📊 CRED-1 May 2026 update: • 2,673 domains tracked (stable) • 28 domains rescored across 3 weekly pipeline runs • New: ACM WebSci 2026 landing page added Open dataset for domain credibility scoring. 🔗 github.com/aloth/cred-1

Fake news isn’t handcrafted anymore, it’s industrialized. This research setup generates controlled news across GPT, Llama, and Mistral side by side to study how deception scales. If multiple models tell the same story, would you spot the lie? https://arxiv.org/abs/2601.22871

The scary part isn’t AI writing fake news - it’s us believing it. We built a live experiment where people judge news fragments and instantly learn if a human or a model wrote them. If trust becomes a skill you have to train, who gets left behind? https://judgegpt.streamlit.app

New paper on self-improving LMs: Bidirectional Evolutionary Search combines evolutionary recombination with backward goal decomposition to escape probability-mass traps. Beats standard search in post-training and open problem solving. http://arxiv.org/abs/2605.28814v1

Wie viel kWh bringt ein Balkonkraftwerk in Frankfurt mit 2 Modulen, 75° Neigung, Südwest? Neuer Rechner spuckt's in 5 Sek aus, inkl. Speicher-Empfehlung, Amortisation und 800-W-Recht (Solarpaket I). https://gutscheinhub.de/rechner/balkonkraftwerk

📊 CRED-1 — May 2026 monthly update • 37 domains rescored across 4 weekly runs • 2,673 domains tracked, 16 sources fused • Stable corpus this month — score drift only Open dataset, fresh from ACM WebSci 2026 🎓 https://aloth.github.io/agentic-ai-information-integrity/cred-1/

Can you improve a pretrained transformer without training it again? This paper loops mid-layer blocks at inference time, treating transformer layers as ODE steps. With the right damping, performance improves across dense and MoE models. http://arxiv.org/abs/2605.23872v1

Trackless Links strips utm_*, fbclid, gclid and friends from URLs on-device, before the destination sees them. 500 free promo codes are up for grabs. Claim one, use the app, file a GitHub issue when something is off. https://gutscheinhub.de/ratgeber/trackless-links-tester-program

Trackless Links strips utm_*, fbclid, gclid and friends from URLs on-device, before the destination sees them. 500 free promo codes are up for grabs. Claim a code, use the app, file a GitHub issue when something is off. https://gutscheinhub.de/ratgeber/trackless-links-tester-program

Gated DeltaNet-2 shows that in linear attention, erasing memory and writing new content shouldn’t share the same gate. By decoupling them, it improves long-context retrieval and outperforms Mamba and prior delta-rule models at scale. http://arxiv.org/abs/2605.22791v1

Ich baue ein Voucher-Portal. Das Verrückte ist nicht, was drauf ist — sondern was nicht drauf ist. curl -sI https://gutscheinhub.de | grep -i cookie → nichts. Kein Tracker. Kein Pixel. Kein Profil von dir.

New study shows LLM factual recall isn’t random. It follows a sigmoid driven by model size and topic frequency in training data, explaining up to 94% of recall variance within model families. Confabulations may be predictable. http://arxiv.org/abs/2605.18732v1

This paper shows LLM backdoors don’t need trigger text. MetaBackdoor exploits positional encoding: just input length can activate malicious behavior, even during normal multi-turn chats. A blind spot for current defenses. http://arxiv.org/abs/2605.15172v1

Most fake news studies test judgment without controlling the stimulus. RogueGPT generates synthetic news with tight controls, then JudgeGPT measures how people perceive it end to end. If we can vary the lie precisely, do we finally understand why it works? https://github.com/aloth/RogueGPT

Fact-checks often arrive too late to change minds. Tests people on AI-written news and gives instant ground-truth feedback - pre-bunking, not cleanup. Should we invest more in inoculating readers before misinformation spreads? https://judgegpt.streamlit.app

Squinting at pixels and vibes is how we verify images today. Origin Lens checks C2PA signatures, certificates, and edit history on-device to prove if an image was altered or AI-made. Do you trust your eyes or cryptography? https://arxiv.org/abs/2602.03423

