Theoretical Limits of Language Model Alignment

• URL: http://arxiv.org/abs/2605.07105v1
• Date: Fri, 08 May 2026 01:32:22 GMT
• Title: Theoretical Limits of Language Model Alignment
• Abstract summary: Language model (LM) alignment improves model outputs to reflect human preferences while preserving the capabilities of the base model.<n>The most common alignment approaches are (i) reinforcement learning, which maximizes the expected reward under a KL-divergence constraint, and (ii) best-of-$N$ alignment, which selects the highest-reward output among $N$ independent samples.<n>We characterize the information-theoretic limits of KL-regularized alignment by deriving the maximum achievable expected reward gain for a fixed KL-divergence budget.
• Score: 9.45142272392467
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Language model (LM) alignment improves model outputs to reflect human preferences while preserving the capabilities of the base model. The most common alignment approaches are (i) reinforcement learning, which maximizes the expected reward under a KL-divergence constraint, and (ii) best-of-$N$ alignment, which selects the highest-reward output among $N$ independent samples. Despite their widespread use, the fundamental limits of reward improvement under a KL budget remain poorly understood. We characterize the information-theoretic limits of KL-regularized alignment by deriving the maximum achievable expected reward gain for a fixed KL-divergence budget. Our first result provides a closed-form expression for the optimal reward improvement, governed by a Jeffreys divergence term rather than the $\sqrt{\texttt{KL}}$ used in prior analyses. We further reformulate this expression as a covariance under the base model, yielding a practical estimator that predicts achievable alignment gains from base model samples alone. We extend our analysis to the proxy reward setting, showing that the gap between ideal and proxy alignment (reward hacking) grows with the magnitude of reward error and when the KL penalty factor decreases. We then prove that reward ensembling mitigates reward hacking, providing a theoretical justification for this technique used in practice. Empirically, we compute the KL-reward Pareto frontier for two tasks for LMs, safety and summarization, and show that best-of-$N$ closely approaches the theoretical limit, while PPO and GRPO remain substantially suboptimal. Our theoretical results shed light on several empirically observed phenomena in the alignment literature and suggest that algorithmic improvements are needed to achieve optimal alignment without high inference costs.

Related papers

• Ratio-Variance Regularized Policy Optimization [64.95520246570446]
  We show that explicitly constraining the policy ratio variance provides a principled local approximation to trust-region constraints.<n>We introduce $bf R2bf VPO$ (Ratio-Variance Regularized Policy Optimization), which implements this constraint via a primal-dual optimization framework.
  arXiv  Detail & Related papers  (2026-05-26T09:53:42Z)
• Binary Rewards and Reinforcement Learning: Fundamental Challenges [7.106986689736826]
  Reinforcement learning with rewards RLVR has become a standard approach for improving reasoning in language models.<n>We provide a structural account of this phenomenon grounded in the properties of binary rewards.
  arXiv  Detail & Related papers  (2026-05-04T09:17:26Z)
• Rethinking KL Regularization in RLHF: From Value Estimation to Gradient Optimization [6.136585583991053]
  Reinforcement Learning from Human Feedback (RLHF) leverages a Kullback-Leibler (KL) divergence loss to stabilize training and prevent overfitting.<n>In methods such as GRPO, its implementation may be guided by principles from numerical value estimation.
  arXiv  Detail & Related papers  (2025-10-02T01:00:02Z)
• FlowRL: Matching Reward Distributions for LLM Reasoning [69.88820066093798]
  We propose FlowRL: matching the full reward distribution via flow balancing instead of maximizing rewards in large language model (LLM) reinforcement learning (RL)<n>We transform scalar rewards into a normalized target distribution using a learnable partition function, and then minimize the reverse KL divergence between the policy and the target distribution.
  arXiv  Detail & Related papers  (2025-09-18T17:56:36Z)
• A Principled Loss Function for Direct Language Model Alignment [0.0]
  We propose a novel loss function derived directly from the RLHF optimality condition.<n>Our proposed loss targets a specific finite value for the logits, which is dictated by the underlying reward, rather than its difference.<n>This inherent stability prevents reward hacking and leads to more effective alignment.
  arXiv  Detail & Related papers  (2025-08-10T01:56:58Z)
• Supervised Optimism Correction: Be Confident When LLMs Are Sure [91.7459076316849]
  We establish a novel theoretical connection between supervised fine-tuning and offline reinforcement learning.<n>We show that the widely used beam search method suffers from unacceptable over-optimism.<n>We propose Supervised Optimism Correction, which introduces a simple yet effective auxiliary loss for token-level $Q$-value estimations.
  arXiv  Detail & Related papers  (2025-04-10T07:50:03Z)
• Correcting the Mythos of KL-Regularization: Direct Alignment without Overoptimization via Chi-Squared Preference Optimization [78.82586283794886]
  $chi2$-Preference Optimization ($chi$PO) is an efficient offline alignment algorithm provably robust to overoptimization.<n>$chi$PO implements the principle of pessimism in the face of uncertainty via regularization.<n>$chi$PO's simplicity and strong guarantees make it the first practical and general-purpose offline alignment algorithm provably robust to overoptimization.
  arXiv  Detail & Related papers  (2024-07-18T11:08:40Z)
• Provably Mitigating Overoptimization in RLHF: Your SFT Loss is Implicitly an Adversarial Regularizer [52.09480867526656]
  We identify the source of misalignment as a form of distributional shift and uncertainty in learning human preferences.<n>To mitigate overoptimization, we first propose a theoretical algorithm that chooses the best policy for an adversarially chosen reward model.<n>Using the equivalence between reward models and the corresponding optimal policy, the algorithm features a simple objective that combines a preference optimization loss and a supervised learning loss.
  arXiv  Detail & Related papers  (2024-05-26T05:38:50Z)
• Asymptotics of Language Model Alignment [27.37118975691123]
  We show that the optimal KL-constrained RL solution satisfies a large deviation principle. We also show that the rate of growth of the scaled cumulants of the reward is characterized by proper Renyi cross entropy.
  arXiv  Detail & Related papers  (2024-04-02T08:40:07Z)
• Theoretical guarantees on the best-of-n alignment policy [110.21094183592358]
  We show that the KL divergence between the best-of-$n$ policy and the reference policy is an upper bound on the actual KL divergence.<n>We propose a new estimator for the KL divergence and empirically show that it provides a tight approximation.<n>We conclude with analyzing the tradeoffs between win rate and KL divergence of the best-of-$n$ alignment policy.
  arXiv  Detail & Related papers  (2024-01-03T18:39:13Z)
• Submodular Reinforcement Learning [38.40138241424851]
  In reinforcement learning (RL), rewards of states are typically considered additive, and following the Markov assumption, they are $textitindependent$ states visited previously. In many important applications, such as coverage control, experiment design and informative path planning, rewards naturally have diminishing returns, i.e., their value decreases in light of similar states visited previously. We propose $textitsubmodular RL$ (SubRL), a paradigm which seeks to optimize more general, non-additive (and history-dependent) rewards modelled via submodular set functions which capture diminishing returns
  arXiv  Detail & Related papers  (2023-07-25T09:46:02Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.