A researcher's blog post describes turning constraint models into playable puzzle games, based on his paper on scaling Sudoku as a constraint problem. It shares MiniZinc models, a repository of 434,201 Sudoku instances, and playable versions of nine puzzle games.
This paper investigates whether tabular foundation models (TFMs) like TabPFN, TabICL, TabDPT, and TabFM produce predictions consistent with any joint distribution. It demonstrates that all evaluated TFMs violate both marginalization and factorization consistency for classification and regression, questioning their Bayesian inference claims.
This paper proposes a unified definition of uncertainty as pointwise posterior risk and introduces a theory-backed benchmark using semi-synthetic datasets to directly compute oracle epistemic and aleatoric uncertainty, enabling fine-grained evaluation beyond proxy tasks.
This paper proposes OG-SPR, a model-free visual RL algorithm that combines latent self-prediction with observation prediction to learn dynamics-aware representations, achieving improved sample efficiency on DeepMind Control Suite tasks.
This paper introduces THBKG, a temporal heterogeneous biomedical knowledge graph covering 110k entities and 11.1M edges with yearly evidence timestamps, designed to predict whether target-disease pairs entering Phase II trials advance to Phase III using only evidence available at the decision time. The graph-based approach outperforms direct-evidence baselines, especially for pairs lacking direct evidence, and the authors release it as a continually updated resource.
This arXiv paper investigates whether simple linear transformations can translate representations across nine heterogeneous text embedding models, finding that shared structure and transferability depend jointly on architecture, training objective, pooling, and data distribution, challenging the notion of universal latent compatibility.
BioM-JEPA introduces a joint-embedding predictive architecture that learns single-cell representations by predicting graph-connected gene blocks instead of individual genes, showing improved efficiency and downstream performance in perturbation-response tasks.
This arXiv paper introduces CohortHijack, a robustness audit that removes non-target cells from single-cell query cohorts to test how annotation tools can be manipulated without altering the target cell's expression profile. It shows that structured removal and search strategies can change refined labels in popular pipelines while preserving the target, identifying query cohort composition as a vulnerability surface.
Presents FormBharo, a voice agent that uses LLMs with rule-based controls to fill structured forms over phone calls for low-literacy Hindi-speaking users in India, piloted with ARMMAN. The paper also introduces FormVoiceAgentBench, a benchmark of 3,760 multi-turn conversation tests, and shows that end-to-end evaluation is necessary since component-level performance does not predict full form completion.
EpiBench is a new closed-book, sequence-based benchmark for evaluating how well LLMs understand epitopes across five antibody-drug-discovery tasks, finding that current models capture partial signals but struggle with antibody-specific reasoning.
This paper surveys clinical communication processing using LLM-generated synthetic data and presents 13 case studies across EMS reports, nurse handoffs, and more, showing that synthetic data can bootstrap clinical NLP systems.
This paper proposes FALM-PINN, an alternating Levenberg-Marquardt training framework for physics-informed neural networks that uses Fourier-enhanced features to address spectral bias and representation-coefficient coupling, achieving up to two orders of magnitude lower errors on high-frequency and nonlinear PDEs.
This paper introduces MERIT, a training-free agent that uses causal episodic memory of past repair outcomes to improve subsequent Text-to-SQL generations, boosting execution accuracy on Spider and BIRD benchmarks.
This paper presents a comparative evaluation of post-hoc analysis methods (Random Forest surrogate, LIME, PCA) for detecting structured patterns in clustering results, using synthetic datasets with injected patterns. It finds that none of the methods consistently detects all pattern types, highlighting a gap in existing explainability tools.
MACRO is a framework that learns task-specific execution routes over frozen LLM layers using Markov chain-based routing, improving reasoning accuracy without modifying model weights. It outperforms prior routing approaches while reducing search time significantly.
Introduces Synthetic Query Probing (SQP), a reference-free method for learning mappings between similarity score distributions across embedding models, enabling threshold portability in RAG systems. Experiments show systematic score distortions and isotonic regression as the best calibration approach.
Introduces Fast Evidential Rule Learning (FERL), a method for interpretable classification that produces evidential outputs and can abstain when uncertain, with theoretical stability guarantees and strong empirical results across tabular and concept-bottleneck benchmarks.
A research paper presenting a learning-to-rank framework for selecting efficient tensor-network contraction plans for GPU-accelerated quantum circuit simulation, using gradient-boosted rankers trained from GPU measurements.
This paper introduces MameLoshnLM, the first open-source 8B-parameter Yiddish language model, along with the Oytser pretraining corpus and Kashes evaluation benchmark. It demonstrates that continued pretraining on high-quality Yiddish data outperforms general multilingual models, highlighting the value of dedicated low-resource language modeling.
This paper proposes predicting task difficulty for LLM agents without running expensive rollouts, studying the problem across 17 agentic benchmarks and showing that token-level entropy is a useful predictive signal.