arXiv subjects

Alex Iacob

Publications and source records attributed to Alex Iacob.

At least 19 recordsLinked to original sources

arXiv2026

Learning to Read Out: Unembedding Dynamics in Language Model Pretraining

When does a language model acquire a capability? Developmental studies usually answer this by tracking when models start producing the correct token. But that prediction passes through the unembedding matrix, or readout, which is itself still being learned during pretraining. We show that a model's hidden states can encode task information before its own readout expresses it, so timelines read from outputs conflate what the model encodes with what its readout expresses. To track the readout directly, we introduce parameter-trajectory crosscoding, which fits one sparse dictionary to every token's readout row at every checkpoint. Unlike crosscoders on activations, it needs only checkpoint weights. We use it to locate a narrow early window, 0.4-1.1% of the way through Pythia's pretraining, in which the readout reorganizes and WordNet word categories become decodable from its rows. By then, linear probes on the hidden states already decode task information that the model's own outputs do not yet show. To test whether such gaps depend on how fast the readout learns, we pretrain small models from scratch that differ only in the readout's learning rate. We find that raising it makes the readout develop earlier and the gap close sooner, with little effect on final loss. Together, these results imply that early in training, where many developmental transitions are dated, a behavioral date can reflect when the readout learns to express task information and not when the model learns it.

cs.CL↗

arXiv2026

The Red Queen Gödel Machine: Co-Evolving Agents and Their Evaluators

Self-improving agents are state-of-the-art on agentic coding benchmarks, yet their search methods assume a stationary evaluation criterion. This ignores a central feature of evolution: species adapt as their environments change with them. We introduce the Red Queen Gödel Machine (RQGM), an evolutionary framework for recursive self-improvement under non-stationary utilities. This allows learned evaluators to improve alongside the agents they guide. On DeepSWE, the RQGM improves over its fixed-evaluator baseline by adding a complementary agent-as-a-judge code-review signal: a co-evolved reviewer grades coder patches to guide search. At low reasoning effort, the RQGM coder passes 82.1% of held-out tasks against the baseline's 75.0%, and nearly matches the GPT-6 Astra model at high effort. In scientific paper writing and reviewing, and Olympiad-level proof writing and grading, co-evolved evaluators provide an evaluation criterion. Anchored to human IMO grades, a co-evolved grader writes its own milestone rubric and exceeds static baselines at a 3x lower search cost, driving the prover to the best mean score. Since the RQGM can modify the search objective across epochs, it can regularize the search. For example, the RQGM reduces self-preference bias via an additional adversarial objective to discover reviewers equally stringent on AI and human work. Guided by these calibrated reviewers, co-evolved writers reach 1.78x-1.86x higher acceptance rates than the baseline under an agent-as-a-judge panel. The RQGM enables self-improving systems where agents and evaluators recursively bootstrap each other beyond static evaluation.

cs.LG↗

arXiv2026

FoMoE: Breaking the Full-Replica Barrier with a Federation of MoEs

Pre-training Large Language Models (LLMs) typically demands large-scale infrastructure with tightly coupled hardware accelerators. Mixture-of-Experts (MoEs) architectures partially decouple model capacity from per-token compute. This efficiency alone does not make MoE training feasible over ordinary Internet links or loosely connected commodity hardware since active expert routing still assumes high-speed datacenter fabrics. Low-communication methods such as DiLoCo and Photon reduce synchronization frequency across distributed sites, mitigating bandwidth constraints, yet still require full model replicas at every site. This creates a mismatch: modern MoEs have sparse data paths, but their distributed training infrastructure remains communication-dense and memory-inefficient, limiting attempts to pool geographically distributed compute. In this work, we introduce FoMoE, a system that breaks the full-replica paradigm by partitioning expert layers across workers and skipping non-resident experts during local training. We demonstrate that FoMoE: (I) reduces communication costs by up to 1.42x over efficient baselines and 45.44x over Distributed Data Parallelism (DDP) via partial expert replication in controlled regimes; (II) achieves empirical throughput speedups of up to 1.4x through the skip-token mechanism; and (III) shows stable routing in the trained regimes and projects the communication/memory benefits to 100B-scale configurations through system modeling.

