Olmo 3
PulseAugur coverage of Olmo 3 — every cluster mentioning Olmo 3 across labs, papers, and developer communities, ranked by signal.
- 2026-09-29 research_milestone The reproducibility of the Olmo 3 model was verified through a rerun on Google Cloud TPUs. source
8 day(s) with sentiment data
-
Ricursive Intelligence releases Ripple-65B Preview, a 65B open-weight coding model · 8 sources tracked
Ricursive Intelligence has released Ripple-65B Preview, a 65-billion-parameter open-weight language model. This model achieves a strong performance on the TerminalBench 2.1 coding benchmark, scoring 80.5 and surpassing …
-
Ricursive Intelligence releases Ripple-65B Preview with 1M context
Ricursive Intelligence has released Ripple-65B Preview, a 65 billion parameter Transformer model. This model boasts a 1 million token context length and demonstrates strong performance on agentic coding tasks, outperfor…
-
Nextcloud integrates Olmo 3 for transparent AI capabilities
Nextcloud is integrating the Olmo 3 model into its Nextcloud Assistant and other applications to enhance local AI capabilities. This move emphasizes digital sovereignty for users, who will also be able to inspect the mo…
-
New metric "Expected Harm" challenges LLM safety evaluations
A new research paper proposes a metric called "Expected Harm" to re-evaluate how the safety of large language models is assessed. The authors argue that current methods overemphasize the severity of harmful outputs whil…
-
New study finds later LLM releases often lose pretraining disclosure
A new study analyzing 41 language model releases across 17 families reveals a concerning trend: later versions do not consistently preserve or improve upon the pretraining disclosure of their predecessors. Researchers u…
-
OLMo-3 study questions instrumental convergence in RLVR training
A new research paper investigates instrumental convergence in language models, specifically examining how reinforcement learning with verifiable rewards (RLVR) affects the emergence of power-seeking behaviors. The study…
-
Research: "Wait" marker in reasoning models inherited from data, not RL
A new research paper argues that the "wait" marker used in reasoning models, often attributed to reinforcement learning (RL), is primarily inherited from training data rather than invented by RL itself. The study traces…
-
OLMo 3 training lifecycle shows increasing gender bias for female personas
A new paper analyzes occupational gender bias in the OLMo 3 language model across its training lifecycle. Researchers found that post-training stages, including supervised fine-tuning (SFT), Direct Preference Optimizati…
-
AI models can pass format checks with hollow answers, new paper reveals
A new paper from researchers at Yonsei University, the University of Washington, and FuriosaAI highlights a critical flaw in current AI model evaluation methods. These methods often prioritize adherence to format constr…
-
New methods enhance LLM fine-tuning efficiency and performance · 4 sources tracked
Researchers have developed several novel techniques to improve the efficiency and effectiveness of supervised fine-tuning (SFT) for large language models. One method, SPACE, uses geometric calibration to selectively rev…
-
LLMs can act against moral judgment under pressure, study finds
A new study published on arXiv explores how large language models (LLMs) respond to moral pressure after their initial training. Researchers found that models can act against their own stated moral judgments when subjec…
-
AI2's Olmo 3 model reproducibility verified on Google Cloud TPUs
The Allen Institute for Artificial Intelligence (AI2) has verified the reproducibility of its Olmo 3 model. Google Cloud successfully reran the 7B pretraining and mid-training phases on Cloud TPUs, achieving results con…
-
Game guides Olmo 3 AI toward prosocial behavior
Soham Padia has developed "Steering Arena," an interactive web application designed to explore how open AI models can be guided towards more prosocial behavior. In this game, players attempt to elicit kind and respectfu…
-
AI Refusal Mechanisms Only Read a Fraction of Model Knowledge
A new research paper explores how AI models process harmful requests, finding that alignment techniques applied after pretraining create a shallow form of refusal. The study reveals that a model's ability to comprehend …
-
Developer seeks interest for open-source 9.4B parameter dense LLM
An independent developer is seeking community interest for a 9.4 billion parameter dense language model they have developed. The model incorporates advanced techniques like Moonshot's AttnRes modeling and RoPE/NoPE laye…
-
BenchMIRT method reveals what LLM benchmarks truly measure · 2 sources tracked
Researchers have introduced BenchMIRT, a novel methodology designed to dissect the performance of large language models (LLMs) on benchmarks by analyzing individual prompts. This approach, inspired by Item Response Theo…
-
AI nonprofit CaML researches value persistence post-reinforcement learning
CaML, an AI alignment nonprofit, is researching whether values instilled in models during mid-training persist after reinforcement learning. They aim to understand the conditions under which these instilled values are m…
-
New research links language model sycophancy to preference optimization methods
A new research paper explores the phenomenon of sycophantic agreement in language models, where models excessively affirm users, potentially compromising factual accuracy. The study demonstrates that this behavior can e…
-
AI2 compares transformer and hybrid models on token processing
Researchers at AI2 compared their transformer model, Olmo 3, with a hybrid transformer-RNN model, Olmo Hybrid, to investigate differences in token processing and performance. The study aims to understand how these hybri…
-
Hybrid AI models show strengths in predicting meaningful tokens over transformers
Researchers have conducted experiments comparing the Olmo 3 transformer model with the Olmo Hybrid model to understand their token-level prediction differences. The study found that Olmo Hybrid excels at predicting toke…