Intel / article

Long-Horizon State Tracking in LLMs: Executing MD5 through a Deep Sequence of Dependent Tool Calls

Source: arxiv.org
Author: Dheeraj Mohandas Pai, Lu Xian
Date: Published Sep 2, 2026

Context

A study uses MD5, a deterministic calculation producing a fixed-length fingerprint, to test whether gpt-oss-120b can preserve exact intermediate values across 196 dependent tool calls. Because every step has a known answer, researchers can locate bookkeeping errors rather than merely mark the final response wrong. Keeping the model’s intermediate reasoning in subsequent was central to successful runs. The paper demonstrates that such long sequences are possible.

Its success rate needs a careful denominator. One reported serving setup produced 21 correct results among 34 runs that returned a fingerprint. Another 71 records without a final result, including aborted development runs, were excluded. This establishes feasibility more clearly than dependable overall completion.

Terms in this piece · Glossary

context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

Read the source arxiv.org