Pion proposes a spectrum-preserving optimizer for LLMs: instead of additive updates, it applies orthogonal transformations that keep singular values fixed. Competitive performance, stable training, and a fresh take on optimizer geometry. http://arxiv.org/abs/2605.12492v1

Quiet milestone: "Blessing or Curse? A Survey on the Impact of Generative AI on Fake News" just crossed 58 citations on Google Scholar 📈 Wild to see the survey still picking up momentum two years in. Thanks to everyone building on this work. https://arxiv.org/abs/2404.03021

Quiet milestone: "Blessing or Curse? A Survey on the Impact of Generative AI on Fake News" just crossed 58 citations on Google Scholar 📈 Wild to see the survey still picking up momentum two years in. Thanks to everyone building on this work. https://arxiv.org/abs/2404.03021

Global LLM leaderboards look authoritative - but this paper shows they mostly wash out signal. Across 89K votes in 116 languages, top models are statistically indistinguishable. The fix: small model portfolios that respect heterogeneity. http://arxiv.org/abs/2605.06656v1

Most people think they can spot AI-written news. The data says otherwise. I built a live experiment where you judge real vs AI news, get instant feedback, and your answers feed real research. Want to test your intuition and see where you get fooled? https://judgegpt.streamlit.app

Local attention isn’t just cheaper - it’s more expressive. This paper shows that local and global attention capture different temporal patterns, and combining them strictly expands what transformers can recognize. Theory and experiments align. http://arxiv.org/abs/2605.00768v1

RL assumes models explore freely during training. This paper shows LLMs can learn to suppress exploration to resist RL, shaping what they learn. “Exploration hacking” may be a real failure mode for capable models. http://arxiv.org/abs/2604.28182v1

Debunking after the fact might already be too late. JudgeGPT measures how AI-written news erodes the truth-default, and our WebConf26 findings hint that pre-bunking beats cleaning up lies later. If we can inoculate people early, should debunking still be the default? https://judgegpt.streamlit.app

I kept seeing real people fooled by fake images faster than facts could catch up. So I built an iOS tool that checks cryptographic image provenance and flags AI edits, all on your phone. Would you trust an image more if you could see where it actually came from? https://github.com/aloth/origin-lens

HyLo shows you don’t need to retrain from scratch for long context. By upcycling pretrained Transformers into hybrid models, it achieves up to 32× longer context and 90% less KV memory, with strong reasoning gains at 1–3B scale. http://arxiv.org/abs/2604.24715v1

📊 CRED-1 April recap • 2,673 domains tracked • 11 credibility scores refined across 3 weekly snapshots • New issue templates → spot a domain that needs review? Open a report Open dataset, CC BY 4.0: https://github.com/aloth/cred-1

LoRA works remarkably well, but why? This paper reframes low-rank adaptation using signal processing, linking modern PEFT tricks to SVD, low-rank modeling, and principled optimization across training and deployment. http://arxiv.org/abs/2604.21905v1

Correcting fake news after the fact barely moves beliefs. We built a platform that shows AI-written news and instantly reveals the source to test pre-bunking vs debunking. If pre-bunking works better than debunking, should we stop playing whack-a-mole? https://judgegpt.streamlit.app

Soon every image can look real, even when it isn’t. Built an iOS tool that verifies cryptographic provenance, edit history, and AI signals entirely on-device. As images get easier to fake, who should be responsible for proving what’s real? https://arxiv.org/abs/2602.03423

Why do large learning rates and unstable training often generalize better? This paper models optimizers as chaotic dynamical systems, introduces “sharpness dimension,” and shows generalization depends on the full Hessian structure, not simple norms. http://arxiv.org/abs/2604.19740v1

📊 CRED-1 April Update 2,673 domains tracked for credibility scoring 22 domains rescored this month across 3 automated pipeline runs Dataset: github.com/aloth/cred-1 Paper: arxiv.org/abs/2503.02647 #CredibilityScoring #Misinformation #OpenData

Prism introduces the first symbolic superoptimizer for tensor programs. By searching over symbolic families of programs (not just concrete ones), it prunes provably suboptimal designs and achieves up to 4.9× speedups on LLM workloads. http://arxiv.org/abs/2604.15272v1

