The middleman problem
Tool calling can flatten precise evidence
A specialist OCR system can return exact text, bounding boxes, reading order, and confidence. If those records pass through repeated tool calls, a large model remains the final author and may paraphrase or drop details.
Interfaze proposes a different boundary: fuse perceptual specialists into the decoder rather than treating them as external tools.
Native fusion
Different specialists, one embedding space
OCR, vision, GUI, and speech adapters project compact vectors into the same embedding space used by the transformer decoder. The decoder can compare semantic context with specialist evidence in one pass.
Browser, scraper, sandbox, web index, and vector storage remain a distinct action foundation. The paper's important distinction is fused perception versus built-in action.
Honest boundaries
Partial activation helps, but fusion has tradeoffs
A deterministic request can activate only the adapter it needs and may bypass the full decoder for fixed-schema output. The paper reports strong results across OCR, grounding, speech, reasoning, and structured output benchmarks.
Those results are paper-reported, not an independent evaluation. The authors also identify encoder coverage, fusion calibration, and cold-start latency as limitations.