machine learning engineers

Abstract: Zero-cost proxies rank architectures cheaply, but their reliability varies across search spaces. We introduce CoRA-NAS (COarse Ranking + Anchor-residual), a two-stage framework combining a static ranking prior with low-cost learning-curve refinement. CoRA-Rank aggregates capacity and structure-at-initialization proxies through an equal-weight log-rank consensus and a target-free consensus gate. CoRA-Refine samples anchors across this prior, extrapolates their early validation curves, and propagates a learned residual correction with an ExtraTrees model. The refinement uses approximately 1% of the cost of fully training the candidate set. Fully trained architecture-accuracy labels are not used to fit the ranker. One configuration is used across spaces, with space-specific architecture encodings. Across NAS-Bench-201, NAS-Bench-101, TransNAS-Bench-101, and NATS-SSS, CoRA-Refine achieves mean Spearman correlations of 0.946, 0.715, 0.786, and 0.894, respectively. Its worst-space correlation of 0.715 is the highest among the compared methods. On NAS-Bench-201/CIFAR-100, its selected architecture reaches 73.32% accuracy, near the reported ground-truth best of 73.37%. On the pure size space, refinement recovers the static prior's shortfall relative to parameter count, while remaining tied with the strongest capacity proxies within noise. The resulting framework combines cross-space ranking robustness with low-cost architecture selection.

Abstract: We introduce AdamX, a first-order optimizer that incorporates cosine similarity as an adaptive mechanism for controlling update magnitudes. The proposed method is scalable, model-agnostic, and straightforward to integrate into existing training pipelines. We further introduce a variance rectification scheme that promotes smoother optimization during the early stages of training. Overall, we provide empirical evidence that AdamX achieves competitive convergence rates across a range of benchmark datasets and architectures. Performance is evaluated in terms of the number of epochs required to reach predefined performance thresholds under a fixed hyperparameter budget. Code and Experiments available at: https://github.com/FranciscoCaldas/adamX.

Abstract: Diffusion and flow-matching schedules control the signal and noise coefficients that mix data and noise along affine probability paths. Minimizing a kinetic action defined on coefficient paths, motivated by optimal transport, helps explain strong baselines but remains model-agnostic and ignores prediction error. Here we introduce a model-aware schedule construction based on fiberwise optimal transport. At a fixed time and state on the probability path, compatible signal/noise decompositions form an affine fiber. We define a fiberwise prediction risk by averaging optimal-transport costs between the true and predictor-induced decompositions within these fibers. On a fixed coefficient curve, combining this risk with coefficient-path kinetic action yields a closed-form optimal time allocation. This construction extends to general linear prediction targets, and the risk profile can be estimated from an early baseline checkpoint. We evaluate DDPMs and flow matching across prediction targets, training configurations, risk-estimation checkpoints, datasets, and architectures. Our model-aware schedules consistently outperform strong baselines, including a 38.6% relative FID reduction for flow matching on CIFAR-10 at 16 function evaluations. Each model-agnostic kinetic baseline determines its own kinetic reference coordinate. In these coordinates, fiberwise-risk profiles from independently trained models in different settings align closely after normalization to unit area. The resulting schedule deformations used in training also align, suggesting empirical universality across the evaluated models and settings. Pretrained-checkpoint diagnostics extend this normalized-risk agreement to larger conditional latent diffusion and 2-RF models. A frozen analytic allocation template retains most of the model-aware improvement without further risk estimation or model-specific fitting.

Abstract: Cross-cultural understanding has become increasingly important in today's highly connected, cross-national world. The success of LLM-based technologies is now driving the development of automated tools to aid understanding for nonnative people trying to succeed in cross-cultural environments. Building such automated tools is often done by leveraging in-thewild text, audio, and video data. This paper presents techniques for improving speech recognition-based transcript creation in multiple languages from videos to better train these automated tools. The focus is on processes and speech tools that can easily be used by cross-cultural tool builders without requiring deep speech processing expertise. Using publicly available videos from YouTube and Whisper-based tools, average transcription error rate across seven languages (Spanish, Japanese, Korean, Mandarin, Turkish, Russian, and Hebrew) of 30% are observed. With a modest amount of fine-tuning data, the average error rate can be reduced to 20% making such output much more usable for downstream processing. Speech and metadata associated with these videos that can be used by the community to further refine these experiments are released as well.

Abstract: Post-training quantization compresses large language models (LLMs) by storing their weights at reduced precision, and each quantized weight introduces an error into the hidden states. Naively, these errors should accumulate with depth and corrupt next-token prediction; randomly initialized models accumulate these discrepancies rapidly, whereas quantized pretrained models accumulate much less hidden-state error and largely maintain downstream task performance, even though they were never trained with quantization noise. This raises the question we address: why does post-training quantization work? Comparing full-precision and quantized forward passes, we identify two mechanisms that characterize pretrained quantization robustness. First, the error a layer newly introduces tends to oppose the error it inherits from the layer's input. The two cancel partially such that the discrepancy between full-precision and quantized passes grows slowly. This counteracting residual interaction develops during pretraining. Our quantitative analysis identifies it as a major factor slowing hidden-error growth. Second, LM-head geometry preferentially preserves the scores and probabilities of high-ranked tokens, which typically represent the model's most confident predictions. Together, these mechanisms explain why quantization error that passes through numerous layers can still produce only small output changes, and we verify the findings across models and quantization settings.

Abstract: In this paper, we investigate the generalization performance of distributed gradient descent algorithms in a reproducing kernel Hilbert space under a robust loss function $l_σ$. By exploiting the spectral characterization of gradient descent together with the intrinsic properties of robust loss functions, we establish optimal learning rates for the distributed kernel-based robust gradient descent (DKRGD) algorithm with an appropriately chosen scale parameter $σ$. The proposed parameter choice of $σ$ simultaneously alleviates the saturation phenomenon and guarantees statistical robustness. A key technical contribution is a novel error analysis that provides substantially sharper bounds for products of operators, thereby significantly relaxing existing restrictions on the maximum number of local machines while retaining optimal learning rates. Finally, we develop a communication-efficient strategy that further improves the convergence performance of DKRGD.

Abstract: We revisit Kashin-decomposition-based weight quantization for large language models and propose an improved algorithm with stronger convergence properties and structured, efficient orthogonal transforms. The method retains the core factorization of each weight into two components -- one with bounded infinity norm and the other with bounded infinity norm after an orthogonal transformation -- but replaces the dense random orthogonal matrix with a sign-randomized Discrete Cosine Transform (DCT), reducing the per-iteration cost from $\mathcal{O}(N^2)$ to $\mathcal{O}(N \log N)$. The proposed greedy algorithm with alternating updates guarantees the four-peak distribution required for stable 2-bit clustering of each factor and admits closed-form initialization of cluster centers, removing the multi-restart k-means bottleneck of prior work. Composed with OPTQ-style sequential error compensation and QuIP-style incoherence preprocessing, the resulting JAX pipeline is competitive with OPTQ, QuIP, QuIP-RG and a fine-tuning- and vector-quantization-free variant of QuIP# at 4-bit per channel on OPT, Llama-2 and Pythia, with favorable wall-clock scaling. The bounded-$\ell_\infty$ factorization is also notably robust: on stress configurations where QuIP variants diverge to four-digit perplexity (Pythia-6.9B) or abort with NaNs in LDL back-substitution (Mistral-7B), Kashin-DCT remains numerically stable and stays close to FP16 baseline. At inference time, each weight decomposes into two 2-bit factor codes per channel that are structurally suited to native-2-bit hardware.

Abstract: Self-evolving runtime harnesses can substantially improve the capabilities of large language model (LLM) agents and provide a promising paradigm for optimizing agent execution. Existing harness evolution methods typically rely on iterative search, repeatedly evaluating and revising candidate harnesses based on execution feedback from task instances. While this paradigm enables continuous harness optimization, it incurs substantial time overhead due to repeated agent executions and code modifications, and may overfit to observed tasks and specific failure patterns, resulting in degraded generalization to unseen tasks. We identify the lack of principled failure diagnosis as a key bottleneck in harness evolution: an observed failure can reflect either model-specific deficiencies or systematic harness deficiencies, and directly optimizing against individual failures can lead to unnecessary model-specific accommodation. We therefore propose Ecdysis, an efficient and effective framework that distinguishes model-specific accommodation from harness-level repair and biases adaptation toward systematic harness deficiencies by identifying recurring cross-task failure patterns. Ecdysis adopts a batch-level cross-instance failure aggregation paradigm to jointly analyze failure evidence from multiple task instances and further introduces Failure-Driven Collaborative Refinement to diagnose failure causes and iteratively refine harness modification specifications. By combining cross-instance failure analysis with multi-role diagnosis, Ecdysis enables more effective harness evolution with lower training time. Experiments show that Ecdysis achieves up to a 1.84x speedup in harness training compared with existing harness evolution methods, while improving the reasoning accuracy of the resulting harnesses by 18.56%.

