arXiv subjects

Dimitrios Dimitriadis

Publications and source records attributed to Dimitrios Dimitriadis.

At least 19 recordsLinked to original sources

arXiv2026

Do VLMs Read or Rewrite? On Transcription Faithfulness in Vision-Language Models

Vision Language Models (VLMs) are increasingly used in place of traditional OCR pipelines for document understanding. In this paper, we show they do not always act as faithful transcribers: when text is imperfect, they often tend to rewrite it into a more plausible form - a behavior that clean-text OCR benchmarks cannot detect. We introduce FaithC4, a multilingual perturbation benchmark of 1,455 single-page documents (English, Chinese, Korean) with three perturbation families: scramble, random substitution, and visually similar substitution. We use the benchmark to evaluate 15 systems spanning general-purpose VLMs, OCR-specialized VLMs, and traditional OCR pipelines. These three categories differ in WER degradation under perturbation: general-purpose VLMs degrade by up to 6.9 points, OCR-specialized VLMs by 0.1-3.4 points, and traditional OCR by less than 0.8 points on English. Probing Qwen3-VL-4B layer-by-layer, we identify a consistent pattern: rewriting fires only when a perturbed word's final layer FFN representation stays close to the original encoding; when the representation diverges sufficiently, the model transcribes faithfully. Word length affects rewriting rate: short words (4-6 characters) are rewritten up to 10% of the time, with a sharp cutoff at 8 characters above which rewriting drops to 0%.

cs.AI↗

arXiv2026

SkillEvolBench: Benchmarking the Evolution from Episodic Experience to Procedural Skills

Large language model (LLM) agents accumulate rich episodic trajectories while solving real-world tasks, but it remains unclear whether such experience can be distilled into reusable procedural skills. We introduce SkillEvolBench, a diagnostic benchmark for evaluating this step from experience reuse to skill formation. It contains 180 tasks across six real-world agent environments, organized into role-conditioned task families with shared latent procedures. Agents learn from acquisition tasks, update an external skill library using compacted trajectories and verifier feedback, and then face frozen deployment tasks testing context shift, adversarial shortcuts, and composition. By comparing self-generated and curated-start skill evolution against no-skill and raw-trajectory controls, SkillEvolBench separates procedural abstraction from base capability, curated prior knowledge, and direct reuse of episodic traces. Across ten model configurations and three agent harnesses, we find that current agents often adapt locally but rarely form robust reusable skills. Skill-based conditions can improve acquisition or replay, and individual models sometimes gain on specific deployment axes, but these gains are unstable under frozen deployment. Raw-trajectory reuse frequently outperforms distilled skills, suggesting that current abstraction procedures discard contextual and procedural cues that remain useful for future tasks. Capacity and cost analyses further show that writing more skills or larger Tier-3 resource libraries is not sufficient: additional updates can improve coverage while introducing episode-specific drift and procedural clutter. These findings position SkillEvolBench as a testbed for measuring when one-off experience becomes durable procedural knowledge rather than task-local memory.

cs.AI↗

arXiv2026

PIVOT: Bridging Planning and Execution in LLM Agents via Trajectory Refinement

Large language model (LLM)-based agents frequently generate seemingly coherent plans that fail upon execution due to infeasible actions, constraint violations, and compounding errors over extended horizons. PIVOT (Plan-Inspect-eVOlve Trajectories) addresses this plan-execution misalignment through a self-supervised framework that treats trajectories as optimizable objects iteratively refined via environment interaction. The framework comprises four stages: PLAN generates candidate trajectories; INSPECT executes them and computes structured losses with textual gradients encoding plan-execution discrepancies; EVOLVE applies these signals to produce improved trajectories; and VERIFY performs a final global check against task constraints. A monotonic acceptance process ensures a non-decreasing solution quality. Empirical evaluations on DeepPlanning and GAIA demonstrate state-of-the-art performance: with human-in-the-loop (HITL) feedback, PIVOT establishes a strong upper bound up to 94% relative improvement in constraint satisfaction, while its fully autonomous variant retains substantial gains, showing that the core trajectory-refinement mechanism remains effective without external supervision. At the same time, PIVOT remains computationally efficient, requiring up to 3x to 5x fewer tokens than competing refinement methods. These findings establish that (self- or human-supervised) feedback-based trajectory optimization is a principled methodology for mitigating plan-execution gaps in autonomous agent systems.

