English · 简体中文
NLP · AI for Math · Crypto
Open source made new things possible for more people.
Research index · Proofs, counterexamples & extensions
Two negative answers from the AIM problem list · 2004 / 2007 · Polynomial zeros
Questions dating back two decades ask whether polynomial transformations preserve properties of their zeros. AIM 23: a counterexample increases nonreal zeros from 4 to 6. AIM 34: from asking whether sufficiently large integer parameters work to ruling out every positive real parameter. Both include Lean sources, historical replay records and independent AI review. Background: Craven–Csordas · Chasse's dissertation
BRW graph polynomials: progress on a 1994 question · Canadian Journal of Mathematics · 1994 · Graph theory
When graphs are assembled, which polynomial-zero properties survive? A uniform result extends particular cycle lengths to every cycle length: replacing one cycle vertex with a graph satisfying the stated condition preserves the relevant real-rootedness property and permits further attachment along an edge. This covers a whole construction family, as a partial result toward the general conjecture.
Stack-sorting: a uniform proof draft for permutation counts · 2026 · Combinatorics
A proof draft explaining a stronger form of log-concavity in counts of permutations sharing a sorting output. This draft's Git commit: November 4. The proof and a comparison with related work by Colin Defant* are public.
* Colin Defant, Harvard postdoctoral fellow; Theorem 1.1 of his November 10 arXiv preprint gives the corresponding conclusion.
More mathematical work · Full index & verification materials
Dense & Hybrid Retrieval · Knowledge Graphs · Retrieval-Augmented QA
Research interests in retrieval and question answering over text–graph knowledge bases, focusing on complementary retrieval channels, relation-aware entity disambiguation, and coordination between retrieval and reranking.
Current work uses STaRK-MAG to explore how entity text, attributes and graph relations can jointly guide candidate selection under both semantic and relational constraints.
Research-to-Decision — an agent workflow that turns research into a choice. Built on BMad Deep Recon and OpenInquire’s cross-platform reading, it clarifies requirements, compares evidence and trade-offs, and explains when to choose an alternative.
Draft engineering · Sentence Transformers: synchronized per-dataset batches #4317
Align tasks across ranks while allowing task-specific per-rank batch sizes; preserve the training schedule across checkpoint resume. 18 real two-rank CPU/Gloo Trainer runs (6 configurations × full / prefix / resume) matched native-objective checks and continuous-versus-resumed trajectories in deterministic fixtures.
Not merged: its dependency Accelerate #4491 is currently closed and unmerged; upstream CI awaits maintainer approval.
Scheduling diagram, fixed evidence & limits
Fixed reproduction evidence · 2 November 2026 includes rank traces, source hashes and JSON-only oracle replay. The 18 jobs were run in an independent validation environment, not upstream CI. The diagram illustrates scheduling, not measured throughput. Scope: fixed two-rank CPU/Gloo, ordinary map-style sampling and deterministic resume fixtures; no GPU/NCCL, arbitrary stochastic-resume or performance claim. AI-assisted implementation, tests and evidence review.
Selected merged contributions:
Hugging Face Tokenizers · #2567 · merged
Rust · Tokenization
Fixed Kimi-K2 tokenization of special Han characters, restoring encoding consistency. The maintainer-reported regression was reviewed and merged.
LightRAG · #4368 · merged
Python · PostgreSQL · pgvector
Fixed missing data after interrupted vector migrations. The maintainer extended the fix and verified duplicate-free migration of 30,000 vectors across two instances. Regression tests also earned maintainer credit and informed their merged fix for data-safe shutdown retries.
Sentence Transformers · #4251 · #4262 · #4264 · #4284
Python · PyTorch · Semantic retrieval
4 merged fixes: corrected retrieval rankings and training negatives, restored reranker scoring, and kept empty results from aborting evaluation.
REA · #2422 · #2448 · #2429
TypeScript · Node.js · Reverse engineering
3 merged contributions: cross-platform path matching, binary diagnostics and installation guidance. The diagnostic fix was independently verified on Linux.
Spec Kit · #4956
PowerShell · Python · Spec-driven development
Prevented plan files from being overwritten on Windows paths containing special characters. Verified on two PowerShell generations and merged.
More merged contributions
- Haystack
· #13422 · RAG data preprocessing
- spdlog
· #3683 · Cross-platform logging reliability
- Resty
· #1250 · #1266 · Go · HTTP streaming & concurrency
- GitHub gitignore
· #4931 · Development environments & version control
- Claude Trading Skills
· #596 · Trading risk & position management
- awesome-python
· #3376 · Web accessibility & navigation
- IPython
· #15435 · Interactive Python file workflows
- Oh My Zsh
· #14184 · Shell workflow documentation
- KOReader
· #16220 · Reading & annotation workflows
- JSON5
· #371 · Data serialization
- Xarray
· #11711 · Documentation workflow reliability
- pydicom
· #2388 · Data conversion & interoperability
- Developer Roadmap
· #10451 · Python learning resources
- Free Programming Books
· #13533 · Open-source collaboration docs


