ml engineers

Abstract: Long-context models repeatedly solve closely related retrieval problems. Attention or global retrieval can retain large stores, but each forward pass again searches for what matters. Recurrent compression avoids this search only by deciding earlier what information to preserve. Neither mechanism directly exploits that useful memories at neighbouring retrieval steps may substantially overlap. We propose an alternative in which retrieval is a persistent local traversal through an address space. We separate a memory's content from its address. Utility is defined by the information revealed by retrieving the content, while geometry determines only how cheaply that information can be reached. This leads to a cost-regularised indexing objective that preserves information value inside local neighbourhoods while charging for every memory made locally accessible. Its marginal-value form gives both attractive and exclusion pressures. Thus, complementary memories should be co-retrievable when the additional value exceeds retrieval cost, whereas redundant memories should not occupy the same retrieval set solely because they are individually useful. As a proof of principle, we test only the simplest fixed-geometry instance on MiniGrid Memory. A persistent local traversal head reaches $0.95$ mean return after $0.59$M training steps with a five-token context, compared with $0.94$M for Gaussian absolute addressing and more than $2$M for categorical absolute addressing. The experiment does not test learned geometry or repeated retrieval. Instead, it motivates the hypothesis that stateful retrieval may be easier to learn than stateless. To test the full proposed method further experimentation is necessary.

Abstract: Supervised Fine-Tuning (SFT) of LLMs suffers from catastrophic forgetting, and prevailing mitigation strategies struggle to reconcile the plasticity-stability dilemma. To address this, we propose Entropic Control for Homeostatic Optimization (ECHO), a closed-loop control framework. By decomposing the KL divergence, we identify output entropy as a dominant observable component of KL variation. This enables us to leverage entropy as a computationally efficient online proxy to monitor forgetting risks in real-time, bypassing expensive reference-model inference. Consequently, ECHO employs an elastic controller that dynamically modulates the gradient weight of replay data via a proportional control law, using deviations from the pre-trained entropic equilibrium as an error signal. Empirical evaluations across 0.6B to 9B parameter architectures demonstrate that ECHO lies on, and in several comparisons extends, the observed plasticity-stability frontier and maintains high operational robustness against varying data mixing ratios.

Abstract: Mixture-of-Experts (MoE) large language models reduce computation through sparse activation but still incur substantial storage costs. Mixed-precision post-training quantization (PTQ) reduces these costs by quantizing different model components at different bit widths, but existing methods often struggle to preserve generation quality under tight bit budgets. We identify two key challenges: (1) nonlinear quantization errors at extremely low precision are underestimated when cross-layer propagation is ignored; (2) direct router fine-tuning may cause routing drift, and calibration gains may not generalize to downstream tasks. To address these issues, we propose ProQuant, a weight-only PTQ framework for MoE models. Prune-or-Prioritize Mixed-Precision Allocation (PoP) jointly optimizes expert pruning and bit-width assignment. It allows higher-precision experts to substitute for severely degraded ones while improving bit-budget utilization. Routing-Consistent RMSNorm Finetune (RoCo) adapts expert input representations through mathematically equivalent transformations. It compensates for quantization errors to reduce routing drift. Extensive experiments demonstrate that ProQuant outperforms existing methods across diverse generative tasks. Code is available at https://github.com/ThrowBug/ProQuant.git.