Abstract: Muon has emerged as a highly effective optimizer for large language model training, often achieving superior convergence and performance compared with the widely adopted Adam and AdamW optimizers. Nevertheless, Muon is prone to training instability due to its spectral flattening, manifested by loss spikes and unbounded growth of model weights. Existing approaches primarily rely on weight or attention-logit clipping, which require architecture-specific modifications and do not directly address instability across all model components. We propose MomentUm SpEctral Clipping (Musec), which replaces Muon's spectral flattening with spectral clipping: rather than setting all singular values of the momentum matrix to approximately one, Musec clips singular values that exceed a threshold while preserving the underlying spectral structure of the momentum. Our strategy provides an optimizer-level, architecture-agnostic mechanism for stabilizing Muon training. We further develop Soft Musec, an efficient implementation that uses a smooth spectral saturation function approximated by coupled Newton-Schulz iterations. Theoretically, we establish convergence guarantees for Musec in nonconvex nonsmooth stochastic optimization. To the best of our knowledge, this is the first convergence guarantee for Muon-type methods in the nonconvex nonsmooth setting. We provide empirical studies to show that Soft Musec consistently improves training stability over existing Muon variants across a wide range of learning rates and model sizes. Notably, Soft Musec remains stable in settings where existing Muon variants diverge, while matching their performance under well-tuned configurations.

Abstract: Modern high-performance GPU computations partition tensors into tiles to exploit data reuse and parallelism. Individual tile computations complete earlier than the full tensor computation, creating opportunities to overlap computation and communication. However, a mismatch between computation and communication progress can limit these opportunities. Communication stalls when no data is ready, and may lag when data arrives in bursts. Communication can also slow computation by consuming shared resources, offsetting the benefits of overlap. We present Entwine, which coordinates tile computation order, fine-grained communication, and SM resource allocation to minimize overall completion time. Entwine reorders tile computation to produce data for communication at a more regular pace. Entwine couples this schedule with fine-grained SM-based communication to process tile results with low latency and low overhead. Since the communication kernel also consumes SM resources, Entwine coordinates their allocation to balance communication progress against computation slowdown. Across representative tensor-parallel LLM workloads, Entwine achieves a geomean speedup of 1.232x (up to 1.433x) over cuBLAS+NCCL, and outperforms state-of-the-art overlap baselines by 3.1-9.8% in geomean. We will open-source our implementation upon publication.

Abstract: Generative Marginalization Models (MaMs) have been recently introduced as efficient neural sampling models for any-order autoregressive modelling of discrete distributions. By learning both the marginal and conditional probabilities of a persistent-block Gibbs sampler, MaMs enable fast posterior evaluation with a single neural network forward pass. While prior work has considered MaMs to be distinct from Generative Flow Networks (GFlowNets), a well-established paradigm for inference in discrete stochastic models, we show that they are equivalent. Then, we also extend MaMs' sampling strategy to non-autoregressive generative processes. In particular, we describe an automatic criterion for full-state rejuvenation of the Gibbs sampler, derived from the Gelman-Rubin statistic, which plays a key role in speeding up learning convergence. Our experiments show that our method, called Particle GFlowNets, markedly accelerates training in large combinatorial spaces.

Abstract: Maximum likelihood (ML) estimation is a principled and statistically efficient approach for learning probabilistic models. However, for unnormalized models, ML estimation requires evaluating the partition function and differentiating through it, which may not always be tractable. Score matching provides a practically viable alternative that circumvents this obstacle by fitting the score in a way that eliminates dependence on the normalizing constant. We derive the generalized score matching objective on a convex subset of $\mathbb{R}^{d}$ constructively starting from Minimum Probability Flow (MPF) learning, and show how classical score matching as well as domain-adapted variants for non-negative data arise naturally within the proposed framework. We show that the resulting objective is a {\it proper local scoring rule} of second-order, which provides the theoretical guarantee that the true density is recovered when the objective is minimized. Furthermore, for a model belonging to the exponential family, we establish convexity of the objective together with consistency of the finite-sample estimator under standard regularity conditions. Our derivation sheds new light on the scope and applicability of generalized score matching in various problem settings. We compare generalized score matching-based estimators on constrained domains, where the partition function is analytically intractable. We provide experimental results on parameter estimation for model densities belonging to the exponential family defined over convex subsets of $\mathbb{R}^{d}$, and a generative modeling use-case to demonstrate broader applicability of the proposed generalized score matching framework.

Abstract: Evolvable-Substrate HyperNEAT (ES-HyperNEAT), a bio-inspired indirect encoding that determines neuron placement and connection weights from spatial coordinates, exhibits a failure mode on MNIST as a diagnostic benchmark. Because input pixels map to a coordinate space centered at the origin, evolved networks converge on a small central cluster of input pixels, a spatial-concentration bias; prior work observed only 21% mean accuracy in this regime. Is this bias an optimization artifact or an architectural ceiling? Inspired by Mixture-of-Experts (MoE) principles, we partition the input into non-overlapping spatial segments, each assigned to a separately evolved specialist network. With 13 such experts, this design reaches 43% mean accuracy, a 106% relative improvement over the baseline. The architectural gain does not depend on data-driven aggregation: equal-weighted averaging, which uses no validation data, already yields a 70% improvement; the gain comes from partitioning, not the weighting. Receptive-field analysis shows the mechanism: partitioning forces evolution to discover features across the entire image, expanding active pixel coverage from 4% to 79%. Absolute accuracy stays below gradient-trained baselines, but the relative gain points to central bias, not the evolutionary search. Two tools are designed to generalize beyond MNIST: a receptive-field diagnostic for silent input-coverage collapse, and a spatial-partitioning remedy that restores coverage.

Abstract: Efficient language learning requires methods to reduce the reliance on large data and computational resources. We investigate structural transfer: First training models on non-language data to induce useful priors for natural language. This approach is a form of weight initialization for multilingual language modeling. We evaluate transfer via next-token-prediction loss, weight shifts in the model, and downstream linguistic benchmarks. Several symbolic data types - notably music, probabilistic grammars, and cellular automata - yield lower language-modeling loss than random initialization. These gains coincide with smaller weight shifts during subsequent language training, suggesting that structural transfer positions models in a more favorable region of the parameter space. However, a lower loss does not translate consistently into better downstream linguistic performance, and transfer from non-language data is less efficient than additional language data. We conclude that non-language data can serve as a partial substitute for language data for the training objective of next-token prediction but does not reliably support broader linguistic generalization.

Abstract: An unlearning audit reads its verdict off numbers that an unlearned model and its retrained reference each publish, and both also ship batch-normalization statistics that no gradient step wrote and no release records. Refitting them on kept data at bit-identical weights moves 47 of 221 released checkpoints past the spread their own release's seeds show, several inside a method whose average does not move: what moves is the checkpoint's property, not its method's. What does the moving is not the removed data surviving in the state: exchanging kept records for removed ones inside a fixed fitting pool moves a published cell by almost nothing, while how far a checkpoint's shipped state has drifted from any refit does track it. The consequence for a published decision is real but narrow: twelve verdicts cross, four clear a measured recalibration budget, two clear it on every replicate, and a population we trained and sited near its own criterion yields none. A release should therefore name the fitting convention beside the number, on the batch-normalized vision models where this channel exists.

Abstract: Many ML datasets are constructed by running a detector, heuristic, or model over candidate pools; accepted items become labels. Dataset precision is then governed by true-positive prevalence in each pool via Bayes, not solely by detector quality. Using one instrument and period, we hold a detector-defined event dataset plus an independent official index labeling every detected item as real or phantom. One detector, three pools yield phantom rates 81.7%, 9.0%, and 0.0%. Transferring precision from the two high-rate pools to the low-rate pool predicts 0.955 versus measured 0.183, a +422% error; the Bayes expression predicts all three within 3.3%. The detected response curve is an exact convex combination of a true-event and a phantom component (residual 1.1e-16), with phantoms outnumbering true events 473 to 308, so contamination is a second signal with detector-inherited shape, not additive noise. Contamination direction depends on the estimator: on identical windows one statistic is diluted and another inflated because its denominator is also contaminated. A common normalization turns the estimator into a mean of ratios whose expectation need not exist; on the same 335 events it returns 0.40 where the well-defined estimator returns 0.10.

