explainx / blog / topics
AI Research
New papers land every day. A few change how models are trained or what they can do: reinforcement learning methods, distillation, reasoning, and AI applied to math and science.
This page collects the research we explained, with links to the papers and code.
118 stories · latest Oct 28, 2026
Start here
Can LLMs Judge Research Novelty? Ai2 Finds Prompt Wording Flips the Verdict
Ai2 built a benchmark from ICLR 2026 peer reviews to test whether LLMs can reliably judge if a research idea is new. The headline: tiny prompt changes can swing accuracy from 93.5 percent to 40.9 percent, more search and reasoning did not reliably help, and purpose-built tools scored lower and cost more.
Why Is AI-Generated Language So Hard to Understand? What the New Paper Found
Many readers feel AI text has become harder to follow. A paper by Thomas Hikaru Clark, Matias Martin and Kyle Mahowald measures it: nine frontier models use more difficult-to-process constructions than humans, and the authors blame optimization for information density.
SageMaker Multi-Turn RL Walkthrough: gpt-oss-20b Goes From 61% to 84% on Wordle for About $339
AWS published a hands-on guide to Multi-Turn Reinforcement Learning on SageMaker AI. Using Wordle, an agent on Bedrock AgentCore and GRPO with LoRA, gpt-oss-20b improved its solve rate for about $339 in training tokens plus roughly $79 of runtime. Here is the pipeline, the cost breakdown and the caveats.
ReMaRk: Train RL on the Request, Not Each Answer, to Cut GRPO Noise
A Nanyang Technological University paper argues that when a request has several valid human answers, reinforcement learning should train on the request rather than each reply. ReMaRk scores every generated list against the union of all replies, cutting reward noise and RL steps.
Rosetta Stone Paper: Align DINOv2 and Qwen3 Without Image-Caption Pairs
Shared Geometry As A Rosetta Stone shows that a vision encoder like DINOv2 and a language model like Qwen3 can be linked by one orthogonal map, with no image-caption pairs at all. Here is the method, the reported results, and what it does and does not mean for builders.
RunningTab: Why File-Reading Agents Drop Facts They Already Saw, and a Fix
A new paper trending on Hugging Face finds that agents working over a folder of files often open the right document, see the right number, and still leave it out of the final report. RunningTab fixes that by keeping a per-task record in the environment, not in the model. Here is how it works and what to copy.
Timeline
October 2026
Oct 28
Cartesia H-Net: Byte-Level Model Matches SmolLM3-3B With 10x Less Compute and Data, ClaimedCartesia Research published a two-part series on H-Nets, language models that replace the tokenizer with learned chunking. The headline is a 10x efficiency claim against SmolLM3-3B; the measured ladder gains are smaller, and the authors say so.
Oct 28
CodeCap-Bench: Can Executable Code Be a Better Video Caption Than Words?A new benchmark asks models to watch a short 2D video and write code that renders it back. The best agent setup keeps about 73 out of 100 of the visual information, and motion is consistently harder than appearance.
Oct 28
Samaya Sterling: A 3B-Active Model That Beats Claude Fable 5.1 on FrontierFinanceSamaya says its Sterling model, a post-trained 30B mixture-of-experts with 3B active parameters, scores 58.9% on its own FrontierFinance benchmark against 55.9% for Claude Fable 5.1, at one eleventh of the cost. Here is the four-stage recipe and what the numbers do and do not show.
Oct 28
TTME: Evolving a Memory Instead of Weights Beats Test-Time Training on Math DiscoveryA UC Santa Barbara team asks whether an agent that keeps lessons in an external memory can match one that retrains its weights at test time. On five math construction tasks it does, at a fraction of the compute. This post covers the method, the numbers and the caveats.