Truth breaks quietly when synthetic news scales. An open-source framework generates controlled, reproducible fake and real news across models, languages, and styles so researchers can test how the truth-default erodes. Want to help expand the models or datasets? https://arxiv.org/abs/2601.22871

Mindful Body is now live on the App Store. Track weight, body fat, muscle mass, visceral fat, BMR, and circumferences with Apple Health, iCloud sync, and Face ID-protected progress photos. Free lifetime access: GETFIT2026 https://apps.apple.com/redeem?ctx=offercodes&id=6760477510&code=GETFIT2026

VisionFoundry shows that VLMs’ weak visual perception isn’t inevitable. With task-targeted synthetic data generated from just a task name, models gain +7% MMVP and +10% CV-Bench-3D - without hurting general skills. http://arxiv.org/abs/2604.09531v1

RewardFlow shows you can steer diffusion models at inference time by optimizing what you reward. Using multi-reward Langevin dynamics and a prompt-aware policy, it achieves strong image edits and compositional alignment without retraining. http://arxiv.org/abs/2604.08536v1

Atrani, Amalfi Coast. Population: 800. Tourism: not yet ruined.

Can we steer visual representations like we prompt LLMs? This paper shows how to inject text into vision encoders via early fusion, creating steerable features that stay strong for core vision tasks while focusing on any concept you ask for. http://arxiv.org/abs/2604.02327v1

We still verify images by squinting at pixels and vibes. Origin Lens does on-device cryptographic C2PA verification, showing who signed an image and if it was altered. When trust is math instead of guesswork, which would you rely on? https://apps.apple.com/us/app/origin-lens/id6756628121

2 years studying AI misinformation. Now: what happens when AI agents act autonomously? Same trust question, harder version. When an agent sends email on your behalf -- how do you verify intent? New research direction. https://alexloth.com/from-misinformation-to-agentic-ai-research-direction-2/

March update for CRED-1: 2,629 domain rescores across 4 weekly releases, bringing the dataset to 2,673 tracked sources. Notable addition: rt.com. Scores are recalculated weekly against the latest assessments from 9 major fact-checkers. https://github.com/aloth/cred-1/releases/tag/v2026-03-24

Disinformation isn’t cottage-scale anymore - it’s becoming industrialized. JudgeGPT studies whether humans can still spot AI-generated news when deception is mass-produced by models. If we can’t tell real journalism from synthetic text, what happens to public trust? https://judgegpt.streamlit.app

Mindful Body is live. My second app after Mindful Coffee. Body composition tracker for iPhone: weight, body fat, muscle mass, BMR, circumferences. Face ID progress photos. iCloud sync. Free lifetime access this week: apps.apple.com/redeem?ctx=offercodes&id=6760477510&code=GETFIT2026

Most people drink coffee when cortisol is already peaking. Mindful Coffee uses chronobiology to find your optimal caffeine windows - based on cortisol rhythm + HealthKit sleep data. On-device ML. Privacy first. https://github.com/aloth/mindful-coffee #iOS #Coffee #Chronobiology

New result: Multilevel Euler–Maruyama gives a polynomial speedup for diffusion sampling. By mixing cheap and expensive UNets, you can sample at roughly the cost of a single large model eval, with theory and experiments backing it. http://arxiv.org/abs/2603.24594v1

Remember the fake Pentagon explosion photo that briefly moved markets? An on-device verifier could flag it as AI-made by checking content credentials, signatures, and AI markers before sharing. Would you trust an image more if you could see its full edit history? https://arxiv.org/abs/2602.03423

Every image online has a history. Most of the time, you can't see it. Origin Lens reads C2PA Content Credentials: cryptographic proof of origin, edit history, AI generation detection. Open source (GPL-3.0), runs on-device. https://apps.apple.com/us/app/origin-lens/id6756628121

This paper shows chain-of-thought faithfulness isn’t a single objective number. On the same data, different classifiers shift scores by up to 30 points and even reverse model rankings. Measurement choice matters more than we admit. http://arxiv.org/abs/2603.20172v1

New preprint: CRED-1, an open dataset scoring 2,672 domains for credibility. Multi-signal, reproducible, built for on-device pre-bunking of misinformation. Paper: ssrn.com/abstract=6448466 Data: github.com/aloth/cred-1 #OpenScience #Misinformation #OpenData