Abstract: Sparse Mixture-of-Experts (SMoE) models improve scalability by routing input tokens to only a subset of expert networks, but they often suffer from representation collapse, in which multiple experts converge toward similar functions and lose their intended diversity. Prior work has addressed this limitation either by improving router architectures or by introducing auxiliary loss terms that implicitly penalize similarity among experts. However, router-based approaches remain vulnerable to routing collapses, and loss-based approaches can conflict with the task's original objective during optimization. In this work, we propose GaP-MoE (Gradient Accumulation and Projection in MoE), which induces expert specialization by directly manipulating expert output gradients rather than modifying the loss function. At each training step, GaP-MoE stochastically selects one of the $TopK$ routed experts as a leader based on normalized routing scores, and projects the gradients of the remaining follower experts onto the orthogonal complement of the subspace spanned by the leader's recent gradients. This removes directional overlap between expert gradients without altering the task loss, allowing specialization to emerge as a direct consequence of the optimization dynamics rather than an externally imposed constraint. We evaluate GaP-MoE in both fine-tuning settings (BERT-based, on GLUE) and from-scratch pretraining settings (GPT-2-based, on WikiText-103, enwik8, and text8), and further confirm its compatibility as a drop-in module for existing SMoE architectures. Experimental results show that GaP-MoE achieves performance comparable to or better than standard SMoE.

Abstract: Frontier coding agents are increasingly framed as end-to-end automated researchers, and ML for science offers a demanding test of that promise: improving a scientific model often requires domain insight beyond routine ML engineering. We introduce Agent4Sci, a long-horizon testbed for measuring how far today's agents can improve the published best models on scientific ML benchmarks, spanning 20 benchmarks across 11 domains. Rather than ranking agents on a leaderboard, Agent4Sci asks what kinds of progress agents can make on these tasks under in-the-wild conditions, and where they fall short. Empirically, today's frontier agents advance the published best on every benchmark, with Claude Opus reaching 10.4% mean relative improvement and no run flagged for academic misconduct. These gains, however, come almost entirely from a bag of ML-engineering tricks rather than research-level insight. Trajectory analysis surfaces three recurring limitations in this setting: novelty miscalibration, resource under-utilization, and rapid collapse into local engineering sweeps. Lightweight plug-in skills for research ideation, experiment management, and scheduled checklists raise Claude Opus from 10.4% to 20.1% and Codex from 4.4% to 8.1%, while yielding more diverse, domain-relevant interventions. Agent4Sci gives the community a way to track and push frontier agents from ML engineering toward domain insight on real-world ML-for-science problems. Code and data are available on GitHub (https://github.com/maszhongming/agent4sci) and Hugging Face (https://huggingface.co/datasets/MingZhong/agent4sci).

Abstract: Low-rank adaptation (LoRA) has become the standard method for parameter-efficient fine-tuning of large pretrained models, yet theoretical explanations for why low-rank updates suffice remain incomplete. Most existing analyses rely on the Neural Tangent Kernel (NTK) regime, which linearizes the network around the pretrained weights. In this work, we take a different approach that requires no linearization: our key observation is that, once the backbone is frozen, the loss sees each adapted weight matrix only through a finite-rank linear map induced by the frozen activations. Trace-norm regularization then forces the optimizer to pick a low-rank matrix among all matrices that produce the same observed output. This yields global minimizers whose ranks are controlled by the rank of this map, which we call the \emph{bottleneck rank}, rather than the sample size. For multi-head attention, we obtain tighter rank upper bounds by exploiting two additional invariances: row-wise softmax is insensitive to additive row constants in the score matrix, and value updates are observable only after attention weighting and frozen output projection. For the factorized LoRA objective with weight decay, we prove a complementary result on the nonconvex side: after an arbitrarily small generic positive semidefinite perturbation, every first-order stationary point is rank-deficient once the LoRA rank exceeds a threshold determined by the bottleneck rank. Finally, we show that global minimizers of the perturbed objective are near-optimal for the original trace-norm problem, and we provide uniform generalization bounds governed by the spectral structure of the frozen features.

