arXiv subjects

Mohammad Rostami

Publications and source records attributed to Mohammad Rostami.

At least 19 recordsLinked to original sources

arXiv2026

Should We Type or Talk to LLM Agents? A Comprehensive Study of Voice and Keyboard Input Perturbations

Human input reaches language models by typing or speaking, and each channel leaves a distinct signature: orthographic noise for keyboards; for voice, disfluency from conventional transcription and restructuring from AI-backed dictation tools. How do they impact an LLM's performance? In this paper we present HIVE (Human Input-Variation Engine), a suite of voice transcription perturbations and QWERTY keyboard perturbations. We use HIVE to evaluate how robust models are to these perturbations. We present seven findings. (i) Voice transcription perturbations lower accuracy across every instruction-tuned model we test, and it is the structure of the transcription rather than its fillers that carries the cost. (ii) QWERTY keyboard perturbations cost less, and a model absorbs a lot of them before accuracy falls away. (iii) Both trace back to one cause, how many of the question's tokens survive the perturbation: destroying a token is what hurts, while adding new ones alongside it costs little. (iv) The gap between the two channels appears only where the answer must be constructed or deduced; on multiple choice there is none. (v) The harm does not solely come from test-set contamination. (vi) It cannot be trained away with lightweight adaptation. (vii) A thinking budget recovers the keyboard channel almost entirely but leaves the spoken registers untouched, and compressed speech is worse with it.

cs.AI↗

arXiv2026

Can You Trust the Confidence? ConfBench for Vision-Language Models on Document Extraction

Intelligent document processing (IDP) with vision-language models (VLMs) hinges on confidence scores trustworthy enough to route extractions between automation and human review. Existing document benchmarks are dominated by clean, high-quality samples, leaving low accuracy regions too sparse for calibration assessment. We introduce ConfBench, the first calibration-specific benchmark for key information extraction (KIE), built by applying 20 controlled degradation pipelines to a diverse document set, yielding 1,346 variants and 70K+ entity-level evaluations spanning the full accuracy spectrum. We evaluate four proprietary and three open-weight VLMs under verbalized and log-probability confidence estimation methods across three input modalities, and find: (i) OCR+Image modality results in more accurate confidence estimates; (ii) model capability is the dominant factor: within the Claude family confidence quality scales monotonically with capability, while across families parameter count is a poor predictor; (iii) calibration quality varies widely across models, from near-perfect to severely overconfident, and per-model post-hoc correction rescales these absolute confidence values for threshold-based routing without altering ranking-based operational metrics; and (iv) log-probability with first-token aggregation consistently outperforms mean-token and margin aggregations. We also introduce ECARB, a review-budget metric translating discriminative gains into operational savings. We release ConfBench publicly to enable systematic study of confidence estimators and calibration methods for trustworthy IDP application deployment.

cs.AI↗

arXiv2026

In-Context Collapse in Vision-Language Models and How to Mitigate it?

Many-shot in-context learning (ICL) lets vision-language models (VLMs) adapt from image--label demonstrations without weight updates, and is widely assumed to improve as more demonstrations are supplied. We show the opposite: as demonstrations accumulate, a subset of VLMs undergo an \emph{in-context collapse}, a sharp, sometimes catastrophic accuracy drop spanning synthetic classification, natural-image classification, and VQA benchmarks, in some models falling below chance while outputs remain well-formed. Across an open VLM panel ($0.5$B--$11$B) and a frontier model (Claude Sonnet 4.5), the collapse is graded. Two capabilities turn out to be dissociable: robustness to accumulating demonstrations and the ability to learn a novel rule in context, their combinations yield three reproducible regimes. A parameter-matched lesion-and-rescue causally localizes the collapse to the vision-language integration pathway: an adapter on the connector and early/mid layers restores genuine learning (remap accuracy $0.39!\rightarrow!0.91$ at 16 shots), while an equal-capacity adapter on the late readout does not. We propose \textsc{CircA}, whose core is a one-time integration vaccine: trained once on one synthetic task, it transfers collapse-resistance to unseen task families (chance$\rightarrow$$0.71$/$0.60$ on CIFAR/Fashion). The layers best for in-context integration are not the layers best for weight-based consolidation, the late readout achieves higher accuracy and less forgetting at fewer parameters. The collapse is an integration failure at the vision--language interface, correctable by a lightweight, transferable intervention.

cs.CV↗

arXiv2026

FlowBank: Query-Adaptive Agentic Workflows Optimization through Precompute-and-Reuse