Abstract: This work investigates a number of approaches to prevent catastrophic forgetting in class incremental learning scenarios for sound event classification tasks. We analyze the problem using architectural and regularization approaches, using FSD50K and AudioSet datasets. We design incremental stages and solutions that selectively protect the kernels of the network from weight updates to prevent catastrophic forgetting, and a dynamic head solution that expands itself each time a new task is learned. The findings show that catastrophic forgetting mainly happens in deeper layers, in particular in the classifier head. For the studied in-domain sound classification problem, the solution that seems to alleviate catastrophic forgetting and is the most efficient is a full freezing of the feature extractor with a fine-tuning of the dynamic head classifier, showing little to no forgetting and great training stability, and a good balance between memory-stability and learning plasticity.

Abstract: We provide methods for calculating the robustness of the predictions of two types of generative classifiers whose underlying distribution is a Probabilistic Graphical Model (PGM): naive Bayes classifiers and generative forests (a probabilistic extension of random forests). Following the paradigm of robustness quantification, we define the robustness of a prediction as the extent to which the distribution of the classifier can be perturbed without changing this prediction. We consider perturbations obtained by varying the local models of the PGMs within general neighborhoods and focus in particular on epsilon-contamination, total variation distance and chi-squared divergence balls. We test our methods on benchmark datasets, demonstrate that the robustness value of a prediction serves as an indicator for its trustworthiness and compare our approach with other such indicators.

Abstract: In shared-genome language-model societies, restricted evidence visibility favors reusable, value-indexed latent packet interfaces, whereas the sole high-performing globally visible model in the parent study learned an episode-entangled code. This companion study asks whether independently trained societies share one packet language, where strict zero-shot transfer fails, and whether inherited interface state helps or harms later learning. First, a leakage-controlled causal interoperability audit over all 30 ordered pairs of six independently trained restricted societies -- under sealed held-out structure and a preregistered raw/orthogonal/linear/nonlinear alignment ladder -- shows the six semantically similar interfaces do not form one raw language: one same-initialization pair is exactly interoperable in both directions, a second shows asymmetric partial compatibility, and all 26 cross-initialization directions fail every frozen alignment rung. Second, within the tested decomposition and a single sealed source formulation, a source-span control localizes strict zero-shot failure to interpretation and execution of the new operator instructions. Third, in a matched adaptation factorial, the globally trained communication interface acts as a severe negative-transfer prior: reinitializing only the packet reader, writer, and mouth raises final depth-three accuracy from 0.169 to 0.857. Fourth, across two restricted checkpoints and two independently frozen target streams each, inherited interfaces never exceeded fresh-interface controls by the preregistered 0.10 margin. All primary conclusions are bounded to a near-transfer 17-state setting; the negative-transfer factorial concerns one globally visible parent-cohort checkpoint, while an appendix adds a post hoc tagged-global twin case study.

Abstract: Determinism and numerical reproducibility are increasingly required of GPU kernels in machine learning systems, yet deterministic implementations of the same kernel can still differ bit for bit. Floating-point reduction order is the primary cause, alongside partial-sum precision, fused multiply-add operations, and rounding placement. These choices may be hand-coded, selected by a block-level language such as Triton, or hidden inside a closed-source library such as cuBLAS or rocBLAS. A tile shape chosen for performance therefore also determines the arithmetic, potentially breaking batch invariance. Preserving a fixed order can cost up to 20 percent, while an autotuner cannot identify which configurations are bitwise equivalent. We characterize the factors determining the bitwise behavior of reductions and general matrix multiplication (GEMM). First, we introduce a descriptor of GEMM reduction order, including the partitioning of K in split-K GEMM. Using it, we perform the first black-box reconstruction of a closed-source library's arithmetic for bit-level correctness. Our family of Triton GEMMs matches NVIDIA cuBLAS in all tested cases on Blackwell and Hopper. For realistic LLM shapes with fused epilogues, it matches or exceeds torch.compile performance. Second, we enforce balanced-tree reduction during Triton lowering and introduce a data-layout optimization that brings 19 of 27 kernels on GB300 and H100 within 10 percent of free-order performance. Third, we develop sound static checkers for bitwise equivalence between compiled GPU kernels, including the first checker spanning NVIDIA PTX and AMD GCN. Integrated into Triton's autotuner, the checker restricts search to a single bit-equivalence class.

Abstract: Erasing concept-specific information from representations has been proven useful for mitigating bias or interpreting model decisions. The joint objective is to transform the original representations such that the target concept becomes unpredictable, while maximally preserving concept-unrelated information. In this work, we revisit the optimal bounds of concept erasure to derive a novel class of erasure functions that naturally induce a deterministic, dual counterfactual mapping. Bridging the gap between theoretical optimality and practical representation learning, we design an implementation that imposes a translational bias on counterfactual trajectories - a constraint that aligns with how many concepts geometrically manifest in modern language models. Our framework enables seamless navigation between concept erasure and counterfactual generation. We empirically demonstrate its efficacy in improving downstream algorithmic fairness and generating counterfactual texts.

Abstract: Multimodal Sentiment Analysis (MSA) remains constrained by modality imbalance, yet the field continues to rely on optimization-based balancing methods that promise more than they deliver. We provide three contributions: 1) a unified evaluation framework testing gradient and loss-based balancing strategies under controlled settings; 2) a theoretical diagnosis explaining why these methods fail, as they conflate fitting speed with discriminative contribution; and 3) a research agenda toward held-out discriminative modality valuation. Experiments on CMU-MOSI and CMU-MOSEI reveal three shortcomings: no strategy reliably outperforms Late Concatenation; performance is sensitive to hyperparameters; and even ratio calibration fails to yield consistent gains. The core issue is fundamental: loss is not utility, and gradients are not importance. Modality imbalance remains unresolved, motivating utility estimation from held-out performance.

Abstract: Test-time adaptation (TTA) offers many ways to update a deployed model without labels, but choosing the wrong update can make a strong source model worse. Recent methods therefore try to predict which adaptation will work from unlabeled test data. We ask a prior question: does the evidence given to the selector contain enough information to determine the best action at all? We show that this is not guaranteed, even with a perfect selector. If an observation channel makes two deployments look the same while their TTA rankings differ, reliable selection is impossible from that channel; richer evidence can restore the decision only when it resolves the relevant ambiguity. We make this boundary exact in a finite-batch Gaussian TTA model, where doing nothing beats mean recentering for small shifts, recentering wins beyond a unique critical shift, and the boundary shrinks as $1/\sqrt n$. Public benchmark studies on CIFAR-100-C and DomainNet-126 show the same failure mode with modern TTA methods: changing only deployment structure can reverse the oracle action while global order-blind evidence remains unchanged. The result is a practical way to separate two failure modes that are usually mixed together: a weak selector versus an information channel that cannot support the desired decision in the first place.

Abstract: Applying knowledge transfer across multiple optimization tasks, multitask optimization (MTO) emerges as a promising approach to solving synergistic optimization tasks simultaneously. However, the development of effective knowledge transfer mechanisms in MTO fundamentally relies on aligning elite solution distributions across tasks. This dependency creates a critical bottleneck in few-shot optimization regimes, as restricted evaluation budgets impede the identification of elite solution distributions required for beneficial transfer. This challenge is exacerbated in multiobjective multitask problems, where each optimizer must approximate a continuous Pareto manifold rather than a single optimal point. This paper introduces Iterative Sequential Transfer (IST) to circumvent this bottleneck. We model MTO as a sequence of sequential transfer optimization problems, concentrating evaluations on a single target per iteration. We propose a likelihood-informed task prioritization mechanism to maximize transfer utility by identifying the task most likely ready for knowledge integration. Empirical results on benchmark and real-world problems verify the effectiveness of the proposed method under tight budgets.

Abstract: We study the expressivity of spiking neural networks, which provide a natural framework for asynchronous, event-driven computation complementary to conventional feedforward neural networks. We consider the time-to-first-spike model in a setting for which the input-output map is continuous and piecewise linear, with affine pieces governed by causal feasibility constraints that determine which presynaptic spikes occur before a neuron fires. We first show that each neuron's firing time admits a maxout-like representation with exponentially many, highly constrained affine pieces. We then formalize causal regions as polyhedral regions with fixed causal sets and derive upper and lower bounds on the maximal number of causal regions in both shallow and multilayer feedforward spiking networks. Our theoretical and experimental results show that spiking networks can generate richer partitions of the input space than conventional feedforward ReLU networks.

