Project2Task: Graph-Guided Project-Level Planning for Autonomous Research
Huirui Xu, Runtao Xu, Shuo Ren, Jiajun Zhang
cs.AI
2026-08-05
Project2Task turns a research brief into graph-routed task contracts: portfolio quality 7.15 vs 4.58, and AutoResearchClaw accuracy 0.536 to 0.759.
Research agents such as AI Scientist, Agent Laboratory, and AutoResearchClaw can already search the literature, write code, run experiments, and draft a paper from a single topic. A real lab project is usually larger than that: several related contributions, some parallel, some waiting on an upstream artifact. Current systems either stuff the whole brief into one oversized job or emit a flat list of vague, overlapping tasks, leaving humans to police boundaries and order.
Project2Task is the missing layer before execution. It turns a macro project brief into a portfolio of bounded, dependency-aware task contracts that do not assume a particular downstream executor.
The planner takes a structured brief, a resource profile, and retrieved literature. Candidate contributions become innovation atoms: each node stores a title, the problem it solves, a novelty claim, a validation path, and evidence. Edges keep only source, target, a rationale, and an optional required artifact. The result is a directed acyclic lineage graph, finer-grained than the final tasks.
Decomposition is not left to free-form prompting. A router scores four templates on the same graph. Horizontal clustering finds modularity communities on the undirected projection, matching separable contribution modules. Vertical assignment follows topological levels, matching staged research chains. Horizontal-then-vertical runs parallel modules into a sink integration task. Vertical-then-horizontal starts from a shared foundation, then branches.
Each template labels dyads as positions that should or should not contain an edge, fits a two-parameter Bernoulli block model, and picks the template with the lowest negative log-likelihood. The analogy to stochastic block models stops at the scoring formula: the system does not infer latent communities from data. Hybrid templates that cannot form the required sink or foundation block are dropped before ranking.
After routing, graph blocks are blueprints, not hard partitions. The planner may merge or split them, then generate tasks. A repair loop checks coverage, claim overlap, literature grounding, dependency feasibility, and execution readiness. Missing evaluations, dropped core claims, or sibling tasks without boundary rules block handoff. Each contract then lists the objective, inputs, expected artifacts, evaluation requirements, must-cover and must-not-cover rules, shared assets, dependencies, and execution order. A thin adapter renders a contract into whatever prompt an executor already understands.
The benchmark has ten project briefs, about thirty tasks. Seeds are seven NanoResearch items and three ARC-Bench items, expanded by an LLM into project-level briefs across vision, language, tabular, time series, graphs, audio, and multimodal work. Portfolio quality is scored from generated manuscripts. Qwen3.6-Plus grades coherence, coverage, overlap control, consistency, and task division from 1 to 10, five independent runs per dimension.
| Setting | Mean portfolio score |
| Full brief given to AutoResearchClaw | 4.58 |
| Planner task titles only | 5.31 |
| Full Project2Task contracts | 7.15 |
| Forced runner-up decomposition | 6.48 |
Coherence is 7.28 against 4.63 and 4.50. Coverage is 7.50 against 5.50 and 5.40. Overlap control reaches 8.06; task division 7.58. Consistency sits at 5.34, the weakest of the five. The selected template beats the runner-up on 9 of 10 queries; query1 is the exception, 6.84 versus 6.68.
Fed into AutoResearchClaw, mean task accuracy moves from 0.536 to 0.759. The baseline records three invalid runs as 0.000; the planned contracts produce a valid result on every task. Dropping those three failures, the mean is still 0.830 versus 0.766, with gains on 8 of 10 tasks. Query 5, lightweight time-series classification on UCI HAR, goes from 0.699 to 0.832 after the contract turns "lightweight and interpretable" into a >30% parameter cut versus a 1D CNN, single-GPU AMP, and a ban on Transformers and multi-GPU training.
The autoresearch literature has been racing single-task execution. This paper isolates the cut that happens before execution: how a project becomes non-overlapping tasks with explicit ownership. The contracts are executor-agnostic, so an orchestration stack can drop them in as a planning plugin. Titles alone lift the score from 4.58 to 5.31; the structured fields are what take it to 7.15.
Treat it as incremental systems work, not a new scientific method. Ten projects, manuscripts as the quality proxy, one executor.
The authors flag Query 6, which falls from 0.747 to 0.542. A contract can over-constrain search or fight the executor's coding habits. Because planning and execution are decoupled, a bad executor still fails.
The evaluation is the larger caveat. The ten briefs are inflated from single-task seeds, not real lab charters. Portfolio quality is judged from generated manuscripts by an LLM, so whether code, data, and experiments actually compose a project is out of scope. The 0.759 average mixes heterogeneous datasets. Consistency at 5.34 means shared terminology and experimental settings still drift. There is no human panel on the portfolios, and no second executor.