Post-Training with Policy Gradients: Optimality and the Base Model Barrier

• URL: http://arxiv.org/abs/2603.06957v1
• Date: Sat, 07 Mar 2026 00:25:53 GMT
• Title: Post-Training with Policy Gradients: Optimality and the Base Model Barrier
• Abstract summary: We study post-training linear autoregressive models with outcome and process rewards.<n>We prove that variants of policy gradient (PG) can achieve likelihood $1 - varepsilon$ with an essentially minimax optimal number of reward queries.
• Score: 27.674563695368665
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: We study post-training linear autoregressive models with outcome and process rewards. Given a context $\boldsymbol{x}$, the model must predict the response $\boldsymbol{y} \in Y^N$, a sequence of length $N$ that satisfies a $γ$ margin condition, an extension of the standard separability to sequences. We prove that on test samples where the base model achieves a non-trivial likelihood $α$, a variant of policy gradient (PG) can achieve likelihood $1 - \varepsilon$ with an essentially minimax optimal number of reward queries $\tilde{O}((α^{-1} + \varepsilon^{-1})/γ^2)$. However, a barrier arises for going beyond the support of the base model. We prove that the overall expected error after post-training with outcome rewards is governed by a property of the base model called the Likelihood Quantile (LQ), and that variants of PG, while minimax optimal, may require a number of reward queries exponential in $N$ to go beyond this support, regardless of the pre-training algorithm. To overcome this barrier, we study post-training with a process reward model, and demonstrate how PG variants in this setting avoid the curse of dimensionality in $N$ via dependence on a token-level LQ. Along the way, we prove that under the margin condition, SGD with adaptive learning rate (LR) achieves a near optimal test error for statistical learning, and PG with adaptive LR achieves a near optimal number of mistakes for online learning while being computationally efficient whenever possible, both of which may be of independent interest.

Related papers

• How Neural Reward Models Learn Features for Policy Optimization: A Single-Index Analysis [53.063298916923976]
  We study the feedback in a Gaussian single-index model with $r*(x) = *(langle *, xrangle)$ and $x sim N(0, I_d)$.<n>We analyze a two-stage neural reward model that first learns the hidden direction $*$ from reward-weighted samples and then fits the readout layer by weighted ridge regression.
  arXiv  Detail & Related papers  (2026-05-23T22:00:38Z)
• 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)
• Breaking the Bias Barrier in Concave Multi-Objective Reinforcement Learning [46.80389197344682]
  We develop a Natural Policy Gradient algorithm equipped with a multi-level Monte Carlo estimator.<n>We prove that this approach achieves the optimal $widetildemathcalO(-2)$ sample complexity for computing an $$-optimal policy.
  arXiv  Detail & Related papers  (2026-03-09T15:49:10Z)
• Transfer Q Star: Principled Decoding for LLM Alignment [105.89114186982972]
  Transfer $Q*$ estimates the optimal value function for a target reward $r$ through a baseline model. Our approach significantly reduces the sub-optimality gap observed in prior SoTA methods.
  arXiv  Detail & Related papers  (2024-05-30T21:36:12Z)
• A Finite Sample Complexity Bound for Distributionally Robust Q-learning [17.96094201655567]
  We consider a reinforcement learning setting in which the deployment environment is different from the training environment. Applying a robust Markov decision processes formulation, we extend the distributionally robust $Q$-learning framework studied in Liu et al. This is the first sample complexity result for the model-free robust RL problem.
  arXiv  Detail & Related papers  (2023-02-26T01:15:32Z)
• Best Policy Identification in Linear MDPs [70.57916977441262]
  We investigate the problem of best identification in discounted linear Markov+Delta Decision in the fixed confidence setting under a generative model. The lower bound as the solution of an intricate non- optimization program can be used as the starting point to devise such algorithms.
  arXiv  Detail & Related papers  (2022-08-11T04:12:50Z)
• Overcoming the Long Horizon Barrier for Sample-Efficient Reinforcement Learning with Latent Low-Rank Structure [9.759209713196718]
  We consider a class of MDPs for which the associated optimal $Q*$ function is low rank, where the latent features are unknown. We show that under stronger low rank structural assumptions, given access to a generative model, Low Rank Monte Carlo Policy Iteration (LR-MCPI) and Low Rank Empirical Value Iteration (LR-EVI) achieve the desired sample complexity of $tildeOleft((|S|+|A|)mathrmpoly(d,H)/epsilon2right)$ for a rank
  arXiv  Detail & Related papers  (2022-06-07T20:39:51Z)
• KL-Entropy-Regularized RL with a Generative Model is Minimax Optimal [70.15267479220691]
  We consider and analyze the sample complexity of model reinforcement learning with a generative variance-free model. Our analysis shows that it is nearly minimax-optimal for finding an $varepsilon$-optimal policy when $varepsilon$ is sufficiently small.
  arXiv  Detail & Related papers  (2022-05-27T19:39:24Z)
• Breaking the Sample Complexity Barrier to Regret-Optimal Model-Free Reinforcement Learning [52.76230802067506]
  A novel model-free algorithm is proposed to minimize regret in episodic reinforcement learning. The proposed algorithm employs an em early-settled reference update rule, with the aid of two Q-learning sequences. The design principle of our early-settled variance reduction method might be of independent interest to other RL settings.
  arXiv  Detail & Related papers  (2021-10-09T21:13:48Z)
• Sample-Efficient Reinforcement Learning for Linearly-Parameterized MDPs with a Generative Model [3.749193647980305]
  This paper considers a Markov decision process (MDP) that admits a set of state-action features. We show that a model-based approach (resp.$$Q-learning) provably learns an $varepsilon$-optimal policy with high probability.
  arXiv  Detail & Related papers  (2021-05-28T17:49:39Z)
• Model-Based Multi-Agent RL in Zero-Sum Markov Games with Near-Optimal Sample Complexity [67.02490430380415]
  We show that model-based MARL achieves a sample complexity of $tilde O(|S||B|(gamma)-3epsilon-2)$ for finding the Nash equilibrium (NE) value up to some $epsilon$ error. We also show that such a sample bound is minimax-optimal (up to logarithmic factors) if the algorithm is reward-agnostic, where the algorithm queries state transition samples without reward knowledge.
  arXiv  Detail & Related papers  (2020-07-15T03:25:24Z)
• Breaking the Sample Size Barrier in Model-Based Reinforcement Learning with a Generative Model [50.38446482252857]
  This paper is concerned with the sample efficiency of reinforcement learning, assuming access to a generative model (or simulator) We first consider $gamma$-discounted infinite-horizon Markov decision processes (MDPs) with state space $mathcalS$ and action space $mathcalA$. We prove that a plain model-based planning algorithm suffices to achieve minimax-optimal sample complexity given any target accuracy level.
  arXiv  Detail & Related papers  (2020-05-26T17:53:18Z)

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.