Training Efficiency & Optimization - Weekly Roundup

Students trained with OPMD learn to mimic teacher reasoning processes, not just outputs, leading to demonstrably improved reasoning capabilities and altered solution strategies.

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.

Students trained with OPMD learn to mimic teacher reasoning processes, not just outputs, leading to demonstrably improved reasoning capabilities and altered solution strategies.

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.

Blackwell GPUs' FP4 Tensor Cores unlock 2.81x output throughput for long-context LLMs by strategically using mixed precision and rank-one smoothing, bypassing attention bottlenecks.

Existing SFT reweighting methods fail to reverse harmful learned features; a new entropy-guided approach enables precise suppression, reversal, and extrapolation of pretrained knowledge.

HeteroFold unlocks $10.7\times$ faster cross-family KV cache transfer for LLMs, eliminating receiver prefill and matching text-based communication performance.

Control accent strength in zero-shot TTS with a single model and inference-time weight, outperforming cascaded systems on out-of-domain accents without sacrificing speaker identity.

Agents can achieve up to 5.8% higher success rates by learning a reusable "experience tree" directly from data, bypassing LLM-generated skills and enabling true generalization.

AI systems can auto-evolve by learning from their own failures, achieving 20.5% better performance with 65.5% fewer rollouts than existing methods.

This work proposes SLCA-GRPO, a framework incorporating Segment-Locked Credit Assignment (SLCA), a framework incorporating Segment-Locked Credit Assignment (SLCA) that decouples advantage estimation at the structural segment level within a single group of rollouts, without requiring additional rollouts from intermediate states.

Recurrent token mixing via GTR achieves 58.9 AP on COCO with sub-2ms latency, offering a practical, efficient alternative to quadratic self-attention for high-resolution dense prediction.

AURA (Angular Update Rate Adaptation), a per-parameter step-size adaptation that can be added on top of any first-order optimizer, and removed from it, without altering its update direction, is introduced.

A rigidity theorem is established: for continuous piecewise linear generators with a finite number of segments, any \(C^3\) function that can be represented in this way is at most quadratic, which demonstrates how generator dynamics and remainder estimates govern depth allocation and approximation rates of deep neural networks.

It is demonstrated that, for any $(f,\kappa)$-robust aggregator with coefficient $\kappa = O(f/n)$, where f is the number of adversarial clients and n is the total number of clients, convergence holds after an explicit sample burn-in.

It is proved that agents'parameters converge almost surely to the same limit points as in the attack-free case over time-varying communication graphs under linear parameterizations of the value and team-reward functions and Byzantine edge attacks.

This work proposes NeRF-aided Camera Ray Regression (NaCR), a unified framework that seamlessly bridges NeRF and CRR at the ray level and introduces a two-stage training curriculum to ensure stable convergence within the highly non-convex image space.

This work extends planning under a Riemannian metric from unconstrained spaces to constraint submanifolds by applying the induced metric in both a sampling-based planner and a trajectory optimizer.

DynaForge combines low-frequency global planning with high-frequency object-centric inverse kinematics across task phases, and applies a residual policy to correct actions during dynamic interaction, showing the ability of DynaForge for sim-to-real transfer.