Training Efficiency & Optimization - Weekly Roundup

Synthetic pre-pretraining boosts LLM efficiency at scale, but not by teaching grammar; it's the retrieval skills that matter.

StreamMAE unlocks self-supervised learning on continuous video streams, matching i.i.d. MAE performance by tackling intra-batch similarity with novel regularization.

LLMs can learn to generate syntactically valid text without sacrificing semantic quality, even with underspecified prompts, by leveraging self-supervised rewards derived from their own likelihoods.

Task-specific LoRA-Experts and prototype-ensemble matching unlock rehearsal-free class-incremental learning with reduced interference and improved discrimination.

Recurrent memories can retain full Fisher information post-writing with an invertible hold, decoupling write and read optimality and achieving near-perfect task-specific performance.

Synthetic pre-pretraining boosts LLM efficiency at scale, but not by teaching grammar; it's the retrieval skills that matter.

StreamMAE unlocks self-supervised learning on continuous video streams, matching i.i.d. MAE performance by tackling intra-batch similarity with novel regularization.

LLMs can learn to generate syntactically valid text without sacrificing semantic quality, even with underspecified prompts, by leveraging self-supervised rewards derived from their own likelihoods.

Task-specific LoRA-Experts and prototype-ensemble matching unlock rehearsal-free class-incremental learning with reduced interference and improved discrimination.

Recurrent memories can retain full Fisher information post-writing with an invertible hold, decoupling write and read optimality and achieving near-perfect task-specific performance.

Forecast errors, not drift alarms, directly signal when to retrain time series models, achieving state-of-the-art performance with a plug-in module.

Moreau--Yosida unadjusted Langevin sampling achieves $\widetilde O(\varepsilon^{-1})$ iterations for $\varepsilon$-accurate convergence to target distributions, bypassing the need for third derivatives.

An RL agent learns to optimize SpTRSV parallelism by intelligently rewriting sparse matrix dependency graphs, achieving substantial performance gains with minimal structural changes.

Howard's policy iteration for deterministic MDPs is not strongly polynomial, exhibiting an exponential iteration complexity that separates it from the simplex method.

Arbitrary multilevel models can now be amortized with neural networks that precisely preserve their full conditional independence structure, outperforming samplers on large-scale problems.

Sparse POT problems can now be solved efficiently with structured regularizers, outperforming entropic methods on cost, sparsity, and speed.

Inductive Gromov-Wasserstein embeddings are now achievable, enabling out-of-sample generalization without recomputing costly optimal transport plans.

Most failed multi-turn agent trajectories contain a single, recoverable pivotal mistake that, if addressed, can restore task success, a phenomenon PivotOPD exploits for significant performance gains.

Multimodal models can achieve significant image generation improvements by learning from their own critiques during test-time compute, effectively acting as their own recursive self-improvement mechanism.

Faster computer-use agents are bottlenecked by their speed, but existing benchmarks are irreproducible; this new standardized framework reveals counterintuitive speed-efficiency trade-offs and that current open-weight models lag behind.

Physics-constrained generation now achieves comparable fidelity to multi-step methods in a single neural transport evaluation.

A single auxiliary variable per group of failures unlocks near-optimal $T^{-4/5}$ convergence for large-scale Cox regression, outperforming stochastic baselines.

An optimal threshold policy for irrevocable matching in growing trees loses at most twice the total-variation error of the true growth law, even when forecasts are misspecified.

Clinical MLLMs can be significantly improved by optimizing for AUROC via pairwise instance rankings, reversing accuracy-based prompt evolution's performance degradation.

Standard image ViTs, when adapted with masked Siamese learning on "super images," unlock state-of-the-art video representation learning with orders of magnitude less pretraining.

A novel decoupled autoencoder design unlocks 32x spatial compression for diffusion models while simultaneously boosting reconstruction fidelity and accelerating training convergence.

Untied embeddings, not weight tying, unlock superior performance and memory efficiency for decoder-only LLMs under DP-SGD.

