Artificial Intelligence 

At CommBank, our commitment to advancing artificial intelligence (AI) is a key part our Technology strategy. As a leading financial institution, we recognise the transformative potential of AI in enhancing our services, improving customer experiences, and driving operational efficiency.

Our participation in the International Conference on Learning Representations (ICLR) 2025 in Singapore as a Gold sponsor underscores this commitment. ICLR is well known for presenting and discussing novel research in AI and machine learning, making it an ideal platform for us to engage with the global AI community, share our insights, and learn about the latest advancements in the field. 

ICLR 2025 featured a range of research with over 11,000 submissions, approximately 32% acceptance rate, and 3,827 papers accepted. Submissions increased by over 60% from the previous year, reflecting the rapid growth of the AI field.

In terms of the main trends, the conference highlighted topics from agentic AI systems and multi-step reasoning to new fine-tuning strategies, advances in foundational models, and a strong emphasis on AI safety, alignment and real-world deployment.

We're excited to share trends and key takeaways from the conference, including insights from a selection of papers, workshops, and talks.

Agentic Systems and Open-ended Learning

Research of agentic AI systems was more prominent compared to NeurIPS 2024. Key talks emphasised open-ended learning approaches for such agents. For example, in his keynote, Tim Rocktäschel (Google DeepMind and University College London) advocated for training agents via “foundation world models” that generate diverse simulated environments, enabling agents to acquire more general and robust behaviours (see Figure 1). This open-ended paradigm aims to produce AI agents that can endlessly generate and tackle novel tasks, moving beyond narrowly optimised solutions. Early results are promising, we’re seeing self-referential self-improvement loops where AI agents can refine themselves through automated prompt engineering, automated red-teaming, and even AI-vs-AI debates.  

The drive toward agentic models matters because next-generation AI will likely need to operate autonomously in complex, changing environments. We should start thinking about how to design and evaluate systems in CommBank that learn from interaction and remain adaptive when facing unforeseen tasks.

Another recommended paper in the field of Agentic would be AFlow: Automating Agentic Workflow Generation by Jiayi Zhang et al. is a framework designed to automate the creation of agentic workflows using Monte Carlo Tree Search (MCTS). It refines workflows through code modifications and execution feedback. Initial experiments show promising results but highlight the need for further research, particularly for complex tasks in banking.

Fine-Tuning and Adaptation Techniques

With foundation models now widespread, a practical challenge is how to adapt these large models efficiently to specific tasks or evolving data. ICLR 2025 highlighted many advances in fine-tuning methods that make adaptation more accessible. A recurring theme was parameter-efficient fine-tuning: instead of retraining massive models from scratch, researchers are devising lightweight adapters (e.g. LoRA modules) that tweak only small portions of the model.

While this method is not new, ICLR 2025 was rich of various improvements over existing LoRA methods. One standout paper introduced LoRA-X, a technique allowing fine-tuning parameters to be transferred between different base models without any retraining. This approach addresses a common problem – when a foundation model is upgraded or replaced, we can carry over the learned task-specific adaptors without needing the original training data or costly re-training. Such innovations mean practitioners can keep models up-to-date and domain-specialised with minimal compute overhead.

More broadly, the conference underscored that adaptability is crucial for real-world AI: future systems should be easy to fine-tune on new tasks or data.

Some additional papers in the fine-tuning space worth highlighting include:

AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models by Junfeng Fang et al. introduces AlphaEdit, a novel method for updating specific knowledge in Large Language Models (LLMs) without disrupting existing information. By projecting parameter updates onto the null space of preserved knowledge, AlphaEdit ensures that new edits are localised. This approach improves the preservation of original knowledge by 36% in GPT and LLama models.

Preserving Diversity in Supervised Fine-Tuning of Large Language Models by Ziniu et al. presents Entropic Distribution Matching (EDM), a technique to mitigate overfitting and loss of output diversity during supervised fine-tuning. EDM aligns the output distribution of the fine-tuned model with that of the pre-trained model, thereby maintaining the model's ability to generate diverse responses. This method enhances generalisation and robustness across various tasks.

Evolving Foundation Models

This year ICLR demonstrated that foundation models remain central topic in AI research, but the focus is shifting from simply scaling up model size to understanding and extending these models’ capabilities. Many works addressed issues like bounding model capabilities, reducing hallucinations, and using synthetic data to overcome training data gaps.

This reflects a maturation of the field: rather than chasing ever-bigger architectures, researchers are consolidating knowledge about how to optimise and reliably deploy the powerful models we already have.

At the same time, foundation models are expanding in scope. New multimodal models unveiled at ICLR can handle new modalities. For example, one presentation showed a hierarchical vision-language-action model (HAMSTER from NVIDIA), that transfers knowledge from simulations to real-world robotics tasks.

Researchers also showcased architecture innovations (such as hybrid Transformer–state-space models) to make large models more efficient in handling long contexts and high throughput.

Notably, a keynote by Danqi Chen highlighted that academia is finding ways to contribute to foundation model development by innovating under resource constraints – for instance, building smaller but capable models, improving training data quality, and layering new fine-tuning methods on open-source models. All these efforts indicate that foundation models are evolving not just in size, but in versatility and accessibility.

We can expect the next generation of foundation models that are more efficient, cover multiple modalities, and be easier to adapt in smaller settings.

Focus on AI Safety and Robustness

Finally, AI safety was a significant theme at ICLR 2025, indicating an increasing awareness that more advanced models need to be reliable and aligned with human intentions. Several invited talks addressed how to build safe and secure AI systems. Zico Kolter’s talk, for example, was focused on building robust AI systems that enforce strict safety constraints on models. He discussed recent techniques to harden models against adversarial exploits and “jailbreak” attacks, such as robustness at the training stage and even data pre-filtering stage to prevent users from manipulating an LLM’s outputs in harmful ways.

As AI systems become more agentic and ubiquitous, there is a greater need to integrate safety by design – from rigorous testing for adversarial robustness to implementing guardrails that align model behaviour with ethical and policy requirements.

ICLR 2025’s focus on safety underscores that future ML systems will be judged not only on performance, but on trustworthiness and safety robustness.

ICLR 2025 Discussions

Many fascinating discussions were about the potential impact of AI model collapse, attributed to insufficient human training and the use of iterations of LLM-generated text as inputs for subsequent versions of LLM models.

This conversation has paralleled emerging concept of ‘the Era of Experience’ introduced by David Silver and Richard S. Sutton’s paper suggesting that we are transitioning from the era of data to the era of learning from environment similarly to Reinforcement learning concepts of AlphaGo era.

Alongside these developments, the ICLR community has increasingly questioned the adequacy of current evaluation methods, noting that existing benchmarks tend to saturate too quickly and may not effectively reflect real-world challenges, emphasising the need for more sophisticated and realistic eval frameworks.

Conclusion

The research directions highlighted at ICLR 2025 paint a picture of an AI field striving for both greater capability and greater responsibility. Trends like agentic models and adaptive fine-tuning point toward AI that is more autonomous and flexible, able to handle open-ended challenges and continuously learn.

At the same time, the evolution of foundation models, concerns of AI model collapse and the emphasis on safety show a commitment to making these powerful frameworks more usable, reliable, and data selective. Overall, ICLR 2025 was a great opportunity for us to understand what to expect next in this ever-evolving field.