Inference-Time Distillation: Cost-Efficient Agents Without Fine-Tuning or Manual Prompt Engineering

2025/12/02 by Vishnu Sarukkai, Sarukkai, Vishnu, Gupta, Asanshay +6
Computer Science · Materials Science · #Explainable Artificial Intelligence (XAI) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning in Materials Science #Topic Modeling

paper · pdf · doi:10.48550/arxiv.2512.02543

openalex publication_date 2025/12/02 · openalex created_date 2025/12/04 · openalex updated_date 2026/07/28

Abstract

Deploying LLM agents at scale typically requires choosing between quality and cost. Existing cost-reduction approaches fail to preserve agility: the ability to iterate rapidly without human time bottlenecks. Prompt engineering is brittle and slows iteration, while fine-tuning requires multi-day training and commitment to fixed designs; both are impractical for iterative workflows and time-sensitive batch jobs. We demonstrate that established inference-time techniques--dynamic in-context learning and self-consistency cascades--can be leveraged to shift the cost-accuracy Pareto frontier while preserving agility. Practitioners run the teacher on a small task subset to collect demonstrations, then immediately deploy a cheaper student on the remainder. At each step, the system retrieves relevant teacher demonstrations as in-context examples. When multiple student samples agree, we proceed; when they diverge, we fall back to the teacher. This requires no prompt engineering or training. On ALFWorld, we match teacher accuracy at 2.5x lower cost (0.059 to 0.024 per episode). On AppWorld, we achieve 3.5x cost reduction while recovering 79% of teacher accuracy. Our empirical analyses provide guidance on key design choices: teacher database size, demonstration set size, retrieval strategy, and cascade thresholds. These analyses highlight inference-time levers for navigating cost-performance tradeoffs without sacrificing human development speed.

Citations

• Self-Generated In-Context Examples Improve LLM Agents for Sequential Decision-Making Tasks
• Towards Enterprise-Ready Computer Using Generalist Agent
• Deeper Insights Without Updates: The Power of In-Context Learning Over Fine-Tuning
• AppWorld: A Controllable World of Apps and People for Benchmarking Interactive Coding Agents
• RouteLLM: Learning to Route LLMs with Preference Data
• Fine-Tuning or Fine-Failing? Debunking Performance Myths in Large Language Models
• Many-Shot In-Context Learning
• Language Model Cascades: Token-level uncertainty and beyond
• AutoGuide: Automated Generation and Selection of Context-Aware Guidelines for Large Language Model Agents
• TRAD: Enhancing LLM Agents with Step-Wise Thought Retrieval and Aligned Decision
• Soft Self-Consistency Improves Language Model Agents
• RAP: Retrieval-Augmented Planning with Contextual Memory for Multimodal LLM Agents
• AgentTuning: Enabling Generalized Agent Abilities for LLMs
• In-Context Learning with Iterative Demonstration Selection
• FireAct: Toward Language Agent Fine-tuning
• DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines
• ExpeL: LLM Agents Are Experiential Learners
• When Does Confidence-Based Cascade Deferral Suffice?
• FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance
• Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes
• Reflexion: Language Agents with Verbal Reinforcement Learning
• SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models
• Distributing Accountability, Not Capability: Phase Separation and the LLM Workflow Quadrant in Autonomous AI Agent Architectures
• Selective Annotation Makes Language Models Better Few-Shot Learners
• Confident Adaptive Language Modeling
• TangoBERT: Reducing Inference Cost by using Cascaded Architecture
• Self-Consistency Improves Chain of Thought Reasoning in Language Models
• What Makes Good In-Context Examples for GPT-3?
• MiniLMv2: Multi-Head Self-Attention Relation Distillation for Compressing Pretrained Transformers
• CascadeBERT: Accelerating Inference of Pre-trained Language Models via Calibrated Complete Models Cascade
• Wisdom of Committees: An Overlooked Approach To Faster and More Accurate Models
• ALFWorld: Aligning Text and Embodied Environments for Interactive Learning
• Distilling the Knowledge in a Neural Network
• AutoManual: Constructing Instruction Manuals by LLM Agents via Interactive Environmental Learning

Related

• Distilling LLM Agent into Small Models with Retrieval and Code Tools
• AgentDistill: Training-Free Agent Distillation with Generalizable MCP Boxes
• Leveraging Zero-Shot Prompting for Efficient Language Model Distillation
• Structured Agent Distillation for Large Language Model
• Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models