2026
10,781 articles · 136 days · 10 categories · avg 79 per day
TADR Diffusion Module Lifts Anomaly Detector F1 on SMD When Data Blocks Are Missing
The TADR diffusion module reconstructs missing sensor data blocks without erasing anomalies. Detector F1 on the SMD dataset rises, and everything runs on a single…
LLM Agent Built a Gearbox in KOMPAS-3D, but Scored Only 50 out of 100 on the Specification
The kompas-mcp plugin lets an LLM agent drive KOMPAS-3D: it built a gearbox and its design documentation. In a homemade benchmark, 3D tasks score 88–100; the…
HyperRepair Restores 93.4% of General Skills to Fine-Tuned LLMs While Retaining 99.5% of Domain Score
Researchers in Shanghai and Queensland can repair general skills that LoRA models forget, without target-domain data: a hypernetwork generates a side repair network…
Developer of Viy-Based AR Horror Game Builds 12 Monsters in Tripo AI and Cuts the Devil to 7 880 Triangles
Claude Opus 5 built simple baselines that closed 95.5% of the gap to DL models from Nature and Science
Steering vectors in LLM fine-tuning improved SFT and distillation across three seeds
FieldGuideRAG checks photos against reference images in a graph and beats SigLIP by up to 16 accuracy points
CoFiBreak Reward Teaches an Attacker LLM Diverse Jailbreaks Without Losing ASR@10
The CoFiBreak reward for GRPO keeps an attacker model from collapsing into identical prompts and preserves ASR@10, the share of intents hit within 10 attempts…
Gemma-3-12b is worse at spotting lies in RAG: detection rate drops from 25% to 10%
When a poisoned claim is repeated in six of seven documents, Gemma-3-12b spots the substitution in 10% of cases, versus 25% when the claim appears only once.
Qwen3-VL-4B Gains 18.33 Percentage Points on the Drone-PUR Benchmark After Fine-Tuning on Drone Imagery
Eight LLMs Lose 35–46% of Strategy Calibration in Free-Form Generation
Disabling 729 Attention Heads in GPT-2-XL Cut Verbatim Memorization by 91.5%
BILAL-Bench: Six of Seven Open Models Confidently Fabricate Sources in 79–95% of Cases
DLC Correction Keeps KL to On-Policy Training Within Two Noise Floors on Qwen3
The authors split sampler-trainer mismatch in LLM RL into two causes. The DLC correction keeps KL to on-policy training within two noise floors on Qwen3-0.6B and…
PathoJev-27B Beats SlideChat by 6.3 Percentage Points in Average Accuracy on Three WSI Benchmarks
PathoJev encodes a histology slide once and outputs answer probabilities directly. On three WSI benchmarks it beats SlideChat in average accuracy by 5.4 pp (9B) and…
Cosmos generated training video for drones: detector mAP@50–95 up 10.3%
Ohio State researchers grew synthetic clips from nine real frames and lifted a drone detector's mAP@50–95 by 10.3%. An expert rejected 87.5% of clips, yet it still…
Faros: median code review time up 441% in teams using AI agents
The "no gold documents" label lost accuracy, from 0.68 to 0.10, on strong retrievers
TJ Digital: 23% of AI-cited pages open with a direct answer, versus 14% of uncited ones
Karpathy builds a YouTube clone with 6454 videos from Hacker News recommendations in two hours
Karpathy used Claude to build HackerTube from 6,454 videos recommended on Hacker News
Andrej Karpathy had Claude build a YouTube alternative from videos recommended on Hacker News since 2007. The result is HackerTube: 6454 videos across 150 topics.
Arca gives AI agents finance tasks lasting up to 10 days, but leaves the trades to humans
Arca, a wealth manager, runs multi-day tasks through agents with their own memory and harness, while advisers keep control of trading and money transfers.
DeepSeek open-sources Super-agent prototype: p95 task completion time of 15.7 minutes vs. 29.5
Bun creator uses Claude Code to port 750K lines from Zig to Rust in 11 days
Microsoft releases 4B model Mage-VL-Cowork, which solves 66.0% of OSWorld-Verified tasks
Meta's Muse assistant keeps dossiers on loved ones and gets root on Ubuntu by default
8 Skills for video in Claude Code: from 244 pixel effects to rough-cut editing
Author @DDJCXX rounded up 8 Skills for Claude Code: 244 pixel effects, 30-second promos, narrated stories and a rough cut in DaVinci. The top repo has about 2700…
Adaptive noise in an SDE sampler improved texture by only 1–2% on Flux 3 Klein
A Habr author tested whether a sampler can choose noise strength from curvature and solver error. On Flux 3 Klein the texture gain was 1–2%, and people did not…
Local models lose to Opus and Codex: 0.21 on a MacBook versus 1.00 for Codex
A Habr author ran one trap-laden task on the cloud, a MacBook, an RTX 5080 laptop and a DGX Spark. Codex scored 1.00; the best MacBook model scored 0.21.
Habr author fine-tunes YOLOv8l on dental X-rays and pushes mAP50 to 0.642
16 models tested on 7 search tools: Exa averaged 10.5 out of 14, SearXNG about 3
Kimi scores 96.2% on a 1C database analytics test and emerges as a fallback if Claude access is blocked
