ACL 2025 workshop notes on KG and scientific AI in ClimateNLP, Knowledgeable-LM and BioNLP

Overview of ACL 2025 and Its Relevance to Scientific AI

The annual meeting of the Association for Computational Linguistics (ACL) has long served as a premier forum for advancements in natural language processing (NLP) and related fields. As we approach ACL 2025, the workshop agenda reveals a remarkable focus on the intersection of scientific artificial intelligence (AI), knowledge graph (KG) construction, and the nuanced application of NLP in scientific domains such as climate science and biomedicine. This reflects a growing recognition of the pivotal role these methodologies play in catalyzing real-world scientific discovery and driving impactful change.

ACL’s relevance to scientific AI is especially pronounced as researchers and practitioners seek innovative ways to harness unstructured text, transform it into actionable insights, and scaffold the advancement of scientific knowledge. The conference workshops such as ClimateNLP, Knowledgeable-LM, and BioNLP target key challenges in extracting, representing, and reasoning over knowledge from scientific literature, which is both vast and exceedingly complex.

One step toward this goal is the development and refinement of knowledge graphs—structured frameworks that map relationships between scientific concepts, datasets, and findings. KGs are foundational in organizing information to unlock new hypotheses and accelerate scientific collaboration. Workshops at ACL emphasize best practices in KG construction, including integration with large language models (LLMs) for improved comprehension and inference. For example, building domain-specific KGs for climate science can help researchers identify patterns in climate data that might otherwise remain hidden.

The relevance also extends to improving the factuality and explainability of scientific AI systems. Recent progress in language modeling has led to the emergence of Knowledgeable-LMs, systems that integrate external scientific resources, databases, and ontologies for more accurate interpretation and generation of scientific texts. These advancements not only improve the reliability of AI-driven scientific assistants, but also support transparent knowledge transfer—critical for high-stakes domains like healthcare and climate policy.

Case in point: BioNLP and related tracks showcase recent tools and benchmarks for parsing biomedical literature, extracting drug-gene interactions, and supporting hypothesis generation in life sciences. Applied examples include leveraging machine reading to rapidly synthesize new COVID-19 research or using automated summarization tools to track trends in genomic medicine, all made possible by advanced NLP techniques discussed at ACL.

Ultimately, ACL 2025 is shaping the future of scientific AI by bridging methodological innovations with the pressing needs of science communities. Its curated workshops and collaborative initiatives reflect a broadening commitment to responsible, domain-aware AI systems that make a tangible impact in the fight against global challenges—from climate change to public health.

Key Themes Explored in the ClimateNLP Workshop

The ClimateNLP workshop at ACL 2025 brought together leading experts from linguistics, AI, and climate science to explore the intersection of language technologies and climate-related research. The workshop highlighted several key themes that are shaping the future of scientific AI, knowledge graphs, and natural language processing (NLP) applications for environmental and climate science.

1. Knowledge Graph Construction for Climate Science

A core focus was the automated construction and expansion of knowledge graphs to systematize and connect vast climate science datasets and literature. Participants explored approaches for extracting entities, relationships, and events directly from scientific texts, enabling linkage of facts about greenhouse gases, extreme weather events, and ecological impacts. For example, using transformer-based models, researchers built knowledge graphs that map how climate change impacts agricultural yields or polar ice melt, supporting both hypothesis generation and policy-making.

Attendees discussed workflows for aligning these graphs with global ontologies such as the UNEP Knowledge Graph, emphasizing the importance of interoperability and data provenance. Strategies for scaling updates with continual ingestion of new research were also emphasized, ensuring knowledge graphs remain up to date with rapidly evolving climate science literature.

2. Advanced Language Modeling for Climate-Specific Tasks

The workshop showcased innovations in customizing large language models (LLMs) for climate research analysis, scenario generation, and technical translation. ClimateNLP teams demonstrated the adaptation of domain-tuned LLMs to answer climate questions, summarize dense scientific texts, and simulate likely impacts of policy interventions using historical weather and emissions data.

Experts provided step-by-step walkthroughs of fine-tuning LLMs with corpora from NASA and IPCC reports, resulting in AI agents that interpret technical climate findings for both policymakers and the public. They also discussed bias mitigation strategies—crucial since climate discourse can be politicized—and outlined ongoing benchmark efforts to rigorously test model reliability in climate-specific applications.

