Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More
Enterprises need to know if the models that power their applications and agents work in real-life scenarios. This type of evaluation can sometimes be complex because it is hard to predict specific scenarios. A revamped version of the RewardBench benchmark looks to give organizations a better idea of a model’s real-life performance.
The Allen Institute of AI (Ai2) launched RewardBench 2, an updated version of its reward model benchmark, RewardBench, which they claim provides a more holistic view of model performance and assesses how models align with an enterprise’s goals and standards.
Ai2 built RewardBench with classification tasks that measure correlations through inference-time compute and downstream training. RewardBench mainly deals with reward models (RM), which can act as judges and evaluate LLM outputs. RMs assign a score or a “reward” that guides reinforcement learning with human feedback (RHLF).
Nathan Lambert, a senior research scientist at Ai2, told VentureBeat that the first RewardBench worked as intended when it was launched. Still, the model environment rapidly evolved, and so should its benchmarks.
“As reward models became more advanced and use cases more nuanced, we quickly recognized with the community that the first version didn’t fully capture the complexity of real-world human preferences,” he said.
Lambert added that with RewardBench 2, “we set out to improve both the breadth and depth of evaluation—incorporating more diverse, challenging prompts and refining the methodology to reflect better how humans actually judge AI outputs in practice.” He said the second version uses unseen human prompts, has a more challenging scoring setup and new domains.
Using evaluations for models that evaluate
While reward models test how well models work, it’s also important that RMs align with company values; otherwise, the fine-tuning and reinforcement learning process can reinforce bad behavior, such as hallucinations, reduce generalization, and score harmful responses too high.
RewardBench 2 covers six different domains: factuality, precise instruction following, math, safety, focus and ties.
“Enterprises should use RewardBench 2 in two different ways depending on their application. If they’re performing RLHF themselves, they should adopt the best practices and datasets from leading models in their own pipelines because reward models need on-policy training recipes (i.e. reward models that mirror the model they’re trying to train with RL). For inference time scaling or data filtering, RewardBench 2 has shown that they can select the best model for their domain and see correlated performance,” Lambert said.
Lambert noted that benchmarks like RewardBench offer users a way to evaluate the models they’re choosing based on the “dimensions that matter most to them, rather than relying on a narrow one-size-fits-all score.” He said the idea of performance, which many evaluation methods claim to assess, is very subjective because a good response from a model highly depends on the context and goals of the user. At the same time, human preferences get very nuanced.
Ai 2 released the first version of RewardBench in March 2024. At the time, the company said it was the first benchmark and leaderboard for reward models. Since then, several methods for benchmarking and improving RM have emerged. Researchers at Meta’s FAIR came out with reWordBench. DeepSeek released a new technique called Self-Principled Critique Tuning for smarter and scalable RM.
How models performed
Since RewardBench 2 is an updated version of RewardBench, Ai2 tested both existing and newly trained models to see if they continue to rank high. These included a variety of models, such as versions of Gemini, Claude, GPT-4.1, and Llama-3.1, along with datasets and models like Qwen, Skywork, and its own Tulu.
The company found that larger reward models perform best on the benchmark because their base models are stronger. Overall, the strongest-performing models are variants of Llama-3.1 Instruct. In terms of focus and safety, Skywork data “is particularly helpful,” and Tulu did well on factuality.
Ai2 said that while they believe RewardBench 2 “is a step forward in broad, multi-domain accuracy-based evaluation” for reward models, they cautioned that model evaluation should be mainly used as a guide to pick models that work best with an enterprise’s needs.
