Papers from the Department of Computing accepted at ICLR 2026

Members of the Department of Computing are among the authors of papers accepted to ICLR 2026, a major global venue for advances in representation learning and deep learning, to be held 23–27 April 2026 in Rio de Janeiro.

We are pleased to announce that papers co-authored by members of the Department of Computing have been accepted to the International Conference on Learning Representations (ICLR 2026), held 23–27 April 2026 at the Riocentro Convention and Event Center in Rio de Janeiro, Brazil.

ICLR is the premier gathering of professionals dedicated to advancing representation learning (deep learning), and provides a leading forum for publishing and discussing the latest research in the field.

Adablock-dllm: Semantic-aware diffusion llm inference via adaptive block size

Authors: Guanxi Lu, Hao Mark Chen, Yuto Karashima, Zhican Wang, Daichi Fujiki, Hongxiang Fan

Diffusion-based large language models (dLLMs) enable parallel decoding and are a promising alternative to autoregressive LLMs. A conventional strategy is blockwise semi-autoregressive (semi-AR) decoding with a fixed block size, which supports KV caching and offers a good accuracy–speed trade-off. However, fixed blocks introduce two issues: they delay decoding of high-confidence tokens outside the current block (late-decoding overhead) and prematurely commit low-confidence tokens inside the block (premature-decoding error). We present the first systematic study of the fixed-block assumption in semi-AR decoding. By analysing confidence dynamics during denoising, we identify a volatility band (VB) region that captures local semantic structure and can guide adaptive block sizing. Building on this, we propose AdaBlock-dLLM, a training-free, plug-and-play runtime scheduler that adjusts block size to align block boundaries with semantic steps. Across diverse benchmarks, AdaBlock-dLLM improves accuracy by up to 5.3% under the same throughput budget, and our confidence-based analysis suggests new directions for semantics-aware dLLM training. Notably, this research stems from a 2025 Summer Undergraduate Research Opportunity Programme (UROP) project led by an undergraduate student which is a rate achievement.

BoGrape: Bayesian optimisation over graphs with shortest-path encoded

Authors: Yilin Xie, Shiqiang Zhang, Jixiang Qing, Ruth Misener, Calvin Tsay

The paper explores how graph-structured systems can be optimised, for example to identify the best network structure and/or node features. Problems of this kind arise in areas such as molecular design, neural architecture search and sensor placement. The authors present a Bayesian optimisation approach that enables efficient global exploration of the graph domain, and demonstrate their method, BoGrape, through several molecular design case studies

Safe Reinforcement Learning via Recovery-based Shielding with Gaussian Process Dynamics Models

Authors: Alex Goodall, Francesco Belardinelli

This demonstration introduces a recovery-based shielding framework for safe reinforcement learning in unknown and nonlinear continuous dynamical systems.

Using Gaussian Process models, the approach provides a provable lower bound on safety while maintaining strong empirical performance across continuous control environments. The work addresses a key barrier to deploying reinforcement learning in safety-critical applications.

ATL*AS: An Automata-Theoretic Approach and Tool for the Verification of Strategic Abilities in Multi-Agent Systems

Authors: Sofia Garcia de Blas Garcia-Alcaide, Francesco Belardinelli

ATLAS presents two novel symbolic algorithms for model checking Alternating-time Temporal Logic (ATL) under both infinite- and finite-trace semantics.

Implemented in the ATLAS model checker, the tool is evaluated on synthetic benchmarks and a cybersecurity scenario, providing a comprehensive environment for verifying strategic abilities in multi-agent systems.

Synthesis of Safety Specifications for Probabilistic Systems

Authors: Gaspard Ohlmann, Edwin Hamel-de le Court, Francesco Belardinelli

The paper introduces a new method for generating counterfactual explanations, or “what-if” scenarios, that help explain the decisions made by AI systems. Most existing approaches generate counterfactuals by solving complex optimisation problems directly in the model’s raw input space. The authors instead approach the problem from a different perspective, lifting the process into a learned, regularised latent space that reflects the classifier’s predictions. By operating in this latent space, the method can generate counterfactuals that satisfy several desirable properties, including validity, proximity, plausibility, robustness and actionability, all within a single framework.