3. Enhancing Interdisciplinary Collaboration through NLP

Another recurring theme was the role of NLP as a bridge between climate science, policy, and public engagement. Workshop speakers highlighted tools that use NLP for automatic translation of climate research into multiple languages and plain-language summaries, increasing accessibility for global audiences. For example, researchers illustrated automated workflows for generating climate risk assessments from scientific publications, bolstering public information campaigns.

Examples included annotation tools for tagging critical evidence in climate papers, and dashboards powered by NLP to monitor climate misinformation across online platforms. These initiatives supported the workshop’s broader vision: leveraging language technologies to break disciplinary silos and promote collaborative, informed action in response to the climate crisis.

4. Benchmarking, Evaluation, and Open Science

The workshop emphasized the growing need for robust benchmarking datasets and transparent evaluation of ClimateNLP tools. Discussions centered on shared tasks and evaluation criteria to compare entity linking, relation extraction, and summarization methods, referencing open datasets such as those from Allen Institute for AI and World Bank Climate Data.

Participants reaffirmed commitments to open science, with most teams releasing code and datasets to the public, enabling broader application and scrutiny. Collaborative efforts were encouraged to standardize evaluation metrics and foster reproducibility—ensuring that insights from NLP and AI can be relied upon by scientists, policymakers, and the general public alike.

Advancements in Knowledge Graphs for Climate Science Applications

Knowledge Graphs (KGs) are increasingly recognized as a transformative technology for climate science, enabling the aggregation, structuring, and analysis of heterogeneous data sources essential for addressing complex climate-related challenges. Recent advancements have demonstrated that integrating knowledge graphs into climate informatics can bridge critical data and knowledge gaps, paving the way for more accurate forecasts and policy-relevant insights.

One of the most prominent applications involves the curation and linking of disparate climate datasets, including satellite imagery, sensor observations, scientific publications, and open data resources. For example, projects like Global Change Information System (GCIS) have pioneered efforts by providing interoperable KGs that connect research outputs, datasets, and metadata, ensuring that climate-related facts and sources are easily discoverable, verifiable, and reusable.

New techniques in graph construction, such as automated entity recognition and relationship extraction powered by natural language processing (NLP), are dramatically accelerating the population and evolution of climate-related KGs. For instance, using tools like SciSpacy, which is designed for extracting entities from scientific text, researchers can efficiently mine large corpora of climate literature to build semantic representations that inform both experts and AI models. This process enhances the creation of interconnected ontologies—standardized vocabularies essential for cross-disciplinary understanding and machine learning applications.

KG-powered climate applications have made notable strides in both research and policy contexts. For policymakers, interactive platforms built on KGs, such as the World Bank Climate Knowledge Portal, enable data-driven decision making by offering integrated views of climate risks, adaptation strategies, and socioeconomic impacts. These knowledge frameworks facilitate transparent analytics, traceable evidence chains, and hypothesis generation for climate interventions.

Furthermore, by leveraging graph-based AI methods, such as Graph Neural Networks (GNNs), researchers can perform predictive modeling on multi-modal climate data, revealing causal relationships and emergent trends that would be difficult to discern with traditional tabular data analysis. A notable example is the use of KGs for climate event linking, where extreme events (like droughts or floods) are systematically associated with potential precursors and impacts, supporting more effective disaster preparedness and response efforts.

Looking forward, the fusion of KGs with Large Language Models (LLMs) and domain-specific ontologies continues to position knowledge graphs as an indispensable infrastructure for explainable, actionable, and trustworthy ClimateNLP systems. As collaborative efforts expand—such as through workshops at ACL 2025—the climate science community is poised to benefit from even richer knowledge ecosystem tools to advance sustainability and resilience research worldwide.

Innovative Approaches in Knowledgeable-LM for Scientific Research

The landscape of scientific research is rapidly evolving with the adoption of advanced Language Models (LMs) that are enriched with domain-specific knowledge. These “Knowledgeable-LMs” are not merely text predictors—they function as dynamic partners in scientific discovery, capable of reasoning, interpreting, and assisting in the generation of new hypotheses. Below, we explore key innovations in this space, their real-world implications, and practical examples from climate science and biomedical research.

Integration of Structured Knowledge Graphs into LMs

One major breakthrough is the integration of Knowledge Graphs (KGs) into language models. KGs organize scientific concepts, relationships, and experimental data, transforming LMs from static text generators into contextual information hubs. For example, in climate science, KGs linking datasets on greenhouse gas emissions, temperature anomalies, and ecosystem responses enable models to offer nuanced answers that reference real scientific findings. This structured foundation mitigates the risk of hallucinations and ensures that generated insights are grounded in vetted, up-to-date scientific evidence as described in MIT’s research on integrating KG data with LMs.

