Young J. Kim

Hai Pham, Young Jin Kim, Subhabrata Mukherjee et al. · cmu, microsoft-research

Mixture-of-experts (MoE) architecture has been proven a powerful method for diverse tasks in training deep models in many applications. However, current MoE implementations are task agnostic, treating all tokens from different tasks in the same manner. In this work, we instead design a novel method that incorporates task information into MoE models at different granular levels with shared dynamic task-based adapters. Our experiments and analysis show the advantages of our approaches over the dense and canonical MoE models on multi-task multilingual machine translations. With task-specific adapters, our models can additionally generalize to new tasks efficiently.

Young Jin Kim, Raffy Fahim, Hany Hassan Awadalla · microsoft-research

Large Mixture of Experts (MoE) models could achieve state-of-the-art quality on various language tasks, including machine translation task, thanks to the efficient model scaling capability with expert parallelism. However, it has brought a fundamental issue of larger memory consumption and increased memory bandwidth bottleneck at deployment time. In this paper, we propose Mixture of Quantized Experts (MoQE) which is a simple weight-only quantization method applying ultra low-bit down to 2-bit quantizations only to expert weights for mitigating the increased memory and latency issues of MoE models. We show that low-bit quantization together with the MoE architecture delivers a reliable model performance while reducing the memory size significantly even without any additional training in most cases. In particular, expert layers in MoE models are much more robust to the quantization than conventional feedforward networks (FFN) layers. In our comprehensive analysis, we show that MoE models with 2-bit expert weights can deliver better model performance than the dense model trained on the same dataset. As a result of low-bit quantization, we show the model size can be reduced by 79.6% of the original half precision floating point (fp16) MoE model. Combined with an optimized GPU runtime implementation, it also achieves 1.24X speed-up on A100 GPUs.

Jongwoo Ko, Sara Abdali, Young Jin Kim et al.

On-policy distillation is pivotal for transferring reasoning capabilities to capacity-constrained models, yet remains prone to instability and negative transfer. We show that on-policy distillation can be interpreted, both theoretically and empirically, as a form of policy optimization, where the teacher-student log-likelihood ratio acts as a token reward. From this insight, we introduce REOPOLD (Relaxed On-Policy Distillation) a framework that stabilizes optimization by relaxing the strict imitation constraints of standard on-policy distillation. Specifically, REOPOLD temperately and selectively leverages rewards from the teacher through mixture-based reward clipping, entropy-based token-level dynamic sampling, and a unified exploration-to-refinement training strategy. Empirically, REOPOLD surpasses its baselines with superior sample efficiency during training and enhanced test-time scaling at inference, across mathematical, visual, and agentic tool-use reasoning tasks. Specifically, REOPOLD outperforms recent RL approaches achieving 6.7~12x greater sample efficiency and enables a 7B student to match a 32B teacher in visual reasoning with a ~3.32x inference speedup.

Minyoung Kim, Young J. Kim

We propose a 3D face generative model with local weights to increase the model's variations and expressiveness. The proposed model allows partial manipulation of the face while still learning the whole face mesh. For this purpose, we address an effective way to extract local facial features from the entire data and explore a way to manipulate them during a holistic generation. First, we factorize the latent space of the whole face to the subspace indicating different parts of the face. In addition, local weights generated by non-negative matrix factorization are applied to the factorized latent space so that the decomposed part space is semantically meaningful. We experiment with our model and observe that effective facial part manipulation is possible and that the model's expressiveness is improved.

Heajung Min, Kyung Min Han, Young J. Kim

Probabilistic volumetric mapping (PVM) represents a 3D environmental map for an autonomous robotic navigational task. A popular implementation such as Octomap is widely used in the robotics community for such a purpose. The Octomap relies on octree to represent a PVM and its main bottleneck lies in massive ray-shooting to determine the occupancy of the underlying volumetric voxel grids. In this paper, we propose GPU-based ray shooting to drastically improve the ray shooting performance in Octomap. Our main idea is based on the use of recent ray-tracing RTX GPU, mainly designed for real-time photo-realistic computer graphics and the accompanying graphics API, known as DXR. Our ray-shooting first maps leaf-level voxels in the given octree to a set of axis-aligned bounding boxes (AABBs) and employ massively parallel ray shooting on them using GPUs to find free and occupied voxels. These are fed back into CPU to update the voxel occupancy and restructure the octree. In our experiments, we have observed more than three-orders-of-magnitude performance improvement in terms of ray shooting using ray-tracing RTX GPU over a state-of-the-art Octomap CPU implementation, where the benchmarking environments consist of more than 77K points and 25K~34K voxel grids.

Daeun Song, Pierre Fernbach, Thomas Flayols et al.

One challenge of legged locomotion on uneven terrains is to deal with both the discrete problem of selecting a contact surface for each footstep and the continuous problem of placing each footstep on the selected surface. Consequently, footstep planning can be addressed with a Mixed Integer Program (MIP), an elegant but computationally-demanding method, which can make it unsuitable for online planning. We reformulate the MIP into a cardinality problem, then approximate it as a computationally efficient l1-norm minimisation, called SL1M. Moreover, we improve the performance and convergence of SL1M by combining it with a sampling-based root trajectory planner to prune irrelevant surface candidates. Our tests on the humanoid Talos in four representative scenarios show that SL1M always converges faster than MIP. For scenarios when the combinatorial complexity is small (< 10 surfaces per step), SL1M converges at least two times faster than MIP with no need for pruning. In more complex cases, SL1M converges up to 100 times faster than MIP with the help of pruning. Moreover, pruning can also improve the MIP computation time. The versatility of the framework is shown with additional tests on the quadruped robot ANYmal.

Jisu Kim, Young J. Kim

In this paper, we propose a novel penetration metric, called deformable penetration depth PDd, to define a measure of inter-penetration between two linearly deforming tetrahedra using the object norm. First of all, we show that a distance metric for a tetrahedron deforming between two configurations can be found in closed form based on object norm. Then, we show that the PDd between an intersecting pair of static and deforming tetrahedra can be found by solving a quadratic programming (QP) problem in terms of the distance metric with non-penetration constraints. We also show that the PDd between two, intersected, deforming tetrahedra can be found by solving a similar QP problem under some assumption on penetrating directions, and it can be also accelerated by an order of magnitude using pre-calculated penetration direction. We have implemented our algorithm on a standard PC platform using an off-the-shelf QP optimizer, and experimentally show that both the static/deformable and deformable/deformable tetrahedra cases can be solvable in from a few to tens of milliseconds. Finally, we demonstrate that our penetration metric is three-times smaller (or tighter) than the classical, rigid penetration depth metric in our experiments.