cs.LG↗

arXiv2026

LoRDO: Distributed Low-Rank Optimization with Infrequent Communication

Distributed training of foundation models via $\texttt{DDP}$ is limited by interconnect bandwidth. While infrequent communication strategies reduce synchronization frequency, they remain bottlenecked by the memory and communication requirements of optimizer states. Low-rank optimizers can alleviate these constraints; however, in the local-update regime, workers lack access to the full-batch gradients required to compute low-rank projections, which degrades performance. We propose $\texttt{LoRDO}$, a principled framework unifying low-rank optimization with infrequent synchronization. We first demonstrate that, while global projections based on pseudo-gradients are theoretically superior, they permanently restrict the optimization trajectory to a low-rank subspace. To restore subspace exploration, we introduce a full-rank quasi-hyperbolic update. $\texttt{LoRDO}$ achieves near-parity with low-rank $\texttt{DDP}$ in language modeling and downstream tasks at model scales of $125$M--$720$M, while reducing communication by $\approx 10 \times$. Finally, we show that $\texttt{LoRDO}$ improves performance even more in very low-memory settings with small rank/batch size.

cs.LG↗

arXiv2026

Photon: Federated LLM Pre-Training

Scaling large language models (LLMs) demands extensive data and computing resources, which are traditionally constrained to data centers by the high-bandwidth requirements of distributed training. Low-bandwidth methods like federated learning (FL) could enable collaborative training of larger models across weakly-connected GPUs if they can effectively be used for pre-training. To achieve this, we introduce Photon, the first complete system for federated end-to-end LLM training, leveraging cross-silo FL for global-scale training with minimal communication overheads. Using Photon, we train the first federated family of decoder-only LLMs from scratch. We show that: (1) Photon can train model sizes up to 7B in a federated fashion while reaching an even better perplexity than centralized pre-training; (2) Photon model training time decreases with available compute, achieving a similar compute-time trade-off to centralized; and (3) Photon outperforms the wall-time of baseline distributed training methods by 35% via communicating 64x-512xless. Our proposal is robust to data heterogeneity and converges twice as fast as previous methods like DiLoCo. This surprising data efficiency stems from a unique approach combining small client batch sizes with extremely high learning rates, enabled by federated averaging's robustness to hyperparameters. Photon thus represents the first economical system for global internet-wide LLM pre-training.

cs.LG↗

arXiv2026

Response-Conditioned Parallel-to-Sequential Orchestration for Multi-Agent Systems

Multi-agent systems can solve complex tasks through collaboration between multiple Large Language Model agents. Existing collaboration frameworks typically operate in either a parallel or a sequential mode. In the parallel mode, agents respond independently to queries followed by aggregation of responses. In contrast, sequential systems allow agents to communicate via a directed topology and refine one another step by step. However, both modes are inadequate for achieving the desired objectives of minimizing communication and latency while simultaneously maximizing the accuracy of the final response. In this work, we introduce a hybrid paradigm called Nexa, a trainable response-conditioned policy that bridges the gap between the two modes. Nexa begins with a parallel execution stage, embeds the resulting responses into a shared semantic space, and then predicts a sparse directed acyclic communication graph. If the graph is empty, the system remains purely parallel; if it is non-empty, the system performs one sequential message propagation. The policy is a lightweight transformer model, and the method avoids the need for external LLM judges or reward models, as well as hand-crafted test-time topology search. We formalize this hybrid execution problem, show that the resulting graph is acyclic by construction, and that the framework strictly subsumes pure parallel execution, and present a training procedure based on policy-gradient optimization. Results demonstrate that the response-conditioned policy learned by Nexa under one setting can be reused when the number of agents, the task, or the underlying agent changes, thus emphasizing the generalizability of the learned communication policy.

cs.CL↗

arXiv2026

Rethinking Data Curation in LLM Training: Online Reweighting Offers Better Generalization than Offline Methods

Data curation is a critical yet under-explored area in large language model (LLM) training. Existing methods, such as data selection and mixing, operate in an offline paradigm, detaching themselves from training. This separation introduces engineering overhead and makes the curation brittle: the entire pipeline must be re-run under model/task shifts. Moreover, offline methods alter data size through hard filtering or resampling, often sacrificing data diversity and harming generalization. We propose to rethink data curation as an online reweighting problem, where sample importance is dynamically adjusted during training via loss weighting rather than static pre-processing. Specifically, we introduce ADAPT (Adaptive Data reweighting for Pretraining and FineTuning), a dynamic online framework that reweights training samples with adaptive per-sample learning rates guided by similarity-based quality signals, without changing the number of training samples. Unlike offline methods that enforce a static data distribution, ADAPT acts as an implicit curriculum learner, progressively shifting focus from coarse-grained patterns to fine-grained semantic distinctions as the model evolves. Experiments on both instruction tuning and large-scale pretraining show that ADAPT consistently outperforms offline selection/mixing and prior online methods, achieving stronger cross-benchmark generalization under equal FLOPs.