Diffusion language models can now generate high-quality text with up to 49% fewer network evaluations by matching output distributions.

Kilometer-scale extreme heat downscaling accuracy scales linearly with climatological distance, not simulation volume, revealing a path to cost-effective urban adaptation.

Decentralized power allocation in complex MANETs is solved with a learned GNN that matches centralized optimization performance using only local, noisy channel data.

Explicitly constraining class geometry with differentiable prototypes significantly boosts tensor neural network performance under extreme label scarcity, particularly through training-time regularization.

Separating policy roles in offline RL unlocks significant performance gains by tailoring updates for value bootstrapping versus direct action execution.

Standard time-series forecasting losses systematically fail on pathological demand regimes, creating a bias irreducible by model capacity alone.

A low-rank residual connection allows frozen LLMs to learn from feedback with minimal parameters, achieving substantial error reduction in downstream tasks.

LLM pre-training power laws are not fixed but dynamically shaped by learning-rate and batch-size schedules, with a novel forcing-memory surrogate accurately predicting loss across these regimes.

End-to-end learned subspace projections shatter prior compression limits, achieving 70% weight reduction in Llama-2-7B with near-dense performance and drastically shrinking KV cache memory.

Zero-shot voice conversion now achieves $9\times$ faster inference with improved perceptual quality by jointly optimizing flow and content encoding.

Exact incremental BPE achieves $O(\log t)$ amortized append time by weighting centroid search intervals, a significant speedup over prior $O(\log^2 t)$ methods.

Neural operators can now learn directly from noisy, low-cost Monte Carlo estimates, achieving $10^4$-$10^5\times$ speedups for particle transport problems without sacrificing accuracy.

Language models can learn new knowledge without forgetting old by dynamically prioritizing replay based on observed forgetting.

Tree search on on-policy trajectories dramatically boosts policy gradient learning by uncovering rare high-return paths, outperforming PPO on Atari and MuJoCo with minimal bias increase.

Source-validation accuracy is a poor proxy for out-of-distribution reliability; re-ranking checkpoints by aggregated source reliability metrics significantly boosts target probability quality without needing target data.

Human password generation exhibits persistent linguistic regularities that can be exploited for a 16% boost in password recovery by adapting pre-trained language models.

A novel Fenchel duality approach allows efficient generative model adaptation to arbitrary utility functions, achieving up to 20x speedup and broader preference support than existing methods.

Recursive reasoning models achieve state-of-the-art algorithmic performance and OOD generalization with a simple optimization recipe, decoupling architectural complexity from recurrent state stabilization.

A novel cross-attention mechanism dynamically clusters queries to achieve sub-quadratic complexity in neural operators for PDEs without sacrificing spatial resolution or model flexibility.

Neural predictors can be made dynamically feasible without ground-truth controls by a frozen, post-hoc operator that enforces learned dynamics manifolds via iterative control correction.

Hybrid RL for bus holding suffers when simulators misrepresent passenger journey completion, leading to worse outcomes than purely offline methods.

Lightweight sentinels can learn to proactively intervene in coding agent actions, boosting task success by up to 14% by preventing costly error propagation.

Distilling expert LLM behavior is now possible without reference solutions by using a fixed expert's next-token predictions on the student's *prefixes* as privileged context.

Bayesian generative models can now estimate densities without invertible transformations, achieving state-of-the-art performance by adaptively focusing computation.

Spherical interpolation between old and new model embeddings recovers backward-compatibility for vision-language retrieval, outperforming orthogonal alignment by finding optimal query directions in the interstitial space.

Adversarial training gains a principled probabilistic interpretation and improved robustness via KL divergence maximization, outperforming standard methods.

State representation quality, not goal representation, is the critical bottleneck in offline navigation GCRL, with simple positional encodings boosting performance significantly.

Random hyperplane arrangements reveal that neural networks can achieve rank lifting with a width that scales independently of input dimensionality, dramatically improving efficiency for separable data.

