简体中文 · English

SWRouter: Similarity-Contractive Window Routing for Multi-Turn Large Language Model Conversations

L1 core mechanism diagnostic. Window segmentation and profile-overlap routing run. The paper's contrastive encoder is not trained locally.

Paper information

Field	Details
Paper	SWRouter: Similarity-Contractive Window Routing for Multi-Turn Large Language Model Conversations
Institution	Shanghai Jiao Tong University (first author affiliation)
First public release	2026-09-10
Original source code	No public author repository found on 2026-09-14.
Adapter/method	swrouter
Local reproduction code	src/auto_research/foundation_latest_20260914.py

Original paper summary

Background and main changes

Contract multi-turn windows by adjacent semantic similarity, then use a contrastive router to select a model. Evaluate context construction separately from routing correctness.

flowchart LR
  I[Input and public state] --> M[SWRouter core mechanism]
  M --> A[Auditable intermediates]
  A --> O[Output or update]

Key figure from the paper

Figure 1: Overall design and components. From the paper; copyright belongs to the authors. Click for the source.

Core operation

The local deterministic reference kernel runs the defining operations and exposes gates, selections, projections, graph edges or state-update statistics. Tests check formula invariants and boundaries.

Offline and online results

The paper reports 16.26% over the best single model and another 8.22% over Conv-ID Context. These are not directly compared with scaled local results, and no online gains are inferred without reported A/B.

Local reproduction

See three-seed metrics. Candidates share public fixtures, budgets and seeds. The run establishes execution of core state transitions only.

Scope

Window segmentation and profile-overlap routing execute. No paper contrastive encoder was trained.