Miniml Research: shaping frontier AI into production reality
Technical papers, experiments, and evaluations
Follow the Miniml team’s latest technical work across reasoning, long context, efficiency, and evaluation. We publish methods, results, and implementation insights built for real-world AI systems.
Fast and Expressive Multi-Byte Prediction with Probabilistic Circuits
July 6, 2026 • Andreas Grivas, Lorenzo Loconte, Emile van Krieken, Piotr Nawrot, Yu Zhao, Euan Wielewski, Pasquale Minervini, Edoardo Ponti, and Antonio Vergari • ICML 2026
MTPC uses probabilistic circuits to model dependent future bytes while retaining efficient parallel prediction and speculative decoding.
Learning GUI Grounding with Spatial Reasoning from Visual Feedback
July 6, 2026 • Yu Zhao, Wei-Ning Chen, Huseyin A. Inan, Samuel Kessler, Lu Wang, Lukas Wutschitz, Fangkai Yang, Chaoyun Zhang, Pasquale Minervini, Saravan Rajmohan, and Robert Sim • ICML 2026
GUI-Cursor turns GUI grounding into an interactive search in which a vision-language model refines cursor positions from visual feedback.
Neural-HSS: Hierarchical Semi-Separable Neural PDE Solver
July 6, 2026 • Pietro Sittoni, Emanuele Zangrando, Angelo A. Casulli, Nicola Guglielmi, and Francesco Tudisco • ICML 2026
Neural-HSS uses hierarchical semi-separable structure to build a parameter- and data-efficient neural solver for broad classes of PDEs.
VLM-RobustBench: A Comprehensive Benchmark for Robustness of Vision-Language Models
July 6, 2026 • Rohit Saxena, Alessandro Suglia, and Pasquale Minervini • ICML 2026
VLM-RobustBench tests vision-language models across 133 corrupted-image settings, exposing particular fragility to spatial distortions.
PiCSAR: Probabilistic Confidence Selection and Ranking for Reasoning Chains
July 2, 2026 • Joshua Ong Jun Leang, Zheng Zhao, Aryo Pradipta Gema, Sohee Yang, Wai-Chung Kwan, Xuanli He, Wenda Li, Pasquale Minervini, Eleonora Giunchiglia, and Shay B. Cohen • Findings of ACL 2026
PiCSAR ranks sampled reasoning chains by their joint reasoning-and-answer likelihood, improving best-of-n selection without training.
Logit-Contribution Scoring Identifies Non-Literal Retrieval Heads
July 1, 2026 • Aryo Pradipta Gema, Beatrice Alex, and Pasquale Minervini • Mechanistic Interpretability Workshop at ICML 2026
LOCOS detects attention heads that retrieve meaning rather than copy tokens by measuring each head’s contribution to the answer logit.
Stuart-Landau Oscillatory Graph Neural Network
June 29, 2026 • Kaicheng Zhang, David N. Reynolds, Piero Deidda, and Francesco Tudisco • The Web Conference 2026
SLGNN uses complex-valued Stuart–Landau dynamics to preserve amplitude and phase information in deep graph neural networks.
SCOPE: Self-Play via Co-Evolving Policies for Open-Ended Tasks
May 29, 2026 • Wai-Chung Kwan, Aryo Pradipta Gema, Joshua Ong Jun Leang, and Pasquale Minervini • arXiv 2026
SCOPE trains language models on open-ended tasks without curated prompts by co-evolving challenger and solver policies through self-play.
Can Retrieval Heads See Images? Multimodal Retrieval Heads in Long-Context Vision-Language Models
May 26, 2026 • Aaron Branson Cigres Li, Zhaowei Wang, Yu Zhao, Yiming Du, Haobo Li, Xiyu Ren, Ginny Wong, Simon See, Lishu Luo, Haodong Duan, Pasquale Minervini, and Yangqiu Song • EMNLP 2026
A multimodal retrieval-head detector reveals the sparse attention circuits that locate textual and visual evidence in long-context vision-language models.
Test-Time Accuracy-Cost Control in Neural Simulators via Recurrent-Depth
April 23, 2026 • Harris Abdul Majid, Pietro Sittoni, and Francesco Tudisco • ICLR 2026
RecurrSim gives neural simulators explicit test-time control over the trade-off between computational cost and predictive accuracy.
Learning and Reasoning on Knowledge and Heterogeneous Graphs in the era of Graph Foundation and Large Language Models
April 22, 2026 • Matteo Zignani, Pasquale Minervini, Roberto Interdonato, and Manuel Dileo • ESANN 2026
A tutorial survey unifies learning and reasoning across heterogeneous graphs, knowledge graphs, foundation models, LLMs, and temporal settings.
Analyzing LLM Instruction Optimization for Tabular Fact Verification
March 24, 2026 • Xiaotang Du, Giwon Hong, Wai-Chung Kwan, Rohit Saxena, Ivan Titov, Pasquale Minervini, and Emily Allaway • Findings of EACL 2026
A systematic study maps how DSPy instruction optimizers affect text-, SQL-, and Python-based approaches to tabular fact verification.
Complex Query Answering with Neural Link Predictors
March 6, 2026 • Miniml
This paper presents a differentiable framework that uses pre-trained neural link predictors to answer complex logical queries on incomplete knowledge graphs with higher accuracy and much less training data.
DeCoRe: Decoding by Contrasting Retrieval Heads
February 28, 2026 • Miniml Research • EMNLP 2025 Findings
DeCoRe is a training-free decoding method that contrasts retrieval heads to curb hallucinations in context-grounded generation.
