TokenRatio: Principled Token-Level Preference Optimization via Ratio Matching

Direct Preference Optimization (DPO) is a widely used RL-free method for aligning language models from pairwise preferences, but it models preferences over full sequences even though generation is driven by per-token decisions. Existing token-level extensions typically decompose a sequence-level Bradley-Terry objective across timesteps, leaving per-prefix (state-wise) optimality implicit. We study how to recover token-level preference optimality using only standard sequence-level pairwise comparisons. We introduce Token-level Bregman Preference Optimization (TBPO), which posits a token-level Bradley-Terry preference model over next-token actions conditioned on the prefix, and derive a Bregman-divergence density-ratio matching objective that generalizes the logistic/DPO loss while preserving the optimal policy induced by the token-level model and maintaining DPO-like simplicity. We introduce two instantiations: TBPO-Q, which explicitly learns a lightweight state baseline, and TBPO-A, which removes the baseline through advantage normalization. Across instruction following, helpfulness/harmlessness, and summarization benchmarks, TBPO improves alignment quality and training stability and increases output diversity relative to strong sequence-level and token-level baselines.

Spectral Flattening Is All Muon Needs: How Orthogonalization Controls Learning Rate and Convergence

Muon orthogonalizes the momentum buffer before each update, replacing its singular values with ones via Newton-Schulz iterations. This simple change lets Muon tolerate far larger learning rates and converge faster than other optimizers, but why? We show that the mechanism is spectral flattening, and develop two results around it. First, we prove that Muon's maximal stable step size scales with the average singular value of the gradient rather than the largest, which bottlenecks standard gradient descent. Second, we recast Muon as a preconditioned gradient method and show, under a Kronecker-factored curvature model, that it improves the effective convergence factor, with the improvement controlled by the spectrum of the gradient covariance. Extensive experiments validate both results: Muon remains stable at learning rates that cause SGD to diverge within the first few iterations, and reaches accuracy milestones several epochs earlier even at identical step sizes. Taken together, our results offer a principled, geometric explanation for Muon's empirical success.

BSO: Safety Alignment Is Density Ratio Matching

Aligning language models for both helpfulness and safety typically requires complex pipelines-separate reward and cost models, online reinforcement learning, and primal-dual updates. Recent direct preference optimization approaches simplify training but incorporate safety through ad-hoc modifications such as multi-stage procedures or heuristic margin terms, lacking a principled derivation. We show that the likelihood ratio of the optimal safe policy admits a closed-form decomposition that reduces safety alignment to a density ratio matching problem. Minimizing Bregman divergences between the data and model ratios yields Bregman Safety Optimization (BSO), a family of single-stage loss functions, each induced by a convex generator, that provably recover the optimal safe policy. BSO is both general and simple: it requires no auxiliary models, introduces only one hyperparameter beyond standard preference optimization, and recovers existing safety-aware methods as special cases. Experiments across safety alignment benchmarks show that BSO consistently improves the safety-helpfulness trade-off.

Establishing Robust Retinal Eye Tracking: A Weakly Supervised Algorithmic Framework

Retinal image-based eye tracking is widely used in ophthalmic imaging and vision science, and is a promising path to deliver higher gaze accuracy than the pupil- and cornea-based approaches commonly used in modern AR/VR devices. Nevertheless, existing retinal tracking algorithms still primarily rely on classical template-matching registration, which can be insufficiently robust to retinal feature variability and real-world imaging conditions. In this work, we propose a novel weakly-supervised, learning-based framework for robust retinal eye tracking. Initial studies demonstrate high accuracy, achieving the 95th-percentile gaze error < 0.45 deg across a cohort of 6 participants.

PhysGraph: Physically-Grounded Graph-Transformer Policies for Bimanual Dexterous Hand-Tool-Object Manipulation

Bimanual dexterous manipulation for tool use remains a formidable challenge in robotics due to the high-dimensional state space and complicated contact dynamics. Existing methods naively represent the entire system state as a single configuration vector, disregarding the rich structural and topological information inherent to articulated hands. We present PhysGraph, a physically-grounded graph transformer policy designed explicitly for challenging bimanual hand-tool-object manipulation. Unlike prior works, we represent the bimanual system as a kinematic graph and introduce per-link tokenization to preserve fine-grained local state information. We propose a physically-grounded bias generator that injects structural priors directly into the attention mechanism, including kinematic spatial distance, dynamic contact states, geometric proximity, and anatomical properties. This allows the policy to explicitly reason about physical interactions rather than learning them implicitly from sparse rewards. Extensive experiments show that PhysGraph significantly outperforms baseline - ManipTrans in manipulation precision and task success rates while using only 51% of the parameters of ManipTrans. Furthermore, the inherent topological flexibility of our architecture shows qualitative zero-shot transfer to unseen tool/object geometries, and is sufficiently general to be trained on three robotic hands (Shadow, Allegro, Inspire).

