Training Efficiency & Optimization - Weekly Roundup

Distilling explicit, reward-guided skills into VLMs without the student model ever seeing the skills themselves unlocks superior multimodal performance.

Generative time series models can now produce multiple plausible future trajectories as fast as quantile models query a single point, unlocking accurate dependency modeling for critical applications.

Restricting LLM fine-tuning to the bottom singular subspace nearly eliminates catastrophic forgetting, even when protecting 99% of weights, while maintaining strong target-task performance.

Gating policy distillation with directional agreement between early and late RL shifts unlocks more effective knowledge transfer, boosting LLM performance on reasoning tasks.

Assigning reward credit to specific tokens, rather than uniformly, dramatically improves multi-reward RL for LLMs.

Distilling explicit, reward-guided skills into VLMs without the student model ever seeing the skills themselves unlocks superior multimodal performance.

Generative time series models can now produce multiple plausible future trajectories as fast as quantile models query a single point, unlocking accurate dependency modeling for critical applications.

Restricting LLM fine-tuning to the bottom singular subspace nearly eliminates catastrophic forgetting, even when protecting 99% of weights, while maintaining strong target-task performance.

Gating policy distillation with directional agreement between early and late RL shifts unlocks more effective knowledge transfer, boosting LLM performance on reasoning tasks.

Assigning reward credit to specific tokens, rather than uniformly, dramatically improves multi-reward RL for LLMs.

Diffusion models can learn new objectives by stochastically routing attention within their existing architecture, achieving 4x faster RL training without sacrificing generation quality.

Regularized importance sampling in LLM RL breaks reward shift invariance, but a covariance-based estimand recovers this property, with distinct centering strategies impacting bias and variance.

Simultaneously evaluate $K$ policies in linear contextual bandits with near-optimal sample complexity by sharing information across evaluations via a unified ridge regression estimator.

Hyperscalers can achieve 1.5x training throughput gains on AMD GPUs by moving beyond direct CUDA porting to hardware-model co-design and agentic optimization workflows.

A single model query can now estimate the full acceptance probability for parallel masked diffusion, slashing rejection sampling costs.

Lossless compression of sparse ReLU activations slashes LLM activation VRAM by up to 31% with near-zero overhead.

Exponential Hopfield networks learn exponentially more archetypes than pairwise models by leveraging softmax attention, but risk memorizing noisy examples over true archetypes.

A simple byte reordering strategy slashes multilingual boundary prediction errors by over 70% in scripts like Korean and Chinese, fundamentally altering how byte-level models perceive codepoint transitions.

Transformers trained with ABNAT learn to mimic clean attention maps, achieving up to 85.77% less attention deviation and 11.28% higher accuracy on noisy analog hardware.

Standard MIL suppresses secondary findings in coexisting lesions; this novel framework recovers them by modeling the dominant-label annotation process.

Deep learning compilers are silently introducing bugs through state retention, with 35 stateful issues found in just four frameworks.

Offline-trained neural networks are mathematically proven to diverge in closed-loop control, but bounded online adaptation offers a practical solution.

A simple, fixed-budget best-arm identification algorithm can be made nearly as adaptive as state-of-the-art methods by optimizing its phase schedule, achieving $11/12$ of the optimal normalized rate for Gaussian arms.

Neural networks preferentially learn low-frequency sinusoidal patterns due to the inherent graded structure of their gradient descent dynamics, explained by an orthogonal polynomial framework.

Fine-grained data balancing in heterogeneous pipelines unlocks 2x training speedups for large models by overcoming memory bottlenecks and coarse workload assignments.

Hundreds of billions of midtraining tokens on a curated software engineering corpus unlock substantial gains in agentic capabilities, outperforming standard SFT alone.

A single weight-space encoder now masters 100 transformer architectures, outperforming specialized models and handling unseen, larger networks with less memory.

Continual learning, aiming to improve deepfake detector accuracy over time, systematically intensifies unfairness across demographic groups.

A single INR can now model complex, periodic deformations across entire image sequences, enabling self-regularization and outperforming prior methods.

A two-bit code and shared reconstruction program slash LLM conditional memory storage by 75%, enabling 4x larger tables on GPU.

A 32.5x smaller grid foundation model achieves superior reactive power prediction and significant active power improvements over its larger teacher, enabling real-time edge inference.

Adaptive planning horizons emerge naturally from policy confidence, outperforming fixed discount factors on challenging credit assignment problems.

A closed-form depth-width law for neural collapse reveals how representation compression scales with the L/N ratio in deep linear networks.

Single-qubit quantum circuits, despite universal function approximation, get stuck in training due to "phase-locking," requiring orders of magnitude more optimization steps than multi-qubit equivalents.

Shared experts in MoEs paradoxically limit the very exploration fine-grained segmentation needs, but MirrorMoE unlocks this potential by adaptively pruning the routing space.

