Is Outcome Reward Models superseded?

Outcome Reward Models (LLM reasoning / chain-of-thought): cited as a baseline — critiqued by newer work, not yet beaten on a benchmark here. 2 paper(s) critique it, 0 beat it on benchmarks — #92 of 772 most-superseded. Sub-problem: cluster led by Outcome Reward Models. Newer alternatives in the same sub-problem include Verifiable Process Reward Models (VPRMs).

LLM reasoning / chain-of-thought

Outcome Reward Models

Superseded baseline#92 of 772 most-superseded

Cited as a baseline — critiqued by newer work, not yet beaten on a benchmark here

2 papers critique it · 0 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites Outcome Reward Models as a baseline.

Traditional outcome-only verifiers (Outcome Reward Models) are limited, evaluating only the final answer and often missing intermediate errors that compromise the reasoning trajectory Wang2024.

— Training Vision-Language Process Reward Models for Test-Time Scaling in Multimodal Reasoning: Key Insights and Lessons Learned

“most applications of RLVR to date focus on outcome-level verification, where the model receives a scalar reward only if the final answer is correct. While such outcome rewards improve overall performance, they provide little guidance on the internal reasoning process itself. As a result, a model may reach a correct conclusion through unsound, inconsistent, or opaque reasoning traces, limiting interpretability and trustworthiness.”

— Beyond Outcome Verification: Verifiable Process Reward Models for Structured Reasoning

What to use instead

Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.

Verifiable Process Reward Models (VPRMs)Beyond Outcome Verification: Verifiable Process Reward Models for Structured Reasoning
Jan 23, 2026