Source
Olmo Team (2025), LIT-130 — the Olmo 3 RL-Zero track.
What the track is
Take the pretrained base, skip the supervised stage on reasoning traces, and run reinforcement learning with verifiable rewards straight away: a rule-based verifier with reference answers for math, test cases for code, an LLM judge for general chat. Olmo 3 releases four such 7B series with their data and checkpoints so that RL algorithms, and the effect of pretraining data on RL, can be studied from a clean start.
What the practice rests on is LIT-130’s statement of the track rather than a measured win over the SFT path: Olmo 3 runs RLVR from the base in four domain-focused series (math, code, instruction following, general chat), decontaminates the RL-Zero data aggressively and checks it with negative controls trained against random rewards, and frames the result as an experimental pathway and an open benchmark for RL, not as its production recipe.
Why this is Proposed
The source itself calls the track experimental and positions it as a benchmark, and the Think models Olmo 3 actually ships go through the SFT stage (SOTA-129). The idea predates Olmo 3: the “R1-Zero” style of training is where the name comes from, and that work is now filed (LIT-164). DeepSeek-R1 is where the pathway was first stated, and the record names it as the origin: R1-Zero applies RL directly to DeepSeek-V3-Base “without relying on supervised fine-tuning (SFT) as a preliminary step”, with a rule-based reward for accuracy and format, and the paper calls it the “first open research to validate that reasoning capabilities of LLMs can be incentivized purely through RL, without the need for SFT.” It too ships the SFT path for its production model — R1 adds cold-start data before RL to fix R1-Zero’s poor readability and language mixing — so neither the origin nor Olmo 3 offers this as the recommended recipe. Olmo 3 remains the source because it is the open, reproducible statement of the pathway.
A second reason to stay Proposed arrived with the literature. LIT-167 finds that DeepSeek-V3-Base already exhibits the “Aha moment” before any RL, and that Qwen2.5 bases reason without a prompt template — so how much of the result belongs to skipping SFT, and how much to what the base model already carried, is unsettled. “Run RLVR directly on the base” is a claim whose answer may depend on which base. LIT-119 cites the pathway to mark the other direction one can go from SOTA-129: it removes the middle stage, where the anti-curriculum (SOTA-123) merges the first two.