Fine-Tuning LMs for Domain-Specific Reasoning

Generic language models often struggle with the special terminology, implicit assumptions, and data complexities of scientific domains. Recent approaches deploy domain-specific fine-tuning, where LMs are retrained on curated corpora from fields like climate science or biomedical research. This process involves:

• Curating and annotating large-scale datasets relevant to a specific scientific problem.
• Supervised training to help the model learn the unique language and logic of the field.
• Rigorous evaluation against expert-annotated benchmarks (e.g., peer-reviewed papers).

As a result, researchers now use knowledgeable LMs to synthesize literature, extract implicit knowledge, and even prioritize novel research directions—capabilities highlighted by Stanford’s pilot projects on scientific AI literacy.

Contextual Retrieval and Enhanced Explainability

Innovative architectures pair LMs with retrieval systems that scan vast scientific databases in real time, ensuring responses are both updatable and verified. This is crucial in fast-moving fields like climate response modeling or biomedical discovery, where the latest insights matter. For instance:

• When prompted, a knowledgeable-LM can cite primary sources or recent publications from repositories like PubMed or arXiv.
• The LM will present a stepwise rationale or provenance for its outputs, improving transparency and user confidence.
• Researchers can query the LM with domain-specific questions, such as “What are the genetic markers associated with drought tolerance in maize?” and receive evidence-backed, reference-linked answers.

Explainable AI approaches, as advocated by Nature Machine Intelligence, are a cornerstone for responsible integration in scientific workflows.

Real-World Applications: From ClimateNLP to BioNLP

Knowledgeable-LMs underpin groundbreaking ClimateNLP tools, which synthesize climate datasets, policy documents, and academic literature, aiding in the identification of mitigation strategies and predicting climate impacts. Similarly, in BioNLP, these models are guiding automated literature review, drug discovery, and biomarker identification, as evidenced in recent biomedical informatics studies.

In summary, the innovative approaches powering next-generation Knowledgeable-LMs are transforming scientific research by making the discovery process more efficient, reliable, and explainable. Their adoption is set to streamline vast knowledge integration and foster novel research across climate science, healthcare, and beyond.

Insights from BioNLP: Bridging Biology and AI

BioNLP, or Biomedical Natural Language Processing, stands at the critical intersection of computational linguistics and life sciences, offering transformative opportunities for both biomedical research and clinical applications. Leveraging advances in large language models (LLMs), knowledge graphs (KGs), and domain-specific AI, the BioNLP community is pushing the boundaries of how unstructured biological and clinical data can be converted into actionable knowledge. This is particularly relevant as the biomedical domain continues to generate vast quantities of scientific literature, patient records, and multimodal data.

Harnessing Knowledge Graphs to Unify Biomedical Data

One striking development discussed at the ACL 2025 workshops is the maturation of knowledge graphs in BioNLP. Knowledge graphs serve as structured repositories that capture relationships among biological entities—such as genes, proteins, diseases, and drugs—enabling sophisticated reasoning and improved data interoperability.

• Example: Projects like PubTator and BioC are building annotated databases and APIs that bridge publication text and structured resources, making it possible to link mentions in literature to entities in standardized ontologies such as UMLS and Gene Ontology.
• Process: Entity recognition followed by normalized linking, relation extraction, and graph construction are foundational steps. For example, identifying a gene-disease association in research literature requires both accurate extraction and mapping to standardized databases.

Language Models in Scientific AI: Beyond General NLP

The application of large language models (LLMs) like BioBERT and SciBERT, which are pre-trained specifically on biomedical and scientific corpora, has propelled BioNLP tasks to new heights. These models outperform general NLP architectures thanks to domain-specific vocabulary and syntax training.

• Use Case: Automatic summarization of research findings is a capability brought to life by these LLMs. For instance, SciSummNet provides annotated datasets for document-level summarization, while others enable extraction of gene-mutation relationships from clinical reports.
• Advantage: Such AI models help mitigate the information overload faced by researchers by facilitating rapid, accurate literature triage and highlighting novel scientific insights hidden within thousands of papers.

Bridging Biology and AI: New Frontiers in Biomedical Knowledge Discovery

The fusion of KGs and advanced LLMs is accelerating hypothesis generation, drug repurposing, and precision medicine. By representing complex relationships found in the biomedical domain, these tools can uncover unexpected links that may otherwise remain unnoticed.

