arXiv subjects

Josif Grabocka

Publications and source records attributed to Josif Grabocka.

At least 19 recordsLinked to original sources

arXiv2026

FondueQA: A Benchmark for Evaluating Parametric Compensation in Multi-Hop QA

Large Language Models (LLMs) show remarkable performance based on parametric knowledge, i.e., their ability to recall facts and associations internalized in a model's weights. However, studying this capability on established multi-hop benchmarks may be complicated by benchmark contamination, as many of these benchmarks have been publicly available long enough to plausibly appear in modern pretraining corpora. As a result, correct answers under missing evidence may reflect either general parametric knowledge or memorization of benchmark-specific instances. To study how models use internal knowledge when retrieval evidence is incomplete, we introduce FondueQA, in which every required multi-hop fact is uniquely localized to a designated evidence document. This enables controlled document-level evidence ablations to study parametric compensation under incomplete retrieval. Our empirical evaluation reveals that parametric compensation strongly depends on both entity popularity and the structural role of missing evidence. Generally, models compensate more successfully for facts involving popular entities. On FondueQA, we show that bridging evidence can be more useful than the answer-bearing document itself. Moreover, models may recall a missing fact in isolation but fail to use it under partial evidence, particularly when the context contains plausible competing entities, which can reduce Exact Match by up to 15 percentage points.

cs.LG↗

arXiv2026

Muffin - A Shared Memory Buffer for Asynchronous AI Agents

The rise of Agentic AI systems has changed the landscape of artificial intelligence by enabling strong reasoning capabilities through specialized multi-agent interaction with structured self-reflection. However, in target-driven reasoning tasks, single- and multi-agent systems are often on par in terms of performance gains. A major challenge is that failures often arise from an ineffective exchange of critical information during inter-agent interaction and coordination. Different from existing multi-agent debate paradigms that rely on synchronous turn-taking and sharing information across different agents, we propose an inter-agent protocol based on shared Memory bUFFer INference - referred to as MUFFIN - in an asynchronous agentic framework. In the first step, an orchestrator routes queries based on their reasoning demands and delegates demanding ones to a multi-agent setting, where a shared memory buffer provides asynchronous exchange of intermediate findings while an agent's private reasoning remains local. Our experiments show that our framework Muffin outperforms established competitors, achieving the best accuracy in 17 out of 20 of our benchmark settings, and provides a better quality-cost trade-off than multi-agent debate: Muffin improves accuracy by 19.2 accuracy points while using 77.4% fewer tokens, without requiring full-trace sharing.

cs.AI↗

arXiv2026

MuPACE: Learning to Pace Muon for LoRA Fine-Tuning

Learning rate schedules are typically fixed before training begins, yet when fine-tuning large language models with low-rank adapters (LoRA), the appropriate scale varies substantially with the model, task, and adapter configuration. Practitioners re-tune a schedule for every new problem, yet a tuned schedule still cannot react to the training dynamics it encounters. We instead amortize learning rate selection across LoRA fine-tuning problems. We leave the base optimizer, Muon, untouched, so that its normalized update fixes the direction, and meta-learn only the global step size. Our method MuPACE is a small recurrent policy, meta-trained across a distribution of fine-tuning runs, that conditions on the observed trajectory and emits bounded adjustments to an interpretable analytic schedule. On 64 fine-tuning problems, MuPACE outperforms standard fixed schedules at a strong shared default on 64-78% of problems and is on par with the same schedules with per-task tuned peaks. Meta-trained only on models with up to 16 layers, MuPACE transfers one-shot to unseen tasks on models as deep as 64 layers, such as Qwen3-32B and Qwen3.8-27B, where a single run remains competitive with swept schedules.

cs.LG↗

arXiv2026

FTFM: Factorized Tabular Foundation Models

Tabular foundation models (TFMs) are becoming a powerful alternative to conventional tabular learning, adapting to new datasets in context without task-specific training. Current architectures, however, face a trade-off: models that preserve cell-level representations are expensive because they repeatedly attend across rows and columns, while faster models collapse each row into a fixed embedding and lose this fine-grained state. We introduce Factorized Tabular Foundation Models (FTFM), which preserve cell representations while factorizing their computation through one contextual latent per row and one per column. Cells no longer attend to one another directly; instead, each cell is updated through the interaction of its row and column latents, reducing repeated attention while retaining feature-specific information. FTFM achieves competitive predictive performance in tabular classification on major benchmarks such as TabArena and BeyondArena.

