olmo-3-5f92ca86·4 events·first seen Aliases: OLMo-3, OLMo-3.1, OLMo 3 32B, OLMo 2 32B
A new arXiv preprint proposes a principled methodology for making valid extractable memorization claims about LLMs, addressing both over- and under-statement problems in prior work. The core contribution is a 'matched comparison' approach that measures generation probabilities of training sequences against comparable non-training sequences to establish a calibrated baseline for predictability. Two formalizations are offered: a conformal test for population-level claims and a census method for single-document claims. Applied to OLMo 2 32B on Wikipedia and Llama 3.1 70B on books, the framework reveals significant false-positive rates in naive extraction studies and supports memorization claims at probability thresholds as low as 1e-27.
Researchers present Soofi S 30B-A3B, a Mixture-of-Experts hybrid Mamba Transformer foundation model for German and English, activating only 3B of 30B parameters per token with near-constant inference cache for long-context efficiency. Pretrained on ~27 trillion tokens with up-weighted German data, it claims top aggregate scores among fully open models in both languages, outperforming OLMo 3 32B and Apertus 70B, and surpasses all European sovereign baselines tested. The model was built on Deutsche Telekom's sovereign HPC infrastructure in Munich and will be released with weights, checkpoints, hyperparameters, training code, and full data accounting under permissive terms.
A new arXiv preprint introduces Logit-Contribution Scoring (LOCOS), a method for identifying attention heads responsible for non-literal retrieval in long-context LLMs — cases where models synthesize answers from meaning rather than copying tokens verbatim. Existing detectors fail at this task because they rely on a literal-copy criterion that misses the output-value (OV) circuit mechanism. Evaluated across Qwen3, Gemma-3, and OLMo-3.1, LOCOS outperforms prior attention-based detectors on the NoLiMa benchmark, with ablation of 50 heads on Qwen3-8B collapsing ROUGE-L from 0.401 to 0.000 while the best baseline retains 0.292. The identified heads are retrieval-specific, leaving parametric recall and arithmetic reasoning unaffected.
SCOPE is a data-free self-play framework for training language models on open-ended tasks without external supervision or frontier-model judges. It co-evolves two policies—a Challenger that generates document-grounded tasks and a Solver that answers via multi-turn retrieval—using a frozen copy of the initial model as a self-judge that writes task-specific rubrics. Across three 7-8B models (Qwen2.5, Qwen3, OLMo-3), SCOPE achieves up to +10.4 points on eight open-ended benchmarks and +13.8 points on seven held-out short-form QA benchmarks, matching or exceeding GRPO trained on ~9K curated prompts. Ablations identify rubric generation quality as the primary bottleneck for self-judging.