cs.AI↗

arXiv2026

MMDeepResearch-Bench: A Benchmark for Multimodal Deep Research Agents

Deep Research Agents (DRAs) generate citation-rich reports via multi-step search and synthesis, yet existing benchmarks mainly target text-only settings or short-form multimodal QA, missing end-to-end multimodal evidence use. We introduce MMDeepResearch-Bench (MMDR-Bench), a benchmark of 140 expert-crafted tasks across 21 domains, where each task provides an image-text bundle to evaluate multimodal understanding and citation-grounded report generation. Compared to prior setups, MMDR-Bench emphasizes report-style synthesis with explicit evidence use, where models must connect visual artifacts to sourced claims and maintain consistency across narrative, citations, and visual references. We further propose a unified, interpretable evaluation pipeline: Formula-LLM Adaptive Evaluation (FLAE) for report quality, Trustworthy Retrieval-Aligned Citation Evaluation (TRACE) for citation-grounded evidence alignment, and Multimodal Support-Aligned Integrity Check (MOSAIC) for text-visual integrity, each producing fine-grained signals that support error diagnosis beyond a single overall score. Experiments across 25 state-of-the-art models reveal systematic trade-offs between generation quality, citation discipline, and multimodal grounding, highlighting that strong prose alone does not guarantee faithful evidence use and that multimodal integrity remains a key bottleneck for deep research agents.

cs.CV↗

arXiv2025

Routing-Based Continual Learning for Multimodal Large Language Models

Multimodal Large Language Models (MLLMs) struggle with continual learning, often suffering from catastrophic forgetting when adapting to sequential tasks. We introduce a routing-based architecture that integrates new capabilities while robustly preserving foundational knowledge. While Multi-Task Learning (MTL) offers a theoretical performance upper bound, it incurs a linearly scaling computational overhead as the number of tasks increases. In contrast, our method maintains fixed data and compute requirements regardless of the task sequence length. Across models ranging from 2B to 8B parameters, we demonstrate that our routing approach performs on par with MTL while retaining the training efficiency of sequential fine-tuning. Beyond merely mitigating forgetting, we observe that token-level routing facilitates cross-modal transfer, leveraging knowledge from one modality to bolster performance in another. Ablation studies confirm the approach's scalability: routing remains robust even with large expert pools and effectively capitalizes on task relatedness. Finally, we show that our method scales favorably, with larger models exhibiting minimal degradation compared to fully specialized fine-tuning.

cs.LG↗

arXiv2025

Leveraging Uncertainty Estimation for Efficient LLM Routing

Deploying large language models (LLMs) in edge-cloud environments requires an efficient routing strategy to balance cost and response quality. Traditional approaches prioritize either human-preference data or accuracy metrics from benchmark datasets as routing criteria, but these methods suffer from rigidity and subjectivity. Moreover, existing routing frameworks primarily focus on accuracy and cost, neglecting response quality from a human preference perspective. In this work, we propose the Confidence-Driven LLM Router, a novel framework that leverages uncertainty estimation to optimize routing decisions. To comprehensively assess routing performance, we evaluate both system cost efficiency and response quality. In particular, we introduce the novel use of LLM-as-a-Judge to simulate human rating preferences, providing the first systematic assessment of response quality across different routing strategies. Extensive experiments on MT-Bench, GSM8K, and MMLU demonstrate that our approach outperforms state-of-the-art routing methods, achieving superior response quality while maintaining cost efficiency.

cs.NI↗

arXiv2024

CroMo-Mixup: Augmenting Cross-Model Representations for Continual Self-Supervised Learning

