Selected research · Technical reports · Open-source projects · Open Research Reports
WSDM 2027 · Accepted · Project overview · Publication record
Across 180 questions and 1,440 matched settings, checking the proposed evidence set reduced mean normalized regret from 0.272 to 0.162 as mean scoring calls rose from 9 to 9.61. Taking the best already-measured set at nine calls gave 0.213. This offline evaluation uses reference answers.
Read the report · 中文伴读 · Reproduce on CPU
When Is Program Structure Required? Certificate-Grounded Calibration of Program-Induction Difficulty
arXiv preprint · November 2026 · Paper · Project overview
Across 120 reference families, changing one demonstration can raise the certified minimum program length from one rule to two. With five shared examples, Qwen3-8B completed 67/120 control tasks and 12/120 separation tasks. Its success on a separate hidden witness pair increased by 9.17 percentage points.
Technical report · 中文伴读 · Hugging Face explorer · GitHub release
AACL-IJCNLP 2026 · Long Paper · Project overview · ACL Anthology
Qwen ranked 148 of 150 matched pairs correctly. A threshold learned on other pairs classified both members correctly in 132, and the separate generation check succeeded on 105.
Read the report · 中文伴读 · Reproduce on CPU
I share completed studies here when I consider their evidence too narrow for the conference I had in mind. Each release includes the manuscript, its measured results and a CPU reproduction path. Browse the collection · 中文目录 · Hugging Face collection
Follow the same RAG question as two passages change order. Across 539 questions, relaxing the passage budget accounts for 4.79 of Mistral’s 5.91 token-F1-point gain.
PDF · Technical report · 中文技术报告 · Hugging Face · Release
Two 23-state observation plans differ in which city their edits share. On atomic text the shared-value plan reduces reconstruction error by 53.4% for BGE and 36.0% for E5.
PDF · Technical report · 中文技术报告 · Hugging Face · Release
On 32 paired worlds, document replay leads closed-book replay by 19.25 percentage points with synthetic attachments and trails by 7.92 points with WikiText. Sixteen fresh worlds preserve both directions.
PDF · Technical report · 中文技术报告 · Hugging Face · Release
- Independent Researcher · Late April 2026 – Present
- Research Intern, Earth and Space Center · July 2025 – Late April 2026