Abstract: The discovery of the EML operator, sufficient to evaluate the standard explicit purely transcendental elementary functions, has led to considerable interest and discussion across multiple scientific disciplines. However, most authors have focused on the binary EML itself, while numerous similar variants with slightly different properties are now known. This article attempts to close this gap by enumerating and classifying them. We also take this opportunity to clarify common misconceptions related to the EML operator. The principal goal, symbolic regression within an architecture as close as possible to proven neural networks which combine matrix multiplication with a single univariate non-linear activation function, remains beyond reach. Instead, we propose a Möbius layer, with rational functions replacing matrix operations, and showcase the recently discovered activation function eml(x,1/x), which allows exp(x) and ln(x) to be recovered separately, and hence all elementary functions to be evaluated within a rational generalization of the neural network.

Abstract: Soft context compression condenses a context into a few memory tokens that a frozen LLM consumes in place of the raw text, but existing compressors fix the compression ratio at training and inference: each deployed ratio requires a separately trained model, and the chosen ratio is applied uniformly to all inputs, whose actual needs vary drastically. We propose FlexComp, a method-agnostic framework that decouples the ratio from both training and deployment: Matryoshka-style training samples the memory budget $K$ per instance, turning one model into an any-ratio compressor, and the budget is then chosen per input by: (1) confidence-based cascade routing or (2) a lightweight learned $K$ predictor. Across ICAE, 500xCompressor, and SAC on MRQA, a single FlexComp model matches separately trained fixed-ratio specialists with minimal degradation. Cascade routing preserves over 98% of the mildest ratio's accuracy at up to 266x average compression; the $K$ predictor, in a single compression-decoding pass, reaches 158-236x within 0.7 F1 of the mildest ratio. At serving-scale batch sizes, the $K$ predictor cuts context KV cache by 50% and improves decoding throughput by 47%.

Abstract: This paper is the second in a two-part investigation of the physics of information geometry. While Part I develops a physical foundation for distributional motion on the probability simplex, the present paper studies how that framework manifests in active inference. The treatment is fully self-contained and does not require familiarity with Part I. We focus in particular on active inference through small distributional steps and the geometric structure induced by such local motion. Starting from an initial distribution, an agent evolves its belief state toward a final target distribution through a sequence of constrained updates. We define a relative free energy functional with respect to the preferred distribution and extend it to a relative potential energy analogous to the Helmholtz/Gibbs free-energy decomposition. The evolution is subject to a per-step kinetic constraint expressed through the Kullback-Leibler (KL) divergence between consecutive distributions, which serves as a discrete kinetic energy on the probability simplex. Using the information-geometric Pythagorean theorem on KL balls, we show that sufficiently small local moves dominate large direct jumps, and that greedy maximization of free-energy reduction is globally optimal under the kinetic constraint. This leads to a sequential variational principle in which the optimal trajectory minimizes the associated Lagrangian of the optimization problem. Similar to classical mechanics, the Lagrangian takes on the form as the difference between the kinetic and potential terms, establishing a least-action principle for distributional motion on the simplex. The resulting optimal update admits a closed form as an exponentially tilted version of the current distribution toward the preferred distribution, parametrized by an inverse-temperature-like multiplier. We further extend the framework to incorporate state-dependent geodesic...

Abstract: Structured pruning of large language models (LLMs) offers hardware-efficient compression, yet existing methods require calibration data, gradient computation, or large auxiliary policy networks at pruning time. LILA (\emph{Latent-Informed Layer Analysis}) scores neuron importance via the Kolmogorov--Smirnov (KS) distance between empirical singular value distributions of the full and neuron-ablated feed-forward network (FFN) weight matrix, providing a closed-form spectral rule requiring no training, calibration data, or auxiliary network. Without any fine-tuning, LILA surpasses PruneNet (45M-parameter RL policy) by 1.57~pp in zero-shot accuracy on LLaMA-2-7B at 25\% sparsity, and outperforms WikiText-2-calibrated SliceGPT by up to 6.0~pp across all sparsity levels, while preserving the original architecture. After one epoch of LoRA recovery fine-tuning, LILA achieves highly competitive performance, matching the heavily calibrated SliceGPT baseline to within a 0.48~pp margin across LLaMA-2-7B and Phi-2, despite using zero calibration data. A Neural Tangent Kernel analysis confirms a 22$\times$ reduction in functional distortion versus random pruning, providing theoretical grounding for the spectral importance criterion. Finally, extending LILA to dynamically allocate sparsity budgets via KS-scores yields state-of-the-art generative preservation at moderate compression, while uncovering fundamental single-layer architectural bottlenecks at higher compression regimes.

Abstract: Model-generated text is finding its way back into training corpora, and there is plenty of evidence that training on such data over and over collapses output diversity. Prior work has studied the phenomenon itself: which protocols and which data mixtures cause collapse. But different models behave very differently under the same process. We fix one recursive contamination protocol and let 13 publicly released checkpoints form an ecosystem that shares a common corpus for five generations. The unique 4-gram outcome after five generations ranges from 0.187 to 0.940 across checkpoints, a roughly five-fold spread: some models are barely touched, others degenerate into repetitive fragments. Changing the composition of the shared pool or mixing in human text keeps the Spearman correlation of the ordering at 0.91--0.97, and changing the random seed keeps it at 0.93--0.98. Whether a model collapses easily under recursive training is, then, a property of the checkpoint itself, and one that has gone largely unexamined. Parameter scale alone does not explain it, since a three-size ladder within one family is not monotonic in size, and none of the static indicators we tested predicts it either. What does work is cheap: let a model iterate on its own output for two or three generations, and its fragility in the larger ecosystem can be inferred from that alone. Collapse speed also responds to intervention. Tightening top-p, which cuts the low-probability tail at generation time, nearly stops collapse within three generations and stabilizes six checkpoints spanning the whole spectrum together, while data-side filtering slows collapse without stopping it.

Abstract: Data curation is a critical bottleneck in industrial-grade LLM development, where petabyte-scale unstructured corpora are scattered across hundreds of physical tables, feature engineering relies on manual, table-centric pipeline orchestration, and data lineage is largely absent. We present OmniTable as an architecture blueprint for a unified wide-table layer built on Logical Unification, Physical Separation, targeting petabyte-scale LLM data curation and exploration. OmniTable makes four contributions: (1) a unified wide-table abstraction that consolidates multi-source heterogeneous data and thousands of derived features under a single logical schema via logical-physical mapping; (2) declarative feature lifecycle management that automates dependency resolution, execution planning, operator fusion, and lineage tracking, replacing manual pipeline orchestration with a "declare-and-execute" paradigm; (3) an adaptive execution engine with autonomous governance that achieves stable PB-scale feature backfill through heterogeneous compute routing (CPU/GPU), adaptive tuning, UDF-level fault tolerance, and automated storage layout optimization; and (4) hybrid-accelerated data exploration combining a global ID index, transparent OLAP offloading, and background materialized views to deliver second-level point lookups and filtered exports exceeding 20 TB/hour. In production, OmniTable manages over 35 PB of training data across web, code, PDF, and SFT domains, reducing the human-in-the-loop curation cycle from approximately 14 days to approximately 2.5 days (5.6x over the pre-OmniTable production workflow), with consistent feature versioning, auditable lineage, and minimal manual intervention.

Abstract: AI-generated text is flowing back into the training corpora of the next generation of models. Recursive training on it drives model collapse, and recent work extends the setting to many models feeding one another -- but almost always with the market split evenly, while real generative AI is an oligopoly. Concentration raises two worries: fewer, more uniform sources may make collapse faster, and later models may be dragged toward the oligarch's output. We test both in controlled ecosystems: 13 open 1--4B models form natural ecosystems of 3 to 13 players, plus an injected probe that pushes the top share to 90%; each generation, every model's output is mixed into a shared pool by market share and every model is retrained on that pool from clean base weights, for five generations. Yet within the range we test, neither worry materializes; what emerges instead is an invariance. Making the split more unequal barely changes the speed of collapse. Destinations move even less: the share and identity knobs shift five-generation endpoints by only a few percent of the drift common to all arms -- the ecosystems collapse to nearly the same place. An extreme share paired with the strongest injected bias still does not guarantee steering, and the topic shifts it does produce leave only a faint trace on the ruler that measures collapse. What sets the speed is who supplies the pool and how readily those suppliers are carried along: with every share held fixed, swapping the members of a K=3 ecosystem changes five-generation drift by 2.8x; a share-weighted index of each member's susceptibility explains the speed differences across nineteen arms with R^2 = 0.68; and replacing half the pool with human text roughly halves drift without changing its course. Within the tested range, concentration sets neither the destination nor the pace of collapse; the pace follows whose text fills the pool.