Continual self-supervised learning (CSSL) learns a series of tasks sequentially on the unlabeled data. Two main challenges of continual learning are catastrophic forgetting and task confusion. While CSSL problem has been studied to address the catastrophic forgetting challenge, little work has been done to address the task confusion aspect. In this work, we show through extensive experiments that self-supervised learning (SSL) can make CSSL more susceptible to the task confusion problem, particularly in less diverse settings of class incremental learning because different classes belonging to different tasks are not trained concurrently. Motivated by this challenge, we present a novel cross-model feature Mixup (CroMo-Mixup) framework that addresses this issue through two key components: 1) Cross-Task data Mixup, which mixes samples across tasks to enhance negative sample diversity; and 2) Cross-Model feature Mixup, which learns similarities between embeddings obtained from current and old models of the mixed sample and the original images, facilitating cross-task class contrast learning and old knowledge retrieval. We evaluate the effectiveness of CroMo-Mixup to improve both Task-ID prediction and average linear accuracy across all tasks on three datasets, CIFAR10, CIFAR100, and tinyImageNet under different class-incremental learning settings. We validate the compatibility of CroMo-Mixup on four state-of-the-art SSL objectives. Code is available at \url{https://github.com/ErumMushtaq/CroMo-Mixup}.

cs.CV↗

arXiv2024

Do Not Design, Learn: A Trainable Scoring Function for Uncertainty Estimation in Generative LLMs

Uncertainty estimation (UE) of generative large language models (LLMs) is crucial for evaluating the reliability of generated sequences. A significant subset of UE methods utilize token probabilities to assess uncertainty, aggregating multiple token probabilities into a single UE score using a scoring function. Existing scoring functions for probability-based UE, such as length-normalized scoring and semantic contribution-based weighting, are designed to solve certain aspects of the problem but exhibit limitations, including the inability to handle biased probabilities and complex semantic dependencies between tokens. To address these issues, in this work, we propose Learnable Response Scoring (LARS) function, a novel scoring function that leverages supervised data to capture complex dependencies between tokens and probabilities, thereby producing more reliable and calibrated response scores in computing the uncertainty of LLM generations. Our comprehensive experiments across question-answering and arithmetical reasoning tasks with various datasets demonstrate that LARS significantly outperforms existing scoring functions, achieving improvements of up to 16\% AUROC score.

cs.CL↗

arXiv2024

Can Synthetic Audio From Generative Foundation Models Assist Audio Recognition and Speech Modeling?

Recent advances in foundation models have enabled audio-generative models that produce high-fidelity sounds associated with music, events, and human actions. Despite the success achieved in modern audio-generative models, the conventional approach to assessing the quality of the audio generation relies heavily on distance metrics like Frechet Audio Distance. In contrast, we aim to evaluate the quality of audio generation by examining the effectiveness of using them as training data. Specifically, we conduct studies to explore the use of synthetic audio for audio recognition. Moreover, we investigate whether synthetic audio can serve as a resource for data augmentation in speech-related modeling. Our comprehensive experiments demonstrate the potential of using synthetic audio for audio recognition and speech-related modeling. Our code is available at https://github.com/usc-sail/SynthAudio.

cs.SD↗

arXiv2024

MARS: Meaning-Aware Response Scoring for Uncertainty Estimation in Generative LLMs

Generative Large Language Models (LLMs) are widely utilized for their excellence in various tasks. However, their tendency to produce inaccurate or misleading outputs poses a potential risk, particularly in high-stakes environments. Therefore, estimating the correctness of generative LLM outputs is an important task for enhanced reliability. Uncertainty Estimation (UE) in generative LLMs is an evolving domain, where SOTA probability-based methods commonly employ length-normalized scoring. In this work, we propose Meaning-Aware Response Scoring (MARS) as an alternative to length-normalized scoring for UE methods. MARS is a novel scoring function that considers the semantic contribution of each token in the generated sequence in the context of the question. We demonstrate that integrating MARS into UE methods results in a universal and significant improvement in UE performance. We conduct experiments using three distinct closed-book question-answering datasets across five popular pre-trained LLMs. Lastly, we validate the efficacy of MARS on a Medical QA dataset. Code can be found https://github.com/Ybakman/LLM_Uncertainity.

cs.CL↗

arXiv2023

Local or Global: Selective Knowledge Assimilation for Federated Learning with Limited Labels