A novel PI-DeepONet implementation slashes training time by up to $3.75\times$ and makes 3D physics-informed learning feasible, overcoming critical computational barriers.

Byte-level LLMs can achieve a 25% jump in math accuracy by patching where text is *new*, not just surprising.

Consolidating specialist LLMs into a single, more accurate generalist is possible with *zero* gradient updates, relying solely on forward passes and input statistics.

An adaptive acquisition function learned online via temporal difference learning discovers optimal perturbation experiments 2x more effectively than fixed methods.

Most self-evolving LLM updates are useless, yet current methods inherit them all; new selection strategies drastically cut harmful inheritance and boost reasoning performance.

A single LR gradient equation explains the shape of learning-rate transfer curves across diverse settings, revealing layer-specific sensitivities and theoretical bounds.

Generative models can inject crucial scientific knowledge into GP posteriors, recovering hidden modes and improving predictive accuracy where traditional GPs fail.

Cost-penalized active feature acquisition is shown to be a direct Lagrangian relaxation, with a computable certificate for greedy suboptimality and a structural guarantee in the near-submodular regime.

In-context learning emerges significantly earlier than specialized prediction, revealing a fundamental trade-off in how LLMs adapt to novel data based on pretraining scale.

Clients can achieve superior worst-decile accuracy in federated learning by stopping local training based on an optimal stopping framework that balances loss reduction against drift, outperforming tuned baselines and existing adaptive methods.

Offline risk-sensitive RL is now provably efficient, with new algorithms and tight bounds for dynamic and static optimized certainty equivalents.

A neural evaluator predicts Lean 4 proof acceptance with 87.8% accuracy and achieves 4.2x the throughput of a multi-worker Lean baseline, drastically accelerating proof discovery.

Full-page document parsing models can now pinpoint and correct local errors by assigning credit directly to tokens, outperforming baselines across six benchmarks.

A single forward pass now reveals which input variables matter for PFN predictions, achieving exact FDR control and outperforming lasso, BART, and other methods on sparse nonlinear regression.

Agentic RL agents can learn to efficiently manage context and subtasks across multiple sessions by estimating the marginal contribution of each session, outperforming naive advantage broadcasting.

Injecting uncertainty samples into DNNs during federated training unlocks robust, near-centralized performance for joint communication and computation resource allocation in XR systems, outperforming state-of-the-art.

Train LLMs with up to $1.455\times$ less memory and enable $3.5\times$ longer context lengths by losslessly compressing tensors on-the-fly.

Early stopping of Randomized Block Coordinate Descent can achieve optimal generalization rates in nonconvex learning, a feat previously unquantified.

Aggressive KV cache compression can be achieved with near-zero accuracy loss by treating the cache as a trainable neural network optimized at test-time.

Spectral decomposition via a learnable Hodge-Laplacian unlocks a 47% reduction in interior MSE for steady-PDE surrogates on coarse meshes.

A single, stage-specific objective per expert in a decoupled diffusion distillation framework enables state-of-the-art video generation with a 10,000x reduction in training cost.

Asynchronous RL training on long-horizon tasks achieves superior performance by leveraging Muon and a novel single-rollout recipe that avoids waiting for complete trajectories.

LLMs can permanently "remember" factual edits and skill gains from context, retaining 99.8% of original performance after 50 sequential updates.

Heterogeneous LLMs can share internal factual states without fine-tuning by exchanging word indices mapped to receiver-local vector directions.

Teachers can be coaxed into recovering from their own errors during distillation, unlocking substantial gains in student model performance.

Language agents can achieve superior sample efficiency and performance by jointly learning to act and forecast multi-step action sequences within a single policy backbone.

Distilling LLM agents by supervising semantic actions, not token sequences, unlocks significant performance gains and bypasses tokenization constraints.

Most self-distillation methods for agentic RL fail to consistently outperform a strong baseline, and impressive final scores often hide unstable training dynamics and substantial computational overhead.

Continuous integration over a piecewise linear memory curve unlocks exact normalized attention with analytical gradients, outperforming softmax and focal attention in causal Transformers.

Sparse human guidance, covering less than 20% of interactions, is sufficient to dramatically improve real-world mobile manipulation performance by reshaping the critic's value landscape.

LLM agents trained on function calling forget nearly 25% of their capability when action names are perturbed, but a simple self-supervised regularization recovers over half of that loss.

A single pretrained tabular foundation model can solve diverse power grid analysis tasks, from power flow to contingency screening, without fine-tuning.

Training-free neural emulators achieve solver-level accuracy and out-of-distribution robustness by directly encoding known polynomial physics.

A single-rollout RLVR method achieves faster training and more stable gradients by using a trajectory's own future as a hindsight context, eliminating the need for group sampling and reference solutions.

Budgeted adaptation under opponent shift can be solved by dual-counterfactual control that explicitly balances immediate adaptation value against future adaptation loss and local strategic risk.

