Clase general (Articles)

R0:aa77bca17589f025d763d708bbb99028-Can LLMs Improve Their Own Harness? A Controlled Study on ARC-AGI-3

Can LLMs Improve Their Own Harness? A Controlled Study on ARC-AGI-3

If a language model can rewrite the software it uses to play a game, how can it tell whether an edit was worth keeping? Jens Lehmann and Sahar Vahdati let each model build a program, play five practice games and revise that program twice. They then compared all three versions on twenty games whose results had been withheld during development. Across 27 independent sequences, the practice scores rose by 8.42 points on average. On the reserved games, the gain shrank to 2.80 points, and twelve sequences finished below their starting score. Cost tells a similarly mixed story. Eighteen final programs were cheaper to run, but only nine of those also played better. By preserving the original programs and their descendants, the study reveals something a final leaderboard cannot show on its own: when an agent already plays well, how much of that ability did it acquire through revision, and how much was present from the start?

Can LLMs Improve Their Own Harness? A Controlled Study on ARC-AGI-3 Leer más »

R0:abbc9fad0e8a3823e93438ffbc99b607-FrontierPhysics: Evaluating Agents for End-to-End Physics Research

FrontierPhysics: Evaluating Agents for End-to-End Physics Research

Can a research agent tell when its own calculation has led it astray? In FrontierPhysics, physicists turned investigations from their own laboratories and research groups into 56 problems. An agent receives a brief, explores the literature, chooses a method and submits a manuscript alongside results that an examiner can check. The strongest configuration passed 46 of 168 trials. It solved 25 problems at least once in three attempts, but only eight on every attempt. The most revealing findings emerged when physicists revisited the experimental records. Some agents had checked a result using the very assumption that produced it. In other cases, a sound answer had failed the examiner’s test. Nine such judgements changed in the detailed audit of one configuration. Together, the repeated trials and the expert review illuminate a practical question for any physicist considering an agent’s work. What evidence would persuade another researcher that the result survives a genuinely independent check?

FrontierPhysics: Evaluating Agents for End-to-End Physics Research Leer más »

R0:6c8a4ce33496b8be6ac636aa4dfd0fef-Proving and Improving the Platonic Representation Hypothesis with JEPAs

Proving and Improving the Platonic Representation Hypothesis with JEPAs

Can a compact text encoder learn representations that agree with vision more closely than those of a much larger language model? This preprint investigates that question through a conditional mathematical theorem and experiments comparing Joint-Embedding Predictive Architectures (JEPAs) with next-token prediction. Its central proposal is to train on different descriptions of the same scene, encouraging models to preserve shared information. A 109-million-parameter text JEPA achieves stronger visual alignment than the evaluated language models, including models with up to 72 billion parameters. The reported gains concern representation geometry, retrieval and selected classification tasks. The paper also combines a JEPA with an LLM decoder to generate image descriptions. This publication examines those contributions alongside the assumptions of the theorem, the meaning of the reported sample-efficiency advantage, the role of generated training captions and the limits of comparisons with general-purpose language models. The findings suggest a practical direction for representation learning while leaving broader claims about reasoning, planning and autonomous world models open.

Proving and Improving the Platonic Representation Hypothesis with JEPAs Leer más »

R0:36cbb300eb0fca3fc0ffa717d93011b5-Explainable deep learning improves human mental models of self-driving cars

Explainable deep learning improves human mental models of self-driving cars

How can a passenger understand why a self-driving car behaves unexpectedly? This study introduces the Concept-Wrapper Network (CW-Net), which connects a learned planner to human-readable concepts and examines whether explanations improve understanding of vehicle behaviour. Its evaluation combines private-track experiments with a real autonomous vehicle, expert and non-expert studies, and an online study using public-road recordings collected while a human drove. Participants receiving concept explanations performed better on selected questions about surprising situations. The accompanying critique distinguishes the architecture in which concepts determine the planning reward from an alternative that provides diagnostic explanations without determining trajectory selection. It also examines concept-recognition errors, the inclusion of caution questions in a prediction measure, and uncertainty about effects during ordinary situations. The publication offers evidence about interpretable planning and human understanding, while explaining why these results do not establish safer autonomous operation on public roads or unrestricted replication of the complete vehicle system.

Explainable deep learning improves human mental models of self-driving cars Leer más »

R0:d8e45fda6267d330d183ccaffd9b1502-LLM-Based Agents for Cybersecurity: A Systematic Review of Architectures, Applications, and Open Challenges

LLM-Based Agents for Cybersecurity: A Systematic Review of Architectures, Applications, and Open Challenges

This systematic review examines how language-model agents are being used in cybersecurity, from penetration testing and threat intelligence to incident response. It classifies 59 studies published between January 2022 and April 2026 by security function, agent architecture, knowledge augmentation, human oversight and evaluation rigor. The taxonomy helps readers compare single-agent tools with coordinated multi-agent systems and identify where evidence remains limited. Penetration testing accounts for just over half of the selected studies; none is classified as a production-grade evaluation. That finding describes this corpus, rather than proving an absence of production evidence across the field. The accompanying editorial critique checks the review’s counts and supporting records, distinguishes architectural prevalence from demonstrated superiority, and examines the subsequent withdrawal of a prominent incident-response result cited by the review. The publication offers a useful map of approaches and open challenges, while showing why laboratory performance, incomplete screening records and inconsistent supplementary versions require careful interpretation.

