Results suggest that geometric regularization of a recurrent bottleneck can act as a useful training-time prior without adding any inference overhead to the inference-time architecture.
Abstract
Regularization can improve generalization by constraining how a model uses its internal representation. In this paper, we study whether algebraic fixed-point constraints applied to the final LSTM hidden state during training can reorganize the recurrent latent space and improve held-out frame prediction. Rather than modifying the inference-time architecture, we introduce four training-time operators, GlobalHouseholder (reflection), GlobalGivens (rotation), Composition, and Lie Algebra, that bias the hidden state toward geometrically structured regions without changing the decoder pathway. Experiments across four datasets (indoor robot sequences, KITTI driving, Flying Shapes 2D, and Moving 3D Shapes) show that lightweight constraints consistently improve prediction on structured scenes, with GlobalGivens achieving up to +1.04 dB PSNR and −11.3% MAE over the unconstrained baseline on held-out Indoor sequences. The latent analysis reveals that the operators that generalize best are not those that compress the representation most aggressively but those that redistribute latent energy while preserving broad dimensional participation. Lie Algebra, despite collapsing activation variance by 81–96%, degrades under latent perturbation and does not match the lighter operators on structured datasets identifying over-constraint as a clear failure mode. These results suggest that geometric regularization of a recurrent bottleneck can act as a useful training-time prior without adding any inference overhead.
Reinforcement learning with verifiable rewards (RLVR) improves the reasoning ability of vision-language models (VLMs), and diversifying the rollouts within each optimization group amplifies its gains. Existing approaches diversify through decoding temperature or pixel-space image distortion; we ask whether the perturbation belongs in the model's latent space instead. We introduce Noise-Contrastive GRPO (NC-GRPO), which injects scale-calibrated Gaussian noise into the last hidden layer of the prompt-encoding pass for half of each rollout group, branching those rollouts from a displaced departure state. Branches that reach the answer despite the displacement are reinforced over those derailed by it, converting sensitivity at the branch point into policy-gradient signal; the objective, reward, and inference protocol are untouched. On Qwen2.5-VL-7B trained on Geometry3K, NC-GRPO significantly improves out-of-domain mathematical reasoning over vanilla GRPO across five held-out benchmarks (pooled McNemar $p \le 0.001$) while also improving in-domain accuracy and hallucination robustness -- the latter an axis on which image-space noise regresses even while posting a larger OOD average on perception-heavy benchmarks. Mechanism ablations indicate that independent stochastic diversity, not noise budget or direction, is the active ingredient, and a noise-scale study exposes a dial between reasoning specialization and general capability. NC-GRPO is designed to be modality-agnostic and integrates into a standard RLVR pipeline as a ~50-line change to the inference engine.
Michael M. Jerge, Joseph Pelczar, J. Downes· 0 citations
A latent world model trains its decoder on latents anchored to observations, then deploys it on the model's own free-running rollout, hundreds of steps past the last observation. Rollout-Decoded Reconstruction (RDR) closes this gap with a single loss term that free-runs the model during training exactly as evaluation will, decodes every rollout latent, and penalizes reconstruction error against ground truth. The term adds no parameters, costs training-time compute only, and reduces to the standard objective at weight zero, so every comparison in this paper is a one-flag A/B. On the chaotic Kuramoto-Sivashinsky equation, RDR raises valid prediction time (the time to first crossing of normalized error 0.5) from $3.87 \pm 0.23$ to $6.97 \pm 0.42$ time units at an identical 193,568 parameters, a $1.80\times$ improvement confirmed on seeds never used in selection and holding in 10 of 10 preregistered configurations at ratios of 1.71-2.50$\times$. The results come from a single system; a sweep in which the advantage grows with latent width is descriptive, and control experiments on two classic tasks are preliminary.
Vision-Language Models (VLMs) exhibit strong zero-shot capabilities, making them an attractive solution for continual learning across diverse tasks. However, during continual adaptation, both catastrophic forgetting and zero-shot degradation occur, severely degrading performance. In this paper, we introduce TASSO, a new paradigm that efficiently preserves the latent space geometry while ensuring network plasticity. We achieve this with two complementary techniques: subspace learning and geometry-aware knowledge distillation. Specifically, we first learn a sequence of task-specific low-rank projectors, which we use to project the latent representations before optimizing cross-entropy. Secondly, we employ a geodesic-distance-based loss that distills knowledge from the previous-task model while effectively preserving the latent space geometry. These design choices not only avoid unnecessary parameter updates along the full embedding dimensions but also improve learning by focusing on task-specific manifolds. Moreover, the geometry-aware distillation provides strong regularization and significantly reduces both catastrophic forgetting and zero-shot degradation throughout the continual learning sequence. Experimental results with the CLIP vision language model in the multi-domain task incremental and class incremental learning benchmarks demonstrate clear improvements over state-of-the-art methods in mitigating forgetting and preserving zero-shot capabilities.
Changming Sun, Francesco Barbato, Matteo Caligiuri et al.· 0 citations
The Interval Denoiser, a theoretically rigorous framework for latent-free generation, derived directly from the flow matching ODE, establishes an exact analytical mapping for intermediate trajectory states and is shown to reside on a low-dimensional manifold across any time interval.
A.M. Zaytsev, Dmitry Baranchuk, Alexander Korotin et al.· 0 citations
Chunked Muon (CMuon) is introduced, a simple yet highly effective strategy that partitions these matrices into independent sub-components prior to orthogonalization, effectively overcoming the late-stage convergence plateaus of vanilla Muon.
The Robust-WAM is a general post-training method for video-generation-based WAMs that preserves the VAE-based generative path and adds a lightweight semantic foresight alignment objective on the action stream to retain the large-scale VGM pretraining while grounding actions in appearance-invariant dynamics.
Haodong Yan, Junfeng Li, Junjie He et al.· 0 citations