cs.LG↗

arXiv2026

High-Throughput LLM-Based Compression and What It Takes: A Systems Study of Inference Engines and Entropy Coding Optimizations

Large language models are attractive text compressors because next-token probabilities can be converted into code lengths. In practice, however, the main bottleneck is often not modeling quality but systems cost: a compressor must obtain logits quickly enough to compete with strong classical baselines. This paper examines systems optimizations for LLM-based compression, including multistream pipelined arithmetic coding and target-interval capture, and evaluates their trade-offs across inference engines, configurations and datasets. We evaluate LLM-based text compression on three different inference engines: Transformers, vLLM, and TensorRT-LLM. Notably, we obtain a speedup of 18.84x compared to prior work's Transformers implementation and a maximum recorded inference speed of 181 kilobytes per second at a compression ratio of 7.17, when running on an NVIDIA A100.

cs.DB↗

arXiv2026

TabMF: Learning Dataset Representations with Tabular Foundation Models

Representations of supervised datasets support higher-level machine-learning tasks such as dataset retrieval and distribution-shift monitoring. However, existing approaches typically rely either on hand-crafted meta-features or on dataset encoders trained specifically for representation learning. We introduce TabMF, a framework that repurposes pretrained tabular foundation models (TFMs) as encoders of supervised datasets. TabMF equips a predictive TFM with an architecture-aware dataset-level readout and adapts the model end-to-end using contrastive post-training on stochastic views of synthetic supervised datasets. We instantiate TabMF on TabICLv2 and TabPFN-2.5 and evaluate the resulting representations across six tasks on unseen real datasets. Simple frozen TFM readouts yield competitive representations, indicating that predictive pretraining already induces substantial dataset-level structure. The two TabMF encoders obtain the strongest aggregate performance among the evaluated learned and classical dataset representations, with particularly strong results in dataset retrieval, compact benchmark construction, and distribution-shift severity assessment.

cs.LG↗

arXiv2026

End-to-End Compression for Tabular Foundation Models

The long-standing dominance of gradient-boosted decision trees for tabular data has recently been challenged by in-context learning tabular foundation models. In-context learning methods fit and predict in one forward pass without parameter updates by leveraging the training data as context for predicting on query test points. While recent tabular foundation models achieve state-of-the-art performance, their transformer architecture based on the attention mechanism has quadratic complexity regarding dataset size, which in turn increases the overhead on training and inference time, and limits the capacity of the models to handle large-scale datasets. In this work, we propose TACO, an end-to-end tabular compression model that compresses the training dataset in a latent space. We test our method on the TabArena benchmark, where our proposed method is up to 94x faster in inference time, while consuming up to 97\% less memory compared to the state-of-the-art tabular transformer architecture, all while retaining performance without significant degradation. Lastly, our method not only scales better with increased dataset sizes, but it also achieves better performance compared to other baselines.

cs.LG↗

arXiv2026

ValueGround: Evaluating Culture-Conditioned Visual Value Grounding in MLLMs

Cultural values are expressed not only through language but also through visual scenes and everyday social practices. Yet existing evaluations of cultural values in language models are almost entirely text-only, leaving it unclear whether culture-conditioned judgments remain stable when response options are visualized. We introduce ValueGround, a benchmark for evaluating culture-conditioned visual value grounding in multimodal large language models (MLLMs). Built from World Values Survey questions, ValueGround uses minimally contrastive image pairs to represent opposing response options while controlling irrelevant variation. Given a country, a question, and an image pair, a model must choose the image that best matches the country's value tendency without access to the original response-option texts. Experiments across six MLLMs and 13 countries show that models perform substantially worse with visualized response options than with the original textual options, with average accuracy dropping from 72.8% to 62.6%. Our benchmark provides a controlled testbed for studying cross-modal transfer of culture-conditioned value judgments.

cs.CL↗

arXiv2026

When is Warmstarting Effective for Scaling Language Models?

