Articles in the Research section on Beckmann.https://beckmann.aien-UShttps://beckmann.ai/en/research/2026-07/gpt-56-disproves-statistics-assumptionhttps://beckmann.ai/en/research/2026-07/gpt-56-disproves-statistics-assumptionWharton statistician Edgar Dobriban has disproved a twenty-year-old assumption of the Benjamini-Hochberg procedure using the AI model GPT-5.6 Sol Pro. The model found a counterexample in about ninety minutes, while the predecessor GPT-5.5 remained unsuccessful after more than twenty hours. The procedure, introduced in 1995 for controlling false discoveries, is considered one of the most influential methods in statistics.Thu, 16 Jul 2026 05:18:00 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/claude-values-language-studyhttps://beckmann.ai/en/research/2026-07/claude-values-language-studyA new evaluation of 309,815 Claude conversations shows: the tone and caution of the responses strongly depend on the language and model version used. The AI model responds warmer in Hindi and Arabic, stricter and more critical in English and Russian. According to Anthropic, the four measured value axes explain only about 15 percent of these differences.Wed, 15 Jul 2026 22:23:40 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/soofi-s-german-open-ai-modelhttps://beckmann.ai/en/research/2026-07/soofi-s-german-open-ai-modelA German research consortium has released Soofi S, an open language model with 31.6 billion parameters for German and English. It was trained on Deutsche Telekom's Industrial AI Cloud in Munich, funded with around 20 million euros from the federal economics ministry. The developers report top scores among open models on combined German-English benchmarks, though these figures have not been independently verified.Wed, 15 Jul 2026 13:03:12 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/google-sensorfm-health-aihttps://beckmann.ai/en/research/2026-07/google-sensorfm-health-aiA new AI model from Google Research called SensorFM delivers better predictions than previous specialized models on 34 of 35 tested health tasks. The system was trained on smartwatch sensor data from five million people, collected over the course of a year. The results come from a study published on July 9, 2026; no public access to the model exists so far.Tue, 14 Jul 2026 05:31:28 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/oak-lab-sutton-ai-agentshttps://beckmann.ai/en/research/2026-07/oak-lab-sutton-ai-agentsReinforcement learning pioneer Richard Sutton has founded the AI company Oak Lab together with Khurram Javed. The two left John Carmack's startup Keen Technologies to do so. Oak Lab develops learning algorithms that train agents directly from experience instead of stored datasets. The long-term goal is a trillion-parameter agent that learns and plans on about twenty watts.Tue, 14 Jul 2026 01:37:15 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/agentic-sts-ai-agent-memory-card-gamehttps://beckmann.ai/en/research/2026-07/agentic-sts-ai-agent-memory-card-gameA research team doubles the win rate of an AI agent in the card game Slay the Spire 2 from three to six out of ten runs with the new memory system AgenticSTS. Instead of a growing chat log, the agent uses five separate memory layers, which keeps token usage constant. With only ten runs per configuration, the authors themselves consider the difference not statistically confirmed.Mon, 13 Jul 2026 06:33:38 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/orca-baai-world-model-roboticshttps://beckmann.ai/en/research/2026-07/orca-baai-world-model-roboticsThe Beijing research institute BAAI has introduced a world model with Orca that learns from 125,000 hours of video and 160 million event descriptions. The model outperforms larger competing systems in text and image tasks in tests and achieves the level of specialized systems in robot control, even though it saw no action labels during pre-training.Sun, 12 Jul 2026 05:37:13 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/gpt-56-sol-ultra-graph-theory-proofhttps://beckmann.ai/en/research/2026-07/gpt-56-sol-ultra-graph-theory-proofOn July 10, 2026, OpenAI published an alleged proof for the fifty-year-old Cycle Double Cover Conjecture, created by the language model GPT-5.6 Sol Ultra in under an hour with 64 subprograms. The prompt instructed the model to start from an existing proof, raising doubts within the expert community about its validity. Independent verification by mathematicians is still pending.Sun, 12 Jul 2026 02:21:52 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/arxiv-2607-08573https://beckmann.ai/en/research/2026-07/arxiv-2607-08573arXiv:2607.08573 – SHAP-Weighted Cross-Modal Expert Fusion for Emotion and Sentiment Recognition: