Ai2

Ai2 compares Olmo 3 and Olmo Hybrid under matched training and splits loss by token type. Hybrids lead on nouns, verbs and adjectives, while transformers match them on verbatim repeats and closing brackets, suggesting exact copying is an attention strength that averages obscure.

Ai2 released MolmoMotion, which predicts the 3D paths of points on an object from a few video frames and a natural-language instruction, tracked in a shared world frame so it stays stable across camera changes. It ships with MolmoMotion-1M and the PointMotionBench evaluation.

Ai2's OlmoEarth v1.1 cuts run cost roughly threefold by tokenizing Sentinel-2 inputs more aggressively, since compute scales quadratically with token count, then repairs the resulting accuracy loss through pretraining changes. The dataset is unchanged from v1, isolating the method.

Ai2's ModSleuth reads papers, model and dataset cards and configs to rebuild a model's family tree. Olmo 3 traces to 89 models and 183 datasets, Nemotron 3 to 273 and 560, with chains up to eight hops deep, exposing hidden license inheritance and train/eval coupling.

Georgia Tech researchers traced Olmo 3 benchmark behavior to pretraining text using influence functions over 5.68M Dolma 3 files grouped into 576 categories. Social reasoning leaned on dialogue-rich writing, and unlearning the top-attributed literature documents cut SocialIQA accuracy more than rand

Ai2 releases OlmoEarth v1.2 open Earth-observation models in four sizes, replacing fixed positional signals with rotary embeddings after partners reported artifacts in embeddings. Ai2 reports consistent kNN and linear-probe gains. Based on the announcement; the tech report was not read.

Ai2 adds olmOCR 2, a compact open vision-language model, to its Playground. It reads hard documents in a single pass, including handwriting, equations, tables and multi-column layouts. Weights and training data are downloadable from Hugging Face.

Ai2 released the complete stack behind its MolmoAct 2 robotics foundation model: fine-tuning scripts, every training dataset, evaluation rollouts, and the open tokenizer training recipe. The model also gains official Hugging Face LeRobot support for existing robot setups.

Ai2 open sources olmo-eval, an evaluation workbench for repeated checkpoint scoring. Models, tools and LLM judges are swappable, only benchmarks that need isolation run sandboxed, results share one schema, and a comparison view diffs checkpoints per question to expose gains and regressions.

Ai2's ArtifactLinker models Hugging Face as a graph with eval scores as edges, uses a GNN to predict unevaluated model and benchmark pairs likely to set SOTA, then has a coding agent run the evals. It reproduces official scores within 80% about 73% of the time, with code and data released.

Ai2's infini-gram counts how often a phrase of any length appears across indexed corpora, and a Stony Brook group used it to dissect machine-written prose. Rare phrasing made up 41.6 percent of suspected AI books against 19.1 percent of award winners, and traced one flagged story to fan fiction.

Ai2 previews TutorMoments, which pauses real tutoring transcripts at teacher-flagged decision points, hands the tutor turn to a model, and scores whether it supported or over-helped. Models told only to tutor well give away the answer; naming the tradeoff helps, but only so far.

Ai2's BenchMIRT applies multidimensional item response theory to 100 models across 16 benchmarks. It recovers reasoning and safety axes without labels, shows BBQ tracks reasoning and ToxiGen adds little, and finds the top 10% of questions largely preserve model rankings. Models predate March 2025.

Ai2 keeps AutoDiscovery in early access until July 31, with 500 hypothesis credits for new accounts. The system analyzes a dataset, generates hypotheses unprompted, writes code to test each, then ranks the most surprising results. Prior runs produced findings later published in peer review.

Danish Foundation Models adapted Ai2's FlexOlmo into FlexMoRE, replacing most full-size modules with compact representations. Its best configuration reportedly matches or beats FlexOlmo with under a third of the parameters, so institutions with unshareable data can contribute modules on commodity ha

Ai2 and Providence researchers ran AutoDiscovery, a multi-agent hypothesis search, on TCGA breast cancer data. An unguided run found nothing actionable. An oncologist-seeded run surfaced a stronger immune signature in lobular carcinoma, then validated it in a second cohort and tissue imaging. Prepri

Ai2's DiScoFormer estimates the score function and density of an arbitrary dataset from one pretrained model, no per-dataset retraining. At 100 dimensions it reports roughly 6.5x lower score error and 37x lower density error than tuned kernel density estimation, and holds up on unfamiliar shapes.

Ai2 updated Asta: AutoDiscovery hypotheses open in DataVoyager with datasets and results preloaded, and paper search defaults to a Deep mode that judges whether results answer the query and keeps searching. Slower per query. No benchmark numbers given in the thread.

A Thai team adapted Ai2's Dolma pipeline into Mangosteen, a 47B-token corpus. Sentence-level dedup failed on Thai, so they kept document and URL dedup and rewrote filters. Filtering removed over 80 percent of Common Crawl while holding or improving Thai benchmark results.

Ai2 announces AIMIP, a CMIP-style intercomparison for AI climate emulators: train through 2014, forecast the held-out decade, and score on averages, trends, El Nino response, variability and warmer-ocean behavior. Eight models from six groups entered; several underestimate warming.

Ai2's ACE2S-SHiELD+ learns to separate sea surface temperature effects from CO2 effects. Earlier ACE emulators had only seen the two rise together and broke on standard probe scenarios, so the team generated training runs that decorrelate them, varying CO2 every 30 days against a warming ocean.

Researchers replaced parts of drug names with invented strings to test whether Olmo 3 answers from the name's suffix or from drug knowledge. For 51 to 59 percent of drugs, real and fake names drew near-identical answers. Rarer drugs in training data showed more affix reliance.

Ai2 retires SciArena after about 3,900 votes from roughly 1,700 researchers. Citation quality ranked first at 23%, ahead of depth and direct answers. o3 led over Claude Opus 4.1, Gemini 3 Pro Preview and DeepSeek-R1. The expert annotations ground evaluation data.

PointCheck checks accessibility the way a user experiences it: tab through the page, screenshot every stop, and have Molmo locate the focused element, with MolmoWeb and Olmo 3 handling the rest. It catches visual failures such as invisible focus rings that markup linting cannot see.