LLM-Based Agents for Cybersecurity: A Systematic Review of Architectures, Applications, and Open Challenges Leer más »

R0:46da84dd896e47151a6ab7d4d611099c-Large language models detect methodological errors that human peer review misses in clinical trials

Large language models detect methodological errors that human peer review misses in clinical trials

Can language models help reviewers find methodological weaknesses in clinical-trial reports? This preprint tests fifteen models on documents containing deliberately inserted errors, first using reports from an earlier human-review experiment and then extending the task to 58 trials published in 2026. It examines detection scores, concerns raised about control reports and the time required to produce reviews. The results suggest potential value for AI-assisted methodological screening, but the comparisons require care: the human benchmark is historical, model coverage differs across trials, and other language models score the reviews against a predefined answer key. The accompanying critique explains why the highest major-error score is not sensitivity across every error type or evidence of superiority over contemporary reviewers on all recent trials. It also distinguishes agreement between evaluators from independent validation. Readers can examine both the reported gains and the limits of a task that does not inspect underlying trial data, analysis code or the complete peer-review process.

Large language models detect methodological errors that human peer review misses in clinical trials Leer más »

R0:4356b5e07c7adfaa1d015773dc56f061-A number with no recorded reason may end up deciding for a patient

A number with no recorded reason may end up deciding for a patient

Transparency in artificial intelligence has become a question of provenance. Yet identifying where content came from does not explain why a decision was made. This editorial examines AI-generated scientific hypotheses, expert assessments recorded as scores, and experiments showing that a model’s explanation may omit influences that changed its answer. It considers how research incentives and simplified headlines can obscure important qualifications. The concern is especially acute when decisions guide medical research and may eventually affect patients. The article argues for preserving the reasons behind judgments, so scientific conclusions remain open to scrutiny rather than becoming an archive of unexplained results.

A number with no recorded reason may end up deciding for a patient Leer más »

R0:10f953633b3db3b4dc07e63a493e9b66-Contemporary AI lacks the imagination to diverge or negate in science

Contemporary AI lacks the imagination to diverge or negate in science

How useful are AI-generated scientific ideas to the researchers who understand the underlying work? This study asks authors of recent preprints to evaluate model-generated hypotheses for novelty, feasibility, plausibility and willingness to adopt them. Its source corpus spans 121,640 preprints, predominantly in biology and medicine, with smaller chemistry and social-science contributions. The reported results compare the diversity of ideas from different model classes, show how scientific preferences vary, and examine whether automated evaluators can reproduce expert ratings. A reward model trained on those ratings is also evaluated. The accompanying editorial critique separates these findings from broader claims about scientific imagination. It examines the restricted disciplinary coverage, self-selected participation and the distinction between a relative shortage of explicit null hypotheses and their complete absence. It also asks what is lost when numerical judgments are retained without the reasons behind them. The publication offers a substantial case study of AI-assisted hypothesis generation and of the evidence needed to assess it.

Contemporary AI lacks the imagination to diverge or negate in science Leer más »

R0:b621309bd8d28cda43682baaeb45707f-An Embarrassingly Simple Graph Heuristic Reveals Shortcut-Solvable Benchmarks for Sequential Recommendation

An Embarrassingly Simple Graph Heuristic Reveals Shortcut-Solvable Benchmarks for Sequential Recommendation

Do standard recommendation benchmarks genuinely require sophisticated models, or can their structure reward simpler shortcuts? This study investigates that question with a graph heuristic that starts from a user’s last one or two interactions, retrieves nearby candidate items and ranks them by feature similarity. Across fourteen datasets, the authors report that the heuristic remains competitive on ten, with model rankings changing as dataset properties change. They examine local transition patterns, similarity between successive items and the value of longer interaction histories as possible explanations. The accompanying critique distinguishes those observed associations from causal evidence: the study compares datasets rather than removing a shortcut within a fixed benchmark. It also notes that the apparently simple method uses representations from a pretrained text encoder and that comparisons depend on the authors’ implementations and chosen settings. The paper is useful for understanding why a strong leaderboard result does not, by itself, establish the advanced capability a benchmark is intended to test.

An Embarrassingly Simple Graph Heuristic Reveals Shortcut-Solvable Benchmarks for Sequential Recommendation Leer más »

R0:21222df107462b80fde54a74d060bc7e-When silence is safer: a review and decision-theoretic framework for LLM abstention in healthcare

When silence is safer: a review and decision-theoretic framework for LLM abstention in healthcare

In healthcare, a fluent answer can be less useful than a justified decision not to answer. This review examines how language models abstain when information is uncertain or when a response could be harmful. It distinguishes uncertainty-driven abstention from safety-driven refusal and finds that many existing approaches depend on additional tools rather than the model alone. The authors report limited evaluation in realistic medical settings and continuing difficulty with prompts that should not receive an unrestricted answer. They then formalize the choice between answering and withholding a response in terms of uncertainty and possible harm. This leads to MedSAFE, a framework for evaluating abstention in clinical dialogue, illustrated through a proof-of-concept pilot based on scenarios drawn from the review. The publication helps readers examine when an answer should be deferred, how abstention can be assessed, and why a pilot framework should be distinguished from evidence of reliable deployment across clinical practice.

When silence is safer: a review and decision-theoretic framework for LLM abstention in healthcare Leer más »