Large Language Model (LLM)-based multi-agent systems are increasingly powerful, but current agentic workflow optimization paradigms make an unsatisfying trade-off. Task-level methods spend substantial offline compute yet deploy only a single workflow, leaving complementary candidates unused, while query-level methods synthesize a new workflow per query at substantial inference cost. Our motivating analysis shows these paradigms are more complementary than competing: workflows discovered during offline search often solve different subsets of queries, and many queries handled by expensive query-level generation can already be solved by cheaper precomputed workflows. This suggests a different objective: rather than searching for one universally best workflow or regenerating one per instance, we should build a compact bank of reusable, complementary workflows and select among them adaptively at inference time. Doing so requires solving three coupled problems: generating complementary rather than redundant candidates, compressing them into a small deployable portfolio, and assigning each query to the right workflow under a performance-cost trade-off. To this end, we present FlowBank, a three-stage framework for portfolio-based agentic workflow optimization. Diversifying proposes DiverseFlow to steer search toward under-covered queries and produce a high-coverage candidate pool. Curating proposes CuraFlow to compress this pool into a compact portfolio with minimal redundancy. Matching casts deployment as edge-value prediction on a query-workflow bipartite graph and routes each incoming query to the portfolio member with the best predicted utility. Across five benchmarks, FlowBank achieves the highest average score among the evaluated methods while remaining cost-competitive, improving over the strongest automated and handcrafted baselines by 4.26% and 14.92% relative, respectively.

cs.LG↗

arXiv2026

What-If World: A Causal Benchmark for General World Models in Embodied Scenarios

Video generation models are increasingly used as world simulators for tasks like driving and robotic manipulation. What matters in these settings is not whether a single video looks right, but whether the model's output changes when its input changes. We test this by giving a model two prompts describing the same scene with one physical detail varied, and checking whether the two videos diverge the way physics predicts. The wording difference between the prompts is small by design, since only one variable is changed, but the correct physical difference is not. A model that misses this can still produce two videos that each look plausible individually, and existing benchmarks score videos one at a time and cannot detect this failure. We introduce What-If World, 319 such prompt pairs built on real frames from nuScenes and DROID, organized by a taxonomy of six physical variables shared across driving and manipulation. Each pair is scored with APEO, a four-part rubric checking whether each video follows its prompt (Adherence), is physically consistent (Physics), preserves the shared scene (Environment), and ends in the correct difference (Outcome). Across nine state-of-the-art models, no system exceeds 52% on the paired score, and open-source models cluster near 28%. Every model tested fails on a large fraction of causal interventions, indicating substantial room before these models can reliably support action-conditioned simulation or model-based planning. Where models do score well, performance appears to track the visual prominence of the intervention rather than the tractability of its underlying physics. Some visually subtle interventions score as low as 14.2%, while visually pronounced ones reach 40.4%.

cs.CV↗

arXiv2026

Hedwig: Dynamic Autonomy for Coding Agents Under Local Oversight

Despite coding agents' advances in handling increasingly complex tasks, their continued tendency to introduce unintended edits, subtle bugs, and scope drift that slip past code review means developers must still decide how much autonomy to grant them. However, existing approaches for setting an agent's level of autonomy, such as static permission settings or instruction files, cannot account for how developers' preferences for agent autonomy can shift across tasks and over time. We conducted a formative survey with 21 software engineers who use coding agents and found that they experience frustration with calibrating autonomy and have evolving preferences for level of oversight. Building on these insights, we present Hedwig, a CLI coding agent that dynamically adjusts its autonomy level based on developer-agent interactions across sessions. Rather than operating on a global, fixed autonomy configuration, Hedwig learns an evolving set of behavioral guidelines from developer decisions and feedback, reducing friction on work for which the agent has earned trust, while tightening oversight when the agent operates outside familiar territory. Hedwig demonstrates the potential of a new paradigm where agents intelligently adapt their level of autonomy based on user trust through active, longitudinal collaboration.

