Every Question Has Its Own Value: Reinforcement Learning with Explicit Human Values
Dian Yu, Yulai Zhao, Kishan Panaganti, Linfeng Song, Haitao Mi, Dong Yu
TL;DR
RLEV tackles the mismatch between uniform correctness rewards and real-world importance by introducing explicit human values into LLM training. It defines a human utility $U(x,y) = v(x) \cdot \mathbf{1}_{\text{correct}}(y)$ and derives a stable surrogate reward $r(x,y)$ that scales with the per-prompt value $v(x)$, via $r(x,y) = s(x) \cdot \mathbf{1}_{\text{correct}}(y)$ with $s(x) = 1 + \min(\alpha v(x),1)$. Empirically, RLEV improves value-weighted accuracy and induces a value-sensitive termination policy, while ablations and gradient analysis confirm these gains arise from explicit value alignment rather than mere reward magnification. The approach remains robust under noisy value signals and generalizes to out-of-distribution benchmarks, offering a practical path to align LLM behavior with prioritized human objectives. Future work could explore dynamic, multi-dimensional value functions and combinations with RLHF to capture both objective correctness and subjective qualities.
Abstract
We propose Reinforcement Learning with Explicit Human Values (RLEV), a method that aligns Large Language Model (LLM) optimization directly with quantifiable human value signals. While Reinforcement Learning with Verifiable Rewards (RLVR) effectively trains models in objective domains using binary correctness rewards, it overlooks that not all tasks are equally significant. RLEV extends this framework by incorporating human-defined value signals directly into the reward function. Using exam-style data with explicit ground-truth value labels, RLEV consistently outperforms correctness-only baselines across multiple RL algorithms and model scales. Crucially, RLEV policies not only improve value-weighted accuracy but also learn a value-sensitive termination policy: concise for low-value prompts, thorough for high-value ones. We demonstrate this behavior stems from value-weighted gradient amplification on end-of-sequence tokens. Ablation studies confirm the gain is causally linked to value alignment. RLEV remains robust under noisy value signals, such as difficulty-based labels, demonstrating that optimizing for an explicit utility function offers a practical path to aligning LLMs with human priorities.