Many existing FL methods assume clients with fully-labeled data, while in realistic settings, clients have limited labels due to the expensive and laborious process of labeling. Limited labeled local data of the clients often leads to their local model having poor generalization abilities to their larger unlabeled local data, such as having class-distribution mismatch with the unlabeled data. As a result, clients may instead look to benefit from the global model trained across clients to leverage their unlabeled data, but this also becomes difficult due to data heterogeneity across clients. In our work, we propose FedLabel where clients selectively choose the local or global model to pseudo-label their unlabeled data depending on which is more of an expert of the data. We further utilize both the local and global models' knowledge via global-local consistency regularization which minimizes the divergence between the two models' outputs when they have identical pseudo-labels for the unlabeled data. Unlike other semi-supervised FL baselines, our method does not require additional experts other than the local or global model, nor require additional parameters to be communicated. We also do not assume any server-labeled data or fully labeled clients. For both cross-device and cross-silo settings, we show that FedLabel outperforms other semi-supervised FL baselines by $8$-$24\%$, and even outperforms standard fully supervised FL baselines ($100\%$ labeled data) with only $5$-$20\%$ of labeled data.

cs.LG↗

arXiv2023

Learning to Specialize: Joint Gating-Expert Training for Adaptive MoEs in Decentralized Settings

Mixture-of-Experts (MoEs) achieve scalability by dynamically activating subsets of their components. Yet, understanding how expertise emerges through joint training of gating mechanisms and experts remains incomplete, especially in scenarios without clear task partitions. Motivated by inference costs and data heterogeneity, we study how joint training of gating functions and experts can dynamically allocate domain-specific expertise across multiple underlying data distributions. As an outcome of our framework, we develop an instance tailored specifically to decentralized training scenarios, introducing \textit{Dynamically Decentralized Orchestration of MoEs} or \texttt{DDOME}. \texttt{DDOME} leverages heterogeneity emerging from distributional shifts across decentralized data sources to specialize experts dynamically. By integrating a pretrained common expert to inform a gating function, \texttt{DDOME} achieves personalized expert subset selection on-the-fly, facilitating just-in-time personalization. We empirically validate \texttt{DDOME} within a Federated Learning (FL) context: \texttt{DDOME} attains from 4\% up to an 24\% accuracy improvement over state-of-the-art FL baselines in image and text classification tasks, while maintaining competitive zero-shot generalization capabilities. Furthermore, we provide theoretical insights confirming that the joint gating-experts training is critical for achieving meaningful expert specialization.

cs.LG↗

arXiv2023

GPT-FL: Generative Pre-trained Model-Assisted Federated Learning

In this work, we propose GPT-FL, a generative pre-trained model-assisted federated learning (FL) framework. At its core, GPT-FL leverages generative pre-trained models to generate diversified synthetic data. These generated data are used to train a downstream model on the server, which is then fine-tuned with private client data under the standard FL framework. We show that GPT-FL consistently outperforms state-of-the-art FL methods in terms of model test accuracy, communication efficiency, and client sampling efficiency. Through comprehensive ablation analysis across various data modalities, we discover that the downstream model generated by synthetic data plays a crucial role in controlling the direction of gradient diversity during FL training, which enhances convergence speed and contributes to the notable accuracy boost observed with GPT-FL. Also, regardless of whether the target data falls within or outside the domain of the pre-trained generative model, GPT-FL consistently achieves significant performance gains, surpassing the results obtained by models trained solely with FL or synthetic data. The code is available at https://github.com/AvestimehrResearchGroup/GPT-FL.

cs.LG↗

arXiv2023

Counterfactual Augmentation for Multimodal Learning Under Presentation Bias

In real-world machine learning systems, labels are often derived from user behaviors that the system wishes to encourage. Over time, new models must be trained as new training examples and features become available. However, feedback loops between users and models can bias future user behavior, inducing a presentation bias in the labels that compromises the ability to train new models. In this paper, we propose counterfactual augmentation, a novel causal method for correcting presentation bias using generated counterfactual labels. Our empirical evaluations demonstrate that counterfactual augmentation yields better downstream performance compared to both uncorrected models and existing bias-correction methods. Model analyses further indicate that the generated counterfactuals align closely with true counterfactuals in an oracle setting.