Oct 27
Can LLMs Judge Research Novelty? Ai2 Finds Prompt Wording Flips the VerdictOct 27
Exowatt: Fine-Tuned Qwen3 Beats Claude Opus 4.7 at 49x Lower Cost on Its Own TasksAn AWS blog post, co-written around customer Exowatt, reports that two small fine-tuned Qwen3 models beat Claude Opus 4.7, Sonnet 4.6 and Haiku 4.5 on narrow solar-fleet tasks, at up to 49x lower cost. We lay out the numbers, the method and the limits of the comparison.
Oct 27
NVIDIA and C-BRAIN: Open-Source AI Co-Scientists for Dementia Research, ExplainedNVIDIA's blog describes how the C-BRAIN consortium uses two open-source agentic tools, OpenScientist and InsightEngine, to speed Alzheimer's and dementia research. The headline claim is analyses in minutes for under $10 a run. Here is what the sources support and what they do not.
Oct 27
Physical Superintelligence Claims 5 Quantum Results: Claimed vs. VerifiedOn October 23, 2026, Physical Superintelligence published five results on open problems in quantum information theory, each with a paper, a Lean formalization and verification code. PSI says AI models and agents did much of the work. Here is what the company claims, what you can check yourself, and what no outside expert has confirmed yet.
Oct 26
Why AI Text Is Exhausting to Read: Six Patterns From an NYU and UT Austin StudyResearchers at NYU and UT Austin compared text from 14 language models with human writing and found six constructions frontier models overuse. Their explanation: training on code and math rewards correctness and density, not clarity. Here is what was reported and what you can do about it.
Oct 25
A 10-Year-Old CNN Beat GPT-6 Astra and Claude Fable 5.1 at Spotting Skin CancerDr Magnus Lynch showed 200 biopsy-proven skin lesions to GPT-6 Astra, Claude Fable 5.1 and a fine-tuned Inception v3 network from 2015. The old network won, 62% to 42%. Here is what the test shows, what it does not, and how to read it.
Oct 24
A Z-Library Paper Cited a TorrentFreak Article That Does Not Exist: What Happened and How to Check CitationsTorrentFreak reports that a journal paper on Z-Library, written with disclosed AI help, cites a TorrentFreak article that was never published. We lay out what the outlet says, what is still unconfirmed, what Hacker News readers made of it, and a practical way to verify any reference list.
Oct 24
Why Is AI-Generated Language So Hard to Understand? What the New Paper FoundOct 23
Ai2 LitPivot: An AI Editor That Tests Your Research Idea Against the Literature, Tested on 17 ResearchersAi2 published LitPivot, a research-ideation system built with the University of Pennsylvania, for UIST 2026. It organizes literature around the part of a proposal you are editing, offers critiques with excerpts, and suggests revisions you can accept or reject. Here is how it works, what the study shows, and its limits.
Oct 23
SageMaker Multi-Turn RL Walkthrough: gpt-oss-20b Goes From 61% to 84% on Wordle for About $339Oct 23
Why AI Voting Advice Sent Left-Leaning Japanese Voters to the Communist PartyA working paper from Stanford and Hoover researchers asked five web-search-enabled AI models to recommend a party to 36,300 synthetic Japanese voters. Left-leaning views converged on one party, and the authors blame what the models can read, not what they were trained on.
Oct 22
Amazon SageMaker Training Sessions: Write Your Own RL Loop Without Managing GPUsAmazon SageMaker AI serverless model customization now includes training sessions, generally available October 22, 2026. You write the loop with forward_backward, optim_step and sample calls; AWS handles GPUs, recovery and checkpoints.
Oct 22
RAG Techniques Do Not Add Up: IBM Study Finds +47% Predicted Became +17%A new paper from IBM and the Weizmann Institute tested whether RAG tricks that help alone still help together. For retrieval, mostly not: three techniques predicted +47% and delivered +17%. The team released Lasagna, an Apache 2.0 framework to test your own.
