Papers from the factory.

Lagrangian methods are a fundamental paradigm of safe reinforcement learning for constrained Markov decision processes, but the choice between one mixed critic and separate dedicated critics has never been analysed theoretically. This paper gives that analysis, and proves the mixed critic structure carries a bias from the target drift of the Lagrange multipliers, while the dedicated critic structure, separate critics for the reward and each constraint, does not. Experiments on a realistic multi-constraint power system environment bear it out: the dedicated structure satisfies the constraints where the mixed structure fails.

Synthetic data is easy to generate and hard to audit. This work introduces an interpretable auditing framework built on explanation shift: by comparing rule-based explanations extracted from real and synthetic data, it surfaces which relationships a generator preserves, which it weakens, which it removes outright, and which it invents. The result is a readable account of what a synthesiser actually did to the structure of your dataset, rather than a single aggregate fidelity score.

DiLoCo-style training cuts communication by letting learner islands train locally before an occasional outer sync, which suits fleets where training shares hardware with latency-sensitive serving. The open question is when a merge is worth its system cost, and whether choosing which windows to defer matters at all. Most scheduling studies omit the control that isolates timing from budget: matched random deferral. We show matched random ties or beats every forecast-free policy we test, then present Workload-Aware DiLoCo and a calibration protocol that determines when matched random can be beaten, then beats it in real vLLM sidecar replay, cutting SLO violations from 6.54% to 5.09%.

Does an optimizer change what a language model learns, or mainly how it stores and reaches it? We study Muon against AdamW on a 124M GPT-2-class model, matching sparse-autoencoder features across independently trained models by firing-pattern correlation and calibrating against within-optimizer seed baselines. Cross-optimizer feature agreement reaches the AdamW seed ceiling, while Muon changes spectra and yields a more seed-reproducible representation. Applied to individual attention sub-circuits, Muon on the value circuit is largely sufficient to recover the loss and reproducibility gains; Muon on the query/key scoring circuit raises stable rank yet recovers neither.

Machine unlearning is well studied; feature unlearning, removing the effect of specific features from an already trained model, is not. This paper presents a direct approach: redistribute the values of the targeted features across the original training set, then fine-tune on the shuffled data. Under mild assumptions the method carries a theoretical guarantee, disrupting the correlation between the unlearned features and the target outcome while preserving the relationships that remain. Experiments across several datasets confirm both the removal and the preservation, at a faster convergence rate.

Long-context serving has been a CUDA moat: the kernels that make million-token attention tractable ship for one vendor. Chunder injects dual-chunk attention into vLLM at runtime through Triton, putting 1M-token Qwen3 serving on AMD MI355X hardware. On the full RULER grid, serving stays exact at 32K, 131K, 512K and 1M, with every needle-in-a-haystack, multi-key and depth-sweep cell scoring 1.00 at 1M. The only degradation appears in model-capability multi-item recall, which plateaus rather than collapses from 512K to 1M.