Abstract: Accurate posterior prediction need not require accurate approximation of Bayesian updates. We prove that an unbounded gap between the update maps can coexist with vanishing predictive KL for every fixed finite $K\ge2$ in a stationary symmetric Gaussian HMM. Exact Bayesian mixing and an explicit deterministic radial filter act on the same $K-1$ belief coordinates. As $q\to0^+$, their separation in centered logits in the worst case grows at least linearly in the natural confidence scale $L_K(q)$, while their categorical $D_{\mathrm{KL}}(\mathrm{exact}\|\mathrm{radial})$ vanishes at the same explicit witness. Along stationary HMM trajectories, the expected terminal KL between filtered posteriors also converges to zero at $H(q)=\lceil-\log(q)/c\rceil+1$. Typical blocks without switches drive both filters into a common confidence cone, where softmax curvature suppresses their disagreement; a single Gaussian maximal event controls adaptive noise. A sweep with equally spaced Gaussians over $K\in\{2,4,8\}$ illustrates the opposing trends, and binary controls at long horizons compare saturating and nonsaturating recurrences. The result isolates two missing links between internal update gaps and predictive cost: the contribution of separating states to expected loss and decoder sensitivity. Thus even an unbounded internal update gap does not by itself certify predictive failure. The construction is fixed in $K$ and does not provide a universal criterion for when compression is harmless or characterize when internal gaps must incur task loss.

Abstract: Reproducibility is essential to scientific progress, yet the growing volume and complexity of scientific publications make exhaustive manual verification increasingly impractical. Although recent advances in large language model (LLM) agents enable automated experiment reproduction, existing evaluations largely focus on final repositories and are typically limited to machine learning (ML). We introduce AgentActionBench, a process-oriented benchmark for evaluating agent-based experiment reproduction across ML and AI4Science domains. Our framework uses an MCP-based Action Recorder to capture agents' behaviour throughout the reproduction process and evaluates the resulting traces with paper-specific rubrics. AgentActionBench contains 150 papers, including 120 ML papers and 30 AI4Science papers. A human-annotated subset covering 10% of the benchmark provides validation data, while model-assisted augmentation expands the full benchmark to more than 10,000 rubric items. Experimental results show that current systems remain limited, with execution as the primary bottleneck. Meanwhile, the strong Pearson and Spearman correlations between model-generated and human-annotated rubrics validate the reliability of our scalable rubric-generation approach.

Abstract: Benchmark researchers and developers of large language models (LLMs) and other AI systems need to find relevant evaluations, locate their benchmark datasets and code, and understand the settings behind reported scores. We present Benchmark Radar, a living database and search engine for retrieval and discovery of AI benchmarks, covering LLM evaluation, agentic and tool-use benchmarks, coding, reasoning, safety, and domain-specific evaluations. The system combines daily discovery of benchmark papers, repositories, datasets, and releases with a searchable benchmark catalog, mentions in model cards and technical reports, and score histories. It retains source identities and citations so readers can inspect candidate benchmarks and their evaluation evidence. Daily discovery draws on 37 sources: 13 direct connectors and 24 first-party research and engineering feeds. The catalog contains 1,283 source records drawn from 4 benchmark catalogs and 12,916 numeric observations on 790 records. We describe collection and retrieval, audit the full catalog, and examine benchmark saturation, adoption trends, and the limits of score comparisons. A worked example walks through a complete prior-art search, showing how to query the catalog and inspect benchmark evidence when designing a new evaluation. We release the web dashboard with a benchmark leaderboard, a Pareto frontier view of score against measured use, saturation and trend views, daily feeds, downloadable evidence, a command-line interface (CLI) for offline queries, and reproducible analysis.

Abstract: Large language models learn similar behaviours, yet it remains unclear what structure they share or how to change one behaviour without disturbing others. The Fisher-Rao geometry of next-token probabilities connects these questions: behaviour determines this geometry up to output-preserving symmetries, whereas activation geometry depends on coordinates. Across transformer, state-space and recurrent models, output geometries agree more strongly than activation geometries, and shared geometry supports semantic-category transfer. Agreement with human word choices increases with predictive accuracy, scale and training, and improves further after model-only calibration. Token probabilities and read-out geometry jointly predict the spectrum and its effective dimension. Controlled language assignments show that geometry follows the language law across architectures. Pretraining corpus statistics predict held-out fact acquisition without recalibration, while randomised experiments show that deeper evidence substantially delays acquisition across every tested architecture and evidence construction. Finally, the geometry prescribes minimum-disturbance local interventions, predicts their relative cost, and supports reusable control: updates learned on donor prompts transfer to unseen prompts while better preserving behaviour on reference prompts than Euclidean control. The same geometric correction improves steering, editing, attribution, dictionary learning and fine-tuning.

Abstract: Tree-structured rollouts give critic-free reinforcement learning with verifiable rewards (RLVR) step-level credit: fork a chain at an intermediate point, and sibling outcome differences estimate step value. Each fork adds sampling cost, so realistic budgets typically allow only a few forks per chain. A fork placed where the outcome is already largely settled yields siblings that mostly agree and provide almost no credit signal; hence, for a given tree size, where forks are placed largely determines how much step-level RL can gain. Most existing mainstream methods place forks by structure, such as fixed lengths, midpoints, and delimiters, or by next-token entropy. We formalize fork placement as locating the \emph{pivots} of the chain's value curve, where the expected outcome turns. We propose \emph{belief-shift branching}: read the model's answer belief at candidate boundaries and fork just before the step where consecutive beliefs diverge most. Three instantiations, none needing step-level supervision, span access levels: a black-box probe, a logit-lens depth profile, and a learned activation direction, which is fit offline and therefore used only in the validation before RL training. The signal only \emph{places} forks, and the probe costs about $1\%$ of step compute on mathematics and under $5\%$ on code when it runs inside the rollout engine. In that validation, against Monte-Carlo value curves, a belief-shift signal ranks first in each of the eight model$\times$benchmark panels, ahead of entropy, structural, and LLM-judge baselines. In RL across three model families and two domains, belief-shift forking leads every mathematics aggregate, on OLMo-3-7B by $+2.6$ aggregate and $+2.9$ on AIME 2026 over the strongest baseline, and sweeps every OLMo code column, by $+6.5$ on LiveCodeBench-medium.

Abstract: No Free Lunch theorem implies that any performance gains achieved by a classifier on a particular image distribution are necessarily offset by a loss of performance over the set of all possible problems; thus, no single model is universally optimal. Selecting the most suitable classifier for image datasets is a critical yet challenging task due to the intrinsic complexity and diversity of images. This paper proposes a meta-learning framework that leverages a comprehensive set of meta-features capturing dataset complexity to predict classifier performance without exhaustive training. By extracting and selecting features using methods such as autoencoders, pre-trained networks, and dimensionality reduction techniques, we train regression models to efficiently estimate classifier accuracies. Additionally, clustering techniques are employed to group classifiers with similar performance patterns, simplifying the recommendation process. The datasets used span a wide range of concepts, including nature, animals, numbers, motorcycles, medical images, and human bodies, to ensure broad generalization. Evaluated on 56 diverse image datasets, our approach achieves an average ranking prediction accuracy exceeding 86%, demonstrating its effectiveness in guiding model selection. This scalable and interpretable framework provides a practical solution to improve classification performance while reducing computational costs.

Abstract: Unlabeled-unlabeled (UU) learning allows us to learn a binary classifier from two sets of unlabeled data with different class-priors. It is a general framework because it includes a wide variety of supervised learning such as positive-unlabeled (PU) learning, noisy label learning, and similarity-based learning. Existing UU learning assumes that the test and training distributions have the same class-conditional densities. However, this assumption rarely holds in practice due to distribution shifts. This paper proposes a distribution shift adaptation method for UU learning that uses UU data in the training distribution and a few UU data in the test distribution. The proposed method is based on the importance weighting, which minimizes the test risk by using training data with estimated importance weights. Although existing importance weighting methods cannot handle UU data, we show that it can be done in a principled manner. Thanks to the generality of UU learning, our method can handle various learning problems such as PU and noisy label learning under distribution shift within a single framework while existing methods are usually tailored to a specific problem. Moreover, it does not require any assumption of the shift types such as covariate shift. We experimentally demonstrate the effectiveness of the proposed method with real-world datasets.

