What happened
The paper, “F-GRPO: Don’t Let Your Policy Learn the Obvious and Forget the Rare,” was submitted to arXiv on February 6, 2026, and revised on September 2, 2026. Its authors are Daniil Plyusov, Alexey Gorbatovski, Boris Shaposhnikov, Viacheslav Sinii, Alexey Malakhov, Daria Korotyshova and Daniil Gavrilov. The work studies reinforcement learning with verifiable rewards, where a training process can check whether an answer is correct and use that result as a reward.
The authors focus on group sampling, a common way to estimate advantages and stabilize policy updates. With practical compute limits, training often uses finite groups of sampled rollouts rather than very large groups. A small group can contain correct answers but miss a rare, valid way of solving the prompt. The authors analyze these prompt-local “tail-miss” events as a function of group size and report non-monotonic behavior. In their categorical abstraction, the mass assigned to correct solutions that were not sampled can shrink even while the total probability assigned to correct answers grows.
F-GRPO is a difficulty-aware scaling coefficient inspired by Focal loss. It down-weights updates on sampled groups with high success, addressing the risk that training over-reinforces the common correct behaviors visible in a small group. The paper evaluates the proposal with categorical simulation, a Maze test with a single solution, and language-model experiments. For the language-model work, it reports a representative GRPO group-size sweep and fixed-group-size transfer tests across GRPO, DAPO and CISPO.
On Qwen2.5-7B with group size N=8, average math pass@256 rose from 64.1 to 70.3 with F-GRPO over GRPO, from 69.3 to 72.5 over DAPO, and from 73.2 to 76.8 over CISPO. The authors also report improved out-of-distribution pass@256 in all three comparisons. They state that these gains came without increasing group size or computational cost. The paper presents F-GRPO as a training method; the arXiv record does not announce a software release or hosted service.
What the archive adds
New Runtime’s earlier coverage described fine-tuning GPT-OSS-20B for multilingual reasoning and argued that method choice should follow a model’s capability gap (Fine-tuning GPT-OSS-20B for multilingual reasoning on Amazon SageMaker HyperPod, October 22, 2026). That note concerns a different model and objective. F-GRPO is a new research result on how to preserve rare correct behaviors during RLVR, not a restatement of that case.
Trend signal
This does not yet fit a pattern New Runtime tracks. The paper studies a specific optimization to policy updates, but it does not establish a broader shift in agent systems, deployment, or organizational practice.
Why it matters
For teams training models against verifiable rewards, the result highlights a tension between compute-efficient sampling and coverage of a task’s solution space. A small rollout group costs less to generate, but it may provide a biased view of which correct behaviors the policy can learn. If updates favor only the successful responses that happen to be sampled most often, optimization can improve aggregate success while narrowing the set of behaviors retained.
F-GRPO’s reported comparison is useful because it tests the same group size across three algorithms, rather than attributing improvement to simply drawing more samples. The result suggests a possible route to better coverage under a fixed rollout budget. It is not evidence that every team should replace its current method: results are limited to the paper’s experiments, and the value of a training change depends on the model, reward design, tasks and evaluation distribution. For practitioners, the practical question is whether rare valid solutions matter in their own workload and whether an evaluation can detect their loss.
Limits and uncertainty
The reported numbers are experimental results from the paper and have not been independently checked here. The abstract identifies Qwen2.5-7B and the pass@256 comparisons, but does not establish how results generalize to other model scales, domains or reward setups. It also does not quantify the method’s implementation overhead or report production deployment outcomes. Better pass@256 is evidence about the evaluated sampling metric, not proof that every individual deployment will retain a broader or more reliable set of solutions.
Skill to build now
If you run RLVR experiments, track not only total pass rate but also which correct solution types appear in sampled rollouts. Keep group size fixed in an experiment comparing your current update rule with a difficulty-aware alternative, and evaluate both in-distribution and on held-out prompts. This makes it easier to distinguish a genuine improvement in learning from gains caused by simply spending more on sampling.
Reader takeaway
F-GRPO targets a specific failure mode in small-group RLVR: updates can reward common sampled successes while overlooking rare correct trajectories. The authors report better pass@256 on Qwen2.5-7B across GRPO, DAPO and CISPO at the same group size. Treat that as a promising research result to test against your own tasks, not a universal fine-tuning recipe.