms and optimizing self-distillation processes.

0 citationsRead paper

Deceptive Grounding: Entity Attribution Failure in Clinical Retrieval-Augmented Generation

Jul 10, 2026

This study addresses the problem of "deceptive grounding" in clinical retrieval-augmented generation (RAG) systems, wherein models erroneously attribute genuine evidence for one drug to another, producing responses that appear plausible yet are substantively misleading. The work provides the first formal definition and quantitative assessment of this issue, introducing a controlled-factor benchmark to systematically evaluate 13 models across 740 drug–disease pairs. Surprisingly, domain-finetuned models exhibit higher error rates. To mitigate this, the authors propose a high-precision entity attribution verification mechanism combining inverse probability weighting correction with human-curated gold standards, achieving 97.0% precision and 98.7% recall under adversarial conditions. Real-world evaluation reveals an overall deceptive grounding rate of 7.8% in deployed systems—rising to 13.6% for novel drugs—highlighting both the severity of the problem and the efficacy of the proposed solution.

0 citationsRead paper

Metrics or Mirage? An Audit of Evaluation Inconsistencies in Colonoscopy Polyp Segmentation Benchmarks

Jul 09, 2026

This study addresses the lack of standardized evaluation protocols in colonoscopic polyp segmentation, which undermines reliable model comparison. Through a systematic audit of 27 papers, we uncover three critical structural issues: the omission of boundary-sensitive metrics such as Hausdorff distance, inconsistent train/test split protocols, and the absence of statistical significance testing. To rectify these shortcomings, we establish a unified evaluation framework, re-evaluating representative models under multiple protocols with rigorous significance analysis, and propose a domain-specific five-item reporting checklist (PSRC). Our findings reveal that conventional metrics often obscure deficiencies in boundary accuracy and recall, and that the ranking of “best-performing” models varies substantially with evaluation settings—demonstrating the unreliability of current leaderboards and providing a methodological foundation for standardized polyp segmentation assessment.

0 citationsRead paper

TheBioCollection: Unified Pre-Training Scale LLM Corpus for Biology

Jul 09, 2026

This work addresses the limitation of existing biological large language models, which suffer from a lack of unified, structured, high-quality pretraining corpora due to the heterogeneous and fragmented nature of biological data. To overcome this, we construct TheBioCollection, a standardized biological corpus comprising 52.6 billion tokens that systematically integrates multimodal data—including small molecules, proteins, genomes, cells, and pathways—and enriches it with computationally derived biological attributes via a bioinformatics toolchain. We further design novel instruction-tuning tasks targeting known model weaknesses and establish a corresponding evaluation benchmark. Using a fixed Gravity-16B-A3B architecture, our model achieves more than a twofold overall performance improvement on TheBioCollection-Eval, with significant gains across all subdomains while preserving its general language capabilities.

0 citationsRead paper

DALPHIN: Benchmarking Digital Pathology AI Copilots Against Pathologists on an Open Multicentric Dataset

May 05, 2026

This study addresses the lack of open, multicenter benchmarks for evaluating the real-world efficacy of AI assistants in routine digital pathology diagnostics. To this end, the authors introduce DALPHIN, the first open multicenter visual question answering (VQA) benchmark, encompassing 1,236 whole-slide images across 130 diseases and 14 subspecialties from six countries, with performance benchmarked by 31 international pathologists. Employing a blinded ground-truth access protocol, the study systematically evaluates GPT-5, Gemini 2.5 Pro, and the authors’ PathChat+ model under both independent and sequential answering paradigms. PathChat+ achieves expert-level performance in four out of six tasks, significantly outperforming competing models. The complete dataset and evaluation platform are publicly released to establish a reliable benchmark for future pathology AI copilot research.

