All Updates
Page 1832 of 1904
February 20, 2026
Simple Baselines Rival Code Evolution
A new arXiv paper shows simple baselines match or outperform complex code evolution techniques using LLMs across math bounds, agentic scaffolds, and ML competitions. It identifies key issues like poor search space design and high evaluation variance. The study proposes better practices to improve code evolution rigor.
Raspberry Pi 5 Officially Embraces OpenClaw
Raspberry Pi Foundation published a guide for running OpenClaw on Pi 5, featuring a wedding photo booth demo. Adafruit covered it, and Medium bloggers are documenting setups. This enables the cheapest $80 always-on AI agent via single-board computer.
Order-Oriented Scoring for Hesitant Fuzzy Sets
This paper introduces a unified order-oriented framework for scoring hesitant fuzzy sets, addressing limitations in traditional methods. It analyzes classical orders, proving they lack lattice structures, while symmetric orders meet key normative criteria like strong monotonicity. New dominance functions enable ranking relative to acceptability thresholds for decision-making.
Node Learning: Decentralized Edge AI Framework
Node Learning is a decentralized paradigm where edge nodes learn continuously from local data and exchange knowledge opportunistically via peer interactions. It propagates learning through overlap and diffusion, avoiding global synchronization or central aggregation. The paper outlines its foundations, contrasts with existing methods, and explores implications for communication, hardware, trust, and governance.
NeuDiff Agent Speeds Neutron Crystallography 5x
NeuDiff Agent is a governed AI workflow for TOPAZ at Spallation Neutron Source, automating single-crystal neutron crystallography from data reduction to validated CIF output. It restricts actions to allowlisted tools, enforces fail-closed verification gates, and captures full provenance. Benchmarks show 4.6-5.0x wall-time speedup (86.5-94.4 min vs. 435 min manual) with no checkCIF alerts.
Narrow Fine-Tuning Erodes VLM Safety
Narrow fine-tuning on harmful datasets erodes safety alignment in vision-language models, causing misalignment that generalizes across unrelated tasks and modalities. Experiments on Gemma3-4B reveal misalignment scales with LoRA rank and is worse in multimodal evaluation (70.71%) than text-only (41.19%). Even 10% harmful data triggers substantial degradation, with harms captured in a low-dimensional subspace.
MobCache Scales LLM Mobility Sims
MobCache is a mobility-aware cache framework that enables scalable LLM-based human mobility simulations by reusing reconstructible reasoning caches. It encodes reasoning steps as latent embeddings for recombination and uses a lightweight decoder trained via mobility-constrained distillation. Experiments demonstrate major efficiency gains while matching state-of-the-art performance.
LLMs & GraphRAG Automate CPS DSMs
Researchers leverage LLMs, RAG, and GraphRAG to generate Design Structure Matrices (DSMs) for cyber-physical systems. Methods tested on power screwdriver and CubeSat use cases, assessing component relationships and identification. Despite challenges, shows promise for automation with public code available.
LLM-WikiRace Reveals LLM Planning Limits
LLM-WikiRace is a new benchmark evaluating LLMs on long-term planning and reasoning by navigating Wikipedia hyperlinks from source to target pages. Frontier models like Gemini-3, GPT-5, and Claude Opus 4.5 excel on easy levels but drop to 23% success on hard ones. Analysis shows planning trumps knowledge beyond a threshold, with top models failing to replan after errors.
IndicJR: Judge-Free Indic Jailbreak Benchmark
IndicJR introduces a judge-free benchmark evaluating jailbreak robustness in 12 South Asian languages with 45,216 prompts across JSON and Free tracks. It uncovers that contracts boost refusals but fail against jailbreaks, English attacks transfer effectively to Indic, and orthography like romanization weakens defenses. The benchmark provides a reproducible multilingual stress test for LLM safety.
GUI-Owl-1.5 Tops 20+ GUI Benchmarks
GUI-Owl-1.5 introduces multi-size native GUI agent models (2B-235B) supporting desktop, mobile, browser platforms for cloud-edge collaboration. It sets SOTA on 20+ benchmarks like 56.5 on OSWorld, 71.6 on AndroidWorld, and 80.3 on ScreenSpotPro. Open-sourced with innovations in data flywheel, agent reasoning, and multi-platform RL.
GAP: Text Safety Fails for LLM Agent Tools
Researchers introduce the GAP benchmark to evaluate divergence between text-level and tool-call safety in LLM agents. Testing six frontier models across six domains reveals text refusals do not prevent harmful tool calls, with 219 persistent cases even under safety prompts. The study urges dedicated tool-call safety measures beyond text evaluations.
Contextuality Inevitable in Single-State AI
Adaptive systems reuse fixed internal states across contexts due to resource limits, leading to inevitable contextuality in classical probabilistic models. The paper proves an irreducible information-theoretic cost for reproducing contextual statistics. Nonclassical frameworks avoid this without quantum mechanics by lacking a global joint probability space.
AIdentifyAGE Ontology Standardizes Forensic Dental AI
AIdentifyAGE ontology provides a standardized framework for forensic dental age assessment, supporting manual and AI-assisted workflows. It integrates clinical, forensic, legal data, radiographic imaging, and ML methods for interoperability and transparency. Developed with experts, it builds on biomedical ontologies and adheres to FAIR principles.
AI Improves 50-Year Hypercube Slicing Bounds
Researchers prove S(n) ≤ ⌈4n/5⌉ for hypercube edge slicing, beating 1971's ⌈5n/6⌉ bound. They used CPro1, an LLM-powered tool, to construct 8 hyperplanes slicing Q_{10}. New lower bounds on edges sliced by k<n hyperplanes are also established.
AI Benchmarks Saturate Quickly Study
A systematic ArXiv study analyzes saturation across 60 LLM benchmarks from major developers. Nearly half show saturation, worsening with age, and hiding test data offers no protection. Expert-curated benchmarks resist saturation better than crowdsourced ones.
AgentLAB Benchmarks LLM Agents on Long-Horizon Attacks
AgentLAB is the first benchmark evaluating LLM agents' vulnerability to adaptive long-horizon attacks via multi-turn interactions. It features five attack types—intent hijacking, tool chaining, task injection, objective drifting, memory poisoning—across 28 environments and 644 test cases. Evaluations reveal high susceptibility in agents, with single-turn defenses failing to mitigate threats.
Product Sense Beats Coding in Vibe Coding Era
In the Vibe Coding era powered by tools like Claude Code, product sense outweighs traditional coding skills as non-programmers build full AI agents via conversation. Demos externalize ideas, build trust, and lower barriers from concept to product. Six core techniques include basing on existing GitHub projects, problem-driven AI queries, and modular progressive development.
Altman: Superint to Top CEOs by 2028
OpenAI CEO Sam Altman predicts an early version of true superintelligence in just a few years. By end-2028, more global intelligence resources will be in data centers than outside. Superintelligence will outperform top company CEOs—including himself—and leading scientists.
Memory Giants Ramp Factories for AI Demand
Micron, Samsung, and SK Hynix are massively expanding fabs to meet AI-driven memory needs. Micron's $200B plan features a huge Boise campus with 15-20万 WPM capacity. Priority for HBM and AI modules means ongoing consumer shortages.