AI Publishing

arXiv draws a line: AI can help write, but fake citations are now the author’s problem

arXiv’s new penalty targets papers where hallucinated references, AI meta-comments, or similar traces show that authors did not verify the text before submission.

AI writing is becoming a test of who actually checked the work

arXiv is not banning AI tools, but it is making the author’s name mean something again when a paper shows obvious signs of unchecked model output.

The clue to these blue cosmic flashes may be where they happen, not how bright they are

A new analysis of LFBOT host galaxies supports a compact-object collision with a Wolf-Rayet star.

A robot fingertip can see before it touches. Now it has to survive the messy work

Researchers at the National University of Singapore and RoboScience have built FingerEye, a compact sensor that keeps visual and tactile signals together from approach to contact.

The Crab Pulsar’s stripes show why plasma and gravity can’t be read apart

A robot hand that knows when not to squeeze harder

FORTE, a University of Texas at Austin robotic hand, reached 91.9% single-trial grasping success on 31 objects by using compliant fingers that measure force and slip.

Invisible dark-matter kicks may explain why dwarf galaxies end up looking alike

Peñarrubia and Nadler propose that dwarf spheroidal galaxies evolve toward an attractor linking stellar radius and velocity dispersion.

Speech AI can keep the sounds and still lose the meaning in tone languages

Paper arXiv:2604.07467 shows that discrete speech units encode lexical tone less reliably than segmental speech structure.

Byte-Level Distillation Cuts Through LLM Tokenizer Mess

A new method ditches the messy heuristics of cross-tokenizer distillation by working at the byte level, offering a shockingly simple fix for a stubborn LLM training problem.

LLMs ace benchmarks yet still fail at common sense

A new study proves LLMs can memorize test answers without understanding the questions—and the gap is measurable.

LLMs Finally Admit They’re Making Things Up

A new arXiv paper treats LLM hallucinations as a classification error—and builds a gate to block them before they escape.

Old support tickets could become the map for the next network outage

An arXiv paper compares fine-tuning, RAG, and a hybrid LLM approach for building an RCA knowledge base from support tickets.

AI’s Blind Refusal Problem: When Safety Becomes Stupidity

A new arXiv study reveals language models refuse to help users bypass rules—even unjust ones—95% of the time.

New Star Class Found

Researchers from the Institute of Science and Technology Austria have made a significant discovery, identifying a new class of stars known as Merger Remnants.

Talking robot guide dogs: AI’s next accessibility stunt?

AI can know the answer forward and still fail the same fact in reverse

A new arXiv study on the reversal curse shows that bidirectional training can help models connect facts in both directions.

IC3-Evolve: AI tunes hardware checks—but who’s buying?

A new arXiv paper automates the finicky tuning of IC3, the algorithm that keeps hardware from melting down—but trust may be harder to verify than code.

LLM failure rates: A new math trick or just better packaging?

AI models are getting too expensive to run; SoLA looks for a softer way to shrink them

SoLA is interesting because it does not promise another smaller model trained from scratch, but tries to compress an existing LLM without extra training or special hardware.

LLMs Learn to Code

Researchers have made a significant breakthrough in teaching Large Language Models to generate consistently correct code, with a new paper on arXiv detailing the approach.

AI models now have to show their work, not just land on the answer

XpertBench introduces rubric-based evaluation for professional domains, which matters more than another general-knowledge leaderboard.

AI’s next test is not more knowledge, but a rule it can prove

The new arXiv work on ARC tasks is worth watching because it does not try to win by scaling, but by combining neural proposals with symbolic verification.

AI that learns from three examples first has to learn what to ignore

SIEVE uses SIEVE-GEN to create synthetic queries from decomposed context and then distills them into model weights.

AI chatbots are entering mental health. This study tests the safety net

Automated evaluation can scale safety checks, but it must not pretend to be diagnosis.

Holos tries to turn AI agents from one-off tools into persistent web infrastructure

A new arXiv preprint introduces the first large-scale multi-agent system built explicitly for the Agentic Web, where heterogeneous agents autonomously interact and co-evolve.

M2-Verify: A benchmark that exposes AI’s multimodal blind spots

Top AI models’ accuracy plunges from 85.8% to 61.6% when tested on M2-Verify’s high-complexity scientific claims—a gap that exposes multimodal reasoning as brittle.

Sven’s pseudoinverse trick: A natural gradient with less hype

Sven’s authors claim their pseudoinverse-based optimizer cuts natural gradient costs to *k*× stochastic overhead—without defining *k* for real-world models.

NASA’s next Earth-hunting telescope has to weigh planets, not just photograph them

The Habitable Worlds Observatory may image an Earth-like planet, but without a precise mass measurement that discovery remains scientifically unfinished.

