Paper Finalized

The paper is finally done. The manuscript for ICIC submission has been generated, and after several days of revisions and formatting, I’m planning to officially submit within the next couple of days. Fingers crossed everything goes smoothly.

The paper centers on the impact of data filtering strategies on SWE-agent LoRA fine-tuning performance. All core experimental results have been completed and written into the main text. Here’s a screenshot of some key results from the paper:

---

Key Experimental Findings

The most important conclusions from the paper:

H1: Gate Effect

ResolvedOnly filtering (keeping only resolved trajectories) shows no statistically significant difference compared to full-pool random sampling at the 500-sample scale. This indicates that a simple binary “resolved or not” gate alone is not sufficient to meaningfully improve fine-tuning quality.

H2: Score Effect

The sanity check comparing TopQ-500 vs BottomQ-500 reveals a statistically significant quality gap, validating the effectiveness of the composite scoring system. Score-based ranking can indeed distinguish “good data” from “bad data.”

H3: Scaling Effect

This is the most robust finding in the study. Doubling the training set from 500 to 1,000 trajectories yields a highly consistent ~12.7% relative loss reduction across all three strategies, with an absolute loss decrease of Δ ≈ −0.060 — approximately 26× larger than the quality filtering effect at 500 samples (Δ ≈ −0.003 for TopQ vs. Random).

Strategy	Loss @500	Loss @1000	Δ Loss	Relative Reduction
Random	0.4737	0.4140	−0.0597	12.6%
TopQ	0.4704	0.4106	−0.0598	12.7%
ResolvedOnly	0.4786	0.4175	−0.0611	12.8%

---

Submission Checklist

• ✅ Paper body and all figures/tables finalized
• ✅ Formatting meets ICIC 2026 submission requirements
• ⬜ Final proofreading pass
• ⬜ Submit to the submission portal

---

Coming Up

• Complete the final proofreading and detail check
• Officially submit to ICIC 2026
• Organize experiment code and data in preparation for potential revisions