Testing quantum and simulated annealers on the drone delivery packing problem
February 16, 2026 • Sara Tarquini, Daniele Dragoni, Matteo Vandelli, and Francesco Tudisco • Quantum Machine Intelligence 2026
Two QUBO formulations expose the capabilities and limitations of quantum annealing for battery-constrained drone delivery packing.
Ambient Physics: Training Neural PDE Solvers with Partial Observations
February 14, 2026 • Harris Abdul Majid, Giannis Daras, Francesco Tudisco, and Steven McDonagh • arXiv 2026
Ambient Physics learns coefficient–solution distributions for PDEs directly from partial observations, without requiring complete training examples.
Tensor factorization for temporal knowledge graph forecasting
January 29, 2026 • Manuel Dileo, Pasquale Minervini, Matteo Zignani, and Sabrina Gaito • Neurocomputing 2026
An extended tensor-factorisation model forecasts temporal knowledge graphs efficiently by encoding unseen times and regularising temporal change.
FLARE: Faithful Logic-Aided Reasoning and Exploration
January 28, 2026 • Miniml Research • Empirical Methods in Natural Language Processing (EMNLP)
FLARE pairs LLM planning with logic programming and simulation to improve faithfulness in multi-step reasoning.
GRADA: Graph-based Reranker against Adversarial Documents Attack
January 28, 2026 • Miniml Research
GRADA defends RAG pipelines by reranking retrieved documents to resist adversarial injections while preserving accuracy.
MMLongBench: Benchmarking Long-Context Vision-Language Models
January 28, 2026 • Miniml Research
MMLongBench evaluates long-context VLMs across tasks and image types, revealing gaps in long-context multimodal reasoning.
Neuro-symbolic Diffusion Models
January 28, 2026 • Miniml Research
NeSyDMs use discrete diffusion to model dependencies among symbols, improving accuracy and calibration for neurosymbolic prediction.
Activation Sparsity and Enterprise AI Efficiency
January 18, 2026 • Miniml • ICLR 2021
Activation sparsity suggests large language models can become more efficient at inference without sacrificing capability.
Low-Rank Compression of Language Models via Differentiable Rank Selection
December 14, 2025 • Sidhant Sundrani, Francesco Tudisco, and Pasquale Minervini • arXiv 2025
A differentiable method learns how much to compress each language-model layer without requiring post-compression fine-tuning.
Graph-Convolutional-Beta-VAE for synthetic abdominal aortic aneurysm generation
December 4, 2025 • Francesco Fabbri, Martino Andrea Scarpolini, Angelo Iollo, Francesco Viola, and Francesco Tudisco • Medical & Biological Engineering & Computing 2026
A graph-convolutional β-VAE generates realistic, diverse abdominal aortic aneurysm geometries from a small clinical dataset.
OpenSIR: Open-Ended Self-Improving Reasoner
November 1, 2025 • Wai-Chung Kwan, Joshua Ong Jun Leang, Pavlos Vougiouklis, Jeff Z. Pan, Marco Valentino, and Pasquale Minervini • arXiv 2025
OpenSIR lets a language model generate and solve increasingly novel problems without annotated data or external verifiers.
Neurosymbolic reasoning shortcuts under the independence assumption
September 23, 2025 • Miniml Research
Why the independence assumption in NeSy predictors can hide uncertainty and lead to shortcut reasoning.
Q-Filters: Leveraging QK geometry for efficient KV cache compression
August 12, 2025 • Miniml Research
Q-Filters compress the KV cache at inference by filtering keys using QK geometry, without training.
Inverse Scaling in Test-Time Compute
July 19, 2025 • Aryo Pradipta Gema, Alexander Hägele, Runjin Chen, Andy Arditi, Jacob Goldman-Wetzler, Kit Fraser-Taliente, Henry Sleight, Linda Petrini, Julian Michael, Beatrice Alex, Pasquale Minervini, Yanda Chen, Joe Benton, and Ethan Perez • TMLR 2026
Longer reasoning can reduce model accuracy by amplifying distraction, framing sensitivity, spurious correlations, and other failure modes.
NOISER: Bounded input perturbations for attributing large language models
April 3, 2025 • Miniml Research • Conference on Language Modeling (COLM)
NOISER estimates token attributions by injecting bounded noise into embeddings to test output sensitivity.
PosterSum: A multimodal benchmark for scientific poster summarization
February 24, 2025 • Miniml Research
PosterSum introduces a large multimodal benchmark for summarizing scientific posters into abstracts.
Contractivity of Neural ODEs: An Eigenvalue Optimization Problem
February 10, 2025 • Nicola Guglielmi, Arturo De Marinis, Anton Savostianov, and Francesco Tudisco • Mathematics of Computation 2025
A new eigenvalue-optimization method analyses and enforces contractive behaviour in neural ordinary differential equations.
Solaris: A Foundation Model of the Sun
November 25, 2024 • Harris Abdul Majid, Pietro Sittoni, and Francesco Tudisco • arXiv 2024
Solaris is a foundation model trained on a full solar cycle of multi-wavelength imagery to forecast the dynamics of the Sun’s atmosphere.
Robust Low-Rank Training via Approximate Orthonormal Constraints
December 10, 2023 • Dayana Savostianova, Emanuele Zangrando, Gianluca Ceruti, and Francesco Tudisco • NeurIPS 2023
A low-rank training method reduces neural-network cost while preserving robustness through approximate orthonormal constraints.
Implicit MLE: Backpropagating Through Discrete Exponential Family Distributions
December 6, 2021 • Mathias Niepert, Pasquale Minervini, and Luca Franceschi • NeurIPS 2021
Implicit MLE enables end-to-end learning through discrete probability distributions and combinatorial optimisation without smooth relaxations.