Salimur Choudhury

We present LastDance, an autoresearch framework that autonomously discovers optimized Retrieval-Augmented Generation pipelines for compact language models under token budget constraints. The framework guides a coding agent (Claude Code paired with Qwen3-Coder-Next as the underlying coder model) through a proposed project structure comprising three instruction files, a modifiable pipeline file, a fixed evaluation script with diagnostic report generation, and a curated research knowledge base. After each experiment, a diagnostic module classifies every evaluation sample into four failure categories and recommends which pipeline stage to modify next. This helps the agent make research decisions based on diagnostic evidence. The methodology discovered by LastDance, KnapsackRAG, formulates passage selection as a 0/1 integer linear program with cardinality, coverage, and pairwise redundancy constraints. On the NQ Open and TriviaQA validation sets, KnapsackRAG achieves 36.9% and 36.2% relative F1 improvements over the strongest top-k baseline with Qwen2.5-1.5B-Instruct as the generator. The discovered configuration transfers to Llama-3.2-3B-Instruct without modification, yielding 99.3% and 17.3% relative F1 gains on the same two datasets.

Intent-Based Networking (IBN) allows operators to specify high-level network goals rather than low-level configurations. While recent work demonstrates that large language models can automate configuration tasks, a distinct class of intents requires generating optimization code to compute provably optimal solutions for traffic engineering, routing, and resource allocation. Current systems assume text-based intent expression, requiring operators to enumerate topologies and parameters in prose. Network practitioners naturally reason about structure through diagrams, yet whether Vision-Language Models (VLMs) can process annotated network sketches into correct optimization code remains unexplored. We present IntentOpt, a benchmark of 85 optimization problems across 17 categories, evaluating four VLMs (GPT-5-Mini, Claude-Haiku-4.5, Gemini-2.5-Flash, Llama-3.2-11B-Vision) under three prompting strategies on multimodal versus text-only inputs. Our evaluation shows that visual parameter extraction reduces execution success by 12-21 percentage points (pp), with GPT-5-Mini dropping from 93% to 72%. Program-of-thought prompting decreases performance by up to 13 pp, and open-source models lag behind closed-source ones, with Llama-3.2-11B-Vision reaching 18% compared to 75% for GPT-5-Mini. These results establish baseline capabilities and limitations of current VLMs for optimization code generation within an IBN system. We also demonstrate practical feasibility through a case study that deploys VLM-generated code to network testbed infrastructure using Model Context Protocol.

Large language models (LLMs) have demonstrated remarkable capabilities in code generation across various domains. However, their effectiveness in generating simulation scripts for domain-specific environments like ns-3 remains underexplored. Despite the growing interest in automating network simulations, existing tools primarily focus on interactive automation over rigorous evaluation. To facilitate systematic evaluation, we introduce SIMCODE, the first benchmark to evaluate LLMs' ability to generate ns-3 simulation code from natural language. SIMCODE includes 400 tasks across introductory, intermediate, and advanced levels, with solutions and test cases. Using SIMCODE, we evaluate three prominent LLMs, Gemini-2.0, GPT-4.1, and Qwen-3, across six prompt techniques. Furthermore, investigating task-specific fine-tuning's impact reveals that while GPT-4.1 outperforms others, execution accuracy remains modest, with substantial room for improvement. Error analysis identifies missing headers and API mismatches as dominant failures. Nevertheless, SIMCODE provides a foundational step toward evaluating LLMs and research in domain-aware generative systems.

In a two-sided matching with preferences, an agent has preferences over the agents from the other side. In a head-to-head election between any two matching, the agent votes for the matching with a better allocation in pairing based on its preference. A matching M is popular when the number of votes for M is not less than the number of votes for any other matching M′ in such an election. A U-popular matching is introduced recently with a two-sided matching model where U and W are the two sets of agents, but only the preferences of the agents from U are emphasized, and the preferences of the agents from W are ignored. The matching which is popular among the votes of the agents from U is defined as U-popular. We consider such a model for a one-to-one matching where the agents from U have an inclusive mindset and want to integrate the voting decision of the agents from W into their voting process. In parallel, the agents from W rely more on their ranking on the preference list of the agents from U than their own preferences due to the uncertainty involved in the construction of their preferences.We define an inclusive voting model with such a predominant-subordinate agent scenario (U as predominant, W as subordinate) and prove that the Boston mechanism matching is U-popular under the model. However, it is possible that the U-popular matching is not the choice of the majority of agents. The choice to be neutral is added in the voting process, and the U-neutral matching type is introduced when the majority of agents vote for neutral. We characterize the U-neutral and max-size U-neutral matching under the inclusive voting model and propose a polynomial-time algorithm to determine a max-size U-neutral matching. The experiments we performed with synthetic instances endorse the algorithm based on the theoretical foundations established.

Linear Programming (LP) problems aim to find the optimal solution to an objective under constraints. These problems typically require domain knowledge, mathematical skills, and programming ability, presenting significant challenges for non-experts. This study explores the efficiency of Large Language Models (LLMs) in generating solver-specific LP code. We propose CHORUS, a retrieval-augmented generation (RAG) framework for synthesizing Gurobi-based LP code from natural language problem statements. CHORUS incorporates a hierarchical tree-like chunking strategy for theoretical contents and generates additional metadata based on code examples from documentation to facilitate self-contained, semantically coherent retrieval. Two-stage retrieval approach of CHORUS followed by cross-encoder reranking further ensures contextual relevance. Finally, expertly crafted prompt and structured parser with reasoning steps improve code generation performance significantly. Experiments on the NL4Opt-Code benchmark show that CHORUS improves the performance of open-source LLMs such as Llama3.1 (8B), Llama3.3 (70B), Phi4 (14B), Deepseek-r1 (32B), and Qwen2.5-coder (32B) by a significant margin compared to baseline and conventional RAG. It also allows these open-source LLMs to outperform or match the performance of much stronger baselines-GPT3.5 and GPT4 while requiring far fewer computational resources. Ablation studies further demonstrate the importance of expert prompting, hierarchical chunking, and structured reasoning.

The introduction of 5G networks has significantly advanced communication technology, offering faster speeds, lower latency, and greater capacity. This progress sets the stage for Beyond 5G (B5G) networks, which present new complexity and performance requirements challenges. Linear Programming (LP), Integer Linear Programming (ILP), and Mixed-Integer Linear Programming (MILP) models have been widely used to model the optimization of resource allocation problems in networks. This paper reviews 103 studies on resource allocation strategies in 5G and B5G, focusing specifically on optimization problems modelled as LP, ILP, and MILP. The selected studies are categorized based on network architectures, types of resource allocation problems, and specific objective functions and constraints. The review also discusses solution methods for NP-hard ILP and MILP problems by categorizing the solution methods into different categories. Additionally, emerging trends, such as integrating AI and machine learning with optimization models, are explored, suggesting promising future research directions in network optimization. The paper concludes that LP, ILP, and MILP models have been widely adopted across various network architectures, resource types, objective functions, and constraints and remain critical to optimizing next-generation networks.