0 citationsRead paper

Recent publications

Latest Papers

Routing Divergence Is Not Evidence of Behavioral Influence in Same-Weight MoE Self-Distillation

Aug 16, 2026

This study addresses the controversy regarding the correlation between routing divergence and behavioral impact in equal-weight MoE self-distillation. We propose a block-level decomposition method to disentangle routing from content components, complemented by residual stream exposure analysis and noise-controlled experiments. Results demonstrate that routing shifts do not equate to behavioral changes; their effects are significantly weaker than those induced by natural contexts and can be replicated by matched noise. Consequently, this work reveals the non-behavioral nature of routing divergence and introduces an innovative "exposure-first" evaluation mechanism. By advocating for exposure assessment prior to intervention, we establish a novel paradigm for interpreting MoE internal mechanisms and optimizing self-distillation processes.

0 citationsRead paper

Deceptive Grounding: Entity Attribution Failure in Clinical Retrieval-Augmented Generation

Jul 10, 2026

This study addresses the problem of "deceptive grounding" in clinical retrieval-augmented generation (RAG) systems, wherein models erroneously attribute genuine evidence for one drug to another, producing responses that appear plausible yet are substantively misleading. The work provides the first formal definition and quantitative assessment of this issue, introducing a controlled-factor benchmark to systematically evaluate 13 models across 740 drug–disease pairs. Surprisingly, domain-finetuned models exhibit higher error rates. To mitigate this, the authors propose a high-precision entity attribution verification mechanism combining inverse probability weighting correction with human-curated gold standards, achieving 97.0% precision and 98.7% recall under adversarial conditions. Real-world evaluation reveals an overall deceptive grounding rate of 7.8% in deployed systems—rising to 13.6% for novel drugs—highlighting both the severity of the problem and the efficacy of the proposed solution.

0 citationsRead paper

Metrics or Mirage? An Audit of Evaluation Inconsistencies in Colonoscopy Polyp Segmentation Benchmarks

Jul 09, 2026

This study addresses the lack of standardized evaluation protocols in colonoscopic polyp segmentation, which undermines reliable model comparison. Through a systematic audit of 27 papers, we uncover three critical structural issues: the omission of boundary-sensitive metrics such as Hausdorff distance, inconsistent train/test split protocols, and the absence of statistical significance testing. To rectify these shortcomings, we establish a unified evaluation framework, re-evaluating representative models under multiple protocols with rigorous significance analysis, and propose a domain-specific five-item reporting checklist (PSRC). Our findings reveal that conventional metrics often obscure deficiencies in boundary accuracy and recall, and that the ranking of “best-performing” models varies substantially with evaluation settings—demonstrating the unreliability of current leaderboards and providing a methodological foundation for standardized polyp segmentation assessment.

0 citationsRead paper

TheBioCollection: Unified Pre-Training Scale LLM Corpus for Biology

Jul 09, 2026

This work addresses the limitation of existing biological large language models, which suffer from a lack of unified, structured, high-quality pretraining corpora due to the heterogeneous and fragmented nature of biological data. To overcome this, we construct TheBioCollection, a standardized biological corpus comprising 52.6 billion tokens that systematically integrates multimodal data—including small molecules, proteins, genomes, cells, and pathways—and enriches it with computationally derived biological attributes via a bioinformatics toolchain. We further design novel instruction-tuning tasks targeting known model weaknesses and establish a corresponding evaluation benchmark. Using a fixed Gravity-16B-A3B architecture, our model achieves more than a twofold overall performance improvement on TheBioCollection-Eval, with significant gains across all subdomains while preserving its general language capabilities.

0 citationsRead paper

DALPHIN: Benchmarking Digital Pathology AI Copilots Against Pathologists on an Open Multicentric Dataset

May 05, 2026

This study addresses the lack of open, multicenter benchmarks for evaluating the real-world efficacy of AI assistants in routine digital pathology diagnostics. To this end, the authors introduce DALPHIN, the first open multicenter visual question answering (VQA) benchmark, encompassing 1,236 whole-slide images across 130 diseases and 14 subspecialties from six countries, with performance benchmarked by 31 international pathologists. Employing a blinded ground-truth access protocol, the study systematically evaluates GPT-5, Gemini 2.5 Pro, and the authors’ PathChat+ model under both independent and sequential answering paradigms. PathChat+ achieves expert-level performance in four out of six tasks, significantly outperforming competing models. The complete dataset and evaluation platform are publicly released to establish a reliable benchmark for future pathology AI copilot research.

0 citationsRead paper