Jiaojiao Jiang

Research profile

Research summary

Jiaojiao Jiang is a researcher at UNSW Sydney whose record, profiled from OpenAlex bibliographic data, shows an unusually intense 2026: 24 works in the last three years, with counts-by-year jumping from 1 work in 2025 to 23 in 2026. The five reviewed papers span four distinct areas. In machine learning, Jiang appears as last author on two neural architecture search papers — RiPPLE and CoRA-NAS — both building performance prediction from zero-cost proxies and learning-curve extrapolation, with CoRA-NAS arguing for label-free cross-space ranking. In theory, Jiang contributes to a Lean-checked characterization of Gold-style language generation in the limit via finite witnesses, and to nearly tight Rademacher complexity bounds for sparsely activated ReLU networks that remove a √n factor with matching lower bounds. In AI security, DRIFT presents a black-box diffusion-watermark removal attack claiming 98–100% success across nine schemes. OpenAlex topics additionally flag steganography and watermarking (10), digital media forensics (9), and adversarial robustness (6), suggesting breadth beyond the reviewed set. Citation indicators remain thin — h-index 1, i10-index 0 — so output and impact are left unscored; this profile is name-resolved from public bibliographic data.

Research keywords

Highlights

• Resolves an open characterization of language generation in the limit

  Using finite witnesses, the paper answers the open question left by Raman et al. and attains every width level up to ω+1; Evolve notes the core proofs are Lean-checked and rates reproducibility high.

• Nearly tight Rademacher bounds for sparsely activated networks

  The work removes the √n factor of ADKM's bound up to a log^{3/2}(2m) term and supplies matching lower bounds; Evolve finds the proofs self-contained, though core revisions were AI-assisted.

• DRIFT: black-box removal of diffusion watermarks

  By deflecting the generative trajectory, the attack claims 98–100% removal success across nine watermarking schemes; Evolve judges the evidence largely convincing but flags a single backbone, no significance tests, and unverified theory premises.

• Label-free two-stage NAS ranking with cross-space robustness

  CoRA-NAS combines coarse ranking with anchor-residual refinement, claiming robust cross-space performance on four vision benchmarks without labels; Evolve finds the evidence largely supportive, though size-space ties best proxies.

Timeline

1. 2026Contributes to a Lean-checked characterization of language generation in the limit via finite witnesses, answering an open question of Raman et al. and reaching width levels up to ω+1.
2. 2026Publishes nearly tight Rademacher complexity bounds for sparsely activated ReLU networks, removing ADKM's √n factor up to a log^{3/2}(2m) term with matching lower bounds.
3. 2026Appears as last author on two NAS papers, RiPPLE and CoRA-NAS, proposing cross-space performance prediction and label-free two-stage ranking from zero-cost proxies and learning-curve extrapolation.
4. 2026Proposes DRIFT, a black-box diffusion-watermark removal attack that deflects the generative trajectory and claims 98–100% success across nine watermarking schemes.

Papers in the Evolve catalog

• 2026-09-11

  RiPPLE: Cross-Space Performance Prediction from Early Training for Neural Architecture Search

• 2026-09-10

  CoRA-NAS: Coarse Ranking and Anchor-Residual Refinement for Neural Architecture Search

• 2026-09-09

  Characterizing Language Generation in the Limit: Finite Witnesses and a Separation-Width Hierarchy

• 2026-09-08

  Nearly Tight Rademacher Bounds for Sparsely Activated Neural Networks

• 2026-09-08

  DRIFT: Removing Diffusion Watermarks by Deflecting the Generative Trajectory

Frequent co-authors