N-gram embeddings can be learned without any additional parameters or GPU memory, unlocking massive scaling and instant domain adaptation for LMs.

A novel PPO variant leverages complementary temporal dynamics of self-teacher signals and critic value predictions to boost reasoning accuracy by up to 2.59pp.

Larger momentum coefficients ($β$) are crucial for preserving optimization gains in LLM post-training with noisy rewards, outperforming conventional choices and boosting accuracy by nearly 2% on large models.

Learning optimal policies from a single RL trajectory can be as sample-efficient as unrestricted exploration.

Deep learning models can achieve near-perfect accuracy, then forget it all, only to relearn it later, defying intuition about monotonic learning curves.

Efficiently tracking a small, switching subset of experts in continuous action spaces is provably harder than theoretically possible, with polynomial-time algorithms facing inherent regret limitations.

Generative representation learning unlocks compute-optimal latent tokenization, achieving superior image quality with drastically fewer FLOPs than prior methods.

Continual learning for molecular communication achieves robust performance estimation across evolving experimental complexities by intelligently synthesizing past data.

Tabular foundation models can achieve state-of-the-art in-context learning by selectively activating feature-specific experts, outperforming dense models on challenging datasets.

Adaptive perturbation radius in SAM, guided by local curvature, unlocks superior generalization by escaping sharp minima more effectively without sacrificing convergence speed.

Quantizing MoE models with mixed precision is significantly improved by calibrating expert error sensitivity using an information-theoretic contrast target, outperforming simple ranking-based methods.

Speculative decoding drafters can achieve over 2x speedup on 256K context by treating distant tokens as shared far memory with sharpened attention.

A simple reconstruction pre-training stage unlocks state-of-the-art one-step generative modeling with 10x less compute.

A closed-form, covariance-aware orthogonal transform for LLM quantization achieves state-of-the-art performance with a third of the memory and time of learned rotation methods.

Feldman's memorization lower bound can be arbitrarily inaccurate, but a corrected version using full-sample posterior probabilities offers an exact replacement.

A hypothesis selection algorithm achieves optimal sample complexity and a near-optimal approximation factor while improving the state-of-the-art runtime by reducing the dependency on failure probability from polynomial to logarithmic.

Generative models for PDEs can now recover fine-scale details and maintain long-horizon stability by learning residual distributions and aligning multiple trajectories.

Average-reward option Q-learning achieves $\tilde{O}(1/\text{iterations})$ convergence, unlocking continual RL tasks previously intractable for option-based methods.

Neural operators learn optimal derivative orders strictly below the PDE order, revealing a data-driven spectral regularization mechanism.

Standard imputation methods fail to predict downstream utility, but task-aware imputation recovers complete-data performance in up to 80% of challenging train-test mismatch scenarios.

Agents can recursively improve their problem-solving by learning to adapt their own policy and memory management strategies, achieving substantial error reduction in program selection.

Spectral convolutional kernels unlock significantly greater accuracy gains from learned resolution allocation than their spatial counterparts, even when spatial area is normalized.

LLM reasoning can be optimized by learning from the geometry of state transitions, achieving reflection-level accuracy with 28% lower latency and 25.8% fewer tokens.

Robot policies can now adapt to unseen terrains and payloads in real-time, boosting performance by up to 80% without needing rewards or retraining.

Actively repelling uncertain features during representation learning slashes extreme segmentation errors in ambiguous anatomical boundaries by 22x, even without sacrificing overall accuracy.

For star-free languages, transformers learn best when attention weights lag behind output layer updates, defying intuitive feature learning theories.

Continual pretraining a ViT encoder with LoRA on prostate MRI segmentation significantly improves downstream task alignment and performance, particularly for segmentation tasks, by inducing targeted representational shifts.

One-step RF signal separation achieves superior BER and MSE by learning a direct clean-endpoint transport via rectified flow matching.

A novel geometric analysis reveals how fairness scaling warps client update distributions, enabling structure-aware Byzantine detection beyond simple magnitude filtering in $q$-fair federated learning.

The minimal width of ReLU networks for approximating functions from $\mathbb{R}^n$ to $\mathbb{R}^m$ is $n+m$ for specific $n$ values, directly linking network capacity to the topological properties of real projective space.

Conflicting findings on model representation convergence are unified by a lattice structure, revealing that "Platonic" global agreement is only one facet of a broader representational landscape.

MoE models slash the performance cost of multilingual expansion by 83% when scaling from 12 to 200 languages, fundamentally altering the trade-offs in capacity allocation.

Poisson noise, not just Gaussian, unlocks powerful generative models for a critical class of inverse problems.

Amortized batch black-box optimization via flow matching generates variable-size batches in constant time, outperforming traditional methods on expensive optimization tasks.

Outperforming traditional model selection, a retrieval-based approach finds the best anomaly detector for unseen time series by leveraging similarity to past data, even in challenging OOD scenarios.