ProSh: Probabilistic Shielding for Model-free Reinforcement Learning

Authors: Edwin Hamel-de le Court, Gaspard Ohlmann, Francesco Belardinelli

ProSh proposes probabilistic shielding via risk augmentation for model-free reinforcement learning under cost constraints.

Under practically achievable assumptions, the approach guarantees safety during training, as demonstrated experimentally. This contributes to the broader goal of developing reinforcement learning systems that are not only high-performing but also safe to deploy.

Synthesising Counterfactual Explanations via Label-Conditional Gaussian Mixture Variational Autoencoders

Authors: Junqi Jiang, Francesco Leofante, Antonio Rago, Francesca Toni

The authors introduce a new method for generating counterfactual explanations - the "what-if" scenarios that explain AI decisions.

Most methods generate coun`terfactuals by solving complex optimization problems directly in the model’s raw input space. We approached the problem from a new angle and lifted this process into a learned, regularised latent space reflecting the classifier’s predictions. By operating in the latent space, we can generate counterfactuals that satisfy many interesting properties (validity, proximity, plausibility, robustness and actionability). All of this within a single, seamless framework.

ZK-APEX: Zero-Knowledge Approximate Personalized Unlearning with Executable Proofs

Authors: Mohammad M Maheri, Sunil Cotterill, Alex Davidson, Hamed Haddadi 

Modern AI systems are trained on huge amounts of data, including personal information. Laws like the “right to be forgotten” say that if someone asks for their data to be deleted, the AI should “unlearn” it — as if that data had never been used. This paper, presents ZK‑APEX, to offer Zero‑knowledge proofs of deletion for AI providers to be able to prove they have honoured data‑deletion requests on personalised models, without seeing the user’s private data and without paying the high cost of full retraining—making stronger privacy and compliance more realistic in real‑world deployments.

Inference-Time Scaling of Discrete Diffusion Models via Importance Weighting and Optimal Proposal Design

Authors: Zijing Ou, Chinmay Pani, and Yingzhen Li

While generative AI is great at creating content, it often struggles to follow specific rules. The authors have developed a new framework that acts as a "smart GPS" for these models, guiding them step-by-step to ensure their output remains both high-quality and strictly controlled. By using a mathematical technique called Sequential Monte Carlo, their method effectively "nudges" the AI toward desired outcomes in real-time. Whether it’s writing better sentences, designing new DNA sequences, or generating precise images, their approach proves that we can make AI more reliable and useful for complex, real-world tasks without sacrificing its creative power.

On the Interaction of Compressibility and Adversarial Robustness

Authors: Melih Barsbey, Antônio H. Ribeiro, Umut Simsekli, Tolga Birdal

The authors theoretically and empirically show that AI models trained to be compressible inherently result in adversarial vulnerability, which can be leveraged by classical attack methods. 

Correlations in the Data Lead to Semantically Rich Feature Geometry Under Superposition

Authors: Lucas Prieto, Edward Stevinson, Melih Barsbey, Tolga Birdal *, Pedro A. M. Mediano *

The authors introduce the Bag-of-Words Superposition (BOWS) framework which reveals that AI models with a 'bottleneck' encode the low rank structure in the data, arranging feature representations according to their co-activation patterns, yielding a geometric latent structure.

HOG-Diff: Higher-Order Guided Diffusion for Graph Generation

Authors: Yiming Huang, Tolga Birdal

Graphs are widely used to model unstructured information such as social networks, molecules and proteins. While many studies aim to generate graphs using state-of-the-art tools such as diffusion modelling, existing approaches often struggle to preserve the topological structure of the data when mapping noise to inputs. The paper introduces HOG-Diff, a principled framework for progressively generating plausible graphs with inherent topological structure. The method follows a coarse-to-fine generation process guided by higher-order topology and implemented using diffusion bridges. The authors also show that their model provides stronger theoretical guarantees than classical diffusion frameworks. Experimental evaluations demonstrate that HOG-Diff achieves state-of-the-art performance in graph generation across multiple data distributions.

