reinforcement learning practitioners

Abstract: In this work, we propose a unified framework for training algorithms across different stages of modern large language model development, based on the perspective of distribution matching. We show that a broad class of language model training algorithms can be interpreted as minimizing a certain divergence between the current model and a target distribution, and that characterizing the properties of this target distribution provides substantial insight into the behavior of the corresponding training algorithms. By analyzing specific forms of the target distribution, we establish that On-Policy Distillation (OPD) enjoys an advantage over classical Knowledge Distillation (KD) when the current model distribution is close to the target distribution, while exhibiting a disadvantage when the two distributions are far apart. Furthermore, when the target distribution itself depends on the model being optimized, we show that the accumulation of estimation variance can induce a deterministic degradation of the target distribution. These results provide a theoretical explanation, at least in part, for the long-term behavior of existing Reinforcement Learning (RL) algorithms. With this intuition we further designed Conditional KL (CKL) regularization for mitigating the degradation of the target distribution. CKL can effectively preserve mode diversity while relaxing the constraint imposed by classical KL regularization on reward improvements. We believe this theoretical perspective can bring some new understanding of language model training algorithms and provide useful guidance for future research.

Abstract: Before a new policy replaces a deployed one, we often want a certificate, computed from logged episodes, that it is better. We ask how many episodes such a certificate needs when the logging policy is known but history-dependent, and the error probability must stay below $δ$ in every environment consistent with the logs. The answer is an exponent $K^\star$: the smallest KL divergence from the logged law to a law with the same logged action probabilities under which the target does not improve. No sound certifier stops in fewer than $\log(1/δ)/K^\star$ expected episodes, at any $δ$, and whenever $0<K^\star<\infty$ a plug-in Kelly certifier attains this rate to first order as $δ\to0$. Coverage, computed by one backward induction, decides whether certification is impossible ($K^\star=0$), possible at a finite rate, or possible in a constant number of episodes ($K^\star=\infty$). By convex duality, the optimal certificate bets on the importance-weighted advantage and also places fair bets on the logged actions, so doubly robust control variates act as hedges rather than estimators. As a consequence, no choice of off-policy estimator changes $K^\star$, a coverage ceiling explains the curse of memory, and structural assumptions can be priced in exponent units. A cushion certifier extends the approach to trees too large to enumerate. In experiments the plug-in's median stopping time is 1.13-1.43x the lower bound on memory tasks, 1.13-1.51x on Tiger ($H\le5$) and 1.17-1.64x on random trees. The best scalar betting baselines need up to 2.65x (memory) and 4.1x (Tiger) the plug-in's episodes; only on a few shallow random trees is a doubly robust bet faster, by at most 18%.

Abstract: We study the minimax estimation of the pseudo-spectral gap $γ_{ps}$ of a Markov chain from the observation of a single trajectory. Extending earlier works on finite-state space chains by Wolfer and Kontorovich (2024), we provide an estimator in the case of a compact, continuous state-space. Our estimator appears to be minimax-optimal on the class of chains with a transition operator with Lipschitz density, up to logarithmic terms. Our estimator, combined with Paulin's version of Bernstein inequality, allows to derive empirical concentration inequalities for Markov chains. This unlocks many applications, from confidence intervals, to active learning and generalization bounds. In particular, for rested bandits with Markov rewards, we provide a version of UCB with a regret bound similar to the one in the iid case, without assuming any knowledge on the pseudo-spectral gap of the chains.

Abstract: Every recurrence of a looped language model adds cost in training, decoding, prefill, and reinforcement learning (RL). The closer recurrent states get to fixed points, the less the path to them matters. This enables truncated backpropagation in training; terminal key-value (KV) sharing for decoding with almost no loss in accuracy; a distilled student that prefills up to 1.79x faster; and RL updates that compute gradients from saved rollout states, 2x faster than backpropagating through the replayed trajectory. We therefore improve the two components of training that shape these fixed points: the depth prior and input injection. Fixed-depth training breaks KV sharing, and Huginn's broad depth prior supports sharing but dilutes supervision at the target depth more than sharing requires; we learn the prior from prediction feedback, with an entropy term that keeps it broad. Existing injection schemes let the state's component along the input amplify or cancel the injection; we remove this component with orthogonal injection. From 100M to 1.6B parameters, the learned prior and orthogonal injection lower perplexity at every scale relative to Huginn's prior and existing injection schemes, respectively. At 1.6B, the learned prior with a 3x smaller KV cache matches the downstream average of fixed-depth training with the full cache.

Abstract: Policy mirror descent (PMD) has a mature finite-time theory in discounted Markov decision processes (MDPs), but less is known in the average-reward setting, a more natural objective for many control applications. We give a finite-time, finite-sample analysis of PMD in ergodic average-reward MDPs built around a single master recursion that governs convergence for any critic, without external regularization. Its specializations yield linear rates for exact, inexact-tabular, and linear function approximation (LFA) updates, with a superlinear regime for exact PMD. We complement these convergence results with end-to-end sample complexities of order $t_{\mathrm{mix}}^3/\varepsilon^2$ in both tabular ($|S||A|$-dependent) and LFA ($d$-dependent) settings. Our LFA sample complexity sharpens the prior best $t_{\mathrm{mix}}^5$ mixing dependence to $t_{\mathrm{mix}}^3$, and matching information-theoretic lower bounds establish that the critic's $t_{\mathrm{mix}}^3/\varepsilon^2$ sample complexity is unimprovable in both settings.

Abstract: Stopgrads are widely used in training machine learning models, but stopgrads can alter the gradient, stationary points and convergence guarantees of the original objective, which can make stopgrad training theoretically ungrounded. We introduce a stopgrad regression principle, which identifies a general template for stopgrad objectives with a closed-form characterization of stationary points and their uniqueness, unifying stopgrad objectives for flow maps, reinforcement learning, and diffusion samplers. We provide theoretical grounding for optimizing stopgrad flow map objectives by showing their unique stationary point is the true flow map, and showing positive convergence results for Eulerian and Lagrangian objectives, including MeanFlow and improved MeanFlow. Remarkably, we show that under functional semi-gradient flow, the learned flow map has a closed-form expression composing the initial flow map and the true flow map. We additionally use our stopgrad regression principle to propose modified stopgrad placements for flow map objectives which reduce training memory by 2x.