Stable 4-bit floating-point (FP4) pretraining is difficult because the E2M1 payload represents only a narrow range of magnitudes. NVIDIA's Transformer Engine \nv{} recipe addresses this with current-tensor scaling, a randomized Hadamard transform (RHT), and bfloat16 (BF16) final layers, adding work outside the FP4 matrix multiplications. We instead pair E2M1 payloads with unsigned E5M3 (\ue{}) block scales. Their wider range permits periodic tensor scaling, while our recipe applies selective stochastic rounding to backward gradients, omits RHT, and uses FP4 in all eligible internal linears. We pretrain a Nemotron-H 8B model for nearly 190 billion tokens. Compared with Transformer Engine \nv{}, the proposed block-16 recipe finishes with lower final-window training loss and, under their respective quantized-inference policies, lower validation loss measured as held-out negative log-likelihood. Its quantized-inference downstream point estimates are also higher on all three reported aggregates. A native \nv{} execution ablation that jointly removes RHT and the BF16 final-block exemption increases measured model-body token throughput by 21.2\%. These results demonstrate end-to-end software-emulated \uefp{} pretraining with a simpler recipe and motivate native support for \ue{} block scaling.
Reinforcement learning with verifiable rewards (RLVR) has emerged as a powerful paradigm for large language model (LLM) post-training, but its reliance on coarse outcome rewards leads to limited guidance on intermediate reasoning processes. Existing approaches such as process reward modeling and on-policy distillation introduce additional constraints, such as reliance on a specialized reward model or assuming identical reasoning patterns between teacher and student. Nevertheless, we observe that once a reasoning process first goes wrong, evaluating the subsequent reasoning provides limited additional information, as it is already conditioned on an invalid prefix. Therefore, we propose Cliff, a reward shaping strategy that utilizes an off-the-shelf LLM as a teacher to identify the first mistake in each rollout. As a result, the rollout is naturally decomposed into two parts: a correct prefix and an incorrect suffix. Cliff then converts this signal into token-level advantages, assigning positive advantages for the correct prefix and negative feedback afterward. Experiments across 12 different scenarios demonstrate that Cliff consistently improves reasoning performance, outperforming on-policy distillation by 15% and standard GRPO by 7%, even with teachers of modest capability. Furthermore, we analyse the role of ``ground truth''in Cliff and investigate its training dynamics. These results establish Cliff as a simple, general and effective approach for improving RLVR with richer, fine-grained supervision.
Peixuan Han, Runnan Wang, Ketan Ramaneti et al.· 0 citations
Tabular foundation models (TFMs) learn to fill in tables the way language models fill in text, and tables are arguably the format in which most physical measurement arrives. Did they learn any physics in the process? They are Bayesian by construction, so the question is what their prior contains. We probe it directly, evaluating four of them (TabPFN-3, TabICLv2, TabDPT and Real-TabPFN-2.5) against six baselines on datasets sampled from 316 physical equations, in and out of domain. TFMs dominate, out of the box and after tuning. But we show that their prior can represent neither a noiseless mechanism nor physical units, which is why they interpolate physics without yet being able to act as physical models.
Low-rank adaptation (LoRA) is the standard way to fine-tune large models, yet when its two factors are trained independently, the update ignores the geometry of the low-rank weight change it induces. We introduce LoRA-TSD, an optimizer that treats every LoRA step as a tangent vector of the fixed-rank matrix manifold and takes the spectral-norm steepest-descent step of Muon inside that tangent space, mapping the result back to the factors through a retraction native to the LoRA parametrization. The step avoids expensive operations on full weight matrices, and its retraction is up to $2.8\times$ cheaper than the truncated-SVD retraction used by prior manifold methods. We prove that the Frobenius-norm version of our surrogate recovers LoRA-Pro, and we identify the tangent-projected gradient, the Riemannian gradient of the manifold, as the stationarity measure natural to LoRA training and computable from the factor gradients alone. Under this measure we give the first global convergence guarantees for both LoRA-Pro and LoRA-TSD, with rates that drive the factor-gradient norms to zero. Across six commonsense and natural-language-inference benchmarks with Llama-3.2-1B, Llama-3.1-8B and Qwen3-32B, LoRA-TSD outperforms every competing LoRA optimizer and stays robust to the adapter rank. Code is available at https://github.com/brain-lab-research/LoRA-TSD.