Oct 22
CRUX Shadow Evaluation: Experts Reject AI Research Papers From GPT-6 Astra and Fable 5.1A Princeton-led project called CRUX asked frontier AI agents to tackle the central question of unpublished NeurIPS papers, then had the original authors grade the results. In July, both papers were unambiguously rejected. An October follow-up with GPT-6 Astra and Claude Fable 5.1 reportedly saw the same verdict with some progress.
Oct 22
Q Labs Dust: Pretraining Transformers Without Backprop, and the 4,000x Claim CheckedQ Labs researchers released Dust, a zeroth-order method that trains transformers without backpropagation, plus a Noise speedrun benchmark. A viral post claims pretraining time fell from 1.76 hours to 1.6 seconds. explainx.ai compares that claim with the paper and the benchmark record tables.
Oct 22
Southampton Historians Used AI to Read 100,000 Pages of Handwritten Port BooksThe University of Southampton used AI to transcribe about 100,000 pages of handwritten port books from the National Archives and built a first-of-its-kind interactive map of Britain's merchant shipping from 1565. We explain the method, the findings, such as Europe taking roughly 80 percent of trade in 1696-97, and what the project says about AI for archives.
Oct 22
VocabX: Specializing a Small Model's Vocabulary Cut Inference Cost 77% in ProductionDharma-AI says a cost-weighted vocabulary rebuild of a fine-tuned Qwen3-4B essay grader cut responses from 350 to 101 tokens and measured inference cost by 77%, with a higher quality score. We explain why generated tokens cost more than read ones, how VocabX picks tokens, and what is still unpublished.
Oct 21
Holon-S1: Sophont Predicts Spatial Gene Expression From Routine Pathology SlidesSophont published the Holon-S1 technical report in October 2026, describing a model that infers spatially resolved gene expression and pathway activity from a routine H&E biopsy image. It reports wins over DeepSpot-M, candidate biomarkers in breast and colorectal cancer, and clear limits: research-only, regional rather than single-cell, retrospective.
Oct 21
FAHR-Face: Mass General Brigham AI Predicts Cancer Survival From a Face PhotoMass General Brigham researchers built FAHR-Face, a foundation model trained on more than 40 million face images, and two tools that estimate biological age and predict cancer survival from a single photograph. Here is what the paper reports, what is still unproven, and why bioethicists are worried.
Oct 20
Electric Plant's Bombadil: An AI Model That Reads Plants' Electrical Signals, Claimed vs VerifiedTechCrunch reported on October 20, 2026 that startup Electric Plant has built Bombadil, a model trained on plant electrical signals from about 125 species, and raised an $11.2 million seed. Founder claims include 20% or more water savings and sensors costing tens of dollars. Here is what is established, what is the company's claim, and the questions that would settle it.
Oct 20
Metis Reasoning: AlphaGo Co-Creator Thore Graepel Launches a London Lab, and What Is Actually VerifiedAlphaGo co-creator Thore Graepel has launched Metis Reasoning, a 12-person London lab pitching AI reasoning that can be checked and needs far less compute than frontier models. Headlines say it will end hallucinations. Here is what the founders actually said, and what nobody has shown yet.
Oct 20
Omnii-1: Radical Numerics Reports the First AI-Designed Phage Given to a PatientOn October 20, 2026 Radical Numerics, founded by members of the team behind Evo and Evo 2, released Omnii-1, a multimodal genome language model, with eight application results and a report that AI-designed phages were given to a drug-resistant infection patient. Here is what is claimed, what is measured, and what independent checks are still missing.
Oct 19
OpenMobile-2: An Open Stack for Training Phone Agents, 79.9 on AndroidWorldA Nanjing University and SenseTime-led team released OpenMobile-2, an open stack for training phone-control agents: simulated commercial apps, trajectories, verifiable RL tasks and two agents. We explain what is in the box, the reported scores, and what you can check yourself.
Oct 19
Scene-Designer: An 8B Model That Designs 3D Rooms and Critiques Its Own WorkZhejiang University researchers released Scene-Designer, a Qwen3-VL-8B based agent that builds 3D indoor scenes and judges them. Code, an 8B checkpoint and a critique benchmark are public. explainx.ai walks through the idea, the numbers, and the caveats.
