Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Interfaze: The Future of AI is built on Task-Specific Small Models

Harsha Vardhan Khurdula, Vineet Agarwal, Yoeven D Khemlani

TL;DR

Interfaze tackles the problem that scaling LLMs alone is insufficient for deployed AI systems that must perceive, index external sources, and act. It proposes a context-centric architecture with three layers: a perception stack of small models, a context-construction layer, and an action layer that orchestrates tools before passing distilled context to a fixed LLM. Interfaze-Beta integrates OCR, chart/diagram parsing, multilingual ASR with diarization, retrieval, and sandboxed code execution behind a single endpoint, and demonstrates competitive or state-of-the-art results on knowledge, math, coding, and multimodal benchmarks. The work highlights that most gains come from the perception/tool stack and context compilation, suggesting practical routes to reduce reliance on bloated large models.

Abstract

We present Interfaze, a system that treats modern LLM applications as a problem of building and acting over context, not just picking the right monolithic model. Instead of a single transformer, we combine (i) a stack of heterogeneous DNNs paired with small language models as perception modules for OCR involving complex PDFs, charts and diagrams, and multilingual ASR with (ii) a context-construction layer that crawls, indexes, and parses external sources (web pages, code, PDFs) into compact structured state, and (iii) an action layer that can browse, retrieve, execute code in a sandbox, and drive a headless browser for dynamic web pages. A thin controller sits on top of this stack and exposes a single, OpenAI-style endpoint: it decides which small models and actions to run and always forwards the distilled context to a user-selected LLM that produces the final response. On this architecture, Interfaze-Beta achieves 83.6% on MMLU-Pro, 91.4% on MMLU, 81.3% on GPQA-Diamond, 57.8% on LiveCodeBench v5, and 90.0% on AIME-2025, along with strong multimodal scores on MMMU (val) (77.3%), AI2D (91.5%), ChartQA (90.9%), and Common Voice v16 (90.8%). We show that most queries are handled primarily by the small-model and tool stack, with the large LLM operating only on distilled context, yielding competitive accuracy while shifting the bulk of computation away from the most expensive and monolithic models.

Interfaze: The Future of AI is built on Task-Specific Small Models

TL;DR

Interfaze tackles the problem that scaling LLMs alone is insufficient for deployed AI systems that must perceive, index external sources, and act. It proposes a context-centric architecture with three layers: a perception stack of small models, a context-construction layer, and an action layer that orchestrates tools before passing distilled context to a fixed LLM. Interfaze-Beta integrates OCR, chart/diagram parsing, multilingual ASR with diarization, retrieval, and sandboxed code execution behind a single endpoint, and demonstrates competitive or state-of-the-art results on knowledge, math, coding, and multimodal benchmarks. The work highlights that most gains come from the perception/tool stack and context compilation, suggesting practical routes to reduce reliance on bloated large models.

Abstract

We present Interfaze, a system that treats modern LLM applications as a problem of building and acting over context, not just picking the right monolithic model. Instead of a single transformer, we combine (i) a stack of heterogeneous DNNs paired with small language models as perception modules for OCR involving complex PDFs, charts and diagrams, and multilingual ASR with (ii) a context-construction layer that crawls, indexes, and parses external sources (web pages, code, PDFs) into compact structured state, and (iii) an action layer that can browse, retrieve, execute code in a sandbox, and drive a headless browser for dynamic web pages. A thin controller sits on top of this stack and exposes a single, OpenAI-style endpoint: it decides which small models and actions to run and always forwards the distilled context to a user-selected LLM that produces the final response. On this architecture, Interfaze-Beta achieves 83.6% on MMLU-Pro, 91.4% on MMLU, 81.3% on GPQA-Diamond, 57.8% on LiveCodeBench v5, and 90.0% on AIME-2025, along with strong multimodal scores on MMMU (val) (77.3%), AI2D (91.5%), ChartQA (90.9%), and Common Voice v16 (90.8%). We show that most queries are handled primarily by the small-model and tool stack, with the large LLM operating only on distilled context, yielding competitive accuracy while shifting the bulk of computation away from the most expensive and monolithic models.
Paper Structure (27 sections, 12 equations, 1 figure, 2 tables)

This paper contains 27 sections, 12 equations, 1 figure, 2 tables.

Table of Contents

  1. Introduction
  2. Literature Survey
  3. Context is key
  4. Where routing falls short
  5. Tools and model composition
  6. Small language models and specialization
  7. Interfaze-Beta Architecture
  8. System overview
  9. Small-model and perception stack
  10. Context construction and backfilling
  11. Offloading large-context work
  12. Action layer and specialized model chain
  13. Perceptual Encoders at the Vision-Audio Interface
  14. ASR with Diarization
  15. Optical Character Recognition and Complex Document Parsing
  16. ...and 12 more sections

Figures (1)

  • Figure 1: Interfaze Architecture