HumanOrbit: 3D Human Reconstruction as 360° Orbit Generation

We present a method for generating a full 360° orbit video around a person from a single input image. Existing methods typically adapt image-based diffusion models for multi-view synthesis, but yield inconsistent results across views and with the original identity. In contrast, recent video diffusion models have demonstrated their ability in generating photorealistic results that align well with the given prompts. Inspired by these results, we propose HumanOrbit, a video diffusion model for multi-view human image generation. Our approach enables the model to synthesize continuous camera rotations around the subject, producing geometrically consistent novel views while preserving the appearance and identity of the person. Using the generated multi-view frames, we further propose a reconstruction pipeline that recovers a textured mesh of the subject. Experimental results validate the effectiveness of HumanOrbit for multi-view image generation and that the reconstructed 3D models exhibit superior completeness and fidelity compared to those from state-of-the-art baselines.

Consistency-Preserving Diverse Video Generation

Text-to-video generation is expensive, so only a few samples are typically produced per prompt. In this low-sample regime, maximizing the value of each batch requires high cross-video diversity. Recent methods improve diversity for image generation, but for videos they often degrade within-video temporal consistency and require costly backpropagation through a video decoder. We propose a joint-sampling framework for flow-matching video generators that improves batch diversity while preserving temporal consistency. Our approach applies diversity-driven updates and then removes only the components that would decrease a temporal-consistency objective. To avoid image-space gradients, we compute both objectives with lightweight latent-space models, avoiding video decoding and decoder backpropagation. Experiments on a state-of-the-art text-to-video flow-matching model show diversity comparable to strong joint-sampling baselines while substantially improving temporal consistency and color naturalness. Code will be released.

Learnable Multi-level Discrete Wavelet Transforms for 3D Gaussian Splatting Frequency Modulation

3D Gaussian Splatting (3DGS) has emerged as a powerful approach for novel view synthesis. However, the number of Gaussian primitives often grows substantially during training as finer scene details are reconstructed, leading to increased memory and storage costs. Recent coarse-to-fine strategies regulate Gaussian growth by modulating the frequency content of the ground-truth images. In particular, AutoOpti3DGS employs the learnable Discrete Wavelet Transform (DWT) to enable data-adaptive frequency modulation. Nevertheless, its modulation depth is limited by the 1-level DWT, and jointly optimizing wavelet regularization with 3D reconstruction introduces gradient competition that promotes excessive Gaussian densification. In this paper, we propose a multi-level DWT-based frequency modulation framework for 3DGS. By recursively decomposing the low-frequency subband, we construct a deeper curriculum that provides progressively coarser supervision during early training, consistently reducing Gaussian counts. Furthermore, we show that the modulation can be performed using only a single scaling parameter, rather than learning the full 2-tap high-pass filter. Experimental results on standard benchmarks demonstrate that our method further reduces Gaussian counts while maintaining competitive rendering quality.

CTPD: Cross Tokenizer Preference Distillation

While knowledge distillation has seen widespread use in pre-training and instruction tuning, its application to aligning language models with human preferences remains underexplored, particularly in the more realistic cross-tokenizer setting. The incompatibility of tokenization schemes between teacher and student models has largely prevented fine-grained, white-box distillation of preference information. To address this gap, we propose Cross-Tokenizer Preference Distillation (CTPD), the first unified framework for transferring human-aligned behavior between models with heterogeneous tokenizers. CTPD introduces three key innovations: (1) Aligned Span Projection, which maps teacher and student tokens to shared character-level spans for precise supervision transfer; (2) a cross-tokenizer adaptation of Token-level Importance Sampling (TIS-DPO) for improved credit assignment; and (3) a Teacher-Anchored Reference, allowing the student to directly leverage the teacher's preferences in a DPO-style objective. Our theoretical analysis grounds CTPD in importance sampling, and experiments across multiple benchmarks confirm its effectiveness, with significant performance gains over existing methods. These results establish CTPD as a practical and general solution for preference distillation across diverse tokenization schemes, opening the door to more accessible and efficient alignment of language models.

Depth Any Panoramas: A Foundation Model for Panoramic Depth Estimation

In this work, we present a panoramic metric depth foundation model that generalizes across diverse scene distances. We explore a data-in-the-loop paradigm from the view of both data construction and framework design. We collect a large-scale dataset by combining public datasets, high-quality synthetic data from our UE5 simulator and text-to-image models, and real panoramic images from the web. To reduce domain gaps between indoor/outdoor and synthetic/real data, we introduce a three-stage pseudo-label curation pipeline to generate reliable ground truth for unlabeled images. For the model, we adopt DINOv3-Large as the backbone for its strong pre-trained generalization, and introduce a plug-and-play range mask head, sharpness-centric optimization, and geometry-centric optimization to improve robustness to varying distances and enforce geometric consistency across views. Experiments on multiple benchmarks (e.g., Stanford2D3D, Matterport3D, and Deep360) demonstrate strong performance and zero-shot generalization, with particularly robust and stable metric predictions in diverse real-world scenes. The project page can be found at: \href{https://insta360-research-team.github.io/DAP_website/} {https://insta360-research-team.github.io/DAP\_website/}