Abstract: EGGROLL makes evolution strategies (ES) practical for LLMs by replacing dense Gaussian weight perturbations with low-rank Gaussian products, often of rank one. This choice is computationally attractive but geometrically severe: each rank-one perturbation lies in a zero-volume subset of the ambient matrix space, despite having identity covariance. We characterize the mean EGGROLL update field at finite rank and nonzero perturbation radii, then analyze the error of its finite-population estimator. The population field is obtained by applying an explicit resolvent to the gradient of the objective smoothed by the perturbations. We show that the resolvent can introduce a nonconservative component and can reverse the local stability of an optimum. EGGROLL is nevertheless exact on every quadratic objective at every rank and radius. For smooth objectives, its first local finite-rank correction is $O(σ^2/r)$, and nonasymptotic bounds control the resulting field error under smoothness assumptions. Under a local affine model, rank-one perturbations increase the variance of the gradient estimator by only $\frac{2(m+n+1)}{mn+1}$ relative to dense Gaussian ES, or $0.098\%$ for a $4096\times4096$ matrix. We then introduce LOO-ROLL, a leave-one-out estimator that preserves the finite-rank population field while replacing EGGROLL's two antithetic evaluations per direction by one. At equal evaluation cost, LOO-ROLL halves estimator MSE in transformer blocks. At matched wall time across ten post-training settings and models up to 8B parameters, LOO-ROLL improves seven outcomes in individual paired tests, with no significant loss. On the GSM8K test set, accuracy increases from $38.1\%$ to $63.0\%$ at 0.6B and from $65.9\%$ to $80.0\%$ at 8B. Transformer measurements recover the predicted finite-rank variance, while the rank comparisons show no reproducible reward-based advantage for rank eight.

Abstract: Associative memory in the Hopfield network is attractor dynamics in a disordered many-body system, and higher-order and exponential extensions turn its retrieval update into softmax attention. The polynomial and exponential regimes have been analyzed by different methods, with no common architecture in which to ask what fixes the storage scale. In this paper we study the bipartite architecture of Krotov and Hopfield, which we call the class $H$, whose model is fixed by a Lagrangian for each layer, taking the hidden neurons as the order parameter of retrieval. At polynomial load the replica method yields the replica-symmetric phase diagrams and closed-form capacities, and the crosstalk moment is common to Ising and spherical visible neurons, so their differences come from the visible entropy. With a softmax hidden layer the load is exponential, and a copy representation maps the thermodynamics onto random-energy-model counting, with paramagnetic, condensed, and frozen phases. Heating destabilizes retrieval by quantized reassignments of attention, and typical Gaussian patterns remain metastable at every load. The regimes differ in their crosstalk statistics, central-limit at polynomial load and large-deviation at exponential load, and the class $H$ splits retrieval into two roles, the visible Lagrangian fixing stability and the hidden one the storage scale, two axes that may also guide the design of new Lagrangians.

Abstract: Data poisoning corrupts training data to degrade a model or to plant attacker-controlled behavior. This study evaluates two representative training-time attacks, label flipping and backdoor poisoning, on MNIST and Fashion-MNIST with three baseline classifiers: Logistic Regression, Linear SVM, and Random Forest. Clean training is compared with poisoning rates of 5%, 10%, and 20% using clean-test accuracy, macro-precision, macro-recall, macro-F1, and, for backdoors, attack success rate. Label flipping caused clear degradation, largest for Logistic Regression and Linear SVM, while Random Forest stayed comparatively stable. Backdoor poisoning reached attack success rates from 0.9667 to 1.0000 on both datasets and all three models while often keeping clean-test performance near baseline. The results separate indiscriminate poisoning, which shows up in standard metrics, from targeted backdoor poisoning, which stays comparatively stealthy while embedding highly effective malicious behavior, and they support security-oriented evaluation beyond conventional clean-test metrics.

Abstract: Soft positional priors can help small Transformers learn retrieval circuits, but it is unclear whether the resulting circuits remain functional once the prior is removed. We test this with an annealable soft-prior Transformer whose attention biases can be learned, faded, or zeroed during training and evaluation. On associative recall, unforced models perform well with the prior active ($0.772 \pm 0.020$) but collapse at zero gate ($0.095 \pm 0.009$). Smooth fade-to-zero training preserves high zero-gate accuracy ($0.734 \pm 0.028$), whereas forced-zero training, hard switching, and post hoc continuation fail to recover the same effect. The pattern also appears on Markov induction. Linear regression ICL provides a boundary case because zero-gate training can learn that task directly. Mechanistic traces show that circuit consolidation occurs after the gate reaches zero, even though the responsible heads vary across seeds. These results suggest that circuit removability in small discrete retrieval tasks depends on the training trajectory, not just the final architecture.

Abstract: Reinforcement learning over a frozen reasoner has become a common recipe for teaching a policy which external tools to invoke. We show that this recipe becomes structurally mismatched in specialist scientific settings where the complete tool-subset space is enumerable. There, a small set of recurring computational capabilities covers the domain, so the space of tool subsets is combinatorial yet small enough to enumerate, and GRPO still estimates an action expectation from a handful of sampled rollouts. Worse, the approximation degrades as training succeeds: as the policy concentrates on preferred subsets it resamples them, sampled rewards collide, and the group-normalized advantage vanishes. On genomic reasoning the fraction of questions yielding no reward signal rises from 0.2% under a uniform reference policy to 20.8% after GRPO training. As a remedy, we introduce FGPO (Full-Group Policy Optimization), which (1) scores every tool subset and optimizes the exact action expectation, so each update sees the complete action space, and (2) precomputes the reward of each question--subset pair into an exhaustive table, removing frozen-reasoner calls from the training loop entirely. Across five frozen reasoners and three genomic benchmarks, FGPO outperforms GRPO in all 15 settings by 6.75 points on average and up to 14.20, while a standard on-demand GRPO schedule would require 2.4 times as many frozen-reasoner reward evaluations and, on GenomeQA, FGPO cuts invoked tools per question from 2.36 to 1.40.

Abstract: Attias, Hanneke and Ramaswami (NeurIPS 2025) asked whether randomization provably reduces the oracle calls needed for online learning when the class is accessible only through an oracle. We study the instance they singled out: transductive online learning of thresholds on an unknown total order of T instances, with a consistency-type ERM oracle that returns a full concept consistent with a queried labeled set (or reports non-realizability). Our main result is a separation for a fixed natural oracle. When the oracle is the minimal-prefix rule (or the maximal-prefix rule), every deterministic learner makes M mistakes and Q calls with $M+Q\ge T-\varepsilon$ on some instance ($\varepsilon\in\{0,1\}$, according to whether the empty prefix is a concept), and the constant is exact; hence $O(\log T)$ mistakes cost $T-\varepsilon-O(\log T)$ calls, whereas that paper's randomized learner achieves $O(\log T)$ expected calls and mistakes under the same rule. The randomized order is optimal: on an explicit hard distribution under the minimal-prefix rule, every learner has expected mistakes at least $((T+1-\varepsilon)\,128^{-\mathbb{E}[Q]}-1)/2$, so $Ω(\log T)$ expected calls are necessary for polylogarithmic mistakes. The separation is governed by the oracle's selection rule, not by the class alone: for a legal feasible-median ERM rule a deterministic learner achieves $O(\log T)$ calls and mistakes, while a global-median rule again forces linear total cost. The same linear bound holds when the oracle's answers are chosen adversarially and then frozen into a memoryless oracle. We add partial tradeoff results for fixed query budgets (the middle regime is open) and an interface contrast: with only a weak consistency oracle, returning a realizability bit, both deterministic and randomized learners need $Θ(T)$ calls.

Abstract: In-context learning (ICL) is widely used in multimodal large language models (MLLMs) and achieves strong performance across a wide range of multimodal tasks. However, existing multimodal ICL methods often rely on surface level imitation of in-context demonstrations, making it difficult for MLLMs to align their responses with the reasoning path required by the given multimodal input. This limitation becomes more pronounced in complex multimodal tasks, thereby restricting further improvements in MLLM performance. To address this issue, we propose a new multimodal ICL framework that combines contrastive demonstration modeling with the self-refinement capability of MLLMs. Specifically, our framework reformulates each demonstration by explicitly contrasting a suboptimal response with a better response under the same input, together with a reasoning path that reveals how the response should be refined. This contrastive formulation makes the reasoning path toward the desired response more explicit and guides the MLLM beyond superficial imitation. Furthermore, because effective refinement depends on the current response, we introduce a response-conditioned retrieval mechanism to select demonstrations whose reasoning paths are more relevant to the current response. In addition, we use a lightweight alignment controller to predict response quality and determine whether further refinement is needed. Experiments on three types of multimodal tasks show that the proposed framework consistently improves MLLM performance, with particularly notable gains on visual question answering (VQA).