Dmitrii Andriianov, Andrey Veprikov, Aleksandr Beznosikov· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
Deploying 3D point cloud models on edge hardware such as the NVIDIA Jetson Orin Nano is severely constrained by compute and memory budgets. Existing compression methods require access to the model's original source code, rendering them inapplicable to the Open Neural Network Exchange (ONNX) binaries commonly distributed by vendors and model repositories. We present \textbf{H3DNAS}, a hardware-aware model compression framework that operates directly on ONNX computational graphs without requiring original source code, architecture class definition, or gradient access during search. H3DNAS makes three contributions: (1) a \textbf{Channel Dependency Graph (CDG)} that classifies ONNX operators into four constraint classes and formally establishes that the free parameter fraction $\rho_f$ is topological invariant, a provable compression ceiling computable in $\mathcal{O}(|V|+|E|)$; (2) a \textbf{Two-Stage Hierarchical Search} that prunes candidate architectures by $L_1$-importance channel selection, ranks them by output fidelity as a zero-shot label-free proxy, and applies GhostConv structural mutation to Pareto-optimal candidates; and (3) the \textbf{first source-code-free compression pipeline for 3D point cloud models}, operating entirely via ONNX graph surgery with no original architecture definition required. On ModelNet40, H3DNAS reduces the number of parameters in PointNet, PointNet++, and PointMLP by $65.5\%$, $43.2\%$, and $49.1\%$, respectively, while achieving $1.99\times$, $1.29\times$, and $1.67\times$ inference speedups with negligible loss in accuracy. The source code is publicly available\footnote{https://github.com/ClarityLab-Org/h3dnas}.
Leech-lattice vector quantization holds the strongest reported 2-bit quality under its own evaluation protocol. Its kernel decodes one shell; we found no implementation of the multi-shell decoder the rate requires. This paper supplies one and measures its serving cost for decode-phase GEMV at batch 1. First, a serving path for the full 301-class codebook: an offline expansion into GPU layouts and a fused dequantize-plus-matvec kernel reading them without warp divergence, verified against f64. Second, the in-VRAM rate is a design axis distinct from the on-disk rate. Four bit-exact layouts timed in one process show binary bit planes beating one-hot masks on size and speed at constant bandwidth (4.80 bits per weight, 2.15x FP16). Below 4.3 bits a second, irregular stream enters; at 3.6 the decode stops being shifts and masks. Third, deployed four-bit (AWQ) and two-bit (QTIP) GEMV kernels run in the same process. The trellis kernel reads 2.40x fewer bytes than our served layout and runs 2.27x faster at near-equal fractions of their byte bounds: the time gap tracks the traffic gap, the price of unfolding a codebook too large for a lookup table. Fourth, the validity envelope: the trellis kernel outruns our no-weights control, so our launch geometry sets that floor, and on a second memory hierarchy every lattice arm falls below FP16. With the output head held identical across arms, the kernel-and-format path gains 1.11x, 1.29x and 1.41x end to end at 4B, 8B and 14B; with an int8 output head the served 4B reaches 87.0 tok/s in 2.60 GB. The quality cost, 1.38x perplexity and 14.7 MMLU points at 4B, shrinks across the three sizes measured.
Planning a large data center is difficult because a facility big enough to matter changes the electricity prices it will pay. Those prices are set by market clearing, a constrained optimization problem solved anew in every operating condition. However, simulating the market tells a planner how a candidate plan performs but not how to improve it. Here we treat market clearing as a differentiable optimization layer: each forward pass solves the market, and reverse-mode automatic differentiation propagates the planning cost back through the cleared prices to the plan. After validating these gradients against finite differences, we apply them to a concrete problem: allocating 50 MW of data-center load across six candidate buses in two synthetic networks, under a fixed cost per active site, evaluated over 36 operating states. Judged against exhaustive enumeration of all site combinations, gradient optimization recovers the continuous allocations almost exactly, with worst-case objective gaps of 2.3\% and 8.5\% of the cost difference between the best and worst single site. Its one systematic error is instructive: near the costs at which a site should close, the smooth relaxation of the discrete site count shrinks the site rather than closing it, so discrete transitions arrive late. Differentiable market clearing thus turns market-aware planning into a problem gradients can search.
Luca Mungo, Maarten P. Scholl, Arnau Quera-Bofarull· 0 citations
Knowledge graphs have become an important source of structured knowledge for Web applications, including search, question answering, and recommender systems. In these applications, link prediction can serve either as a prediction task itself or as a means to enrich incomplete knowledge graphs for downstream tasks. Interestingly, different link prediction models, or even different training runs of the same model, can produce substantially different predictions for the same query. This suggests a variability in the capture of the underlying knowledge by models, thus raising a fundamental question: to what extent do different models capture complementary knowledge, and how much of this knowledge could be recovered by combining them? We propose to measure model complementarity through the performance of an oracle that, for each query, selects the best prediction among a considered set of models, hence providing an upper bound on the performance achievable through model combination. Across several architectures and benchmarks, we find a substantial gap between individual models and their oracle, revealing that different models capture complementary knowledge. Yet, this complementarity rapidly saturates as more models are added, leaving a persistent subset of queries unsolved even by a large number of models. These findings reveal both the potential of model complementarity and a fundamental limit to what current link prediction models can collectively recover; thereby highlighting the need for further research to build robust Web applications.
