CodeSpecWeave.
CODESPECWEAVE / RESEARCH 2026

When
requirements
interact.

Scaling Long-Horizon Coding
via Interdependent Specifications

Real software is a web of constraints. We build training environments that teach coding agents to reason across them.

Read the paper Explore the idea
+ INDEPENDENT RULES → SHARED DECISIONS
interface compatibility
state consistency
error recovery
execution order
functional semantics
✳One coherent
implementation.
INTERDEPENDENT BY DESIGN
Yishuo Cai1,2,*‡Minzheng Wang1,*Liang Zhao1Linghao Zhang1Lei Li1,3Shicheng Li1Xu Sun2,†Fuli Luo1,†
1 LLM Core, Xiaomi2 Peking University3 University of Hong Kong

* Equal contribution † Corresponding authors‡ Work done during an internship at Xiaomi

4,670.executable training tasks
2,865.real-world repositories
100.held-out benchmark tasks
1,044.benchmark specifications
01 THE CORE IDEA

One change.
Many consequences.

A new feature can change state, break an interface, or alter an error path. Our tasks connect these requirements, making joint reasoning essential.

EXPLORE A REAL TASKskycoin / skycoin

Share fairly. Stay compatible.

Allocate transaction coin hours across multiple recipients,
while preserving the existing single-recipient behavior.

SHARED DECISION SPACE3 connected requirements
Requirements share one transaction builder. Select a requirement to see its connections. Proportionality Change allocation Minimum allocation Output order Compatibility transactionbuilder
THE INTERACTION

A fair split needs a shared policy.

Allocate hours in proportion to coin amounts. This decision must also respect minimum allocations and deterministic rounding in request order.

A locally correct ratio is not enough to satisfy the whole transaction contract.
Click a requirement to trace its connections.Illustration from the paper’s Task Example · not an execution trace

We weave requirements around shared state, execution paths, and lifecycle constraints—so implementation decisions have to work together.

02 THE FRAMEWORK

From real requirements
to reliable environments.

A synthesis agent builds coupled tasks and executable tests. Multi-model trials and trajectory review check whether the tests reward the right behavior.

01 / GROUND⌘

Start with real engineering.

Extract and deduplicate developer requirements into a reference of specification patterns.

REAL REPOSITORIES · REAL CONSTRAINTS
02 / SYNTHESIZE↔

Weave tasks with tests.

Connect requirements through shared decisions. Co-develop each task and its verifier in the repository.

ITERATIVE SYNTHESIS · EXECUTABLE FEEDBACK
03 / VALIDATE⤴

Test the tests.

Review solver trajectories for false positives and false negatives. Repair selected tasks and rerun checks.

MULTI-MODEL REVIEW · REVALIDATION
⌁ See the complete synthesis & validation pipeline+
Paper pipeline: task and verifier synthesis, baseline check, solver trials, difficulty filtering, trajectory review, and repair followed by revalidation.
Original method figure from the paper. Task requirements and test cases have many-to-many coverage.
03 TRAINING RESULTS

Better environments.
Stronger coding agents.

Reinforcement learning on CodeSpecWeave improves both MiMo backbones, with the largest gains on demanding long-horizon engineering tasks.

MODEL BACKBONE
EVALUATE ON
DeepSWE-v1.1
+8.56pp

with CodeSpecWeave RL

BASE MODEL54.42%+ CODESPECWEAVE62.98%
Training settingSuccess rate 0–100%ScoreΔ vs. base
Base model · pp = percentage pointsFull results table ↓

Paper Table 2 · three samples per task · 512K context

Full results across four benchmarks+
Task success rates (%). Average is the unweighted mean of the four benchmarks; best values within each backbone are bold.
Training dataCodeSpecWeave-BenchDeepSWE-v1.1SWE-bench VerifiedSWE-bench ProAverage
01

Gains on both backbones. Highest four-benchmark average among the compared training datasets.

02

Transfer to long-horizon work. DeepSWE-v1.1 improves from 12.39% to 25.30% on MiMo-V2.5.

03

More sustained verification. Flash-SFT edit–test loops increase from 2.96 to 4.90 per trajectory.

04 CODESPECWEAVE-BENCH

How well do frontier agents
hold it all together?

100 tasks. 100 repositories. 1,044 specifications.
A benchmark for engineering changes that satisfy the entire contract.

HIGHEST REPORTED PASS RATE
74.0%

GPT-6 Astra · Codex

PERFORMANCE SPREAD
30.0pp

74.0% to 44.0% across configurations

Frontier model comparison

Reported pass rates across 10 model–framework configurations
Click column headers to sort
Model and agent framework results from the paper. Bars share a 0 to 100 percent scale. Gap is percentage points below the overall leader, GPT-6 Astra at 74.0 percent.
#Agent frameworkGap to best
10 configurationsShared 0–100% scale · Swipe to compare →

Paper frontier comparison. Colors identify agent frameworks; configurations use different frameworks. These scores are separate from the three-sample RL evaluation above. Gap is measured against GPT-6 Astra’s reported 74.0%.

Cite this work

Download BibTeX 
@misc{cai2026requirementsinteractscalinglonghorizon,
      title={When Requirements Interact: Scaling Long-Horizon Coding via Interdependent Specifications},
      author={Yishuo Cai and Minzheng Wang and Liang Zhao and Linghao Zhang and Lei Li and Shicheng Li and Xu Sun and Fuli Luo},
      year={2026},
      eprint={2610.40256},
      archivePrefix={arXiv},
      primaryClass={cs.SE},
      url={https://arxiv.org/abs/2610.40256},
}