cs.LG↗

arXiv2022

Federated Multilingual Models for Medical Transcript Analysis

Federated Learning (FL) is a novel machine learning approach that allows the model trainer to access more data samples, by training the model across multiple decentralized data sources, while data access constraints are in place. Such trained models can achieve significantly higher performance beyond what can be done when trained on a single data source. As part of FL's promises, none of the training data is ever transmitted to any central location, ensuring that sensitive data remains local and private. These characteristics make FL perfectly suited for large-scale applications in healthcare, where a variety of compliance constraints restrict how data may be handled, processed, and stored. Despite the apparent benefits of federated learning, the heterogeneity in the local data distributions pose significant challenges, and such challenges are even more pronounced in the case of multilingual data providers. In this paper we present a federated learning system for training a large-scale multi-lingual model suitable for fine-tuning on downstream tasks such as medical entity tagging. Our work represents one of the first such production-scale systems, capable of training across multiple highly heterogeneous data providers, and achieving levels of accuracy that could not be otherwise achieved by using central training with public data. Finally, we show that the global model performance can be further improved by a training step performed locally.

cs.CL↗

arXiv2022

Efficient and Light-Weight Federated Learning via Asynchronous Distributed Dropout

Asynchronous learning protocols have regained attention lately, especially in the Federated Learning (FL) setup, where slower clients can severely impede the learning process. Herein, we propose \texttt{AsyncDrop}, a novel asynchronous FL framework that utilizes dropout regularization to handle device heterogeneity in distributed settings. Overall, \texttt{AsyncDrop} achieves better performance compared to state of the art asynchronous methodologies, while resulting in less communication and training time overheads. The key idea revolves around creating ``submodels'' out of the global model, and distributing their training to workers, based on device heterogeneity. We rigorously justify that such an approach can be theoretically characterized. We implement our approach and compare it against other asynchronous baselines, both by design and by adapting existing synchronous FL algorithms to asynchronous scenarios. Empirically, \texttt{AsyncDrop} reduces the communication cost and training time, while matching or improving the final test accuracy in diverse non-i.i.d. FL scenarios.

cs.LG↗

arXiv2022

Invariant Aggregator for Defending against Federated Backdoor Attacks

Federated learning enables training high-utility models across several clients without directly sharing their private data. As a downside, the federated setting makes the model vulnerable to various adversarial attacks in the presence of malicious clients. Despite the theoretical and empirical success in defending against attacks that aim to degrade models' utility, defense against backdoor attacks that increase model accuracy on backdoor samples exclusively without hurting the utility on other samples remains challenging. To this end, we first analyze the failure modes of existing defenses over a flat loss landscape, which is common for well-designed neural networks such as Resnet (He et al., 2015) but is often overlooked by previous works. Then, we propose an invariant aggregator that redirects the aggregated update to invariant directions that are generally useful via selectively masking out the update elements that favor few and possibly malicious clients. Theoretical results suggest that our approach provably mitigates backdoor attacks and remains effective over flat loss landscapes. Empirical results on three datasets with different modalities and varying numbers of clients further demonstrate that our approach mitigates a broad class of backdoor attacks with a negligible cost on the model utility.

cs.LG↗

arXiv2022

Distribution inference risks: Identifying and mitigating sources of leakage

A large body of work shows that machine learning (ML) models can leak sensitive or confidential information about their training data. Recently, leakage due to distribution inference (or property inference) attacks is gaining attention. In this attack, the goal of an adversary is to infer distributional information about the training data. So far, research on distribution inference has focused on demonstrating successful attacks, with little attention given to identifying the potential causes of the leakage and to proposing mitigations. To bridge this gap, as our main contribution, we theoretically and empirically analyze the sources of information leakage that allows an adversary to perpetrate distribution inference attacks. We identify three sources of leakage: (1) memorizing specific information about the $\mathbb{E}[Y|X]$ (expected label given the feature values) of interest to the adversary, (2) wrong inductive bias of the model, and (3) finiteness of the training data. Next, based on our analysis, we propose principled mitigation techniques against distribution inference attacks. Specifically, we demonstrate that causal learning techniques are more resilient to a particular type of distribution inference risk termed distributional membership inference than associative learning methods. And lastly, we present a formalization of distribution inference that allows for reasoning about more general adversaries than was previously possible.

cs.CR↗