The Education University of Hong Kong (EdUHK) has emerged as a major force in the global race to build more reliable and socially responsible artificial intelligence after researchers from the university and partner institutions secured three full-paper acceptances at the 43rd International Conference on Machine Learning (ICML) 2026. The papers were produced in collaboration with the University of Technology Sydney, Wuhan University, and Huazhong University of Science and Technology, and address three of the most urgent problems confronting modern AI: how to improve reasoning in large language models, how to make recommender systems fairer, and how to ensure that machine-generated confidence scores reflect reality. One of the three papers was selected for an oral spotlight presentation, a distinction reserved for only 168 papers from 23,918 submissions. The achievement places EdUHK researchers at the centre of a rapidly expanding movement focused not merely on making AI more powerful, but on making it more dependable, transparent, and aligned with human interests.
The research was led by Professor Xu Guandong, Chair Professor in EdUHK’s Department of Mathematics and Information Technology, together with Dr Yu Yang, an Assistant Professor at the university’s Centre for Learning, Teaching and Technology, and Research Assistant Mr Nitin Bisht. Their work reflects a shift taking place across machine learning research. For years, progress was largely measured by benchmark accuracy, model size, or computational performance. Increasingly, however, researchers are asking whether a system knows when it might be wrong, whether it distributes opportunities fairly, and whether its predictions can be trusted by people making consequential decisions. The three papers accepted at ICML approach these questions through different technical routes, including modified optimisation objectives, Bayesian uncertainty estimation, adaptive calibration, and theoretical performance guarantees. Together, they offer a broad framework for developing AI systems that can reason with appropriate uncertainty, distribute attention more equitably, and communicate their predictions more honestly.
The oral spotlight paper, “Don’t Force the Fit: Bounded Log Likelihood Loss for Enhanced Reasoning in Large Language Models”, targets a subtle weakness in the way large language models are trained. Most language models rely on log likelihood loss, a standard objective that penalises a system when it assigns low probability to the correct next token. In principle, this encourages the model to predict accurately. In practice, however, extremely rare or surprising tokens can generate disproportionately large penalties. The model may then learn to avoid uncertainty at all costs, favouring familiar patterns and statistically convenient shortcuts rather than developing robust reasoning strategies. Professor Xu’s team proposes a Bounded Log Likelihood, or BLL, loss that places an upper limit on the penalty associated with unlikely tokens. The modification is mathematically simple, but its implications are significant: the model is no longer pushed into treating every unexpected continuation as catastrophic evidence of failure.
By capping the loss, BLL preserves a controlled level of uncertainty during training. This matters because correct reasoning often requires a system to explore possibilities that are initially improbable. A model that is excessively confident may produce fluent answers while relying on superficial correlations, a phenomenon commonly described as shortcut reasoning. Such behaviour can allow an AI system to perform well on familiar examples but fail when a problem is rephrased, combined with unfamiliar information, or presented in a new context. According to the research, the bounded objective improves performance across arithmetic, commonsense, and symbolic reasoning benchmarks. The result suggests that uncertainty is not simply a weakness to be eliminated; when managed correctly, it can be a resource that supports deeper inference. In high-stakes settings such as healthcare, law, and education, an AI system that retains calibrated doubt may be safer and more useful than one that produces confident answers without recognising the limits of its knowledge.
The second ICML paper, “CORAL: Uncertainty Aware Regulation of Exposure Concentration in Recommender Systems”, examines the way online platforms decide which items, creators, products, or news stories receive visibility. Recommendation engines are often optimised for immediate engagement, and this can cause exposure to concentrate around a small group of already popular items. Once an item attracts attention, the resulting data can make it appear even more attractive to the algorithm, creating a feedback loop in which popularity generates further popularity. New or less-established creators may then struggle to receive a fair opportunity, even when their content has genuine potential. CORAL addresses this problem by introducing a dynamic exposure allocation policy. The system reduces the amount of additional exposure given to highly popular items when their quality is already estimated with high certainty, while directing regulated exposure toward uncertain newcomers whose potential has not yet been fully measured.
