Hi, and thanks for maintaining this list — it's a great resource. I'd like to suggest a recent paper for the 🏆 Reward Learning (Process Reward Models) section, alongside neighbors like Math-Shepherd and PRMBench:
Verifiable Rewards Beyond Math and Code: Lightweight Corpus-Grounded Process Supervision for Factual Question Answering (CorVer)
CorVer is a lightweight, plug-in process reward for RL fine-tuning (GRPO) of LLMs on factual QA. Instead of a neural verifier (NLI / LLM-as-judge / retrieve-and-grade), it derives a corpus-grounded signal from Wikipedia co-occurrence counts via an infini-gram index: a small (0.5B) extractor turns each generated sentence into spans whose corpus co-occurrence gives sentence-level credit, which is mapped to token-level advantages. Across 30 cells (six instruction-tuned models, 3B–14B; five QA benchmarks) it beats the raw baseline in every cell (+4.1 pp avg on TriviaQA) and outperforms four neural-verifier baselines in 18/20 cells at 4.8–8.4× lower training cost.
Suggested entry (matching the existing format):
* Verifiable Rewards Beyond Math and Code: Lightweight Corpus-Grounded Process Supervision for Factual Question Answering [[Paper]](https://arxiv.org/abs/2605.29648)  [[Code]](https://github.com/shichengf/CorVer)
Happy to open a PR instead if you prefer. Thanks!
Hi, and thanks for maintaining this list — it's a great resource. I'd like to suggest a recent paper for the 🏆 Reward Learning (Process Reward Models) section, alongside neighbors like Math-Shepherd and PRMBench:
Verifiable Rewards Beyond Math and Code: Lightweight Corpus-Grounded Process Supervision for Factual Question Answering (CorVer)
CorVer is a lightweight, plug-in process reward for RL fine-tuning (GRPO) of LLMs on factual QA. Instead of a neural verifier (NLI / LLM-as-judge / retrieve-and-grade), it derives a corpus-grounded signal from Wikipedia co-occurrence counts via an infini-gram index: a small (0.5B) extractor turns each generated sentence into spans whose corpus co-occurrence gives sentence-level credit, which is mapped to token-level advantages. Across 30 cells (six instruction-tuned models, 3B–14B; five QA benchmarks) it beats the raw baseline in every cell (+4.1 pp avg on TriviaQA) and outperforms four neural-verifier baselines in 18/20 cells at 4.8–8.4× lower training cost.
Suggested entry (matching the existing format):
Happy to open a PR instead if you prefer. Thanks!