Srinivasan Manoharan

Fangbo Tu, Junhua Zhao, Chi Liu et al.

Demand for low-precision inference, including NVFP4-based approaches, has grown as large language models are increasingly deployed in latency and cost constrained production environments. Quantization-aware distillation (QAD) helps recover accuracy lost under low bit quantization by training a quantized student to match the output distribution of a frozen higher precision teacher via a KL-divergence loss. In this work, we first provide a representation level diagnosis of QAD: output matching alone can mask internal degradation, because many intermediate activation geometries can yield similar teacher-aligned logits. Using CKA, we show that KL-only QAD can reduce layerwise representational similarity relative to the BF16 teacher, with especially severe drift in RL-post-trained models. This drift correlates with downstream bottlenecks on reasoning and coding tasks, suggesting that low bit recovery requires preserving internal geometry rather than matching outputs alone. Motivated by this finding, we propose \textbf{CKA-QAD}, a CKA-guided representational alignment method for NVFP4 QAD and low bit LLM accuracy recovery. The method adds a lightweight regularizer that preserves internal representational geometry during distillation by aligning layerwise Gram matrices through CKA. Across Nemotron 3 Nano and Qwen3-4B-Thinking-2507, CKA-QAD substantially improves representational alignment and improves downstream reasoning and coding accuracy with modest training overhead. Our findings position CKA-guided representational alignment as a practical complement to output matching for quantized LLM recovery.

Srinivasan Manoharan, Dilipkumar Nallusamy, Sachin Kumar et al.

Deploying frontier large language models (LLMs) for domain-specific structured evaluation tasks often incurs substantial latency, cost, and data privacy overhead. We present a hybrid framework that combines a fine-tuned small language model (LLaMA 3.1 8B, with only 2.05% trainable parameters via LoRA) and a deterministic rule-based post-processing layer. Trained on just 219 curated examples, the system is applied to multi-label compliance evaluation of conversational transcripts spanning 18 heterogeneous output fields. In blind evaluation on 53 previously unseen production transcripts, it achieves 100% JSON structural validity, 83.0% human-validated overall accuracy, and 100% accuracy on the most critical classification field. The proposed approach formalizes a hybrid neural-symbolic decomposition and introduces targeted hard-negative augmentation to improve performance on critical decision boundaries. Running on a single NVIDIA A100 GPU, inference completes in approximately 2 seconds, which is 2-5x faster than frontier-model APIs. The system costs only $0.013 per evaluation compared with $0.025-$0.055 for proprietary alternatives, resulting in 46-76% cost savings. These results demonstrate that domain-adapted small language models, when combined with deterministic post-processing, can match frontier-model accuracy for structured compliance evaluation while substantially reducing operational cost, latency, and privacy risk. Keywords: small language models, parameter-efficient fine-tuning, LoRA, domain adaptation, hybrid inference, compliance evaluation, structured output.

Sudhanshu Garg, Andrew Wang, Chaitanya Kulkarni et al.

We present the development and optimization of PayPal's Commerce Agent, powered by NEMO-4-PAYPAL, a multi-agent system designed to revolutionize agentic commerce on the PayPal platform. Through our strategic partnership with NVIDIA, we leveraged the NeMo Framework for LLM model fine-tuning to enhance agent performance. Specifically, we optimized the Search and Discovery agent by replacing our base model with a fine-tuned Nemotron small language model (SLM). We conducted comprehensive experiments using the llama3.1-nemotron-nano-8B-v1 architecture, training LoRA-based models through systematic hyperparameter sweeps across learning rates, optimizers (Adam, AdamW), cosine annealing schedules, and LoRA ranks. Our contributions include: (1) the first application of NVIDIA's NeMo Framework to commerce-specific agent optimization, (2) LLM powered fine-tuning strategy for retrieval-focused commerce tasks, (3) demonstration of significant improvements in latency and cost while maintaining agent quality, and (4) a scalable framework for multi-agent system optimization in production e-commerce environments. Our results demonstrate that the fine-tuned Nemotron SLM effectively resolves the key performance issue in the retrieval component, which represents over 50\% of total agent response time, while maintaining or enhancing overall system performance.

Chi Liu, Xin Chen, Xu Zhou et al.

Large Language Models (LLMs) have achieved remarkable success, underpinning diverse AI applications. However, they often suffer from performance degradation due to factors such as catastrophic forgetting during Supervised Fine-Tuning (SFT), quantization, and pruning. In this work, we introduce a performance recovery framework based on Self-Distillation Fine-Tuning (SDFT) that effectively restores model capabilities. Complementing this practical contribution, we provide a rigorous theoretical explanation for the underlying recovery mechanism. We posit that an LLM's generative capability fundamentally relies on the high-dimensional manifold constructed by its hidden layers. To investigate this, we employ Centered Kernel Alignment (CKA) to quantify the alignment between student and teacher activation trajectories, leveraging its invariance to orthogonal transformations and scaling. Our experiments demonstrate a strong correlation between performance recovery and manifold alignment, substantiating the claim that self-distillation effectively aligns the student's high-dimensional manifold with the optimal structure represented by the teacher. This study bridges the gap between practical recovery frameworks and geometric representation theory, offering new insights into the internal mechanisms of self-distillation.

Sarat Mudunuri, Jian Wan, Ally Qin et al.

Large Language Models (LLMs) with tool-calling capabilities have demonstrated remarkable potential in executing complex tasks through external tool integration. The Model Context Protocol (MCP) has emerged as a standardized framework for connecting LLMs to diverse toolsets, with individual MCP servers potentially exposing dozens to hundreds of tools. However, current implementations face a critical scalability challenge: providing all available tools to the LLM context results in substantial token overhead, increased costs, reduced accuracy, and context window constraints. We present a semantic tool discovery architecture that addresses these challenges through vector-based retrieval. Our approach indexes MCP tools using dense embeddings that capture semantic relationships between tool capabilities and user intent, dynamically selecting only the most relevant tools (typically 3-5) rather than exposing the entire tool catalog (50-100+). Experimental results demonstrate a 99.6% reduction in tool-related token consumption with a hit rate of 97.1% at K=3 and an MRR of 0.91 on a benchmark of 140 queries across 121 tools from 5 MCP servers, with sub-100ms retrieval latency. Contributions include: (1) a semantic indexing framework for MCP tools, (2) a dynamic tool selection algorithm based on query-tool similarity, (3) comprehensive evaluation demonstrating significant efficiency and accuracy improvements, and (4) extensibility to multi-agent and cross-organizational tool discovery.

Ally Qin, Jian Wan, Sarat Mudunuri et al.

We evaluate speculative decoding with EAGLE3 as an inference-time optimization for PayPal's Commerce Agent, powered by a fine-tuned llama3.1-nemotron-nano-8B-v1 model. Building on prior work (NEMO-4-PAYPAL) that reduced latency and cost through domain-specific fine-tuning, we benchmark EAGLE3 via vLLM against NVIDIA NIM on identical 2xH100 hardware across 40 configurations spanning speculative token counts (gamma=3, gamma=5), concurrency levels (1-32), and sampling temperatures (0, 0.5). Key findings: (1) gamma=3 achieves 22-49% throughput improvement and 18-33% latency reduction at zero additional hardware cost; (2) acceptance rates remain stable at approximately 35.5% for gamma=3 across all conditions; (3) gamma=5 yields diminishing returns (approximately 25% acceptance rate); (4) LLM-as-Judge evaluation confirms fully preserved output quality; and (5) speculative decoding on a single H100 matches or exceeds NIM on two H100s, enabling 50% GPU cost reduction.