Evidence and Limits investigates whether an explainable weighting of text, audio, and video experts based on TreeSHAP attributions performs similarly well in emotion and sentiment recognition as classical fusion methods – with an additional gain in traceability.Fri, 10 Jul 2026 14:30:33 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/arxiv-2607-08625https://beckmann.ai/en/research/2026-07/arxiv-2607-08625arXiv:2607.08625 – The complexities of patient-centred conversational artificial intelligence – An arXiv preprint shows based on over 2,000 real patient conversations and 1,164 clinically assessed cases that the communication style alone – regardless of the medical content – significantly shifts the urgency level assigned by AI chatbots.Fri, 10 Jul 2026 12:38:20 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/arxiv-2607-08652https://beckmann.ai/en/research/2026-07/arxiv-2607-08652arXiv:2607.08652 – Formal Mechanisms for Market Stability in Self-Interested Agent Societies: A Marketplace Simulation Study – An arXiv preprint tests eight rule sets for trading markets of AI agents and finds that a neutral mediation mechanism (mediation) works best and withstands even targeted attacks without the market collapsing.Fri, 10 Jul 2026 11:19:04 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/arxiv-2607-08716https://beckmann.ai/en/research/2026-07/arxiv-2607-08716arXiv:2607.08716 – Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents describes a separate memory agent that actively decides when a stored memory should be fed into the next action step for long-running AI agents. In tests with Terminal-Bench and τ²-Bench, this increases the success rate by 6.8 to 8.3 percentage points.Fri, 10 Jul 2026 09:12:59 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/arxiv-2607-08734https://beckmann.ai/en/research/2026-07/arxiv-2607-08734arXiv:2607.08734 – The Illusion of Equivalency: Statistical Characterization of Quantization Effects in LLMs: The preprint shows that post-training quantization of language models changes the specific response behavior even with moderate bit reduction – even when classical accuracy or perplexity metrics show no change.Fri, 10 Jul 2026 08:40:51 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/arxiv-2607-08740https://beckmann.ai/en/research/2026-07/arxiv-2607-08740arXiv:2607.08740 – Workflow as Knowledge: Semantic Persistence for LLM-Mediated Workflows: The preprint proposes a concept where LLM workflows, their running instances, and associated model and decision datasets are stored as persistent, queryable knowledge objects instead of ephemeral logs – with a clear distinction between deterministic computation ('derive') and model-mediated judgment ('infer').Fri, 10 Jul 2026 08:14:39 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/arxiv-2607-08745https://beckmann.ai/en/research/2026-07/arxiv-2607-08745The arXiv preprint arXiv:2607.08745 – AUTOPILOT VQA: Benchmarking Vision-Language Models for Incident-Centric Dashcam Understanding presents a benchmark with over 600 dashcam clips and 6,000+ question-answer pairs that tests how well AI systems understand safety-critical traffic incidents – with significant performance variation among the tested approaches.Fri, 10 Jul 2026 06:37:27 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/arxiv-2607-08758https://beckmann.ai/en/research/2026-07/arxiv-2607-08758arXiv:2607.08758 – Ideas Have Genomes: Benchmarking Scientific Lineage Reasoning and Lineage-Grounded Idea Generation: The preprint presents IdeaGene-Bench, a benchmark with 1,961 lineages of scientific ideas from 10 fields. It measures how well AI models trace which prior works a new idea builds upon, repairs, or recombines – current systems achieve only 27.3% exact accuracy.Fri, 10 Jul 2026 06:21:26 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/arxiv-2607-08748https://beckmann.ai/en/research/2026-07/arxiv-2607-08748arXiv:2607.08748 – Using AI-based Learning Assistants in Higher Education: A Large-Scale Descriptive Analysis: A large-scale evaluation of objective usage data from 77,543 distance learners shows how the AI learning assistant Syntea is used in everyday study life and where usage differences appear by gender, age group, study cluster, degree, and study format.Fri, 10 Jul 2026 04:58:59 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/arxiv-2607-07321https://beckmann.ai/en/research/2026-07/arxiv-2607-07321arXiv:2607.07321 – From Atomic Actions to Standard Operating Procedures: Iterative Tool Optimization for Self-Evolving LLM Agents: The arXiv preprint 2607.07321 presents the EvoSOP framework, which allows LLM agents to autonomously construct, merge, evaluate, and discard reusable standard operating procedures (SOPs) from recurring action sequences – without the need to retrain the underlying model.Fri, 10 Jul 2026 01:11:14 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/arxiv-2607-08602https://beckmann.ai/en/research/2026-07/arxiv-2607-08602arXiv:2607.08602 – Towards Precision Therapy in Hepatocellular Carcinoma: A Clinical-Reasoning LLM for Risk Stratification and Treatment Guidance – An arXiv preprint presents HCC-STAR, a language model trained on patient records that provides stage, therapy recommendations, and individual survival prognosis for liver cancer, achieving better results than established guidelines and GPT-5 in a study with 6,668 patients.Fri, 10 Jul 2026 00:00:00 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/arxiv-2607-08681https://beckmann.ai/en/research/2026-07/arxiv-2607-08681arXiv:2607.08681 – SolarChain-Eval: A Physics-Constrained Benchmark for Trustworthy Economic Agents in Decentralized Energy Markets introduces a simulation environment where autonomous AI agents control decentralized energy markets and are evaluated on market utility, physical safety, market distortion, fairness, and traceability.Fri, 10 Jul 2026 00:00:00 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/openai-swe-bench-pro-brokenhttps://beckmann.ai/en/research/2026-07/openai-swe-bench-pro-brokenOpenAI has audited the widely used coding benchmark SWE-Bench Pro (provider: Scale AI) and estimates that about 30 percent of the 731 public test tasks are faulty – due to overly strict tests, underspecified tasks, too lenient evaluations, or misleading descriptions. One example: a task required a space in the visible text, but the hidden test required two, causing correct solutions to fail. OpenAI retracts its previous recommendation for using SWE-Bench Pro and calls for new benchmarks designed by experienced developers. The figures come from OpenAI's own audit on July 8, 2026, and have not yet been independently verified externally.Fri, 10 Jul 2026 00:00:00 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/arxiv-2606-15943https://beckmann.ai/en/research/2026-07/arxiv-2606-15943arXiv:2606.15943 – Graphical-Probabilistic Modeling of Generative Flows in LLM-Native Software Systems. An arXiv preprint proposes "Generation Networks": a graphical-probabilistic model that describes LLM-based software systems as data-dependency graphs and Bayesian networks to formally capture their stochastic behavior and quantitatively compare design decisions.Thu, 09 Jul 2026 22:05:34 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/arxiv-2606-15954https://beckmann.ai/en/research/2026-07/arxiv-2606-15954arXiv:2606.15954 – Green SARC: Predictive Cost and Carbon Governance for Agentic AI Systems. The paper presents Green SARC: a governance framework that limits the financial and ecological costs of AI agents not retrospectively on a dashboard, but through a calibrated prediction gate before each action. A distribution-free confidence procedure guarantees an upper limit for the probability of exceeding the budget.Thu, 09 Jul 2026 20:21:12 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/arxiv-2606-15956https://beckmann.ai/en/research/2026-07/arxiv-2606-15956arXiv:2606.15956 – You Don't Need Strong Assumptions: Visual Representation Learning via Temporal Differences. The paper presents Temporal Difference in Vision (TDV): an image and a motion encoder are trained together so that the representation of the current video frame plus the encoded motion predicts the representation of the next frame – entirely without cropping, masking, or augmentation. TDV achieves a quality comparable to established methods like DINO or iBOT in tasks requiring fine spatial understanding.Thu, 09 Jul 2026 19:43:41 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/arxiv-2606-15963https://beckmann.ai/en/research/2026-07/arxiv-2606-15963arXiv:2606.15963 – PreLort: Prefix-Nested LoRA for Federated Fine-Tuning under Rank Heterogeneity. PreLort is a new federated learning method that trains and aggregates LoRA adapters of different sizes over a prefix hierarchy, allowing low-power devices to benefit more from more powerful devices – with higher accuracy than previous methods.Thu, 09 Jul 2026 17:52:26 GMTResearchBrian Beckmannhttps://beckmann.ai/en/research/2026-07/arxiv-2606-15959https://beckmann.ai/en/research/2026-07/arxiv-2606-15959arXiv:2606.15959 – Quantifying the Impact of Lossy Compression on Neural Generative Surrogate Modeling investigates how much training data for neural surrogate models in science can be lossy compressed. The authors demonstrate storage savings of 23.7 to 39 times with hardly measurable loss of quality in two simulation applications.Thu, 09 Jul 2026 00:00:00 GMTResearchBrian Beckmann