Di Wang

Research profile

Research summary

Di Wang is a researcher at King Abdullah University of Science and Technology whose recent work, as captured in Evolve's catalog, concentrates on making large language models more efficient, more faithful, and more verifiable. The record splits evenly between two research programs — machine learning (federated learning and LLM efficiency) and neuro-symbolic AI for LLM reasoning verification, each 50% of the field distribution — with an additional natural language processing paper on decoding. Three 2026 arXiv preprints anchor the profile: FedSubMuon, which cuts federated fine-tuning upload cost to O(r²) via structured subspace Muon; LogicTrack, which audits chain-of-thought trajectories through LLM autoformalization and Z3 backtracking; and SFAD, a speculative decoding method pairing a DPO-trained faithful draft with friction-gated logit steering for a reported 2.48x speedup. Di Wang is listed last on two of the three papers and in the middle on one, consistent with a senior supervisory role. Bibliometric signals are early-stage: 10 distinct coauthors with a collaboration score of 48, and a reproducibility mean of 55 across two assessed papers, all at medium readiness, with Evolve flagging missing significance tests and unproven formalization fidelity. Output, impact and activity dimensions lack sufficient data and are not scored.

Research keywords

Highlights

• Formal-logic auditing of LLM reasoning traces

  Last-authored 2026 preprint audits chain-of-thought trajectories via LLM autoformalization with Z3 backtracking search; Evolve credits broad verifiability gains but flags unproven formalization fidelity and solver-versus-judge attribution. Reproducibility readiness: medium.

• Subspace Muon for communication-efficient federated LLM fine-tuning

  Last-authored 2026 preprint compresses federated LLM fine-tuning upload cost to O(r²) via structured subspace Muon; strong matched-budget wins are reported, though Evolve notes tiny headline margins without significance tests. Reproducibility readiness: medium.

• Factuality-aware speculative decoding (SFAD)

  Middle-author 2026 preprint pairs a DPO-trained faithful draft model with friction-gated logit steering, reporting large faithfulness gains at a 2.48x decoding speedup; Evolve notes remaining latency inconsistencies. Reproducibility readiness: medium.

Timeline

1. 2026Last-authored arXiv preprint LogicTrack: auditing LLM reasoning trajectories with formal logic solvers (autoformalization plus Z3 backtracking).
2. 2026Last-authored arXiv preprint FedSubMuon: communication-efficient federated LLM fine-tuning via structured subspace Muon, cutting upload cost to O(r²).
3. 2026Middle-author arXiv preprint SFAD: speculative factuality-aware decoding combining a DPO-trained faithful draft with friction-gated logit steering (2.48x speedup).
4. 2026All three Evolve-reviewed preprints land in a single year across federated LLM efficiency, reasoning verification and faithful decoding, each at medium reproducibility readiness.

Papers in the Evolve catalog

• 2026-09-18

  LogicTrack: Auditing Reasoning Trajectories of Large Language Models with Formal Logic Solvers

• 2026-09-05

  FedSubMuon: Communication-Efficient Federated LLM Fine-Tuning via Structured Subspace Muon

• 2026-09-01

  SFAD: Speculative Factuality-Aware Decoding

Frequent co-authors