# Summary: 2026-08-03_02-22-02Z_Post_TrainingonOfficeWorkImprovesSoftwareEngineeri.md
Saved: 2026-08-04 00:24
Source: 2026-08-03_02-22-02Z_Post_TrainingonOfficeWorkImprovesSoftwareEngineeri.md
Model: None

---

## Summary  
The paper investigates whether long‑horizon post‑training on office‑work tasks can enhance a large language model’s performance on software‑engineering benchmarks, proposing that the improvement stems from strengthened goal‑directed execution (GDE) behaviors. By applying Qwen3.5-122B-A10B to 363 Long‑Horizon Multi‑Tool Agent (LHMTA) tasks drawn exclusively from office workflows—none of which are software‑engineering problems—the authors observe a 5.8‑point increase in SWE‑Bench Pro pass@1, indicating cross‑domain transfer beyond mere factual recall. The study also provides a behavioral account, showing that the model’s four GDE behaviors (goal selection, state construction, fidelity maintenance, and verification) improve uniformly across both office and code domains.

## Semantic links
- Summary: 2026-08-02_18-15-49Z_Cluster_AwareOver_the_AirFederatedLearningwithEner.md — 3 title terms overlap; 8 backlinks; 8 summary/topic terms overlap

## Key Contributions  
- [The post‑training on office tasks yields a 5.8‑point improvement in SWE‑Bench Pro pass@1.]  
- [All four GDE behaviors (goal selection, state construction, fidelity maintenance, verification) are enhanced after the transfer.]  
- [Aggregate statistics reveal related changes in information gathering, implementation, and verification processes.]

## Methodology  
The authors approached the problem by constructing a controlled experiment that isolates domain‑independent behavioral mechanisms. They selected 363 LHMTA tasks from office workflows—tasks that require sustained state tracking and goal alignment but lack any code generation. The model Qwen3.5-122B-A10B was first fine‑tuned on these tasks, then evaluated on SWE‑Bench Pro to measure pass@1 accuracy. Trajectory analysis was performed to compare the behavior patterns before and after post‑training, focusing on the four GDE components across both office and software domains.

## Results  
The main experimental results show a statistically significant rise in SWE‑Bench Pro pass@1 from 48.2 % to 54.0 %, corresponding to a 5.8‑point gain. The trajectory analysis confirms that each of the four GDE behaviors improves: goal selection becomes more precise, state construction is richer and more consistent, fidelity maintenance increases (fewer drift events), and verification accuracy rises (more correct completion checks). When aggregated across all tasks, the model’s information‑gathering score improves by 0.9 points, implementation score by 1.2 points, and verification score by 0.7 points.

## Significance  
These findings matter because they demonstrate that long‑horizon post‑training can produce measurable gains in downstream tasks without any direct exposure to those tasks, suggesting a deeper cognitive mechanism rather than superficial memorization. The behavioral interpretation provides a new lens for evaluating cross‑domain transfer and informs strategies for building more adaptable AI agents capable of maintaining coherent goals across diverse environments.

## Related Concepts  
- Goal‑directed execution (GDE) – the four-behavior framework for long‑horizon tasks.  
- Long‑horizon tasks – requiring sustained state and goal alignment.  
- Multi‑tool agent (LHMTA) – a benchmark for multi‑domain sequential reasoning.  
- Domain generalization – transfer of performance across unrelated domains.  
- Post‑training fine‑tuning – applying additional training data to an already trained model.  
- SWE‑Bench Pro – a suite of software‑engineering evaluation benchmarks.