gabriel

i1: A Fully Open Recipe for Competitive Text-to-Image Diffusion Models Using Only Public Data

The paper presents a systematic empirical study of text-to-image diffusion models through 300+ controlled experiments, yielding a fully open 3B-parameter model (i1) trained on publicly available datasets. Key findings include that equal weighting is a strong default for mixing curated datasets, and larger text encoder adapters improve performance with minimal parameter overhead. The model is competitive with leading closed-weight and open-weight models across five benchmarks, and outperforms the best existing fully open model by an average of 29.5 absolute percentage points. All weights, code, and data processing pipelines are released.

Integrated Tactile Sensor with Deformable Pad Enables Slip-Aware In-Hand Manipulation

This paper presents a novel, deformable tactile sensor for in-hand manipulation, uniquely integrating velocity, force/torque, and pressure map sensing within a single compliant structure. Designed for rapid, low-cost fabrication, the sensor demonstrates robust tracking of diverse surfaces and is intended for slip-aware control applications. Comprehensive experiments are detailed to evaluate its capabilities and limitations.

Bounded Solutions and Critical Thresholds in Quasilinear Reaction-Diffusion Equations

This paper investigates a quasilinear reaction-diffusion equation in $N\geq3$ dimensions. For initial data satisfying specific conditions related to its decay at infinity, the authors prove that the solution remains uniformly bounded. This boundedness is shown to be a critical behavior dependent on the parameter $\sigma$, distinguishing it from cases where solutions grow. Additionally, the paper establishes precise rates for the expansion of the solution's support and its decay in regions outside this support.

Structure-Aware Distractor Modeling Drives MCQ Difficulty Estimation Gains

The paper demonstrates that explicitly modeling distractors (wrong answer choices) as separate, structured inputs significantly improves automatic difficulty estimation for multiple-choice questions, compared to naive baselines that concatenate all text or ignore distractors. On Chilean educational datasets (4,114 MCQs), an order-aware architecture with positional tags achieves R²=0.83 (Natural Sciences) and R²=0.71 (Social Sciences), while an order-invariant variant retains nearly identical accuracy with half the parameters. This suggests that distractor content, not just stem-and-key text, is the primary driver of difficulty signal, and that even simple structural representations yield substantial gains.

Arithmetic Terms Exhibit Undecidability, Turing Completeness, and Proof Verification Capabilities

This paper demonstrates the surprising computational power of arithmetic terms, which are finite compositions of basic arithmetic operations on natural numbers. It is shown that determining if a one-variable arithmetic term evaluates to zero, or if two such terms are equivalent, is undecidable. Furthermore, arithmetic terms can represent Turing-complete systems, enabling the expression of chaotic functions like the Logistic Map and even a "wise" term capable of verifying proofs within a formalized theory up to a certain bound.

AGN Feeding-Feedback Cycle Traced Through Spatially Resolved Stellar Populations

A spatially resolved near-IR spectroscopic analysis of three local AGN hosts (NGC 3884, 3C 293, CGCG 012-070) reveals that their inner kiloparsecs are dominated by intermediate-to-old stellar populations with super-solar metallicity, yet all show clear signatures of recent 'rejuvenation' via young stellar components. A central drop in metallicity across nearly all fits suggests metal-poor gas inflows fuel both circumnuclear star formation and AGN activity, while non-stellar components (featureless continuum and hot dust) decrease radially outward. The results support a self-regulating feeding-feedback loop where gas inflows trigger star formation, and subsequent stellar mass loss sustains the AGN.

Multiband superconductivity in the kagome-lattice superconductor Re2Zr

A Causal Probabilistic Framework for Perception-Informed Closed-Loop Simulation of Autonomous Driving

Evidence Markets

Affordance-Based Hierarchical Reinforcement Learning for Quadruped Pedipulation

Continual Learning Bench: Evaluating Frontier AI Systems in Real-World Stateful Environments

Early Results from GLASS-JWST. XXVI. Spatially Resolved Star Formation and Balmer Decrements at $1.1<z<2.3$ from NIRISS Slitless Spectroscopy

Unified formulas for conditional quantities and transportation functionals

PolyGraphPy: A unified Python framework for atomistic simulation and machine learning-driven polymer design

`pandemonium`: High Dimensional Analysis in Linked Spaces

Commutator-Induced Uncertainty in VAEs

IyàwóBench: A Benchmark for Evaluating Large Language Model Clinical Triage Accuracy on Undifferentiated Febrile Illness in Nigerian Primary Health Settings

A comparison between Galactic magnetic field models and polarized synchrotron emission with C-BASS at 4.76 GHz and S-PASS at 2.3 GHz

Linear spaces of rational integrable 1-forms

Do Factual Recall Mechanisms Carry over from Text to Speech in Multimodal Language Models?

