A LLM (Large Language Model) is an artificial intelligence model characterized by its vast size in terms of parameters and training data. These models are typically based on deep learning architectures, such as Transformers, and are trained on extensive datasets to learn the statistical relationships and patterns within natural language.

Posts

Efficient Semantic Communication Through Transformer-Aided Compression

Transformers, known for their attention mechanisms, have proven highly effective in focusing on critical elements within complex data. This feature can effectively be used to address the time-varying channels in wireless communication systems. In this work, we introduce a channel-aware adaptive framework for semantic communication, where different regions of the image are encoded and compressed based on their semantic content. By employing vision transformers, we interpret the attention mask as a measure of the semantic contents of the patches and dynamically categorize the patches to be compressed at various rates as a function of the instantaneous channel bandwidth. Our method enhances communication efficiency by adapting the encoding resolution to the content’s relevance, ensuring that even in highly constrained environments, critical information is preserved. We evaluate the proposed adaptive transmission framework using the TinyImageNet dataset, measuring both reconstruction quality and accuracy. The results demonstrate that our approach maintains high semantic fidelity while optimizing bandwidth, providing an effective solution for transmitting multiresolution data in limited bandwidth conditions.

Latency-driven Execution of LLM-generated Application Code on the Computing Continuum

Latency-critical applications demand quick responses. Ideally, detailed insights are preferable for the best decision making and response actions. However, in situations when detailed insights cannot be provided quickly, even basic information goes a long way in tackling the situation effectively. For example, in marine security application, it is critical to immediately notify as soon as an unauthorized vessel is seen. Hence, timely response may be prioritized over the response based on entire details. To address such latency-critical situations, in this paper, we propose a novel system called DiCE-EC, which leverages LLM to generate distributed code with speculative execution on Edge (fast and simple response using resource constrained hardware) and Cloud (detailed response using powerful hardware, but may be fast or slow depending on network conditions). DiCE-EC breaks down application into smaller components and executes them asynchronously across the edge and cloud computing continuum. As network conditions vary, we show through real-world marine security application, that DiCE-EC is effective in dynamically choosing detailed insights from cloud when received within latency-constraint, or falling back to simple response from edge to guarantee timely alert delivery. Without such dynamic selection of response from edge or cloud, existing systems either always provide simple responses or drop alerts. We perform real network measurements in the Gulf of Pozzuoli in Naples, Italy along accessible areas (inland and in a Ferry) and generate 1 million realistic measurements across four inaccessible regions, and demonstrate that DiCE-EC never misses an alert, while baseline misses up to ?4% alerts with real data and up to ?1% (10,000 alerts) with generated data.

Position Really Matters: Towards a Holistic Approach for Prompt Tuning

Prompt tuning is highly effective in efficiently extracting knowledge from foundation models, encompassing both language, vision, and vision-language models. However, the efficacy of employing fixed soft prompts with a predetermined position for concatenation with inputs for all instances, irrespective of their inherent disparities, remains uncertain. Variables such as the position, length, and representations of prompts across diverse instances and tasks can substantially influence the performance of prompt tuning. We first provide a theoretical analysis, revealing that optimizing the position of the prompt to encompass the input can capture additional semantic information that traditional prefix or postfix prompt tuning methods fail to capture. Then, we present a holistic parametric prompt tuning strategy that dynamically determines different factors of prompts based on specific tasks or instances. Experimental results underscore the significant performance improvement achieved by dynamic prompt tuning across a wide range of tasks, including NLP, vision recognition, and vision-language tasks. Furthermore, we establish the universal applicability of our approach under full-data, few-shot, and multitask settings.

DISC: Dynamic Decomposition Improves LLM Inference Scaling (SSI-FM)

Inference scaling methods often rely on decomposing problems into steps, followed by sampling and selecting the best next steps. However, these steps and their sizes are typically fixed or depend on domain knowledge. We propose dynamic decomposition, a method that adaptively and automatically breaks down solution and reasoning traces into manageable steps during inference. By allocating compute more effectively, particularly by subdividing challenging steps and sampling them more frequently, dynamic decomposition significantly enhances inference efficiency. Experiments on benchmarks such as APPS, MATH, and LiveCodeBench demonstrate that dynamic decomposition outperforms static approaches, including token-level, sentence-level, and single-step decompositions. These findings highlight the potential of dynamic decomposition to improve a wide range of inference scaling techniques.

Humanizing the Machine: Proxy Attacks to Mislead LLM Detectors

The advent of large language models (LLMs) has revolutionized the field of text generation, producing outputs that closely mimic human-like writing. Although academic and industrial institutions have developed detectors to prevent the malicious usage of LLM-generated texts, other research has doubt about the robustness of these systems. To stress test these detectors, we introduce a humanized proxy-attack (HUMPA) strategy that effortlessly compromises LLMs, causing them to produce outputs that align with human-written text and mislead detection systems. Our method attacks the source model by leveraging a reinforcement learning (RL) fine-tuned humanized small language model (SLM) in the decoding phase. Through an in-depth analysis, we demonstrate that our attack strategy is capable of generating responses that are indistinguishable to detectors, preventing them from differentiating between machine-generated and human-written text. We conduct systematic evaluations on extensive datasets using proxy-attacked open-source models, including Llama2-13B, Llama3-70B, and Mixtral-8×7B in both white- and black-box settings. Our findings show that the proxy-attack strategy effectively deceives the leading detectors, resulting in an average AUROC drop of 70.4% across multiple datasets, with a maximum drop of 95.0% on a single dataset. Furthermore, in cross-discipline scenarios, our strategy also bypasses these detectors, leading to a significant relative decrease of up to 90.9%, while in cross-language scenario, the drop reaches 91.3%. Despite our proxy-attack strategy successfully bypassing the detectors with such significant relative drops, we find that the generation quality of the attacked models remains preserved, even within a modest utility budget, when compared to the text produced by the original, unattacked source model.