cs.HC↗

arXiv2026

SHRED: Retain-Set-Free Unlearning via Self-Distillation with Logit Demotion

Machine unlearning for large language models (LLMs) aims to selectively remove memorized content such as private data, copyrighted text, or hazardous knowledge, without costly full retraining. Most existing methods require a retain set of curated examples to prevent catastrophic degradation of general model utility, creating an extra data dependency that complicates deployment. We propose SHRED (Self-distillation via High-surprisal-only Retain-set-free Entropy Demotion), a retain-set-free unlearning method built on a key insight: not all tokens within a forget set instance carry memorized information equally. High-information tokens concentrate the model's memorized knowledge, while low-information tokens reflect general language competence. SHRED operates in two stages. (1) Selection: We perform a forward pass on a forget set instance, collect per-token autoregressive probabilities, and select the bottom (lowest probability, highest Shannon information) as forget positions; the remaining positions are retained as benign anchors. (2) Training: We construct modified KL targets that demote the memorized token's logit at forget positions while preserving the original distribution at benign positions. The model is then trained via a single top KL self-distillation objective that simultaneously drives forgetting and utility preservation. We evaluate SHRED across four standard unlearning benchmarks and demonstrate that it establishes a new Pareto-optimal trade-off between forget efficacy and model utility, outperforming retain-set-dependent methods. Our analysis shows that SHRED is robust against relearning attacks and membership-inference attacks, and it maintains stable utility even after many sequential unlearning runs.

cs.LG↗

arXiv2026

Expert Personas Improve LLM Alignment but Damage Accuracy: Bootstrapping Intent-Based Persona Routing with PRISM

Persona prompting can steer LLM generation towards a domain-specific tone and pattern. This behavior enables use cases in multi-agent systems where diverse interactions are crucial and human-centered tasks require high-level human alignment. Prior works provide mixed opinions on their utility: some report performance gains when using expert personas for certain domains and their contribution to data diversity in synthetic data creation, while others find near-zero or negative impact on general utility. To fully leverage the benefits of the LLM persona and avoid its harmfulness, a more comprehensive investigation of the mechanism is crucial. In this work, we study how model optimization, task type, prompt length, and placement can impact expert persona effectiveness across instruction-tuned and reasoning LLMs, and provide insight into conditions under which expert personas fail and succeed. Based on our findings, we developed a pipeline to fully leverage the benefits of an expert persona, named PRISM (Persona Routing via Intent-based Self-Modeling), which self-distills an intent-conditioned expert persona into a gated LoRA adapter through a bootstrapping process that requires no external data, models, or knowledge. PRISM enhances human preference and safety alignment on generative tasks while maintaining accuracy on discriminative tasks across all models, with minimal memory and computing overhead.

cs.AI↗

arXiv2026

Semi-Supervised Masked Autoencoders: Unlocking Vision Transformer Potential with Limited Data

We address the challenge of training Vision Transformers (ViTs) when labeled data is scarce but unlabeled data is abundant. We propose Semi-Supervised Masked Autoencoder (SSMAE), a framework that jointly optimizes masked image reconstruction and classification using both unlabeled and labeled samples with dynamically selected pseudo-labels. SSMAE introduces a validation-driven gating mechanism that activates pseudo-labeling only after the model achieves reliable, high-confidence predictions that are consistent across both weakly and strongly augmented views of the same image, reducing confirmation bias. On CIFAR-10 and CIFAR-100, SSMAE consistently outperforms supervised ViT and fine-tuned MAE, with the largest gains in low-label regimes (+9.24% over ViT on CIFAR-10 with 10% labels). Our results demonstrate that when pseudo-labels are introduced is as important as how they are generated for data-efficient transformer training. Codes are available at https://github.com/atik666/ssmae.

cs.CV↗

arXiv2026

Finetune-Informed Pretraining Boosts Downstream Performance

Multimodal pretraining is effective for building general-purpose representations, but in many practical deployments, only one modality is heavily used during downstream fine-tuning. Standard pretraining strategies treat all modalities uniformly, which can lead to under-optimized representations for the modality that actually matters. We propose Finetune-Informed Pretraining (FIP), a model-agnostic method that biases representation learning toward a designated target modality needed at fine-tuning time. FIP combines higher masking difficulty, stronger loss weighting, and increased decoder capacity for the target modality, without modifying the shared encoder or requiring additional supervision. When applied to masked modeling on constellation diagrams for wireless signals, FIP consistently improves downstream fine-tuned performance with no extra data or compute. FIP is simple to implement, architecture-compatible, and broadly applicable across multimodal masked modeling pipelines.

cs.LG↗

arXiv2026

CageDroneRF: A Large-Scale RF Benchmark and Toolkit for Drone Perception

We present CageDroneRF (CDRF), a large-scale benchmark for Radio-Frequency (RF) drone detection and identification built from real-world captures and systematically generated synthetic variants. CDRF addresses the scarcity and limited diversity of existing RF datasets by coupling extensive raw recordings with a principled augmentation pipeline that (i)~precisely controls Signal-to-Noise Ratio (SNR), (ii)~injects interfering emitters, and (iii)~applies frequency shifts with label-consistent bounding-box recomputation for detection. The dataset spans a wide range of contemporary drone models, many of which are unavailable in current public datasets, and diverse acquisition conditions, derived from data collected at the Rowan University campus and within a controlled RF-cage facility. CDRF is released with interoperable open-source tools for data generation, preprocessing, augmentation, and evaluation that also operate on existing public benchmarks. It enables standardized benchmarking for classification, open-set recognition, and object detection, supporting rigorous comparisons and reproducible pipelines. By releasing this comprehensive benchmark and tooling, we aim to accelerate progress toward robust, generalizable RF perception models.