Generalised Flow Maps for Few-Step Generative Modelling on Riemannian Manifolds

Authors: Oscar Davis, Nicholas Boffi, Michael Albergo, Michael Bronstein, Joey Bose

This paper introduces the first few-step generative model for Riemannian manifolds. Typically training generative models on structured data like proteins, climate data, robotics, etc... requires contending with very exotic geometric operations to simulate the generative process. The authors have found a way to speed this up by learning how to take big jumps along this process to cut down simulation time and lead to faster samples—-which is exactly the method they introduce called Generalised Flow Maps.

Planner Aware Path Learning in Diffusion Language Models Training (Oral)

Authors: Fred Zhanzhi Peng, Zachary Bezemek, Jarrid Rector-Brooks, Shuibai Zhang, Anru R Zhang, Michael Bronstein, Alexander Tong, Joey Bose

Diffusion language models generate text by repeatedly “cleaning up” a corrupted sequence, often in parallel for speed. A key ingredient to their success at generation time is a planner: a strategy that decides which positions to clean next at each step. Many planners don’t clean positions randomly—they make smarter, targeted choices. New planners choose which parts to denoise next, but this creates a train–test mismatch: training assumes random denoising paths, while inference follows planner-chosen (non-random) paths. This work shows the standard diffusion ELBO no longer fits under planning, derives a Planned ELBO (P-ELBO) that accounts for the planner, and proposes Planner Aware Path Learning (PAPL) to align training with planned inference. PAPL improves results across protein, text, and code generation.

OXtal: An All-Atom Diffusion Model for Organic Crystal Structure Prediction

Authors: Emily Jin, Andrei Cristian Nica, Mikhail Galkin, Jarrid Rector-Brooks, Kin Long Kelvin Lee, Santiago Miret, Frances H Arnold, Michael Bronstein, Joey Bose, Alexander Tong, Cheng-Hao Liu

Crystal structure prediction (CSP) asks whether, given a molecule’s 2D graph, we can predict the 3D crystal structure it forms in the real world. This is important for applications such as pharmaceuticals and organic electronics, where the way molecules pack together determines key material properties. The paper introduces OXtal, an approximately 100-million-parameter all-atom diffusion model that learns, end to end, how molecular conformations and crystal packing co-vary within periodic structures. Rather than hard-coding symmetry or equivariance into the architecture, the model scales through data augmentation and a new lattice-free training method, Stoichiometric Stochastic Shell Sampling, which captures long-range interactions without explicitly parameterising the crystal lattice. Trained on around 600,000 experimental crystal structures, including flexible molecules, co-crystals and solvates, OXtal substantially outperforms previous machine learning approaches to CSP while being far less computationally expensive than quantum-chemical pipelines. It closely recovers experimental structures and achieves over 80% packing similarity, representing a major step forward for organic crystal prediction and hinting at a breakthrough moment for the field comparable to what AlphaFold achieved in protein structure prediction.

Efficient Regression-Based Training of Normalizing Flows for Boltzmann Generators

Authors: Danyal Rehman, Oscar Davis, Jiarui Lu, Jian Tang, Michael Bronstein, Yoshua Bengio, Alexander Tong, Avishek Joey Bose

Many modern generative models, such as diffusion models, are powerful but slow to run, which can be problematic for scientific tools such as Boltzmann Generators that require fast sampling and efficient likelihood evaluation. The paper revisits classical normalising flows, which provide fast likelihoods but are often difficult and unstable to train using standard maximum-likelihood methods. The authors propose RegFlow, a new regression-style approach for training flows. Instead of optimising likelihood directly, the flow is trained to map simple prior samples to “target” samples provided by optimal transport pairings or a pre-trained continuous normalising flow. The method also introduces several stabilising components, including a forward–backward self-consistency regulariser, to make training more robust. In molecular equilibrium sampling tasks, including alanine dipeptide and longer peptides, RegFlow trains more reliably, supports a wider range of architectures, and improves performance and computational cost compared with maximum-likelihood training. An earlier version of this work received the Best Paper Award at the ICML 2025 GenBio Workshop under the title “FORT: Forward Only Regression Training of Normalizing Flows”.

