Nonsense Helps: Prompt Space Perturbation Broadens Reasoning Exploration

Reinforcement learning with verifiable rewards, particularly Group Relative Policy Optimization (GRPO), has significantly advanced the reasoning capabilities of Large Language Models (LLMs). However, in complex tasks, GRPO frequently suffers from the ``zero-advantage problem'': when all sampled rollouts for a query fail, the relative advantage collapses to zero. Consequently, the model loses effective training signals for these questions, wasting the training data and computational budget. While simply increasing the sampling budget for these questions is a common remedy, the static sampling policy inherently constrains reasoning exploration, limiting the success rate. In this paper, we propose Lorem Perturbation for Exploration (LoPE), a simple yet effective training framework to break this exploration bottleneck. We posit that task-irrelevant prompt-space perturbations can shift the model's output distribution enough to unlock orthogonal reasoning pathways for hard questions. Specifically, LoPE prepends seq

PDF

ChatDashboard: A Framework to collect, link, and process donated WhatsApp Chat Log Data

In this paper, we present ChatDashboard, a framework for collecting, linking, and processing donated WhatsApp chat log data. The framework consists of the WhatsR R package for parsing, anonymizing, and preprocessing donated WhatsApp chat logs, the ChatDashboard R Shiny web app for uploading, reviewing, and securely donating WhatsApp chat logs, and DashboardTester, an automated script for testing the correct setup of the framework by simulating participants. With ChatDashboard, researchers can set up their own data collections to gather transparently donated WhatsApp chat log data from consenting participants and link them to survey responses. It enables researchers to retrospectively collect highly granular data on interpersonal interactions and communication without building their own tools from scratch. We briefly discuss the advantages of donated WhatsApp chat log data for investigating social relationships and provide a detailed explanation of the ChatDashboard framework. Additionally, we provide a step-b

14 citations

PDFHTMLAltmetric

Who Benchmarks the Benchmarks? Towards Comprehensive Evaluation of Commonsense Reasoning Benchmarks

Commonsense reasoning is a key language model capability, as it is purportedly a prerequisite for many basic tasks, unlike specific factual knowledge. It is often measured with multiple-choice questions (MCQ) benchmarks, e.g. HellaSwag and PIQA. Some of these benchmarks, however, are outdated and contain numerous validity issues. We illustrate some typical validity issues with a case study on HellaSwag, one of the most popular and problematic benchmarks for commonsense reasoning. The issues we find range from basic ungrammaticality and numerous typos to misleading prompts or equally correct options. We show that if we remove question prompts or replace them with "Lorem ipsum dolor...", about 68% of model predictions do not change. We argue that this occurs due to inner flaws in the benchmark, not mere contamination that might be present in some models. Since benchmark scores are an essential part of model selection in both research and commercial applications, these issues can have severe consequences. Based

PDF

Document versioning system

Glavna cilja magistrske naloge sta pregled obstoječih sistemov za verzioniranje podatkov in njihovih funkcionalnosti ter izdelava lastne spletne aplikacije – sistema za verzioniranje dokumentov. Verzioniranje dokumentov je postopek hranjenja dokumenta in vseh njegovih sprememb, ki so se nad dokumentom zgodile skozi čas. Za potrebe izdelave spletne aplikacije, ki jo uporabniki poganjajo znotraj spletnega brskalnika, smo najprej zgradili podatkovni model. Le tega smo nato implementirali znotraj relacijske podatkovne baze PostgreSQL, v kateri se hranijo podatki, potrebni za delovanje spletne aplikacije. Razvita spletna aplikacija DRS omogoča izdelavo poljubne podatkovne datotečne strukture – poljubna hierarhija map s pripadajočimi dokumenti. Registriranim uporabnikom omogoča dodeljevanje različnih dostopnih pravic na posamezni mapi, omejitev tipa datotek, ki jih uporabniki lahko shranjujejo znotraj aplikacije, in verzioniranje dokumentov. Uporabniki lahko dokumente znotraj izbrane mape pregledujejo v dveh različ

0 citations

PDFHTML

Learning Long-Range Representations with Equivariant Messages

Machine learning interatomic potentials trained on first-principles reference data are becoming valuable tools for computational physics, biology, and chemistry. Equivariant message-passing neural networks, including transformers, achieve state-of-the-art accuracy but rely on cutoff-based graphs, limiting their ability to capture long-range effects such as electrostatics or dispersion, as well as electron delocalization. While long-range correction schemes based on inverse power laws of interatomic distances have been proposed, they are unable to communicate higher-order geometric information and are thus limited in applicability. To address this shortcoming, we propose the use of equivariant, rather than scalar, charges for long-range interactions, and design a graph neural network architecture, LOREM, around this long-range message passing mechanism. We consider several datasets specifically designed to highlight non-local physical effects, and compare short-range message passing with different receptive fi

PDF

"L'extinction est la règle, la survie est l'exception": Regroupement de maliciels selon leurs comportements

0 citations

HTML

Long-Range QKD without Trusted Nodes is Not Possible with Current Technology

A recently published patent (https://www.ipo.gov.uk/p-ipsum/Case/PublicationNumber/GB2590064) has claimed the development of a novel quantum key distribution protocol purporting to achieve long-range quantum security without trusted nodes and without use of quantum repeaters. Here we present a straightforward analysis of this claim, and reach the conclusion that it is largely unfounded.

9 pages, 2 figures and 1 table

PDFAltmetric

Long-Range Machine Learning Interatomic Potentials for Defect Energetics in SrTiO$_3$

Machine learning interatomic potentials (MLIPs) have advanced rapidly in recent years, yet the majority of models remain semilocal in nature and neglect long-range electrostatic interactions. A growing number of long-range models have emerged to address this limitation, but systematic benchmarks comparing their performance on physically realistic systems remain scarce. In this work, we assess three long-range MLIPs -- MACE-POLAR, MACE-LES, and LOREM, against a short-range baseline MACE model, on their ability to capture long-range Coulombic interactions, progressing from idealised synthetic point-charge systems to first-principles data on realistic crystal structures. This progression allows us to isolate model's capability to learn electrostatic interactions from the many competing energy contributions, such as dielectric screening and exchange-correlation, that are present in DFT total energies. As a physically motivated benchmark system, we focus on strontium titanate SrTiO$_3$ with Sr--O Schottky vacancy

20 pages, 13 figures

PDF