简体中文 · English

TOPAS: Prefix-state scheduling for workflows¶

Core mechanism mini-suite. The paper-specific control flow or objective runs; deterministic task success is not paper benchmark reproduction.

Paper information¶

Field	Details
Paper	arXiv 2608.25523
Institution	University of Science and Technology of China (first listed affiliation of the first author)
First public release	2026-08-26 (arXiv v1)
Author code	No public author code found on 2026-08-28.
Local adapter or method key	`topas`
Local reproduction code	`src/auto_research/agent_research/latest_20260827.py`

Original paper summary¶

Background and main changes¶

Under a shared KV-cache budget, jointly estimate the longest remaining workflow path and downstream prefix-reuse benefit. Account for prefix movement, preemption and aging rather than optimizing only local hit rate or one request's progress.

flowchart LR
  I[Task and current state] --> M[Critical path, prefix state and aging]
  M --> O[Policy action or training objective]
  O --> E[Auditable metrics and feedback]
  E --> M

Key figure from the paper¶

Figure 2: Overall workflow, stages and data flow. From the paper; copyright belongs to the authors. Click for the source.

Core equations¶

\[ \Phi(S)=\sum_j w_j\,CP_j(S)-\lambda B_{reuse}(S)+\mu C_{move}(S). \]

Offline and online results¶

Mean/p99 JCT falls by up to 39.8%/49.4% on synthetic workflows and 22.0%/26.6% on MetaGPT-TL.

Local reproduction¶

L2 isolated capability evaluation (primary result)¶

In the stateful tool environment, TOPAS safely ranks and compresses candidates, then recovers based on call feedback. Three-seed ToolRoute-L2.1 joint success is 0.7556, plan-step F1 0.7887 and average cost 4.5571.

See ToolRoute metrics. PlanBench-mini below is only a mechanism diagnostic.

Comparison: On 120 PlanBench-mini episodes, joint success is 1.0000 and average cost 0.4098. Only prefix hits, critical-path updates and aging are simulated.

See PlanBench metrics and batch index.

auto-research agent-eval --method topas --benchmark planbench-mini --episodes 120 --seed 42

Scope¶

The public deterministic candidate/plan mini-suite isolates mechanism differences. The paper checkpoint, private training data, external services and full benchmark are not reproduced.