cs.LG↗

arXiv2025

LLM Unlearning via Neural Activation Redirection

The ability to selectively remove knowledge from LLMs is highly desirable. However, existing methods often struggle with balancing unlearning efficacy and retain model utility, and lack controllability at inference time to emulate base model behavior as if it had never seen the unlearned data. In this paper, we propose LUNAR, a novel unlearning method grounded in the Linear Representation Hypothesis and operates by redirecting the representations of unlearned data to activation regions that expresses its inability to answer. We show that contrastive features are not a prerequisite for effective activation redirection, and LUNAR achieves state-of-the-art unlearning performance and superior controllability. Specifically, LUNAR achieves between 2.9x and 11.7x improvement in the combined unlearning efficacy and model utility score (Deviation Score) across various base models and generates coherent, contextually appropriate responses post-unlearning. Moreover, LUNAR effectively reduces parameter updates to a single down-projection matrix, a novel design that significantly enhances efficiency by 20x and robustness. Finally, we demonstrate that LUNAR is robust to white-box adversarial attacks and versatile in real-world scenarios, including handling sequential unlearning requests.

cs.LG↗

arXiv2025

MT-DAO: Multi-Timescale Distributed Adaptive Optimizers with Local Updates

Training large models with distributed data parallelism (DDP) requires frequent communication of gradients across workers, which can saturate bandwidth. Infrequent communication strategies (e.g., Local SGD) reduce this overhead but, when applied to adaptive optimizers, often suffer a performance gap relative to fully synchronous DDP. We trace this gap to a time-scale mismatch: the optimizer's fast-moving momentum, tuned for frequent updates, decays too quickly to smooth gradients over long intervals, leading to noise-dominated optimization. To address this, we propose MT-DAO, a family of optimizers that employs multiple slow- and fast-moving first momenta or the gradient to track update dynamics across different time scales, for which we provide the first convergence guarantees. Empirically, for language-model pre-training, this eliminates the performance gap with DDP, outperforming infrequent-communication baselines in perplexity and reducing iso-token wall-clock time by 6-27% on Ethernet interconnects. At the 720M scale, MT-DAO reaches a target perplexity in 24% fewer steps and 35% less time than the single-momentum DDP baseline. MT-DAO enables effective cross-datacenter training and training over wide geographic areas.

cs.LG↗

arXiv2025

AbbIE: Autoregressive Block-Based Iterative Encoder for Efficient Sequence Modeling

We introduce the Autoregressive Block-Based Iterative Encoder (AbbIE), a novel recursive generalization of the encoder-only Transformer architecture, which achieves better perplexity than a standard Transformer and allows for the dynamic scaling of compute resources at test time. This simple, recursive approach is a complement to scaling large language model (LLM) performance through parameter and token counts. AbbIE performs its iterations in latent space, but unlike latent reasoning models, does not require a specialized dataset or training protocol. We show that AbbIE upward generalizes (ability to generalize to arbitrary iteration lengths) at test time by only using 2 iterations during train time, far outperforming alternative iterative methods. AbbIE's ability to scale its computational expenditure based on the complexity of the task gives it an up to \textbf{12\%} improvement in zero-shot in-context learning tasks versus other iterative and standard methods and up to 5\% improvement in language perplexity. The results from this study open a new avenue to Transformer performance scaling. We perform all of our evaluations on model sizes up to 350M parameters.

cs.LG↗

arXiv2025

DES-LOC: Desynced Low Communication Adaptive Optimizers for Training Foundation Models