cs.CV↗

arXiv2025

Large Language Models Can Perform Automatic Modulation Classification via Discretized Self-supervised Candidate Retrieval

Identifying wireless modulation schemes is essential for cognitive radio, but standard supervised models often degrade under distribution shift, and training domain-specific wireless foundation models from scratch is computationally prohibitive. Large Language Models (LLMs) offer a promising training-free alternative via in-context learning, yet feeding raw floating-point signal statistics into LLMs overwhelms models with numerical noise and exhausts token budgets. We introduce DiSC-AMC, a framework that reformulates Automatic Modulation Classification (AMC) as an LLM reasoning task by combining aggressive feature discretization with nearest-neighbor retrieval over self-supervised embeddings. By mapping continuous features to coarse symbolic tokens, DiSC-AMC aligns abstract signal patterns with LLM reasoning capabilities and reduces prompt length by over $50$\%. Simultaneously, utilizing a DINOv2 visual encoder to retrieve the $k_\text{NN}$ most similar labeled exemplars provides highly relevant, query-specific context rather than generic class averages. On a 10-class benchmark, a fine-tuned 7B-parameter LLM using DiSC-AMC achieves $83.0$\% in-distribution accuracy ($-10$\,to\,$+10$\,dB) and $82.50$\% out-of-distribution (OOD) accuracy ($-11$\,to\,$-15$\,dB), outperforming supervised baselines. Comprehensive ablations on vanilla LLMs demonstrate the token efficiency of DiSC-AMC. A training-free $7$B LLM achieves $71$\% accuracy using only $0.5$\,K-token prompt,surpassing a $200$B-parameter baseline that relies on a $2.9$K-token prompt. Furthermore, similarity-based exemplar retrieval outperforms naive class-average selection by over $20$\%. Finally, we identify a fundamental limitation of this pipeline. At extreme OOD noise levels ($-30$\,dB), the underlying self-supervised representations collapse, degrading retrieval quality and reducing classification to random chance.

cs.LG↗

arXiv2025

Towards Improving Long-Tail Entity Predictions in Temporal Knowledge Graphs through Global Similarity and Weighted Sampling

Temporal Knowledge Graph (TKG) completion models traditionally assume access to the entire graph during training. This overlooks challenges stemming from the evolving nature of TKGs, such as: (i) the model's requirement to generalize and assimilate new knowledge, and (ii) the task of managing new or unseen entities that often have sparse connections. In this paper, we present an incremental training framework specifically designed for TKGs, aiming to address entities that are either not observed during training or have sparse connections. Our approach combines a model-agnostic enhancement layer with a weighted sampling strategy, that can be augmented to and improve any existing TKG completion method. The enhancement layer leverages a broader, global definition of entity similarity, which moves beyond mere local neighborhood proximity of GNN-based methods. The weighted sampling strategy employed in training accentuates edges linked to infrequently occurring entities. We evaluate our method on two benchmark datasets, and demonstrate that our framework outperforms existing methods in total link prediction, inductive link prediction, and in addressing long-tail entities. Notably, our method achieves a 10\% improvement and a 15\% boost in MRR for these datasets. The results underscore the potential of our approach in mitigating catastrophic forgetting and enhancing the robustness of TKG completion methods, especially in an incremental training context

cs.AI↗

arXiv2025

TNG-CLIP:Training-Time Negation Data Generation for Negation Awareness of CLIP

Vision-language models (VLMs), such as CLIP, have demonstrated strong performance across a range of downstream tasks. However, CLIP is still limited in negation understanding: the ability to recognize the absence or exclusion of a concept. Existing methods address the problem by using a large language model (LLM) to generate large-scale data of image captions containing negation for further fine-tuning CLIP. However, these methods are both time- and compute-intensive, and their evaluations are typically restricted to image-text matching tasks. To expand the horizon, we (1) introduce a training-time negation data generation pipeline such that negation captions are generated during the training stage, which only increases 2.5% extra training time, and (2) we propose the first benchmark, Neg-TtoI, for evaluating text-to-image generation models on prompts containing negation, assessing model's ability to produce semantically accurate images. We show that our proposed method, TNG-CLIP, achieves SOTA performance on diverse negation benchmarks of image-to-text matching, text-to-image retrieval, and image generation.