• Example: The NIH Bridge2AI program exemplifies large-scale collaborative efforts to integrate multimodal data and AI tools, aiming to enable reproducible and explainable biomedical decision-making.
• Stepwise Process:
  1. Aggregate data from heterogeneous sources (literature, omics, clinical notes).
  2. Automate entity extraction using BioNLP pipelines.
  3. Link entities into comprehensive knowledge graphs.
  4. Deploy LLMs to query, reason, and summarize across the integrated dataset.
  5. Validate findings through expert curation and iterative refinement.

In summary, BioNLP’s synergy with scientific AI is reshaping how biological data is managed and interpreted. The continued evolution of domain-specific models, annotated corpora, and interoperable knowledge graphs is paving the way for more interpretable, transparent, and impactful biomedical discoveries. For readers interested in diving deeper, the BioNLP Special Interest Group at ACL offers access to a host of resources, workshops, and research papers at the forefront of this rapidly evolving field.

Emerging Tools and Datasets Introduced at ACL 2025 Workshops

The ACL 2025 workshops showcased a range of groundbreaking tools and datasets that are set to transform research and applications in domains like ClimateNLP, Knowledgeable Language Models, and BioNLP. These introductions not only promise to accelerate scientific discovery but also to make advancements more accessible and impactful for interdisciplinary researchers.

New Datasets: Fueling Robust and Transparent Research

One of the most significant developments from ACL 2025 is the introduction of domain-specific datasets purpose-built for advancing NLP in critical areas:

• ClimateNLP Datasets: The ACL Anthology highlighted large-scale annotated climate science corpora, including the Climate Change AI dataset portal. These resources aggregate scientific literature, policy documents, and real-world sensor data, enabling machine learning models to better extract actionable knowledge for climate action.
• BioNLP Corpora: Expanding on the legacy of PubMed and UMLS, new biomedical text datasets include richer annotations for complex relations, negation, and tacit knowledge inherent in medical literature. These datasets are vital for training and evaluating models that can handle the nuance and unpredictability of biomedical texts.
• Knowledgeable-LM Benchmarks: Workshops introduced the AllenAI Knowledge QA suite, which focuses on multi-hop reasoning and fact-checking across scientific and general knowledge domains. Datasets are engineered to evaluate reasoning over linked graphs and diverse sources, providing deeper insights into model performance and reliability.

Emerging Tools: Bridging Gaps Between Data and Discovery

Alongside new datasets, researchers introduced cutting-edge tools designed to streamline the conversion of complex scientific information into actionable insights:

• Automatic Knowledge Graph Construction Platforms: Enhanced frameworks like MetaMind enable researchers to convert unstructured text into structured knowledge graphs with minimal manual intervention. These platforms now integrate domain ontologies, entity normalization, and cross-document coreference, offering unprecedented depth for climate and biomedical research.
• Explainable AI Toolkits: With the focus shifting to transparency and trustworthiness, toolkits that provide stepwise explanations of model decisions are now embedded in the latest NLP APIs. For example, the AllenNLP Interpret suite allows researchers to visualize the contributions of input data to model outputs—critical for scientific reproducibility and regulatory compliance.
• Augmented Data Annotation Systems: Tools employing interactive annotation, such as those powered by human-in-the-loop AI, were demonstrated to reduce labeling fatigue and increase data quality. These systems also support active learning, where the model suggests the most informative samples to annotate next, streamlining the creation of robust training sets in domains with rare entities or relations.

Practical Examples: Empowering Scientific Discovery

These emerging tools and datasets are already powering new avenues of research. For instance:

• ClimateNLP tools have been used to automatically summarize IPCC reports, helping policymakers quickly grasp the latest projections and recommendations.
• Biomedical knowledge graphs constructed with the new BioNLP datasets are aiding in rare disease research by making connections between disparate case reports, clinical trials, and foundational literature.
• Reasoning benchmarks from Knowledgeable-LM workshops aid in the evaluation of factual accuracy in AI-generated scientific texts, ensuring critical findings are not lost in translation.

For those eager to delve deeper, the ACL Anthology provides direct access to papers, code, and dataset documentation from the workshops, ensuring that the global research community can build upon this year’s foundational work.

Challenges and Opportunities in Domain-Specific NLP

Domain-specific NLP—particularly within fields such as climate science, life sciences, and specialized knowledge graphs—faces unique hurdles and rich opportunities as demonstrated in recent workshops at ACL 2025. The intricacies of scientific language, evolving terminologies, and the demand for precise, actionable insight all contribute to both the challenges and advancements in this area.