DISC: Dynamic Decomposition Improves LLM Inference Scaling (DL4C)

Inference scaling methods often rely on decomposing problems into steps, followed by sampling and selecting the best next steps. However, these steps and their sizes are typically fixed or depend on domain knowledge. We propose dynamic decomposition, a method that adaptively and automatically breaks down solution and reasoning traces into manageable steps during inference. By allocating compute more effectively—particularly by subdividing challenging steps and sampling them more frequently—dynamic decomposition significantly enhances inference efficiency. Experiments on benchmarks such as APPS, MATH, and LiveCodeBench demonstrate that dynamic decomposition outperforms static approaches, including token-level, sentence-level, and single-step decompositions. These findings highlight the potential of dynamic decomposition to improve a wide range of inference scaling techniques.

DWIM: Towards Tool-aware Visual Reasoning via Discrepancy-aware Workflow Generation & Instruct-Masking Tuning

Visual reasoning (VR), which is crucial in many fields for enabling human-like visual understanding, remains highly challenging. Recently, compositional visual reasoning approaches, which leverage the reasoning abilities of large language models (LLMs) with integrated tools to solve problems, have shown promise as more effective strategies than end-to-end VR methods. However, these approaches face limitations, as frozen LLMs lack tool awareness in VR, leading to performance bottlenecks. While leveraging LLMs for reasoning is widely used in other domains, they are not directly applicable to VR due to limited training data, imperfect tools that introduce errors and reduce data collection efficiency in VR, and challenging in fine-tuning on noisy workflows. To address these challenges, we propose DWIM: i) Discrepancy-aware training Workflow generation, which assesses tool usage and extracts more viable workflows for training; and ii) Instruct-Masking fine-tuning, which guides the model to only clone effective actions, enabling the generation of more practical solutions. Our experiments demonstrate that DWIM achieves state-of-the-art performance across various VR tasks, exhibiting strong generalization on multiple widely-used datasets.

TimeCAP: Learning to Contextualize, Augment, and Predict Time Series Events with Large Language Model Agents

Time series data is essential in various applications, including climate modeling, healthcare monitoring, and financial analytics. Understanding the contextual information associated with real-world time series data is often essential for accurate and reliable event predictions. In this paper, we introduce TimeCAP, a time-series processing framework that creatively employs Large Language Models (LLMs) as contextualizers of time series data, extending their typical usage as predictors. TimeCAP incorporates two independent LLM agents: one generates a textual summary capturing the context of the time series, while the other uses this enriched summary to make more informed predictions. In addition, TimeCAP employs a multi-modal encoder that synergizes with the LLM agents, enhancing predictive performance through mutual augmentation of inputs with in-context examples. Experimental results on real-world datasets demonstrate that TimeCAP outperforms state-of-the-art methods for time series event prediction, including those utilizing LLMs as predictors, achieving an average improvement of 28.75% in F1 score.

NEC Labs America Attends the 39th Annual AAAI Conference on Artificial Intelligence #AAAI25

Our NEC Lab America team attended the Thirty-Ninth AAAI Conference on Artificial Intelligence (AAAI-25), in Philadelphia, Pennsylvania at the Pennsylvania Convention Center from February 25 to March 4, 2025. The purpose of the AAAI conference series was to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. Our team presented technical papers, led special tracks, delivered talks on key topics, participated in workshops, conducted tutorials, and showcased research in poster sessions. The team greeted visitors at Booth #208 and was there Thursday through Saturday.

RAG-check: Evaluating Multimodal Retrieval Augmented Generation Performance

Retrieval-augmented generation (RAG) improves large language models (LLMs) by using external knowledge to guide response generation, reducing hallucinations. However, RAG, particularly multi-modal RAG, can introduce new hallucination sources: (i) the retrieval process may select irrelevant pieces (e.g., documents, images) as raw context from the database, and (ii) retrieved images are processed into text-based context via vision-language models (VLMs) or directly used by multi-modal language models (MLLMs) like GPT-4o, which may hallucinate. To address this, we propose a novel framework to evaluate the reliability of multi-modal RAG using two performance measures: (i) the relevancy score (RS), assessing the relevance of retrieved entries to the query, and (ii) the correctness score (CS), evaluating the accuracy of the generated response. We train RS and CS models using a ChatGPT-derived database and human evaluator samples. Results show that both models achieve ~88% accuracy on test data. Additionally, we construct a 5000-sample human-annotated database evaluating the relevancy of retrieved pieces and the correctness of response statements. Our RS model aligns with human preferences 20% more often than CLIP in retrieval, and our CS model matches human preferences ~91% of the time. Finally, we assess various RAG systems’ selection and generation performances using RS and CS.