Model growth from a given checkpoint aims to accelerate training of a larger model, offering potential resource savings. Despite recent interest, warmstarting has seen limited practical adoption in large-scale training. We attribute this to two underexplored factors: (1) an overemphasis on preserving the smaller model's performance at initialization, which constrains operator design for new architectures, and (2) insufficient analysis of how growth interacts with hyperparameters and scaling behavior, compounded by inconsistent growth factors across the literature. We show that preserving the base model's initial post-growth performance is not necessary for strong final performance, and that simple, architecture-agnostic growth strategies can outperform more complex warmstarting operators. Crucially, we empirically identify an upper bound on the growth factor $g$ beyond which training from scratch is more efficient. We observe this across multiple ablation setups. Notably, this limit is also present, but unreported, in prior published results. Across our experiments on dense MLPs and dense language models, we find that a $2\times$ growth factor is the most reliable in yielding convergence speedups, with gains most pronounced under 20 tokens/parameter budgets and diminishing as budget increases. We fit scaling laws over these observations to provide predictive guidance for practitioners deciding when and how much to grow. Together, our analysis provides practical guidelines and empirical limits for model growth.

cs.LG↗

arXiv2026

POP: Prior-Fitted First-Order Optimization Policies

Gradient-based optimizers are highly sensitive to design choices in their adaptive learning rate mechanisms. To address this limitation, we introduce POP, a meta-learned Reinforcement Learning (RL) policy that predicts adaptive learning rates for gradient descent, conditioned on the contextual information provided in the optimization trajectory. Our method introduces a novel RL reward formulation, a new function-scaling strategy for in-distribution generalization, and a novel prior that is used to sample millions of synthetic optimization problems. We evaluate POP on an established benchmark including 43 optimization functions of various complexity, where it significantly outperforms gradient-based methods. Our evaluation demonstrates strong generalization capabilities without task-specific tuning.

cs.LG↗

arXiv2026

Learning to Order: Task Sequencing as In-Context Optimization

Task sequencing (TS) is one of the core open problems in Deep Learning, arising in a plethora of real-world domains, from robotic assembly lines to autonomous driving. Unfortunately, prior work has not convincingly demonstrated the generalization ability of meta-learned TS methods to solve new TS problems, given few initial demonstrations. In this paper, we demonstrate that deep neural networks can meta-learn over an infinite prior of synthetically generated TS problems and achieve a few-shot generalization. We meta-learn a transformer-based architecture over datasets of sequencing trajectories generated from a prior distribution that samples sequencing problems as paths in directed graphs. In a large-scale experiment, we provide ample empirical evidence that our meta-learned models discover optimal task sequences significantly quicker than non-meta-learned baselines.

cs.LG↗

arXiv2025

Zhyper: Factorized Hypernetworks for Conditioned LLM Fine-Tuning

Large Language Model (LLM) conditioning refers to instructing an LLM to generate content in accordance with the norms and values of a specific culture, beliefs of a particular political orientation, or any desired text-specified semantic conditioning. Unfortunately, prompt engineering does not ensure that LLMs behave in accordance with a desired conditioning due to the inductive bias of the pre-training and alignment datasets. Prior works have focused on fine-tuning LLMs by directly conditioning the LoRA weights; however, such methods introduce a large number of parameters. As a remedy, we propose Zhyper, a parameter-efficient factorized hypernetwork framework that generates context-aware LoRA adapters from textual descriptions. Experiments on multiple benchmarks show that Zhyper achieves competitive performance with up to 26x fewer parameters than the state-of-the-art baselines. Furthermore, we extend Zhyper to cultural alignment, demonstrating improved generalization to out-of-domain settings and a better capturing of fine-grained contextual values.

cs.CL↗

arXiv2025

Tabular Data: Is Deep Learning all you need?

Tabular data represent one of the most prevalent data formats in applied machine learning, largely because they accommodate a broad spectrum of real-world problems. Existing literature has studied many of the shortcomings of neural architectures on tabular data and has repeatedly confirmed the scalability and robustness of gradient-boosted decision trees across varied datasets. However, recent deep learning models have not been subjected to a comprehensive evaluation under conditions that allow for a fair comparison with existing classical approaches. This situation motivates an investigation into whether recent deep-learning paradigms outperform classical ML methods on tabular data. Our survey fills this gap by benchmarking seventeen state-of-the-art methods, spanning neural networks, classical ML and AutoML techniques. Our empirical results over 68 diverse datasets from a well-established benchmark indicate a paradigm shift, where Deep Learning methods outperform classical approaches.