HGNet: Scalable Foundation Model for Automated Knowledge Graph Generation from Scientific Literature

Authors: Devvrat Joshi and Islem Rekik

The paper proposes HGNet, a two-stage framework for automatically constructing structured knowledge graphs from scientific papers in order to better organise the rapidly expanding research landscape. In the first stage, the system identifies complex, multi-word scientific concepts across domains without requiring task-specific retraining. In the second stage, it extracts relationships between these concepts and models their hierarchical structure. By enforcing logical consistency and encoding abstraction levels along a learnable geometric axis, the method produces more coherent and interpretable graphs than existing approaches, including large language models. The authors also introduce SPHERE, a large-scale benchmark, and demonstrate substantial improvements, particularly in zero-shot settings, highlighting stronger robustness, scalability and cross-domain generalisation. This work also introduces the SPHERE benchmark as a new resource for evaluating scientific knowledge graph construction at scale.

This ICLR paper marks the culmination of Devvrat’s five-month research journey at the BASIRA Lab (I-X), representing his fourth accepted publication during this period. Remarkably, he had already published three papers (a MICCAI workshop oral paper, EMNLP, and NeLaMKRR 2025), making this achievement even more impressive.

A Structured, Tagged, and Localized Visual Question Answering Dataset with Full Sentence Answers and Scene Graphs for Chest X-ray Images

Authors: Philip Müller, Friederike Jungmann, Georgios Kaissis, Daniel Rueckert

Visual Question Answering (VQA) lets users interact with AI by asking questions about medical images, such as chest X-rays, and receiving specific answers that support a more intuitive and flexible use. Training such systems, however, requires large, annotated datasets. Additionally, for thorough interpretation, we need detailed answers accompanied by visual grounding, such as bounding boxes that highlight relevant regions. Existing datasets do not provide this level of detail, and manually annotating such large-scale data is extremely costly. To address this, the authors developed an automatic pipeline that extracts structured information from radiology reports, from which we generate our new CXR-QBA dataset. It contains 42 million question–answer pairs with multi-part answers, bounding boxes, and structured tags, making it the largest and most sophisticated chest X-ray VQA dataset to date and enabling more accurate, interpretable, and interactive AI tools for medical imaging. The authors thoroughly assess the quality of the dataset and make it publicly available for researchers.

Beyond Uniformity: Regularizing Implicit Neural Representations through a Lipschitz Lens

Authors: Julian McGinnis, Suprosanna Shit, Florian A. Hölzl, Paul Friedrich, Paul Büschl, Vasiliki Sideri-Lampretsa, Mark Mühlau, Philippe C. Cattin, Bjoern Menze, Daniel Rueckert, Benedikt Wiestler

Neural networks that represent signals like images, shapes, or MRI scans as continuous functions, known as implicit neural representations, face a fundamental trade-off: they must be flexible enough to capture fine details, yet smooth enough to generalize well and avoid generating artifacts. This work introduces a principled framework for managing this trade-off by controlling how much each component of the network is allowed to change its output in response to small input changes (i.e., the "Lipschitz budget"). Rather than applying the same rigid constraint everywhere, the method derives a meaningful total budget from task-specific knowledge, such as how much tissue can physically stretch in medical imaging, and then distributes it non-uniformly across different architectural components. Experiments on 3D shape reconstruction, lung registration, and image inpainting show that this flexible budgeting strategy yields smoother, more accurate results than conventional approaches, offering practitioners an interpretable tool for designing better neural representations.

Temporal superposition and feature geometry of RNNs under memory demands. 