Scaling foundation model training with Distributed Data Parallel (DDP) methods is bandwidth-limited. Existing infrequent communication methods like Local SGD were designed to synchronize only model parameters and cannot be trivially applied to adaptive optimizers due to additional optimizer states. Current approaches extending Local SGD either lack convergence guarantees or require synchronizing all optimizer states, tripling communication costs. We propose Desynced Low Communication Adaptive Optimizers (DES-LOC), a family of optimizers assigning independent synchronization periods to parameters and momenta, enabling lower communication costs while preserving convergence. Through extensive experiments on language models of up to 1.7B, we show that DES-LOC can communicate 170x less than DDP and 2x less than the previous state-of-the-art Local ADAM. Furthermore, unlike previous heuristic approaches, DES-LOC is suited for practical training scenarios prone to system failures. DES-LOC offers a scalable, bandwidth-efficient, and fault-tolerant solution for foundation model training.

cs.LG↗

arXiv2025

DEPT: Decoupled Embeddings for Pre-training Language Models

Language Model pre-training uses broad data mixtures to enhance performance across domains and languages. However, training on such heterogeneous text corpora requires extensive and expensive efforts. Since these data sources vary significantly in lexical, syntactic, and semantic aspects, they cause negative interference or the ``curse of multilinguality''. To address these challenges we propose a communication-efficient pre-training framework, DEPT. Our method decouples embeddings from the transformer body while simultaneously training the latter on multiple data sources without requiring a shared vocabulary. DEPT can: (1) train robustly and effectively under significant data heterogeneity, (2) minimize token embedding parameters to only what the data source vocabulary requires, while cutting communication costs in direct proportion to both the communication frequency and the reduction in parameters, (3) enhance transformer body plasticity and generalization, improving both average perplexity (up to 20%) and downstream task performance, and (4) enable training with custom optimized vocabularies per data source. We demonstrate DEPT's potential via the first vocabulary-agnostic federated pre-training of billion-scale models, reducing communication costs by orders of magnitude and embedding memory by 4-5x.

cs.LG↗

arXiv2025

SparsyFed: Sparse Adaptive Federated Training

Sparse training is often adopted in cross-device federated learning (FL) environments where constrained devices collaboratively train a machine learning model on private data by exchanging pseudo-gradients across heterogeneous networks. Although sparse training methods can reduce communication overhead and computational burden in FL, they are often not used in practice for the following key reasons: (1) data heterogeneity makes it harder for clients to reach consensus on sparse models compared to dense ones, requiring longer training; (2) methods for obtaining sparse masks lack adaptivity to accommodate very heterogeneous data distributions, crucial in cross-device FL; and (3) additional hyperparameters are required, which are notably challenging to tune in FL. This paper presents SparsyFed, a practical federated sparse training method that critically addresses the problems above. Previous works have only solved one or two of these challenges at the expense of introducing new trade-offs, such as clients' consensus on masks versus sparsity pattern adaptivity. We show that SparsyFed simultaneously (1) can produce 95% sparse models, with negligible degradation in accuracy, while only needing a single hyperparameter, (2) achieves a per-round weight regrowth 200 times smaller than previous methods, and (3) allows the sparse masks to adapt to highly heterogeneous data distributions and outperform all baselines under such conditions.

cs.LG↗

arXiv2024

The Future of Large Language Model Pre-training is Federated

Generative pre-trained large language models (LLMs) have demonstrated impressive performance over a wide range of tasks, thanks to the unprecedented amount of data they have been trained on. As established scaling laws indicate, LLMs' future performance improvement depends on the amount of computing and data sources they can leverage for pre-training. Federated learning (FL) has the potential to unleash the majority of the planet's data and computational resources, which are underutilized by the data-center-focused training methodology of current LLM practice. Our work presents a robust, flexible, reproducible FL approach that enables large-scale collaboration across institutions to train LLMs. We propose a scalable deployment system called Photon to enable the investigation and development of this new training paradigm for LLM pre-training. We show that Photon can be used by organizations interested in collaborating with their private data sources and computational resources for pre-training LLMs with billions of parameters. This paradigm would mobilize more computational and data resources while matching or potentially exceeding centralized performance. We further show the effectiveness of the federated training scales with model size and present our approach for training billion-scale federated LLMs using limited resources. Thus far, we have used Photon to train LLM models to the size of 7B parameters and anticipate larger models being completed in the near future. Finally, we show that LLM training is highly resilient to the classical challenges of federated statistical and hardware heterogeneity. Furthermore, we show that convergence is robust to partial participation, opening the avenue for compute-efficient collaborative training. Photon will help data-rich actors to become the protagonists of LLMs pre-training instead of leaving the stage to compute-rich actors alone.