cs.LG↗

arXiv2025

Regularized Neural Ensemblers

Ensemble methods are known for enhancing the accuracy and robustness of machine learning models by combining multiple base learners. However, standard approaches like greedy or random ensembling often fall short, as they assume a constant weight across samples for the ensemble members. This can limit expressiveness and hinder performance when aggregating the ensemble predictions. In this study, we explore employing regularized neural networks as ensemble methods, emphasizing the significance of dynamic ensembling to leverage diverse model predictions adaptively. Motivated by the risk of learning low-diversity ensembles, we propose regularizing the ensembling model by randomly dropping base model predictions during the training. We demonstrate this approach provides lower bounds for the diversity within the ensemble, reducing overfitting and improving generalization capabilities. Our experiments showcase that the regularized neural ensemblers yield competitive results compared to strong baselines across several modalities such as computer vision, natural language processing, and tabular data.

cs.LG↗

arXiv2025

Multi-objective Differentiable Neural Architecture Search

Pareto front profiling in multi-objective optimization (MOO), i.e., finding a diverse set of Pareto optimal solutions, is challenging, especially with expensive objectives that require training a neural network. Typically, in MOO for neural architecture search (NAS), we aim to balance performance and hardware metrics across devices. Prior NAS approaches simplify this task by incorporating hardware constraints into the objective function, but profiling the Pareto front necessitates a computationally expensive search for each constraint. In this work, we propose a novel NAS algorithm that encodes user preferences to trade-off performance and hardware metrics, yielding representative and diverse architectures across multiple devices in just a single search run. To this end, we parameterize the joint architectural distribution across devices and multiple objectives via a hypernetwork that can be conditioned on hardware features and preference vectors, enabling zero-shot transferability to new devices. Extensive experiments involving up to 19 hardware devices and 3 different objectives demonstrate the effectiveness and scalability of our method. Finally, we show that, without any additional costs, our method outperforms existing MOO NAS methods across a broad range of qualitatively different search spaces and datasets, including MobileNetV3 on ImageNet-1k, an encoder-decoder transformer space for machine translation and a decoder-only space for language modelling.

cs.LG↗

arXiv2024

Warmstarting for Scaling Language Models

Scaling model sizes to scale performance has worked remarkably well for the current large language models paradigm. The research and empirical findings of various scaling studies led to novel scaling results and laws that guides subsequent research. High training costs for contemporary scales of data and models result in a lack of thorough understanding of how to tune and arrive at such training setups. One direction to ameliorate the cost of pretraining large models is to warmstart the large-scale training from smaller models that are cheaper to tune. In this work, we attempt to understand if the behavior of optimal hyperparameters can be retained under warmstarting for scaling. We explore simple operations that allow the application of theoretically motivated methods of zero-shot transfer of optimal hyperparameters using μTransfer. We investigate the aspects that contribute to the speedup in convergence and the preservation of stable training dynamics under warmstarting with μTransfer. We find that shrinking smaller model weights, zero-padding, and perturbing the resulting larger model with scaled initialization from μP enables effective warmstarting of $\mut{}$.

cs.LG↗

arXiv2024

Interpretable Mesomorphic Networks for Tabular Data

Even though neural networks have been long deployed in applications involving tabular data, still existing neural architectures are not explainable by design. In this paper, we propose a new class of interpretable neural networks for tabular data that are both deep and linear at the same time (i.e. mesomorphic). We optimize deep hypernetworks to generate explainable linear models on a per-instance basis. As a result, our models retain the accuracy of black-box deep networks while offering free-lunch explainability for tabular data by design. Through extensive experiments, we demonstrate that our explainable deep networks have comparable performance to state-of-the-art classifiers on tabular data and outperform current existing methods that are explainable by design.

cs.LG↗

arXiv2024

Ensembling Finetuned Language Models for Text Classification

Finetuning is a common practice widespread across different communities to adapt pretrained models to particular tasks. Text classification is one of these tasks for which many pretrained models are available. On the other hand, ensembles of neural networks are typically used to boost performance and provide reliable uncertainty estimates. However, ensembling pretrained models for text classification is not a well-studied avenue. In this paper, we present a metadataset with predictions from five large finetuned models on six datasets, and report results of different ensembling strategies from these predictions. Our results shed light on how ensembling can improve the performance of finetuned text classifiers and incentivize future adoption of ensembles in such tasks.

cs.CL↗