Enterprises have to know if the fashions that energy their functions and brokers work in real-life eventualities. This kind of analysis can typically be advanced as a result of it’s onerous to foretell particular eventualities. A revamped model of the RewardBench benchmark appears to be like to offer organizations a greater concept of a mannequin’s real-life efficiency.
The Allen Institute of AI (Ai2) launched RewardBench 2, an up to date model of its reward mannequin benchmark, RewardBench, which they declare supplies a extra holistic view of mannequin efficiency and assesses how fashions align with an enterprise’s objectives and requirements.
Ai2 constructed RewardBench with classification duties that measure correlations by means of inference-time compute and downstream coaching. RewardBench primarily offers with reward fashions (RM), which might act as judges and consider LLM outputs. RMs assign a rating or a “reward” that guides reinforcement studying with human suggestions (RHLF).
Nathan Lambert, a senior analysis scientist at Ai2, informed VentureBeat that the primary RewardBench labored as supposed when it was launched. Nonetheless, the mannequin setting quickly advanced, and so ought to its benchmarks.
“As reward fashions turned extra superior and use circumstances extra nuanced, we rapidly acknowledged with the group that the primary model didn’t totally seize the complexity of real-world human preferences,” he mentioned.
Lambert added that with RewardBench 2, “we got down to enhance each the breadth and depth of analysis—incorporating extra various, difficult prompts and refining the methodology to replicate higher how people truly choose AI outputs in apply.” He mentioned the second model makes use of unseen human prompts, has a more difficult scoring setup and new domains.
Utilizing evaluations for fashions that consider
Whereas reward fashions take a look at how effectively fashions work, it’s additionally vital that RMs align with firm values; in any other case, the fine-tuning and reinforcement studying course of can reinforce dangerous conduct, resembling hallucinations, scale back generalization, and rating dangerous responses too excessive.
RewardBench 2 covers six totally different domains: factuality, exact instruction following, math, security, focus and ties.
“Enterprises ought to use RewardBench 2 in two alternative ways relying on their software. In the event that they’re performing RLHF themselves, they need to undertake the perfect practices and datasets from main fashions in their very own pipelines as a result of reward fashions want on-policy coaching recipes (i.e. reward fashions that mirror the mannequin they’re making an attempt to coach with RL). For inference time scaling or knowledge filtering, RewardBench 2 has proven that they’ll choose the perfect mannequin for his or her area and see correlated efficiency,” Lambert mentioned.
Lambert famous that benchmarks like RewardBench provide customers a strategy to consider the fashions they’re selecting primarily based on the “dimensions that matter most to them, somewhat than counting on a slim one-size-fits-all rating.” He mentioned the concept of efficiency, which many analysis strategies declare to evaluate, could be very subjective as a result of an excellent response from a mannequin extremely is dependent upon the context and objectives of the consumer. On the similar time, human preferences get very nuanced.
Ai 2 launched the primary model of RewardBench in March 2024. On the time, the corporate mentioned it was the primary benchmark and leaderboard for reward fashions. Since then, a number of strategies for benchmarking and bettering RM have emerged. Researchers at Meta’s FAIR got here out with reWordBench. DeepSeek launched a brand new method known as Self-Principled Critique Tuning for smarter and scalable RM.
How fashions carried out
Since RewardBench 2 is an up to date model of RewardBench, Ai2 examined each present and newly educated fashions to see in the event that they proceed to rank excessive. These included quite a lot of fashions, resembling variations of Gemini, Claude, GPT-4.1, and Llama-3.1, together with datasets and fashions like Qwen, Skywork, and its personal Tulu.
The corporate discovered that bigger reward fashions carry out finest on the benchmark as a result of their base fashions are stronger. Total, the strongest-performing fashions are variants of Llama-3.1 Instruct. By way of focus and security, Skywork knowledge “is especially useful,” and Tulu did effectively on factuality.
Ai2 mentioned that whereas they consider RewardBench 2 “is a step ahead in broad, multi-domain accuracy-based analysis” for reward fashions, they cautioned that mannequin analysis ought to be primarily used as a information to choose fashions that work finest with an enterprise’s wants.