Oct 16
LeanGraphSearch: Using Mathlib's Dependency Graph to Find the Lemmas a Lean 4 Prover NeedsA research team led by authors including Kai-Wei Chang and Nanyun Peng released LeanGraphSearch, a retrieval method that spreads relevance through the dependency links of Mathlib so a prover finds supporting lemmas that do not resemble the target theorem. The authors report up to 12 percent better recall and up to 40 percent better pass rates than a vanilla prover. Here is how it works, what the numbers mean, and where the evidence is thin.
Oct 16
MAI-Gym: Alibaba Tongyi Opens a GUI Agent Gym and a 9B ModelAlibaba's Tongyi-MAI team published MAI-Gym, an open platform that uses one rollout pipeline to evaluate GUI agents and train them with online reinforcement learning. Its 9B model reports 78.4% on MobileWorld and 76.2% on OSWorld-Verified. The weights are out; the framework code is still marked coming soon.
Oct 15
AI21 Post-Trains Nemotron 3.5 Super VL: Beats Opus 5.5 at 33x Lower Cost? Claimed vs VerifiedAI21 post-trained NVIDIA Nemotron 3.5 Super VL for Gong, Deel and Alice in three weeks and reports near-frontier results at a fraction of the cost. Social posts spread an 88.5 percent payroll figure, but AI21 own post says 72 percent. Here is what is claimed and what is verified.
Oct 15
Into the Unknown: 114 of 135 AI-Written Follow-Up Papers Fail an Integrity Check, DOE Lab Scientists FindInto the Unknown asks AI agents to write follow-up papers to real, unpublished research directions supplied by US national lab scientists. Across 135 submissions from nine frontier agents, none satisfied every condition, and 114 claimed work that was never done or numbers that contradict the evidence.
Oct 15
Radical Numerics Halcyon: Text-to-Image Without Language Pretraining, 275x Less ComputeRadical Numerics says text-to-image generation does not need a pretrained language model. Its Halcyon family, trained only on 100M image-text pairs, beats two baselines with up to 275x less pretraining compute. Here is what was measured and what the headline hides.
Oct 15
ReMaRk: Train RL on the Request, Not Each Answer, to Cut GRPO NoiseOct 15
TRL Tree Packing for AsyncGRPO: The 26x Claim vs. the 3x the Pull Request MeasuresHugging Face TRL merged tree packing for its AsyncGRPO trainer on October 13, 2026. Social posts cite 26x less trainer compute on MiMo Envs, while the pull request itself reports up to 3.1x on a microbenchmark. Here is the mechanism, the numbers, and how to try it.
Oct 14
FineEnvs Multilingual: 12M Open RL Tasks for OCR and Speech in 102 LanguagesFineEnvs, a Hugging Face community project by Adithya S. Kolavi, released two OpenEnv environments with about 12 million tasks: document OCR over 22 languages and speech recognition over 102. Gemma 4 trained on each cut Kannada character errors by 16 percent and 45 percent. Here is what is in it, what the numbers mean, and how to use it.
Oct 14
HiRO-ACEv2: Ai2 Simulates Global Extreme Weather at 3 km for 1/300th of the CostAi2 published HiRO-ACEv2 on October 14, 2026: a two-stage AI emulator that turns a 100 km climate run into 3 km hurricanes, wind and rain for roughly 1/300th of the cost of the physics model it imitates. We explain how it works, the results, what is open, and the limits.
Oct 14
Leash Bio Minos: Measured Chemistry Worlds Where AI Agents Hunt for Drug BindersLeash Bio has released Minos, a framework of premeasured experimental worlds for training and testing AI agents on drug-binder search, with a preprint, an open dataset and thousands of agent reasoning traces.
