HarnessMM reuses settled agent decisions to cut HumanEval token use by 87.93%
Researchers from Tsinghua and Microsoft Research Asia report that reusing resolved decisions lifted DeepSeek Harness success on 98 tasks from 71.43% to 85.71%.
Researchers at Tsinghua University and Microsoft Research Asia report that HarnessMM cut the token cost of DeepSeek Harness on HumanEval from 96.28k tokens per task to 11.62k, an 87.93% saving. Success on 98 tasks rose from 71.43% to 85.71%. The method keeps the underlying model's weights fixed. It works on the harness, the scaffolding that runs a model against tools, by reusing decisions the agent has already settled rather than asking the model again.
The problem it targets is repetition in agent loops. After tool feedback arrives, a harness often sends the model another request, even when an earlier call already settled the next action. HarnessMM stores those settled decisions as what the authors call Executable Decision Graphs. These record the exact action, its literal arguments, the causal dependencies between steps and public guards. A stored decision runs again when the observed tool outputs satisfy its guards. The authors' Certified Callback Substitution executes it through the native interface and records that the slow callback was removed. Anything new, any evidence the guards do not support, and any case where the selector abstains goes back to the original model.
The gains are not uniform across harnesses. Codex held 90.82% success on HumanEval while its tokens fell 29.54%, from 66.72k to 47.02k per task. The abstract attributes the DeepSeek Harness result to guarded reuse together with recursive self-improvement, which tunes companion harness configurations using training feedback. It does not separate how much of the accuracy change comes from each. Paid model calls on DeepSeek Harness fell from 954 to 415. The authors also report a separate MBPP ablation, which measured 26.91% lower cost than single-use graphs and 19.52% lower cost than retained callbacks.
The savings depend on provider cache accounting and on how many actions are already resolved. The same note says that authoring, selection and recovery costs can outweigh the avoided calls.
Why it matters
Repeated model calls in agent harnesses are a direct cost, so teams may save money before they need a larger model. The authors caution that savings depend on provider cache accounting and on how many actions are already resolved.
What you can do
Count repeated decisions in existing agent logs to estimate how much model spend is avoidable before changing harness design.
Original source
Where we saw it
DAIR.AI (Elvis Saravia)
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works