Subnormalizers and character correspondences in $p$-solvable groups

From Arbitrage Removal to Density Extraction: A Model-Free Framework for Short-Dated Options

ML-Driven Quasar Catalog from DES DR2 Achieves High-Precision Classification and Photo-z Estimation at Cosmological Scales

A cross-match of DES DR2 photometry with SDSS DR16 spectroscopic labels yielded a training set of 168,738 point-like objects, enabling a KNN classifier to separate quasars from stellar contaminants at 0.99 precision and 0.77 recall using only four-band PSF magnitudes. Photometric redshifts were estimated via a hybrid ensemble combining ANNz's Boosted Decision Tree with a scikit-learn Decision Tree Regressor, with a stacked outlier classifier added to suppress catastrophic redshift failures. The resulting cleaned catalog contains 675,683 objects suitable for large-scale structure studies over 0 < z < 3, with a broader 872,372-object sample extending reliable coverage to z ≈ 4.

Kelpie: Zero-Query Adversarial Mimicry Attacks Expose Fundamental Weaknesses in ML-Based Binary Function Classifiers

Kelpie is a black-box, query-free adversarial framework that executes targeted mimicry attacks against ML-based binary function classifiers — a significantly stronger threat model than prior evasion work, which typically required iterative classifier queries. By applying functionality-preserving code transformations, Kelpie causes malicious payloads (e.g., keyloggers, wipers) to be misclassified as a chosen target class, all without any direct interaction with the victim model. The framework was validated against six state-of-the-art classifiers spanning different architectures, making it the first demonstrated mimicry attack in a true zero-query setting and raising serious questions about the robustness guarantees of current binary analysis ML pipelines.

RL Fine-Tuning Encodes Reusable Combinatorial Solvers into LLM Weights, Beating Repeated Sampling by 91x

Rather than treating combinatorial optimization as a repeated inference-time search problem, this paper demonstrates that GRPO-based RL fine-tuning of a code LLM (Qwen2.5-Coder-14B-Instruct) can internalize solver construction into model weights, producing a reusable algorithm generator. On Synergistic Dependency Selection (SDS) — a constrained Quadratic Knapsack variant — the RL-trained policy reduces the gap to the Virtual Best Solver from ~28.7% (Best-of-64 baseline) to 5.0%, while cutting post-generation execution cost by 91x. The trained model converges to a correct constraint-aware Simulated Annealing implementation in 99.8% of feasible outputs, fixing a systematic Metropolis acceptance rule misimplementation present in the base model. Transferability to Job Shop Scheduling is positive but narrow, and the recipe is sensitive to reward normalization and feasibility gate design.

Sub-€40 Open-Source Wearable Achieves 0.77 F1 Four-Stage Sleep Classification Without Gel Electrodes

OSSMM is a fully open-source sleep research platform built from 3D-printed parts and COTS components for under €40, capturing EOG, putative EEG, movement, and pulse via a gel-free forehead headband with wireless connectivity. A 15-night single-subject proof-of-concept achieved a Macro F1 of 0.770 and accuracy of 0.776 for Wake/Light/Deep/REM staging against a PSG-validated reference monitor (κ=0.63 with PSG). Critically, classification is driven primarily by spectral features from just two frontal electrodes using reusable conductive thermoplastic polyurethane (CTPU) from commercial chest straps — no dedicated ground reference required — suggesting that minimal electrode configurations may be sufficient for practical sleep staging.

A Comparative Study of Model Selection Criteria for Symbolic Regression

Development and validation of a forward 0.7--4 MeV quasi-monoenergetic neutron capability at the CN Van de Graaff of LNL

Empirical scaling laws in balanced networks with conductance-based synapses

AOI-SSL: Self-Supervised Framework for Efficient Segmentation of Wire-bonded Semiconductors In Optical Inspection