Oct 14
Next-Concept Modeling: Text Generation Up to 31x Faster Than Diffusion Language ModelsNext-Concept Modeling encodes text as a hierarchy of discrete codes and generates one level at a time, predicting every code in a level in parallel. The authors report 10x to 31x speedups over diffusion language models at matched quality, and released code, weights and datasets.
Oct 14
Ultramarine 1 Flash: A Small Team Claims Fable-Class Work at 60x Lower CostPaper Instruments, a small San Francisco lab, launched Ultramarine 1 Flash on October 14, 2026: an open-model release for finance, law, consulting and operations work, with an API now and weights promised. We lay out its benchmark table and cost claim and separate vendor numbers from independent ones.
Oct 14
VRSD: A Teacher-Only Prompt That Stops Self-Distillation From Killing ReasoningOct 14
WTF Wasserstein-Tilted Flow Maps: Reward Fine-Tuning With Up to 280x Less ComputeOct 13
Thinking Inertia: Why LLMs Keep Reasoning When Told Not ToOct 12
David Silver Explains Ineffable: An RL "Superlearner" as the Alternative to LLMsOct 11
AI Assistance Reduces Persistence: What the 10-Minute Study FoundOct 11
arXiv Caps Authors at Two Submissions a Month as AI Papers Flood the ServerOct 11
Brain-IT: Weizmann AI Reconstructs What You See From One Hour of fMRIOct 9
Alzheimer's Translation Challenge: Prima Mente Opens an AI Competition on a 150M-Cell AtlasOct 9
TermGrade: 1,004 Open RL Environments for Terminal Agents, ExplainedOct 8
On-Policy Distillation Can Collapse: A New Paper Explains Why (and the Two Fixes)Oct 8
Perplexity pplx-embed-v2-late: Multimodal Embeddings Beyond One VectorOct 8
RunningTab: Why File-Reading Agents Drop Facts They Already Saw, and a FixOct 6
Gemma 4 + BOTANIC-1: AI Ranks a Melon Gene Variant First of 2,494Oct 3
Adaption Labs Invent API: Generate Post-Training Data From a Prompt, With No Seed ExamplesOct 3
Neuralink Pretrains Brain-Computer Interface Encoders on 50,000 Hours of Unlabeled Neural DataOct 1
AGMAI's Rules for Releasing AI Math — A Checklist, Not a Ban
September 2026
Sep 28
How Can a Language Model Solve Math Problems and Fight Disease?Sep 28
Thinking Fast and Slow in AI (2021): Metacognition on Hacker News AgainSep 26
Harvard exoplanet radio signals: what MeerKAT actually found (and why AI feeds scream aliens)Sep 18
Infinite-Parameter LLMs: Generating Weights From Live Data, Not the PromptSep 17
Stop Using LLMs as Classifiers. Use Them as Feature GeneratorsSep 17
Training a 4B Model to Beat Postgres Query Plans by 81% With RLSep 17
Xiaomi MiMo-V2.6: Livestreaming a Trillion-Parameter RL Training RunSep 16
He Co-Invented ChatGPT. Now He Says It Was a "Weird Detour."Sep 16
Google's "AI in Science" Report: 7 Hours Saved, With a Big AsteriskSep 13
Recurrent Looped Transformer: "Infinite" Reasoning Depth, ExplainedSep 12
flybody: The Fruit Fly Body Model DeepMind Built for MuJoCoSep 11
Why Everyone Is Posting Fruit Fly Brain Videos This WeekSep 10
The 7 Millennium Prize Problems: What AI Has Actually SolvedSep 9
A Startup Built a Neural Network From a Fruit Fly Brain — Then Scrambled ItSep 8
Nature Review: AI Is Now Designing Physics Experiments, Not Just Analyzing ThemSep 7
