# Binary-gated GRPO on GDPval: 79% of GRPO groups had no passing rollout, and pass@1 rose from 31.9% to 40.7% while plain GRPO barely moved (Qwen3.8-27B) — retained record

> Retained reddit record from reddit:reinforcementlearning.

Canonical: https://highsignal.app/data/records/eec2e3ea68207c89

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:reinforcementlearning)  published 2026-11-19 16:42 ingested 2026-11-20 02:39

[original source ↗](https://www.reddit.com/r/reinforcementlearning/comments/1y0e2vu/binarygated_grpo_on_gdpval_79_of_grpo_groups_had/)

## retained text



```
I work at Snorkel AI, and we wanted to share new training work from my colleague that may be of interest here. The write-up is linked at the end of the post.

The setting is RL on long-horizon professional tasks, where a single attempt averages about 50,000 response tokens and success is rare. Qwen3.8-27B was trained on 1,105 tasks from the GDPval+ series. Each task has an expert-built rubric and reference solution, and each group has eight rollouts, about 102K rollouts in total. An attempt passes when at least 90% of rubric items are satisfied.

With a binary signal, GRPO gives zero advantage to any group where every rollout fails. Here, 79% of generated groups had no passing rollout. Dynamic sampling skips those groups and refills the batch, but it does not create new signal. In the groups GRPO discarded, the median best rollout scored 78% on the rubric, and 66% of them had a rollout scoring at least 70%. The attempts in these groups differed substantially.

Binary-gated GRPO works as follows.

* If a group has at least one pass, the advantages are the standard GRPO advantages on pass/fail.
* If no rollout passes, the advantages come from a fallback on the rubric score. Scores at or below 40% map to zero, and scores above that scale linearly up to the 90% pass threshold. The fallback is weighted by 0.85.

The fallback is used only when there is no pass in the group, so partial credit never competes with a pass.

On the 185 GDPval gold tasks, with eight attempts per task, the results were as follows.

* Groups carrying a training signal rose from 21% under GRPO to 95% under binary-gated GRPO.
* Pass@1 rose from 31.9% to 40.7%. GRPO on the same data and a similar rollout count barely moved from the base model.
* In LLM-judged head-to-head comparisons against human experts, deliverables were rated better in 44.0% of cases, up from 36.4%.
* Attempts that exhausted the response budget fell from 17.2% to 7.0%.

Each method was trained once. The judge is an LLM and is noisy, which is why the pass threshold is set at 90%. The results cover GDPval-style tasks only.

Write-up: [https://snorkel.ai/blog/gdpval-agents-rare-success/](https://snorkel.ai/blog/gdpval-agents-rare-success/)

Earlier post on partial credit and completion: [https://snorkel.ai/blog/when-partial-credit-opposes-completion/](https://snorkel.ai/blog/when-partial-credit-opposes-completion/)
```



document fetched 2026-11-20 00:40 · canonical https://reddit.com/r/reinforcementlearning/comments/1y0e2vu/binarygated_grpo_on_gdpval_79_of_grpo_groups_had