The technical foundation of CORAL is the combination of Bayesian uncertainty estimates with an exposure-amplification mechanism. Instead of treating every recommendation decision as equally certain, the method distinguishes between items whose performance is well understood and those about which the system has limited information. This allows the platform to conduct what is effectively a controlled fair trial: promising newcomers receive enough visibility to generate meaningful evidence, while popular items remain available without monopolising attention. The researchers establish a regret bound indicating that the uncertainty-aware policy can approach long-term fairness without sacrificing recommendation accuracy. In machine learning, regret measures the gap between the results achieved by an algorithm and those that could have been achieved by an ideal strategy with complete knowledge. By proving that CORAL’s long-term performance remains close to this ideal while reducing exposure concentration, the paper links fairness to a formal statistical guarantee rather than treating it as an informal design preference. The approach could help reduce filter bubbles, broaden cultural and commercial representation, and support more sustainable digital ecosystems.
The third paper, “CARE: Adaptive Calibration for Reliable Recommendations”, focuses on a different but closely related source of distrust: the gap between a system’s confidence and the probability that its prediction will actually come true. A recommendation platform might, for example, estimate that a job candidate has a 95 per cent chance of matching a position, even though candidates with similar profiles receive interviews only 70 per cent of the time. When predictions are systematically overconfident or underconfident, users cannot interpret them properly, and organisations may make poor decisions based on numbers that appear precise but lack statistical meaning. CARE introduces an adaptive post-processing algorithm that learns separate calibration maps for different user or item subgroups. Calibration means that predictions labelled with a given probability should occur at approximately that frequency in reality. A system is well calibrated when events assigned a 70 per cent probability happen roughly 70 per cent of the time.
The challenge is to achieve subgroup-specific calibration without allowing the estimates to become unstable because of insufficient data. CARE addresses this tension by adapting the granularity of its calibration maps as evidence accumulates. It aims to capture meaningful differences among groups while maintaining the statistical reliability needed for trustworthy predictions. The method is model agnostic, meaning it can be applied after training to a wide range of existing recommender systems rather than requiring the underlying model to be rebuilt. It also adds negligible inference overhead, making it practical for platforms that must generate predictions at scale. The researchers report theoretical guarantees for reducing expected calibration error within groups as more data becomes available. That feature could be particularly important in employment, credit, healthcare information, and other settings where a numerical prediction can influence access to opportunities or services. By making confidence scores more faithful to observed outcomes, CARE could help users distinguish between a strong recommendation and a merely plausible one.
The three ICML papers share a central idea: responsible AI depends on understanding uncertainty rather than hiding it. BLL prevents language models from being trained into brittle overconfidence; CORAL uses uncertainty to distribute attention more fairly; and CARE makes uncertainty legible by aligning predictive scores with real-world frequencies. The work also illustrates how theoretical machine learning can produce practical consequences. Loss functions influence the reasoning habits that models develop, regret bounds provide a way to measure the long-term cost of fairness interventions, and calibration guarantees define whether a probability can be trusted. Professor Xu described the broader goal in clear terms: “The most exciting breakthroughs are those that make our models not just more powerful, but more worthy of the trust we place in them.” That philosophy is increasingly relevant as AI systems move from laboratories into classrooms, workplaces, financial services, public information networks, and clinical environments.
EdUHK’s momentum extends beyond ICML. Professor Xu, Dr Yang, and Dr Yin Nan, an Assistant Professor at the Centre for Learning, Teaching and Technology, have also published four papers at the 32nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD) 2026, one of the leading international forums for data science and artificial intelligence. Although the individual KDD paper titles are not specified in the available information, the combined record reinforces the university’s expanding presence in machine learning research. The ICML results, together with the KDD publications, position EdUHK as a contributor to a new generation of AI research concerned with robustness, fairness, calibration, and social responsibility. At a time when viral AI breakthroughs are often defined by scale or spectacle, these studies point to a quieter but potentially more consequential revolution: systems that know when to hesitate, share attention more fairly, and give people confidence scores they can actually believe.
Subject of Research: Not applicable
References: “Don’t Force the Fit: Bounded Log Likelihood Loss for Enhanced Reasoning in Large Language Models”; “CORAL: Uncertainty Aware Regulation of Exposure Concentration in Recommender Systems”; “CARE: Adaptive Calibration for Reliable Recommendations”; 43rd International Conference on Machine Learning (ICML) 2026; 32nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD) 2026
Keywords: artificial intelligence, machine learning, large language models, AI reasoning, uncertainty-aware AI, recommender systems, algorithmic fairness, exposure concentration, Bayesian uncertainty, model calibration, trustworthy AI, human-centred AI, EdUHK, ICML 2026, KDD 2026