Can AI Talk to Animals? Crow Vocalizations, a $10M Prize, and the Real ScienceSep 7
AlphaFold Plus Brain Organoids: A Reported UCSF Map of Autism-Linked ProteinsSep 7
NEAR AI Reportedly Solved Putnam Bench for $111 — 250x CheaperSep 7
Universal Geometry of Embeddings: Why "Safe" Vector Databases Aren’tSep 6
In-Browser LLM Fine-Tuning: Why Training on WebGPU Is a Bigger Deal Than InferenceSep 6
"LLMs as a Cognitive Virus": A New Paper Models AI Dependence as an EpidemicSep 6
Meta's AIRA₃ Wins Gold in a Live NVIDIA Kaggle CompetitionSep 5
No, Zhi-Wei Sun Did Not Set a Prime-Gap Record With GPT-5.6 SolSep 4
Google and Janelia Mapped a Fruit Fly Brain — With AI Doing the ReconstructionSep 3
"Open-Source RL-as-a-Service": What That Phrase Actually Buys YouSep 1
Physical Superintelligence Raises $58M to AI-Discover Physics BreakthroughsSep 1
Tsinghua Breaks Dijkstra's 41-Year Shortest-Path Record
August 2026
Aug 31
Prefix Sliding: Stanford Method Makes Reasoning 3× FasterAug 22
Meteoric Drones: What the Solar Cloud-Clearing Test ProvesAug 22
Wireless Quantum Network: What Stony Brook Sent Across 21 kmAug 20
Moderna and Merck's AI-Designed mRNA Cancer Vaccine Just Won Phase 3Aug 19
Sentence Transformers v6.0: MultiVectorEncoder Brings ColBERT Late Interaction to RAGAug 18
When Answers Get Cheap, Trust Becomes the JobAug 16
AI Drug Discovery Has an Evidence Problem — And a Benchmark Lesson for Everyone ElseAug 16
Is AI Out-Thinking Mathematicians, or Just Out-Remembering Them?Aug 5
"LLMs Can't Jump": The ICML Paper Arguing AI Can't Do AbductionAug 5
RLSVR/SpyRL: Turning "Who Is the Spy?" Into RL Training SignalAug 3
Paul Graham: Why LLMs Crush Math but Lag at WritingAug 3
Thariq: Jevons Paradox in Math — Demand for Mathematicians Rises
July 2026
Jul 30
AI Unicorns Barely Publish Research: What the Science Analysis ShowsJul 28
Top 10 Closed-Source and Open-Source Embedding Models (2026)Jul 28
What Is an Embedding? Plain-English Examples (2026)Jul 27
Intelligence Ownership: $500 Fine-Tune Beats FrontierJul 26
Can AI Cure Cancer? A Research-Backed Reality CheckJul 21
What Is the Jacobian Conjecture? Fable 5 Counterexample ExplainedJul 20
AI Advice Kills "I Don't Know": Cognitive Surrender in a PsyArXiv StudyJul 14
Richard Sutton’s Oak Lab — New Algorithms for AGI Beyond Static LLM TrainingJul 12
AI Boosts Scientist Careers but Flattens Discovery — Evans Nature Study ExplainedJul 7
Ternlight: 7 MB Embedding Model That Runs in the Browser (WASM SIMD Guide)Jul 3
Grounding vs RAG vs fine-tuning vs prompt engineering: which fix, when (a 2026 decision guide)
June 2026
Jun 27
Will AI replace mathematicians? What IEEE’s “Big Mathematics” debate means for proofs, Lean, and your careerJun 25
Vesuvius Challenge Reads an Entire Herculaneum Scroll — First Time in 2,000 YearsJun 23
94.3 on AIME 2026: VibeThinker-3B and the Case for Small Models With Frontier ReasoningJun 17
Anthropic Research: Domain Expertise Beats Coding Background in Agentic Programming (2026)Jun 16
AI vs Machine Learning vs Deep Learning — What's Actually Different?Jun 16
What Are Embeddings? Vector Search and Semantic AI Explained (2026 Guide)Jun 16
What Is Fine-Tuning an LLM? A Complete Guide for 2026