Complexities of Domain Terminology and Jargon

Scientific fields are constantly evolving with their own vocabularies, acronyms, and nuanced meanings. For instance, the term “KG” (Knowledge Graph) in machine learning differs greatly from its general use. Models often struggle to distinguish highly specialized contexts. Developing robust NLP solutions requires not only extensive domain-specific corpora but also close collaboration with subject matter experts.

• Step 1: Build comprehensive domain glossaries and annotated datasets by leveraging expert input.
• Step 2: Apply and adapt transfer learning techniques from general-purpose LMs to specialized corpora, as outlined by the ACL Anthology in their latest research reports.
• Step 3: Continuously update models as scientific language and knowledge rapidly progress.

Scarcity and Quality of Labeled Data

Unlike general NLP, where large, labeled datasets abound, scientific NLP faces significant data scarcity. Annotating biomedical or climate data often requires domain expertise, making large-scale annotation both costly and time-consuming. This leads researchers to explore few-shot, weak supervision, and unsupervised approaches.

• Example: The UMLS (Unified Medical Language System) and Europe PMC provide foundational resources for BioNLP, but adapting these for new challenges often needs custom curation and extension.
• Opportunity: Synthetic data generation and data augmentation using large language models can help fill gaps, though this raises its own challenges around fidelity and bias.

Interpretable and Trustworthy AI in Scientific Domains

Trust, transparency, and explainability are fundamental in climate and biomedical AI applications. Stakeholders must understand model decisions, particularly when these will inform scientific research or public policy. Building interpretable systems remains a major focus, as stressed in recent Nature Machine Intelligence articles.

• Knowledge graphs and rule-based approaches provide transparency but often lack flexibility, while black-box neural networks deliver performance but are harder to interpret.
• Balancing these two approaches—sometimes through hybrid models—presents a key opportunity for innovation in domain-specific NLP.

Cross-domain Generalization and Adaptation

Scientific research is inherently interdisciplinary. For NLP models, transferring insights from one domain (e.g., climate science) to another (e.g., epidemiology) remains complex. Domain adaptation and multi-task learning, as discussed in the Climate Change AI community, help models generalize across datasets, but require careful tuning to avoid propagating errors or domain mismatches.

Opportunities for Impact and Collaboration

The intersection of advances in Knowledgeable-LMs, KG-based reasoning, and domain-specific resources shows enormous promise. Interdisciplinary initiatives, open-source platforms, and shared benchmarks—such as those initiated by Allen Institute for AI—provide a bedrock for rapid innovation. Active collaboration across institutions, combining algorithmic expertise with deep subject knowledge, is essential for moving the field forward.

Overall, the challenges in domain-specific NLP are intricately linked to its greatest opportunities. By addressing data scarcity, embracing interpretability, and fostering open collaboration, researchers can propel NLP toward empowering scientific discovery, policy, and public understanding.

Collaborative Research Initiatives Highlighted During the Workshops

At ACL 2025, the workshops on Knowledge Graphs (KG) and scientific AI in ClimateNLP, Knowledgeable-LM, and BioNLP brought together prominent researchers and practitioners to push the boundaries of natural language processing (NLP) for scientific advancement. These sessions revealed pioneering collaborative research initiatives, emphasizing the real-world impact of interdisciplinary partnerships and shared data resources.

One standout initiative focused on building shared knowledge graphs to integrate scientific literature and climate data. This collaborative effort unites data scientists, climate researchers, and NLP experts to extract structured knowledge from complex scientific texts. By employing advanced information extraction techniques, teams transform unstructured publications into interconnected knowledge graphs, enabling inter-institutional data sharing. Projects like Semantic Scholar Open Research Corpus (S2ORC) and MIT Climate Portal’s climate data projects serve as foundational resources for building these collaborative datastores.

Another collaborative highlight was the organization of shared tasks and challenges to drive the development of powerful language models in specialized scientific domains. Through initiatives like the BioASQ Challenge in biomedical NLP and the newly-launched ClimateNLP Shared Task, research teams work together to establish standardized benchmarks, annotate large-scale datasets, and fairly compare model performances. These competitions foster a spirit of cooperative rivalry, allowing teams to collectively address pressing questions such as identifying emerging research trends, synthesizing scientific knowledge, and tracking misinformation related to climate or health.