Abstract: Large language models remain fragile against malicious fine-tuning, motivating training-time defenses against harmful persona drift. Preventative Steering injects undesirable-trait persona vectors during fine-tuning and removes them at evaluation time, yet the mechanism behind its lasting protection remains unclear. Analyzing its temporal optimization dynamics, we find that the defense emerges from an early compensatory adaptation phase followed by a steady-state phase where the corrective signal decays; in parameter space, attention output projections emerge as the dominant residual-write route for defensive updates. Through Intervention Delta Preservation (IDP) and IDP Continuation experiments, we further show that preserving or reinjecting the weight offset fails to maintain protection, indicating that preventative steering relies on active adaptation rather than a static defense. Motivated by this finding, we propose Progressive Intensity Scheduling (PIS), which starts with a moderate injection strength and increases it after static-strength alignment begins to decay. Across the evaluated Qwen2.5 and Gemma-3 models, PIS improves safety robustness over static-strength steering while reducing harmful trait expression.

Abstract: Large language and sentence-embedding models provide rich semantic representations, but their high dimensionality poses a challenge for near-term quantum machine learning (QML), where quantum circuits can process only a limited number of input features. We investigate a hybrid quantum-classical pipeline that transforms high-dimensional sentence embeddings into compact representations for variational quantum classification. The workflow combines a pretrained sentence-embedding model, dimensionality reduction, angle encoding, a variational quantum circuit (VQC), and a classical decision layer. We systematically compare principal component analysis (PCA), neighborhood components analysis (NCA), and linear discriminant analysis (LDA), covering both unsupervised and supervised dimensionality reduction. Using the TREC question-classification dataset, we study the relationship between representation dimensionality, information retention, qubit count, and classification performance. Preliminary PCA experiments reveal a strong information bottleneck: reducing 768-dimensional embeddings to 3, 4, 5, and 8 dimensions retains about 8.2%, 10.2%, 11.9%, and 16.4% of the variance, with corresponding classification accuracies of 50.3%, 51.2%, 57.9%, and 63.4%. In contrast, supervised reduction is substantially more efficient. LDA reaches 85.3% accuracy and NCA reaches 83.1% using only 5 dimensions, under a leakage-free cross-validation protocol, compared with 85.1% for a full 384-dimensional classical baseline. These results indicate that supervised dimensionality reduction can preserve task-relevant information far more effectively than variance-based compression, making compact representations a promising route toward practical hybrid quantum-classical NLP models.

Abstract: Preference alignment for flow and diffusion models now spans online reinforcement learning and offline preference optimization, but the relation between these methods remains unclear. In particular, existing forward-process alignment methods require fresh samples from the current model, while offline methods based on fixed preference pairs rely primarily on positive-only fine-tuning or DPO-style likelihood-ratio surrogates. We organize these approaches through a divergence-based framework and introduce FlowCPO, an offline forward-KL objective that uses both preferred and dispreferred samples without online rollouts. For linear interpolation, we show under explicit regularity conditions that the forward-KL objective is bounded by a contrastive flow matching loss, yielding a tractable surrogate on fixed data. We further show that this loss is nonnegative, whereas the signed regression loss of simplified FlowDPO can be unbounded below. In the in-domain setting, FlowCPO achieves higher mean GenEval and OCR scores than the evaluated baselines, reaching 0.84 and 0.87 versus 0.81 and 0.74 for FlowDPO at CFG 3.0. In the out-of-domain setting, the results are mixed, with the best GenEval result but lower reward scores than RFT on several metrics.

Abstract: Federated clients that perform several local optimization steps can return parameter displacements with widely different magnitudes. The quadratic regularization of FedProx grows linearly with displacement and therefore offers limited control over the contrast between ordinary and unusually large client movements. We here introduce HiFedProx, which replaces the quadratic penalty with a scale-matched power-type regularizer indexed by $p\geq2$. All powers have the same regularization-gradient magnitude at a reference displacement $R$, while every $p>2$ gives a weaker response below $R$ and a stronger response above it. An exact affine reference calculation shows that increasing $p$ compresses relative displacement disparities, although very large powers approach fixed-radius behavior and increase local curvature. HiFedProx combines this geometry with finite-budget stochastic client optimization and same-minibatch Armijo backtracking. In paired five-seed experiments on a frozen 60-writer FEMNIST subset, a common-parameter study over $p\in\{2,3,4,5,6,7,8\}$ shows similar clean-training performance but substantial gains under composite stress. The lowest moderate- and severe-stress losses occur at $p=7$ and $p=6$, improving over $p=2$ by $11.44\%$ and $23.16\%$, respectively. Although displacement-tail ratios continue to decrease through $p=8$, predictive performance peaks in an intermediate range and Armijo trial cost increases with $p$. These results indicate that the exponent should be calibrated rather than maximized. In our experiments, $p=5$--$7$ provides the most useful range.

Abstract: Large language models often repeatedly select the same answer even when many alternatives are plausible. Prior work treats this concentration as bias, but it does not distinguish stable model preferences from responses that depend on a particular prompt wording. We introduce a bias depth score that measures both how strongly a model prefers its top answer under direct prompting and whether that answer survives scenario reframing. Across 4,442 opinion prompts and four large language models, only about a quarter of the concentrated preferences survive reframing. We call these persistent cases Deep biases, and the remaining prompt-dependent cases Shallow biases. Our results show that Deep biases are more often inherited from pretraining and preserved through SFT. Under both continued fine-tuning and prompt-based debiasing for diversity, Deep biases are consistently harder to remove than Shallow biases. Bias depth therefore separates stable learned biases from prompt-wording artifacts that single-prompt metrics conflate. Code, models, and data are available at deepbias.github.io.

Abstract: Depth pruning reduces large language model (LLM) inference cost by removing complete Transformer blocks. Activation-based methods collect hidden states through forward passes on calibration data, while existing forward-free methods score each Transformer block separately without measuring similarity between blocks. We propose Weight-Redundancy Pruning (WRP), a forward-free depth-pruning method that estimates inter-layer redundancy from checkpoint weights to select blocks without calibration data or model forward passes. WRP compares attention output and MLP down-projection weights across layers and combines their pairwise similarities with relative projection-scale information. The resulting all-pairs similarity matrix guides layer grouping and block selection. Across multiple pruning settings, model families, and downstream tasks, WRP consistently outperforms existing forward-free magnitude pruning and approaches the performance of activation-based methods.

Abstract: Post-training quantizers select finite code changes using reconstruction proxies or local loss approximations, but the utility of a quantization move depends on the state through which it is executed. We identify two sources of this contextual dependence. First, the displacement of the move matters: evaluating the gradient at the move midpoint captures curvature accumulated along the move that a current-state linearization omits. Across frozen two-bit moves from Llama-3.2 models, midpoint evaluation predicts the direction of exact endpoint loss changes substantially more accurately than current-state gradients. Second, moves interact: exhaustive lattices of legal quantized states are well approximated by quadratic pseudo-Boolean functions, yet their small pairwise components can determine Pareto fronts and cause different evaluation functionals to prefer opposite directions. These effects explain failures of reconstruction-optimal code re-selection and additive composition. Reading each move at its own midpoint repairs the local selection step and improves downstream accuracy and held-out perplexity, while larger supports require evaluating exact endpoints from the state actually reached. Exact-endpoint beam search finds sparse changes that dominate much larger one-shot updates, and repricing the same moves after intervening changes produces widespread sign reversals. These results show that quantization utility is contextual at the granularity of a few moves: reliable construction must evaluate finite changes along their own paths and compose them from the evolving quantized state.

Abstract: Balanced k-shot sampling draws exactly k labeled examples per class. We show that it induces an exact, provable degeneracy in a family of small-sample discriminant estimators. Under balanced sampling, the within-class scatter operator of Kernelized Linear Principal Component Discriminant Analysis (KLPCDA) is not merely rank-deficient but exactly a scaled orthogonal projector. We derive the consequences in closed form: two of KLPCDA's seven variants have every signal eigenvalue exactly equal, so their eigenvector selection criterion is provably indifferent rather than ill-conditioned, and a third has a provably void objective. This follows from the estimators' construction, not any dataset; we confirm it on frozen sentence embeddings and, separately, on residual-stream activations from a decoder-only generative model. An in-formula tie-break repairs the two repairable variants, with recovery gated by class count: the residual subspace constraint costs 5x more on few-class than many-class datasets (p=0.000001). We then evaluate the repaired framework on few-shot text classification on frozen LLM embeddings (n much smaller than d, up to 4096), across four datasets, three embedding sizes, and three trained baselines (SetFit, LoRA, in-context learning). A properly cross-validated logistic-regression probe still beats every KLPCDA variant on three of four datasets, at every embedding size; guidance carried from pixel, vibration-signal, and gene-expression data does not directly generalize to this feature space. Three independent geometric separability metrics fail to explain why one high-dimensional decoder-based embedding model underperforms smaller bidirectional encoders, ruling out anisotropy; the gap is substantially an estimation-efficiency effect, not a permanent ceiling, closing by more than 80% when the support set grows from k<=10 to k=30-50 (p=0.00195, both many-class datasets).