Authors: Pratyaksh Sharma, Alexandra Maria Proca, Lucas Prieto, Pedro A. M. Mediano

(Oral presentation)

Recurrent neural networks (RNNs) must hold information in memory over time, which creates a storage problem: too many things to remember, too little space. This paper shows RNNs handle this similarly to how they handle too many input features -- by overlapping representations. It identifies a clever geometric trick RNNs use: parking things-to-be-forgotten in a "blind spot" where the network's output mechanism can't accidentally read them out.

Correlations in the Data Lead to Semantically Rich Feature Geometry Under Superposition.

Authors: Lucas Prieto, Edward Stevinson, Melih Barsbey, Tolga Birdal, Pedro A. M. Mediano

(Poster presentation)

Neural networks pack more concepts into their internal representations than they have space for -- a phenomenon called superposition. Previous work assumed this always causes harmful "crosstalk" between concepts. This paper shows that when concepts naturally co-occur (like "December" and "Christmas"), the crosstalk can actually be helpful, and explains why language models naturally group related concepts together in clusters and circles.

VITA: Zero-Shot Value Functions via Test-Time Adaptation of Vision-Language Models (ICLR 2026)

Authors: Christos Ziakas, Alessandra Russo

Description: Vision-Language Models (VLMs) can estimate how close a robot is to completing a task based on visual observations and a text description. VITA enhances this capability by enabling VLMs to adapt in real time to improve their progress estimates as the robot executes the task. VITA updates a small part of the model at each time step through an adaptation mechanism that is learned to improve progress estimation while incorporating past information in temporal order. As a result, VITA can act as an effective reward model to train generalist robots to act across different tasks, environments, and robot types.

A Structured, Tagged, and Localized Visual Question Answering Dataset with Full Sentence Answers and Scene Graphs for Chest X-ray Images

Authors: Philip Müller, Friederike Jungmann, Georgios Kaissis, Daniel Rueckert

Visual Question Answering (VQA) lets users interact with AI by asking questions about medical images, such as chest X-rays, and receiving specific answers that support a more intuitive and flexible use. Training such systems, however, requires large, annotated datasets. Additionally, for thorough interpretation, we need detailed answers accompanied by visual grounding, such as bounding boxes that highlight relevant regions. Existing datasets do not provide this level of detail, and manually annotating such large-scale data is extremely costly. To address this, the authors developed an automatic pipeline that extracts structured information from radiology reports, from which we generate our new CXR-QBA dataset. It contains 42 million question–answer pairs with multi-part answers, bounding boxes, and structured tags, making it the largest and most sophisticated chest X-ray VQA dataset to date and enabling more accurate, interpretable, and interactive AI tools for medical imaging. The authors thoroughly assess the quality of the dataset and make it publicly available for researchers.

Beyond Uniformity: Regularizing Implicit Neural Representations through a Lipschitz Lens

Authors: Julian McGinnis, Suprosanna Shit, Florian A. Hölzl, Paul Friedrich, Paul Büschl, Vasiliki Sideri-Lampretsa, Mark Mühlau, Philippe C. Cattin, Bjoern Menze, Daniel Rueckert, Benedikt Wiestler

Neural networks that represent signals like images, shapes, or MRI scans as continuous functions, known as implicit neural representations, face a fundamental trade-off: they must be flexible enough to capture fine details, yet smooth enough to generalize well and avoid generating artifacts. This work introduces a principled framework for managing this trade-off by controlling how much each component of the network is allowed to change its output in response to small input changes (i.e., the "Lipschitz budget"). Rather than applying the same rigid constraint everywhere, the method derives a meaningful total budget from task-specific knowledge, such as how much tissue can physically stretch in medical imaging, and then distributes it non-uniformly across different architectural components. Experiments on 3D shape reconstruction, lung registration, and image inpainting show that this flexible budgeting strategy yields smoother, more accurate results than conventional approaches, offering practitioners an interpretable tool for designing better neural representations.

The acceptance of these papers highlight the Department’s strength in translating foundational advances in argumentation, formal verification, and reinforcement learning into practical, deployable systems. Congratulations to all authors on this outstanding achievement.