Open-source collaboration platforms also featured prominently, with workshops promoting the use of communal repositories and collaborative annotation tools. Examples include the Knowledgeable-LM GitHub repository and datasets made available by respective workshop teams. These platforms enable seamless co-development of models and resources, facilitating widespread community involvement and reproducible research. Participants are encouraged to contribute code, suggest improvements, and build upon one another’s work, showcasing the spirit of collaborative progress in contemporary NLP.

Additionally, the workshops highlighted the importance of interdisciplinary collaboratives in addressing ethical, legal, and social considerations in scientific NLP. Discussions often turned to the need for multidisciplinary governance frameworks, drawing on expertise from law, social sciences, and computer science to ensure responsible AI development and equitable access to scientific knowledge. Initiatives modeled after approaches like those described in the Nature article on ethical AI in science underscore the role of community-wide agreements and shared standards in guiding responsible AI research.

In summary, the ACL 2025 workshops underscored that collaboration—across disciplines, institutions, and borders—is foundational to advancing scientific AI in ClimateNLP, Knowledgeable-LM, and BioNLP. These initiatives are not only elevating the technical frontiers of knowledge extraction and language modeling but are also setting new standards for open, responsible, and impactful research networks.

Future Directions in Scientific AI from the ACL 2025 Community

The ACL 2025 workshops—spanning ClimateNLP, Knowledgeable-LM, and BioNLP—offered a compelling look into where the scientific AI landscape is heading. The convergence of knowledge graphs (KGs) and advanced language models is reshaping how researchers, policymakers, and practitioners leverage large-scale data to address global challenges in climate science and biomedical domains.

Toward Richer Knowledge Graph Integration

One of the most prominent motions across workshops was the deepening integration of knowledge graphs with language models. This pairing is enabling models to move from text-level understanding to reasoning over complex scientific relations. For instance, in ClimateNLP, KGs are being used to map the intricate causal relationships among climate variables, interventions, and outcomes. This allows for:

• Enhanced evidence discovery: Automated agents can draw on graph-based reasoning to summarize vast datasets, identify knowledge gaps, and inform targeted climate action. See recent examples in MIT Climate Research.
• Robust fact-checking and citation: KGs support fact verification in generated scientific claims, helping to reduce misinformation and improve trust in AI-driven climate reports.

Knowledgeable Language Models in Scientific Discovery

The Knowledgeable-LM community highlighted the importance of grounding language models in up-to-date, curated scientific knowledge. The next steps for the field include:

• Dynamic context integration: Ongoing efforts are focused on teaching models to seamlessly reference KGs or external databases during generation, akin to what retrieval-augmented generation achieves. This means AI can provide not just plausible-sounding answers, but ones backed by authoritative sources.
• Model transparency and interpretability: Researchers are developing techniques so users can trace a model’s claims directly to the underlying scientific assertions and entities in KGs. This work addresses the notorious “black box” issue in AI by bringing transparency to decision pathways, critical for high-stakes applications in medicine or climate policy.

BioNLP and the Future of Biomedical Language Understanding

BioNLP workshops showcased state-of-the-art approaches for handling biomedical texts and data. The way forward prioritizes:

• Comprehensive entity and relation extraction: Advanced models are being trained to recognize not only diseases and drugs but also subtle interactions between genes, proteins, and environmental exposures. Examples include the Unified Medical Language System and its role in powering biomedical KGs.
• Cross-domain adaptation: Techniques like few-shot and zero-shot learning, highlighted at ACL, enable transfer of AI capabilities from well-studied diseases to emerging outbreaks or under-researched conditions. This accelerates discovery and adapts rapidly to new biomedical crises.
• Collaborative platforms: The emergence of collaborative open platforms—such as NIH Bioinformatics Resources—encourages multidisciplinary teams to share annotated corpora, pre-trained models, and benchmarks, accelerating community-wide progress.

Bridging Challenges and Setting the Research Agenda

Across all these fronts, the ACL 2025 community identified several pressing challenges and directions:

• Scaling and updating KGs to keep pace with real-time scientific discoveries, ensuring models do not rely on stale or biased data.
• Ethics and responsible AI: Ensuring AI systems used in sensitive domains—especially climate science and healthcare—are robust, fair, and aligned with public good. The importance of frameworks like the AAAI Code of Ethics was widely discussed.
• Human-in-the-loop AI: There is growing consensus on the need for seamless collaboration between domain experts and AI systems, for model oversight, correction of errors, and continual learning from expert feedback.

The workshop conversations made clear that future advances will demand not only algorithmic prowess but also interdisciplinary partnerships and open science practices. This evolution, energized by the ACL community, holds promise for faster, more transparent breakthroughs across scientific frontiers.