cs.LG↗

arXiv2024

Worldwide Federated Training of Language Models

The reliance of language model training on massive amounts of computation and vast datasets scraped from potentially low-quality, copyrighted, or sensitive data has come into question practically, legally, and ethically. Federated learning provides a plausible alternative by enabling previously untapped data to be voluntarily gathered from collaborating organizations. However, when scaled globally, federated learning requires collaboration across heterogeneous legal, security, and privacy regimes while accounting for the inherent locality of language data; this further exacerbates the established challenge of federated statistical heterogeneity. We propose a Worldwide Federated Language Model Training~(WorldLM) system based on federations of federations, where each federation has the autonomy to account for factors such as its industry, operating jurisdiction, or competitive environment. WorldLM enables such autonomy in the presence of statistical heterogeneity via partial model localization by allowing sub-federations to attentively aggregate key layers from their constituents. Furthermore, it can adaptively share information across federations via residual layer embeddings. Evaluations of language modeling on naturally heterogeneous datasets show that WorldLM outperforms standard federations by up to $1.91\times$, approaches the personalized performance of fully local models, and maintains these advantages under privacy-enhancing techniques.

cs.LG↗

arXiv2024

Pollen: High-throughput Federated Learning Simulation via Resource-Aware Client Placement

Federated Learning (FL) is a privacy-focused machine learning paradigm that collaboratively trains models directly on edge devices. Simulation plays an essential role in FL adoption, helping develop novel aggregation and client sampling strategies. However, current simulators cannot emulate large-scale systems in a time-efficient manner, which limits their utility and casts doubts on generalizability. This work proposes Pollen, a novel resource-aware system for speeding up simulations. Pollen addresses two limiting factors from existing simulators: (a) communication inefficiency derived from pull-based client execution and (b) inadequate load balance when using heterogeneous hardware. Pollen executes high-throughput FL simulations at scale by (a) using a push-based client placement system, (b) learning how an adaptable scheduling of clients based on hardware statistics (c) estimating the optimal number of concurrent workers per GPU. We evaluate Pollen on four representative FL tasks and show that Pollen's placement model increases GPU utilization and reduces idle time. We compare Pollen to Flower, Flute, FedScale, Parrot, and pfl and show experimental speed-ups of days or weeks.

cs.DC↗

arXiv2024

FedAnchor: Enhancing Federated Semi-Supervised Learning with Label Contrastive Loss for Unlabeled Clients

Federated learning (FL) is a distributed learning paradigm that facilitates collaborative training of a shared global model across devices while keeping data localized. The deployment of FL in numerous real-world applications faces delays, primarily due to the prevalent reliance on supervised tasks. Generating detailed labels at edge devices, if feasible, is demanding, given resource constraints and the imperative for continuous data updates. In addressing these challenges, solutions such as federated semi-supervised learning (FSSL), which relies on unlabeled clients' data and a limited amount of labeled data on the server, become pivotal. In this paper, we propose FedAnchor, an innovative FSSL method that introduces a unique double-head structure, called anchor head, paired with the classification head trained exclusively on labeled anchor data on the server. The anchor head is empowered with a newly designed label contrastive loss based on the cosine similarity metric. Our approach mitigates the confirmation bias and overfitting issues associated with pseudo-labeling techniques based on high-confidence model prediction samples. Extensive experiments on CIFAR10/100 and SVHN datasets demonstrate that our method outperforms the state-of-the-art method by a significant margin in terms of convergence rate and model accuracy.

cs.LG↗

arXiv2023

Privacy in Multimodal Federated Human Activity Recognition

Human Activity Recognition (HAR) training data is often privacy-sensitive or held by non-cooperative entities. Federated Learning (FL) addresses such concerns by training ML models on edge clients. This work studies the impact of privacy in federated HAR at a user, environment, and sensor level. We show that the performance of FL for HAR depends on the assumed privacy level of the FL system and primarily upon the colocation of data from different sensors. By avoiding data sharing and assuming privacy at the human or environment level, as prior works have done, the accuracy decreases by 5-7%. However, extending this to the modality level and strictly separating sensor data between multiple clients may decrease the accuracy by 19-42%. As this form of privacy is necessary for the ethical utilisation of passive sensing methods in HAR, we implement a system where clients mutually train both a general FL model and a group-level one per modality. Our evaluation shows that this method leads to only a 7-13% decrease in accuracy, making it possible to build HAR systems with diverse hardware.

cs.LG↗