Dependency-Aware Trajectory Refinement for Efficient Multi-Turn Agent Fine-Tuning

简体中文

L1 core-mechanism diagnostic. Runs dependency pruning and merging on auditable fixture DAGs, not an LLM annotator.

Paper information

Field	Details
Paper	Dependency-Aware Trajectory Refinement for Efficient Multi-Turn Agent Fine-Tuning
Organization	First-author affiliation: ShanghaiTech University
First public release	2026-09-16 (arXiv v1)
Authors' code	Alibaba-NLP/VLLM-KB is public.
Adapter / method	dependency-refinement
Local reproduction code	src/auto_research/agent_research/latest_20260919.py

Original paper

Motivation and method

Represent multi-turn trajectories as a turn-level dependency DAG. Prune leaves, merge under strict rules, then merge under looser rules to remove redundant messages without deleting final-answer dependencies.

flowchart LR
  I[Public input and current state] --> M[dependency-refinement mechanism]
  M --> A[Auditable intermediates]
  A --> O[Output or state update]

Key figure from the paper

Figure 2: Overview of the paper's method and its key components. From the original paper; copyright remains with the authors. Click the image for the source.

Core computation

The local reference kernel preserves the paper's decisive gates, weights, state transitions, or scheduling rules and records intermediates in the metrics artifact. Implementation functions and invariant assertions in tests map the formulas and variables.

Paper results

Reported online results, benchmarks, training efficiency, or inference speed are original paper results. The local three-seed artifact checks core mechanisms, shapes, and state invariants without comparing directly to the paper's numbers.

Reproduction boundaries

Dependency pruning and merging run locally. The DAG comes from an auditable fixture and is not produced by an LLM annotator.