Abstract: Massive activations, a phenomenon in which a small number of hidden channels exhibit exceptionally large magnitudes, are pervasive in large language models (LLMs). However, the mechanism by which a token develops massive activations as it propagates through a pretrained LLM remains poorly understood. In this paper, we find that the emergence of massive activations is controlled by a single channel in the input embedding to a spike feed-forward network (FFN). The position of this channel is fixed for a particular LLM. We name this channel the massive activation gating channel (MAGC). When the value of the MAGC is sufficiently large (or small, depending on the LLM), the output of the spike FFN exhibits massive activations. Examining six LLMs across four model families and different model sizes, we verify the existence and effect of MAGC. We further provide a theoretical explanation of the mechanism by which MAGC induces massive activations. When the value of MAGC is sufficiently large (or small), the output of a spike FFN asymptotically reduces to a quadratic form that mixes a few columns of the down-projection matrix of the FFN. Since these columns exhibit the shape of massive activations, the output therefore exhibits massive activations.

Abstract: Evaluating vision encoders requires metrics that reliably predict their downstream performance in multimodal large language models (MLLMs). Although recent studies have shown that cross-modal metrics can better capture such performance, unimodal metrics remain the dominant choice in practice. In this work, we revisit cross-modal evaluation of vision encoders through large-scale experiments. We identify important limitations in both the experimental design and methodological formulation of prior approaches. After addressing these limitations and introducing simple improvements, we propose RAVEL, a training-free method based on cross-modal nearest-neighbor retrieval. Despite its simplicity, RAVEL achieves state-of-the-art performance across our experiments, outperforming prior methods by a substantial margin. Our results demonstrate that simple cross-modal metrics, when evaluated under a careful and comprehensive setup, can provide a strong basis for evaluating vision encoders for MLLMs.

Abstract: Approximate machine unlearning seeks to remove the influence of a forget set from a trained model without full retraining. Existing gradient-based methods require data-dependent hyperparameter search, struggle when forget and retain knowledge are entangled, and offer little insight into where unlearning actually happens inside the network. We recast unlearning through the lens of task arithmetic: if finetuning produces a merged task vector $τ_m$ that combines learning on forget and retain sets, unlearning is the inverse operation that subtracts a learned forget component $τ_F$ to recover the retain task vector $τ_R$. The forget signal is concentrated: at every layer, forget activations lie in a subspace spanned by a handful of dominant directions, so we factorize $τ_F$ in a low-rank forget basis, which is faithful up to a small tail-eigenvalue residual and limits how far the correction can perturb retain. We then optimize three intuitive goals (match the merged vector inside the forget span, suppress leakage into the retain span, and bound the correction size) that provably bound forget leakage and retain damage in activation space. The resulting algorithm, Unmerge, is fast and powerful: on class-level unlearning with ResNet-50 on CIFAR-100 and Tiny ImageNet, it improves Tug-of-War by up to ~24% over a baseline of comparable runtime and by up to ~18% over stronger baselines that run ~5x slower, keeps membership-inference exposure at the level of retraining, and shrinks the feature-distribution gap to the retrained model, where relabeling methods leave forget features cleanly separable. Further studies show that Unmerge also applies to ViT-S/16 and scales to Llama-3.2-3B. The per-layer basis geometry that drives the algorithm also serves as a layerwise diagnostic for when and where unlearning becomes structurally hard.

Abstract: Policies trained from the same base model can appear to solve different problems. An oracle that chooses the best policy for each problem may therefore appear much stronger than any single policy. Selecting the largest estimated success rate also selects favorable sampling errors. We study this effect through reliability coverage, the fraction of problems whose success probability reaches a chosen threshold. Our first test redistributes stored correctness outcomes across policies within each problem. A second also preserves each policy's total successes, accounting for overall quality differences under a specified statistical model. For five training seeds of a seven-billion-parameter vision-language model, redistribution reproduces 0.096 of an estimated 0.113 oracle gap at threshold 0.10. Neither test finds significant evidence at this threshold. Small advantages remain unresolved. Mixtures, routers, voting, and weight averaging show no detectable improvement over their corresponding single-policy baselines. Training policies on different datasets shows little detectable specialization under light post-training, and no router gain. A stronger recipe does create it, both tests detect it, and a router gain appears only at the high thresholds where the specialists separate. In a control with predictable specialization, a router recovers about half the oracle gap. Coverage bounds explain why even a genuine oracle advantage need not yield a deployment gain. The tests assess apparent specialization from stored responses before investment in routing. Code: https://github.com/KurbanIntelligenceLab/oracle-gaps