One TD update is sufficient for linear convergence in regularized MDPs, and off-policy sample complexity can be achieved without trajectory resets or generative models.

A single backward pass can fundamentally alter training outcomes based on how it reuses low-precision forward-pass tensors, defying standard accuracy checks.

Finite-sample performance guarantees for label-noise transition matrix estimation are now achievable by bypassing class-posterior estimation through one-sided selective classification.

Physical degradation models can now learn reusable representations that generalize across entirely unseen system types, significantly improving prognostic accuracy with limited downstream data.

Low-precision matrix multiplication, even when repaired for accuracy, can break LLM prefix invariance and alter downstream task performance, necessitating a new certification approach.

A single video tokenizer can achieve superior semantic alignment and fidelity compared to much larger models by directly conditioning on frozen visual features.

Streaming world models can now maintain consistent representations across minutes-long revisits without quadratic attention costs by leveraging a novel retrieval-based memory bank.

Exploiting local parameter perturbations of a teacher model unlocks complementary supervision signals, boosting LLM mathematical reasoning accuracy beyond standard on-policy self-distillation.

Sparse rewards in multi-objective RL are no longer a bottleneck; a simple density correction unlocks faster, more robust learning.

Two shared scalar signals are sufficient to dynamically govern complex spectral memory, outperforming prior adaptive recurrences and accelerating training.

A single inference engine now unifies 15 sparse attention methods, unlocking over 10x throughput gains for long-context LLMs through intelligent KV cache management.

Weight decay, not just learning rate, dictates the peak magnitude and lifespan of massive activations in transformers.

Perfectly rank-matched tensor factorizations still suffer from quartically flat optimization valleys whenever the underlying Fourier slice ranks are nonuniform.

GRPO wastes roughly half its rollouts on zero-variance batches, but reparameterizing pass rates onto the Bernoulli Fisher–Rao arc length turns prompt curriculum design into a closed-form Gaussian selection problem that cuts sampling compute in half.

Probe-space preconditioning compresses zero-order LLM optimization from 100,000 steps down to 70, outperforming standard backpropagation on OPT-13B while slashing training memory requirements by 90%.

Online learning's long-standing five-expert prediction problem is finally solved: the adversary's optimal finite-horizon policy is $(1,0,1,0,0)$, rigorously refuting the global optimality of the COMB strategy via a machine-verified Lean proof.

Model retraining timing can achieve minimax-optimal $\tilde{O}(\sqrt{T})$ regret independent of the number of candidate intervals by exploiting the nested feedback inherent in multi-step degradation.

Splitting models into fully decoupled, communication-free shards cuts zero-order gradient variance by roughly $1/N$, enabling backprop-free pretraining to beat standard monolithic baselines using significantly fewer perturbations.

Native 8-bit LLM training silently destabilizes at billion-parameter scale due to softmax gradient drift, but restoring the zero-row-sum invariant completely eliminates the FP8 attention gap without architecture modifications.

Optimizers can dramatically alter step-by-step training trajectories, but a conserved geometric invariant ($2\alpha_r + \alpha_t = 1$) locks every algorithm from SGD to Adam and Muon into the exact same $D^{-1/3}$ optimal data scaling wall.

A single trained neural network can harbor vastly different internal computations and representations within its local solution space, a diversity standard optimization misses.

Non-factorized diffusion language models achieve superior few-step generation by leveraging a discrete latent space learned via Mixture-of-Experts routing, bypassing posterior collapse issues.

Independent chunk scoring in autoregressive video generation unlocks sustained visual quality over extended horizons, overcoming the temporal consistency trade-offs of prior methods.

On-policy LLM post-training's generalization advantage stems from its dynamic parameter update direction, a principle transferable to SFT for improved performance and efficiency.

Optimizing soft memory directly for user behavior, not text reconstruction, unlocks state-of-the-art long-context personalization.

Large-scale rewards can unfairly suppress learning signals from smaller rewards in multi-objective RL; CorrGRPO fixes this by normalizing reward covariances.

