Behavior Cloning is Not All You Need: The Optimality of On-Policy Distillation for Noisy Expert Feedback
Authors: Ved Sriraman, Peihan Liu, Daniel Hsu, Adam Block
Organizations: Department of Computer Science, Columbia University · Department of Electrical Engineering, Columbia University
Abstract
Imitation Learning is a natural framework for learning in sequential decision-making systems and has emerged as the dominant paradigm through which we understand language model training. A central puzzle is that, while in theory offline IL can be horizon-free and optimal, in practice online methods such as on-policy distillation often outperform offline methods such as supervised fine-tuning. We propose a noisy expert model to explain this gap, in which the learner only has access to a noisy version of the expert's policy, but wishes to compete against the reward achieved by a clean expert, motivated by the fact that in many applications, e.g. training language models to perform long chains of thought, the expert is often imperfect. In this setting, we show a sharp separation between offline and online IL. Offline learning from noisy trajectories is fundamentally hard: to compete with the clean expert, the sample complexity must grow exponentially, in contradistinction to the clean expert setting where no explicit horizon dependence exists. In contrast, we prove that online interaction with the noisy expert via a novel variant of OPD enables polynomial dependence on the horizon in general. We further show that, under a natural condition on the expert noise distribution, which we show to be necessary for any horizon-free sample complexity, one can obtain such a guarantee, although our proposed algorithm sacrifices statistical efficiency in its dependence on the size of the policy class. Our analysis leads to an alternative loss function that is commonly considered empirically for LM training. We further provide algorithms and lower bounds, and extend our results to the more realistic setting of unknown corruption when the clean expert is deterministic, thereby providing a theoretical foundation for why OPD can outperform SFT when training language models from imperfect teachers.
Online imitation learning (IL), particularly on-policy distillation, has emerged as a strong LLM post-training approach, often outperforming offline supervised fine-tuning (SFT). Yet a principled understanding of when and why online interaction helps remains unclear. In this work, we challenge the view that error accumulation is the main source of online IL's advantage, and instead show that the benefits of online interaction depend critically on whether the setting is realizable, i.e., whether the student policy class can represent the expert policy. Under realizability, we empirically find that offline IL already matches expert performance. In contrast, in non-realizable (misspecified) settings, we prove that offline IL encounters an information-theoretic bottleneck even when horizon H=1, and propose a structural characterization of misspecification relative to the reward, under which online IL provably achieves high performance despite a large distributional mismatch between the expert and student policies.
Imitation learning (IL)---training an agent to replicate expert behavior from demonstrations---underpins applications from robotics to language model training. Standard approaches such as Behavior Cloning (BC) are known to suffer from compounding errors and performance plateaus, particularly when the learner cannot perfectly represent the expert's policy (as is typical, e.g., in distillation). Two interventions are widely understood empirically to improve performance: querying the expert interactively along the learner's own trajectories, and using value function estimation en route to generating a policy rather than directly fitting the expert's full action distribution. We investigate the nature of these improvements and their potentially surprising interplay. Our main finding is that expert interaction relaxes the representational demands on the learner: one only needs a model capable of realizing the expert's value function, bypassing the (often stricter) requirement of realizing the expert's policy itself. Concretely, we introduce OVI, an interactive on-policy IL algorithm that is statistically efficient whenever the learner can represent the expert's value function and computationally efficient given access to a linear maximization oracle. We complement this with a negative result showing that interaction is necessary. Namely, without stronger assumptions beyond expert-value realizability alone, any offline IL algorithm must scale with the complexity of the expert policy class. Our findings bear out empirically. OVI outperforms offline policy-based (BC), interactive policy-based (DAgger), and offline value-based IL methods, with the largest gains when the learner network is substantially less expressive than the expert's.
Prior work on imitation learning from suboptimal demonstrations typically relies on compressed supervision signals such as confidence estimates, discriminator scores, or importance weights. These scalar signals are inherently limited, as they cannot explicitly express intermediate reasoning about task progress, failure modes, or corrective actions. We propose a language-critique framework for imitation learning from suboptimal demonstrations that instead leverages natural language as a structured supervision signal, avoiding the collapse of expressive feedback into scalars. Our method first constructs language labels from demonstrations that explicitly describe current progress, identify suboptimal behaviors, and provide fine-grained corrective guidance. We then introduce a language-critique loss that directly trains policies using these structured signals without reducing them to scalars, and instantiate it for both behavior cloning and diffusion policies, yielding LC-BC and LC-DP. We further provide a theoretical result showing that the proposed objective upper-bounds the expert performance gap under standard assumptions. Empirically, we evaluate on diverse continuous control tasks spanning navigation, manipulation, and gameplay, where our methods consistently outperform strong imitation learning and offline reinforcement learning baselines. These results demonstrate that language can serve as a powerful and structured form of supervision for learning robust policies from suboptimal data.