Guillaume M\'erou\'e, Fabien Gandon, Pierre Monnin· 0 citations
Simulation-based science often requires a distribution over simulator parameters whose push-forward reproduces a set of real observations: this is the source distribution estimation (SDE) problem. Existing methods fit the source against a likelihood surrogate trained once from a fixed proposal prior. Their objective is therefore stated only in terms of the surrogate instead of the true simulator, which may fail for inaccurate areas in parameter space where the surrogate was never trained. We instead solve SDE by expectation maximization: an E-step trains an amortized posterior on fresh simulations from the current source estimate, and an M-step refits the source to the average of that posterior over the observed data. We give two parameterizations, (1) separate source and posterior flows and (2) a single shared conditional flow. We evaluate our method on three benchmark tasks under both broad and misspecified initial priors. Both improve on existing fixed surrogate approaches and on iterated variants of each, most clearly on Lotka--Volterra, where no baseline falls below 0.96 data-space C2ST while our methods reach 0.64-0.68 in three of four initial-prior settings.
Machine-learnt corrections can complement numerical weather prediction only if they adapt to the evolving model state while preserving dynamical consistency and numerical stability. To test this within a global forecasting model, we couple the Met Office (UKMO) Unified Model (UM) with distributed RL agents through rank-local tensors. A DDPG actor shares weights across the 70 vertical model levels of each atmospheric column and applies bounded potential-temperature corrections to the model tendencies. Across ten nudged training forecasts, nudging calculations towards the UKMO operational analysis provides an immediate counterfactual target. The frozen policy is then evaluated in a non-nudged forecast for inference. The coupled workflow successfully completes training and remains numerically stable in the evaluated case. Relative to a matched native UM forecast at +6 h, the learnt policy reduces Z$_{500}$ MAE in four of six latitude bands, including reductions of 45.8% and 40.8% in the northern and southern tropics. MSLP error too decreases in three bands, with a maximum reduction of 27.3% at 0-30{\deg}N. This single-case experiment demonstrates significant promise and feasibility of distributed online learning followed by non-nudged inference, laying the groundwork for RL-based bias correction and parametrisations within operational systems.
Pritthijit Nath, Sebastian Schemm, Peter Haynes et al.· 0 citations
Diagnosing collective anomalies from urban trajectories is increasingly important for traffic governance, as it reveals what happened, who was involved, and where and when the event occurred. Existing detectors efficiently produce scores or labels, whereas vision--language pipelines provide richer semantics; neither couples verifiable diagnosis with low-latency monitoring. The central challenge is to recognize collective patterns and recover exact event details from the source trajectories without running the full diagnostic pipeline for every monitored window. We therefore separate always-on screening from on-demand diagnosis: screening raises alerts, while diagnosis releases only source-verified what--who--where--when records. We present TrajMind, a fast-and-slow framework that switches three role-specialized LoRA adapters over one frozen vision--language backbone. Its slow path, \textit{TrajMind$_{\text{slow}}$}, chains canvas-based typing, type-conditioned localization over serialized trajectories, and executable verification, yielding structured, evidence-backed diagnoses. Additionally, the fast path, \textit{TrajMind$_{\text{fast}}$}, screens each window in a single text-only pass, delivering efficient structured alerts. Extensive experiments show that, TrajMind$_{\mathrm{slow}}$ outperforms the strongest baselines by at least $15.3$ percentage points in anomaly typing and $13.8$ percentage points in localization. These gains persist under cross-city transfer, and TrajMind$_{\mathrm{fast}}$ reduces latency by $41.1\%$ and maintains binary balanced accuracy of at least $93.5\%$. Together, TrajMind delivers accurate, evidence-backed diagnoses across cities and efficient front-line monitoring.
Jiahao Wu, Zhenqun Yang, Chen Jason Zhang et al.· 0 citations
Graph construction is a critical but underexamined design choice in deep reinforcement learning for power grid control. We present a controlled experimental comparison of different graph representations, including physical topology, electrical-sensitivity, and hybrid variants for topology control in the Learning to Run a Power Network (L2RPN) environment. Our findings indicate that matching graph complexity to task granularity is more important than maximizing representational richness, and highlight the importance of controlled representation studies at scale.
Adrian Degenkolb, Qiong Huang, Benjamin Sch\"afer· 0 citations
A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.
MIT News · Artificial Intelligence· news.mit.eduAug 24, 2026
A new method for surgically removing training examples from a model reveals that as datasets grow, the link between what a model learns and what it produces dissolves.