Abstract: Mixed-precision weight quantization is commonly formulated as a Multiple-Choice Knapsack Problem (MCKP), yet existing solvers rely on scalar sensitivity proxies that collapse each weight matrix's Hessian into a single number and treat every module independently. We prove that even the optimal scalar proxy incurs multiplicative distortion up to $\sqrt{κ(\mathbf{A})κ(\mathbf{B})}$ relative to the full activation-aware quadratic, where $κ(\mathbf{A})$ and $κ(\mathbf{B})$ denote the condition numbers of the input- and output-side Hessian factors. This bound varies from $10^1$ to $10^{13}$ for typical LLM modules, making inter-module sensitivity ranking unreliable. To address these limitations, we propose Cross-layer Activation-aware Sensitivity Allocation (CASA), a two-phase method. In Stage 1, the scalar proxy is replaced by an activation-aware metric derived from the Kronecker-factored Hessian, reducing the MCKP to a form whose continuous relaxation admits a closed-form solution. In Stage 2, a cross-layer-aware local search evaluates bit-width updates using the end-to-end model loss. Experiments on multiple LLMs across different bit budgets show that CASA achieves lower perplexity than the latest scalar-proxy baselines, especially at ultra-low bit-widths ($<3$ bits per weight). Moreover, the performance gain in zero-shot accuracy tracks the per-model average condition-number over modules, confirming the distortion bound as a practical indicator of scalar-proxy failure.

Abstract: Membership inference attacks (MIAs) are widely used to audit the privacy disclosure risk of machine learning models, however current state-of-the-art attacks require training computationally expensive shadow models, making large-scale privacy evaluation impractical. In this work, we investigate whether inexpensive spectral metrics derived from the heavy-tailed self-regularisation framework can serve as proxies for MIA vulnerability. We evaluate several WeightWatcher spectral metrics on image and tabular classification tasks and compare their relationship with MIA privacy leakage against conventional measures of generalisation. Across datasets, stable rank exhibits a strong positive correlation with overall MIA success, while Log alpha-Norm shows a consistent negative correlation with MIA vulnerability at the low false-positive regime. These associations are observed to be stronger than those obtained using the generalisation gap. The results indicate that neural network spectra may contain information about privacy leakage that is not fully captured by conventional measures of overfitting, motivating spectral analysis as a promising direction for scalable privacy auditing.

Abstract: Benchmark scores describe what a checkpoint can do now, but they do not determine how it will respond to the next training episode. We measure this missing state by branching four short, standardized, target-independent micro-interventions from the same checkpoint and recording their effects in a common capability space. Together with current capability, these responses form L-State; its pulse block supports a flexible direct readout and a structure-preserving operator readout. Under smooth local dynamics, the operator construction admits an end-to-end cross-family bound with explicit source- and target-family coordinate heterogeneity. In three-family leave-one-family-out development, both pulse readouts reduce source-standardized MSE by 39.4% relative to capability alone, while separating the best response and direction estimates. On sealed GLM-4-9B, the direct and operator readouts reduce MSE by 71.8% and 78.3%, respectively, and the operator readout raises sign balanced accuracy from 0.366 to 0.754. On sealed Granite-3.1-8B, the direct readout reaches RMSE 0.544 and a development-fitted action-wise selector reaches 0.554, compared with 1.172 for capability alone. A five-family audit finds that the operator coordinate varies by action and family, and that modeling these deviations improves retrospective held-trajectory prediction. Target-independent interventions therefore expose training-response information that current capability misses, with direct and structured readouts covering complementary transfer regimes.