NeurIPS 2026 论文列表

This paper concerns the problem of scaling agents to large action spaces demanding pretrained world knowledge. What type of representation pretraining would enable policies such as language models to generalize rewards to appropriate actions? Contrastive learning aligns similar inputs and otherwise enforces distancing of representations, while non-contrastive alternatives omit distancing terms and so learn anisotropic representations that collapse distinctive clusters. Though contrastive pretraining has been proven to promote supervised learning efficiency, non-contrastive pretraining nevertheless remains canonical for language agents. Here we examine the extent to which these pretraining types support on-policy finetuning, in which efficiency is coextensive with online performance. Broadly, we argue that contrastive pretraining optimizes efficient exploration of large spaces. We prove that anisotropic representations collapse distinct actions and necessitate complex policies, whereas contrastive pretraining retains distinctions and enables reward transfer through spectral regularization. We validate our conclusions empirically by providing spectral analyses of language data and simulations involving multimodal personalization policies as well as self-improving reasoning agents finetuned online using a policy gradient. Our latter simulations suggest contrastive pretraining for language agents exploring diverse reasoning paths.

Intrinsic rewards for exploration in reinforcement learning condition on different contexts: lifelong rewards score each transition against accumulated experience but ignore within-rollout redundancy; episodic rewards penalize intra-trajectory repetition but discard lifetime progress. Hybrid methods combine both signals through heuristic weights or require Gaussian-process dynamics that do not scale beyond low-dimensional state spaces. Trajectory-level information gain decomposes into per-step terms that condition on the replay buffer and rollout prefix simultaneously, but remains intractable for deep models. We derive the Conditional Information Gain (CIG) reward as a tractable surrogate: a log-determinant objective over an ensemble disagreement kernel whose Cholesky factorization yields causal per-step rewards that retain both conditioning sets while scaling to high-dimensional state spaces. We instantiate CIG in a model-based setting, where rollouts are short and within-rollout corrections remain largely unexplored. Across twelve tasks spanning discrete (MiniGrid) and continuous control (OGBench), in both clean and stochastic-distractor settings, CIG outperforms or matches prior exploration methods while remaining robust to stochastic distractors.

Recent work on recursive architectures has shown that tiny neural networks can be surprisingly powerful on structured reasoning tasks. The trick is to model reasoning trajectories with a latent dynamical system. We argue that the inference-time behaviour of these architectures is best understood as approximate inference over latent reasoning trajectories, with deterministic recursion as the one-particle, zero-noise limit. We make this view operational through guided stochastic exploration: stochastic perturbations of the reasoning dynamics propose neighbouring trajectories, and the model's existing early-stopping head reweights them online. The framework yields three label-free diagnostics: local stability, guide alignment, and cloud-token entropy. These predict, from inference traces alone, whether the procedure will help and which of its outputs to trust. On Sudoku-Extreme it lifts exact-solve accuracy from $85.9\%$ to $98.0\%$ without retraining; on Maze-Hard the diagnostics flag a misaligned guide, as validation performance later confirms. The same machinery thus characterises both when recursive reasoning has room to improve at the trajectory level and when the model's internal guide can recover it.

We study online learning in the adversarial injection model [GHMS23], where each instance is either drawn i.i.d. from an unknown distribution or is an arbitrary adversarial injection. The learner never knows which rounds are adversarial, but may abstain without penalty on adversarial injections. The learner's goal is to make few total misclassifications and few abstentions on i.i.d. instances. This model lies between classical statistical learning, governed by the VC dimension, and adversarial online learning, governed by the Littlestone dimension. Unlike prior work, which has largely focused on the realizable setting, we consider an agnostic formulation that allows arbitrary labels. We present an online algorithm that guarantees sublinear excess misclassification error and sublinear abstention error, against oblivious and adaptive adversaries, for function classes that admit bounded bracketing number witnessed by brackets of bounded VC dimension. Our algorithm is distribution-independent and does *not* require prior knowledge of the distribution. As corollaries, we obtain new learning guarantees for halfspaces (linear classifiers) and richer classes such as feed-forward neural networks with threshold activations.