Abstract: Low-bit quantization can achieve high recall on some vector representations and fail sharply on others, while average distortion and global rank correlation do not explain the difference. We study quantized vector search at the level of the comparisons consumed by ranking and graph-pruning algorithms. Our first result is a distribution-free decomposition: the probability that a comparison flips is bounded by the probability mass of exact margins near zero plus the tail probability of the calibrated residual. We then account for dependence between residuals that share a query or graph node, and derive covariance-aware second-moment identities and tail bounds under a joint MGF proxy. For a frozen candidate permutation, we prove a deterministic coupling theorem for Vamana neighbour selection: the approximate replay returns the exact neighbour list exactly when all candidate-level pruning actions agree on the frozen exact states. We connect these results to representation geometry through an exact Gaussian oracle, establish a strict correlation gain from a deterministic magnitude bit in an aligned bilinear model, and give a rare-contamination construction showing why marginal Gaussian diagnostics do not imply the required residual tails. When analytical assumptions are unavailable, a held-out block certificate bounds the selective failure risk of a frozen quantized rule. Across learned, classical, and synthetic embeddings, standardized exact margins predict held-out ranking and pruning flip rates substantially better than global rank correlation. The framework applies to coordinate binary codes, RaBitQ, Lucene BBQ, and product quantizers through a common decision interface.

Abstract: The DIRECT algorithm is a deterministic global optimization method known for its versatility and balanced exploration-exploitation strategy. However, DIRECT-type algorithms are primarily effective for low-dimensional problems and often exhibit slow convergence as dimensionality increases, limiting their applicability to more complex optimization tasks. To address this limitation, this paper introduces X-DTC-GL, a novel DIRECT-type algorithm that incorporates dynamic partitioning and hybridization techniques. The dynamic partitioning approach adaptively refines the search space based on local one-dimensional surrogate models, enabling rapid subdivision of promising hyper-rectangles. The hybridization strategy selectively employs a hill-climbing method to exploit promising regions identified by the surrogate models. Extensive experiments on four diverse benchmark suites demonstrate that X-DTC-GL significantly outperforms existing DIRECT-type baselines, achieving improvements of ~12% in solvability and ~27% in solution quality. Performance-profile analyses indicate the fastest convergence on up to ~40% of instances, the best runtime performance on ~17% of problems, and competitive overall execution times. By improving performance within the partition-based framework, these advances strengthen the algorithm's competitiveness in state-of-the-art black-box optimization.

Abstract: Diffusion language models generate text by refining a fixed-size block of token positions through many forward passes, a loop that does not match the per-forward execution unit used by most LLM serving systems. A dense MoE runtime binds all work to the refinement-iteration clock: it rebuilds similar routing structure on every forward, recomputes expert outputs for positions whose logits are already dead, and sends those positions through dense expert-parallel collectives. This paper presents \sys{}, a serving system that treats the diffusion block as a compilation unit. \sys{} compiles a small \emph{block plan} for the block-clock structure of one diffusion block and refreshes every value that can affect a live decode decision on the iteration clock. \sys{} realizes this plan along three dense axes of an MoE forward: \atlas{} compiles a coverage-driven active expert support per layer while recomputing gate logits every iteration; \lsp{} keeps full sequence shards as model state but routes only live, newly decoded, and refresh-required positions through fresh routed-expert computation; \freshlane{} carries this fresh token--expert worklist through expert-parallel dispatch, kernels, and combine, then restores the dense logical shard at the layer boundary. We implement \sys{} on 8 NVIDIA H100 GPUs and evaluate it on three open-weight block-diffusion MoE models (LLaDA-MoE, LLaDA2.0-mini, and LLaDA2.0-Flash, spanning 7B to 100B total parameters) across GSM8K, HumanEval, MGSM, and MT-Bench. \sys{} improves end-to-end execution time by up to 2.7$\times$ over the strongest surviving baseline under the same 8-GPU placement and remains feasible at the largest batch sizes where multiple baselines run out of memory, while preserving task quality relative to the dense reference.

Abstract: Supervised fine-tuning (SFT) applies a uniform cross-entropy loss to all target tokens, even though different tokens provide unequal learning signals for mathematical reasoning. This uniform treatment can over-sharpen already mastered tokens while amplifying learning pressure on uncertain, low-confidence tokens, leading to suboptimal training dynamics. We propose Trimmed Logit-Gap SFT (TrimSFT), a simple token-level reweighting method that scales the SFT loss according to the logit gap between the gold token and its strongest competitor. TrimSFT trims supervision away from both extremes: tokens already mastered (large logit gap) and tokens weakly supported by the current model (small or negative logit gap), concentrating learning within an intermediate logit-gap region between them. We instantiate this principle with a Gaussian weight centered at margin m with bandwidth τ, requiring no reference model or additional forward pass. We evaluate TrimSFT on six base models from the Llama, Qwen, and DeepMath families across five mathematical reasoning benchmarks. TrimSFT consistently improves over standard SFT, achieving the best average performance on five out of six models, with gains of up to +26.9 points over SFT on MATH500. Further analyses show that the bandwidth τ matters more than the exact margin location, and that half-trim variants that remove supervision pressure from only one side yield inferior trade-offs. A token-level logit-gap distribution analysis suggests that TrimSFT reshapes model confidence in a more balanced way than uniform SFT or monotonic reweighting methods. These results suggest that reasoning SFT can benefit from trimming both extremes rather than treating all tokens uniformly.

Abstract: Candidate decision correctness and rationale grounding are different objectives. We examine correctness-gated multi-teacher distillation in a fixed experiment. Eight arms share 4,330 sources, a 63.9M-parameter student, 12,990 optimization rows, 406 updates, evidence inputs, and a decoder; seven teacher-based arms use one fixed three-response pool. Three seeds are evaluated on 267 held-out examples. Relative to unfiltered distillation, the correctness-weighted arm differed in accuracy by +0.1660 (95% observed-matrix interval [0.0670, 0.2455]), five-label macro-F1 by +0.1323 ([0.0916, 0.1731]), and task-defined conditional unsafe-action rate by -0.4979 ([-0.5926, -0.3686]). These shifts do not imply uniformly better behavior. Source-label SFT had the highest mean macro-F1 (0.586). The weighted arm had zero Refuted recall in every seed, and two seeds assigned NotEnoughInfo to all 167 claim examples. In an availability-amended audit at one reference seed, weighted and unfiltered outputs had 0/20 versus 1/20 evidence-supported positives and 20/20 versus 19/20 positives containing unsupported material. Samples were non-paired, source overlap was not serialized, and the amendment followed automatic summarization but preceded annotation. The audit therefore cannot estimate a common-source grounding effect and is inconclusive about system-level improvement or harm. Hard filtering already achieved 0.660 accuracy, 0.530 macro-F1, and 0.135 conditional unsafe rate. The implemented weighted arm showed no demonstrated incremental decision benefit over hard filtering. This fixed-matrix failure analysis shows decision redistribution with lost label functionality; the available human audit does not establish a grounding gain.

Abstract: We study fast algorithms for sparse-PCA certification. Given a positive semidefinite matrix $M$, the problem asks either to rule out a large $k$-sparse quadratic form or to return a high-value (relaxed) witness. The standard semidefinite relaxation provides such certificates, but existing general-purpose solvers require $Ω(d^4)$ time. We give a bicriteria algorithm running in $O(d^2+d k^{O(\log k)})$ time: if some $k$-sparse unit vector has quadratic form greater than $2$, it returns either an $O(k^2)$-sparse unit vector or an SDP-feasible matrix of value at least $1$. For $k\leq\exp(O(\sqrt{\log d}))$, this running time is $O(d^2)$. We also go below the quadratic barrier in the sample-access model: Given $n=d^{o(1)}$ samples, our algorithm obtains a related one-sided certificate in $d^{2 - Ω(1)}$ time for $k=\mathrm{polylog}(d)$, without forming the empirical covariance matrix. As an application, these certificate routines yield the first quadratic and subquadratic-time algorithms for robust sparse estimation for broad families of distributions. Our sparse-PCA algorithm reduces a high-value sparse direction to a bounded-radius set in the graph of large correlations and searches the resulting candidate supports. The subquadratic implementation constructs this graph using fast correlation detection.