cs.CV↗

arXiv2025

Multi-modal Synthetic Data Training and Model Collapse: Insights from VLMs and Diffusion Models

Recent research has highlighted the risk of generative model collapse, where performance progressively degrades when continually trained on self-generated data. However, existing exploration on model collapse is limited to single, unimodal models, limiting our understanding in more realistic scenarios, such as diverse multi-modal AI agents interacting autonomously through synthetic data and continually evolving. We expand the synthetic data training and model collapse study to multi-modal vision-language generative systems, such as vision-language models (VLMs) and text-to-image diffusion models, as well as recursive generate-train loops with multiple models. We find that model collapse, previously observed in single-modality generative models, exhibits distinct characteristics in the multi-modal context, such as improved vision-language alignment and increased variance in VLM image-captioning task. Additionally, we find that general approaches such as increased decoding budgets, greater model diversity, and relabeling with frozen models can effectively mitigate model collapse. Our findings provide initial insights and practical guidelines for reducing the risk of model collapse in self-improving multi-agent AI systems and curating robust multi-modal synthetic datasets.

cs.LG↗

arXiv2025

Plug-and-Play AMC: Context Is King in Training-Free, Open-Set Modulation with LLMs

Automatic Modulation Classification (AMC) is critical for efficient spectrum management and robust wireless communications. However, AMC remains challenging due to the complex interplay of signal interference and noise. In this work, we propose an innovative framework that integrates traditional signal processing techniques with Large-Language Models (LLMs) to address AMC. Our approach leverages higher-order statistics and cumulant estimation to convert quantitative signal features into structured natural language prompts. By incorporating exemplar contexts into these prompts, our method exploits the LLM's inherent familiarity with classical signal processing, enabling effective one-shot classification without additional training or preprocessing (e.g., denoising). Experimental evaluations on synthetically generated datasets, spanning both noiseless and noisy conditions, demonstrate that our framework achieves competitive performance across diverse modulation schemes and Signal-to-Noise Ratios (SNRs). Moreover, our approach paves the way for robust foundation models in wireless communications across varying channel conditions, significantly reducing the expense associated with developing channel-specific models. This work lays the foundation for scalable, interpretable, and versatile signal classification systems in next-generation wireless networks. The source code is available at https://github.com/RU-SIT/context-is-king

cs.LG↗

arXiv2025

Hybrid Learners Do Not Forget: A Brain-Inspired Neuro-Symbolic Approach to Continual Learning

Continual learning is crucial for creating AI agents that can learn and improve themselves autonomously. A primary challenge in continual learning is to learn new tasks without losing previously learned knowledge. Current continual learning methods primarily focus on enabling a neural network with mechanisms that mitigate forgetting effects. Inspired by the two distinct systems in the human brain, System 1 and System 2, we propose a Neuro-Symbolic Brain-Inspired Continual Learning (NeSyBiCL) framework that incorporates two subsystems to solve continual learning: A neural network model responsible for quickly adapting to the most recent task, together with a symbolic reasoner responsible for retaining previously acquired knowledge from previous tasks. Moreover, we design an integration mechanism between these components to facilitate knowledge transfer from the symbolic reasoner to the neural network. We also introduce two compositional continual learning benchmarks and demonstrate that NeSyBiCL is effective and leads to superior performance compared to continual learning methods that merely rely on neural architectures to address forgetting.

cs.LG↗

arXiv2025

DenoMAE2.0: Improving Denoising Masked Autoencoders by Classifying Local Patches

We introduce DenoMAE2.0, an enhanced denoising masked autoencoder that integrates a local patch classification objective alongside traditional reconstruction loss to improve representation learning and robustness. Unlike conventional Masked Autoencoders (MAE), which focus solely on reconstructing missing inputs, DenoMAE2.0 introduces position-aware classification of unmasked patches, enabling the model to capture fine-grained local features while maintaining global coherence. This dual-objective approach is particularly beneficial in semi-supervised learning for wireless communication, where high noise levels and data scarcity pose significant challenges. We conduct extensive experiments on modulation signal classification across a wide range of signal-to-noise ratios (SNRs), from extremely low to moderately high conditions and in a low data regime. Our results demonstrate that DenoMAE2.0 surpasses its predecessor, Deno-MAE, and other baselines in both denoising quality and downstream classification accuracy. DenoMAE2.0 achieves a 1.1% improvement over DenoMAE on our dataset and 11.83%, 16.55% significant improved accuracy gains on the RadioML benchmark, over DenoMAE, for constellation diagram classification of modulation signals.

cs.LG↗