AI Smells the Difference—But Can It Tell Chanel from Cheetos?

Researchers tested 21 language models on 1,010 smell-related questions—and found even top performers floundering like overcaffeinated truffle pigs.

A third dark matter-free galaxy strengthens violent collision theory

NGC 1052-DF9’s stars move at speeds implying virtually no dark matter—yet the galaxy remains intact, defying a core tenet of astrophysics.

CAMP: AI’s First Case-Adaptive Clinical Panel

ArXiv 2604.00085v1 replaces flat majority voting with a dynamically assembled specialist panel that scores 12 points higher on disputed cases.

E-STEER: Emotion as a Knob for LLMs—Not Just Another Paper

A new arXiv study introduces E-STEER, the first framework to embed emotion as a steerable variable in LLM hidden states—not just a surface-level style.

Google’s Willow quantum processor: Hype or hardware leap?

Google’s Willow quantum processor is now a gated playground for researchers—with a May 15 deadline to prove they’re worthy of entry.

Neuro-symbolic AI tries to fix process monitoring’s blind spots

Logic Tensor Networks just became the rare AI method that cares more about your hospital’s protocols than its own accuracy metrics.

EEG emotion recognition’s cross-dataset problem just got a patch

Cross-dataset EEG emotion recognition just got a prototype-driven upgrade—on paper, at least, with PAA-L’s local alignment outpacing global adversarial methods in early arXiv tests.

KGWAS Upgrade

The KGWAS framework has been upgraded to incorporate contextual information, aiming to improve detection power and provide mechanistic insights.

Multilingual speech translation’s hidden architecture war

A new arXiv study exposes how uniform architectural sharing in multilingual speech models creates representation conflicts that stall low-resource language performance by up to 40%.

Knowledge graphs get real—or just another AI hype cycle?

The arXiv paper’s authors admit what KG vendors won’t: 90% of the world’s textual data is still *unstructured noise*—and no one’s cracked the cost-efficient way to turn it into actionable graphs.

AI Depression Detectors Cheat by Reading the Interviewer

A new study reveals AI depression detectors ace benchmarks by cheating—memorizing interviewer scripts instead of patient symptoms.

Care home AI speakers: Safety first, hype second

Supervised trials in care homes—where 184 reminder-containing interactions became potential failure points—reveal the gap between AI’s demo fluency and its real-world reliability.

AI’s New Report Card: Grading Models on How They Cheat

A dismantles accuracy as a meaningful AI benchmark by scoring models on *how* they fail—not just whether they do.

AI Medical Benchmarks Just Got Smarter—But Who’s Counting?

A new study claims CAT frameworks can evaluate 38 LLMs for a tenth of the cost of static benchmarks—if the medical item bank holds up.

JWST’s redshift record rewrites early-universe timelines

LLMs’ Confidence Problem Gets a Reality Check

AI self-improvement hits a human-data ceiling

A new paper argues AI self-improvement will stall when human-written data runs out.

A tennis robot points to the harder question: how messy can robot training data be?

LATENT achieved a 96.5% success rate on a Unitree G1 returning tennis balls within 2.5 meters of the target.

Provably accurate or just provably overpromised?

A new continual-learning paper claims to eliminate forgetting with fixed embeddings—but the demo ends where real-world challenges begin.

Physics-inspired kernels are elegant - but are they useful?

Neural Matter Networks replace standard blocks with a single geometrically grounded kernel.

Data Gold

Researchers have long been puzzled by the paradox of tabular machine learning, where high-dimensional, collinear, and error-prone data yield state-of-the-art performance.

RLHF’s blind spot: can P-GRPO fix the preference echo chamber?

P-GRPO tries to keep personalized gradients intact instead of flattening feedback into one global average.

Reasoning-Based LLM Unlearning Targets Model Safety Gaps

New reasoning-based LLM unlearning method cuts model bias 40% by surgically removing unsafe knowledge—without full retraining.

Meta’s NLLB-200 isn’t just translating—it’s mapping how languages think

A new arXiv study shows NLLB-200 partly tracks language phylogeny, suggesting deeper linguistic patterns.

SkillNet: AI’s Skill Library Finally Grows Up

SkillNet’s arXiv debut marks the first serious attempt to turn AI’s ‘reinventing the wheel’ problem into a scalable infrastructure.

AI may not need to read every long document like a full universe

Cheaper in AI often means dumber. This proposal is interesting because it tries to be cheaper more intelligently.

The AI That Knows You Better May Push Back Less

Nine frontier LLMs show that tailoring responses to user traits increases emotional agreement but weakens factual pushback in peer-like interactions.

Banks want AI agents to find dirty money, but regulators will want the receipts