Key contribution not extracted.

Co-evolved robot communication protocols fail to transfer robustly to the physical world, even with sensory calibration, revealing that ecological and navigational context is as crucial as the signaling process itself.

Outcome-only RL can train highly effective process supervision models by rewarding accurate outcome prediction, even without explicit step-level feedback.

Static parallelism layouts leave massive post-training efficiency on the table, but dynamically re-slicing distributed model states across GPUs cuts RL step latency by 28% with less than 0.1% transition overhead.

Transformer row representations fail to form an autonomous thermodynamic critical class during training, meaning popular physical scaling analogies collapse unless explicit optimizer memory and readout error budgets are accounted for.

Billion-parameter LLMs no longer require global backpropagation to pretrain competitively: sharing a single delayed readout across decoupled modules matches standard zero-shot accuracy while yielding up to 1.44x pipeline throughput.

In regimes where high-accuracy gradients are brutally expensive to compute, minimizing a convex objective costs asymptotically no more than evaluating its gradient just once.

Adam's complex adaptive behavior boils down to an invariant, heavy-tailed moment ratio that enables zero-overhead 4-bit optimizer compression and formally unifies Adam with sign-based momentum.

Interleaving expensive Muon spectral steps with cheap Lion sign descent not only cuts distributed training wall-clock time by 33%, but actually beats full Muon and AdamW on final perplexity.

Black-box optimizers routinely discard the situational context of their failures, but framing past search trajectories as structured episodic memory allows algorithms to dynamically adapt their operators using negative evidence rather than just chasing historical winners.

Manual tuning of DRO uncertainty sets can be completely automated without statistical penalty, matching exhaustive grid-search sample complexity via nested minimax optimization.

Context compaction in agentic RL can run up to 5x faster simply by streaming the KV cache instead of flushing it—accidentally turning standard LLMs into recurrent agents that preserve evicted context purely through RL.

Learning conditional expectation operators nonparametrically no longer requires pre-selected kernel bases or risking bad local optima—every nondegenerate local minimum over centered $L^2$ spaces is provably globally optimal.

Adversarial imitation learning can achieve fast $\widetilde{O}(1/\varepsilon)$ sample complexity across both expert demonstrations and online environment interactions—provided you jointly regularize the policy and the reward.

One-sided matrix normalization leaves performance on the table: equilibrating both rows and columns accelerates LLM pretraining convergence while completely shedding AdamW's second-moment memory footprint.

Dynamic MoE load balancing can run at every single microbatch with zero CPU synchronization by having GPUs deterministically solve bitwise-identical expert replication plans locally, boosting cluster throughput by up to 11.4%.

Backing a four-candidate evolutionary algorithm with an on-demand Adam fallback cuts hyperparameter and architecture screening runtime by over 3x while preserving configuration ranking fidelity.

Diffusion models are completely blind to class frequencies at high signal-to-noise ratios, acquiring relative mode weights exclusively within a narrow $\mathcal{O}(1)$ SNR window known as the speciation time.

Standard Soft Actor-Critic suffers from an irreducible policy tracking error under conventional Gibbs targets, but switching to a policy mirror descent formulation enables direct control of target drift to guarantee provable $\mathcal{O}(N^{-1/5})$ convergence.

Boundary singularities usually cause learned optimizers to fail on constrained problems, but coupling primal-dual shifts with learned diagonal preconditioners slashes exact interior-point solver iterations by up to 67% without ever computing a Hessian.

Distillation's success hinges more on KL direction and learning rate than on-policy rollouts, challenging conventional wisdom.

A novel training strategy allows video segmentation models to learn from hours-long sequences, overcoming memory and gradient limitations to achieve 10x speedup over SOTA.

Predictive pretraining for parametric PDEs unlocks a 51.4% generalization improvement on unseen governing parameters by explicitly modeling latent dynamics geometry.

A novel kernel function circumvents gradient starvation in off-policy RLVR, enabling learning from previously discarded positive trajectories.