Dacheng Tao

Research profile

Research summary

Dacheng Tao is a researcher at Nanyang Technological University whose cataloged output shows an unusually wide spread across four fields, each contributing 0.25 of the reviewed record: submodular optimization and online algorithms, robotics and embodied AI, high-dimensional statistics and learning theory, and AI security. All four papers in the Evolve catalog are 2026 arXiv preprints, and the researcher is the last-listed author on three of them — a position typically associated with senior or supervisory roles — and a middle author on the remaining survey. The theoretical work stands out: one preprint resolves a STOC 2025 open problem by proving that consistent submodular maximization is capped at 2−√2 ≈ 0.586 under polynomial queries and constant recourse, and another gives a negative answer to the COLT 2015 heavy-tailed restricted-eigenvalue open problem through pathwise constant-width counterexamples and a sharp VC occupancy law. On the systems side, a survey proposes a Plausible→Controllable→Actionable ladder for embodied world models, while TrojanWorld demonstrates backdoor attacks on world-model reinforcement learning agents via imagination steering. Collaboration metrics show 25 distinct coauthors and a collaboration score of 65; the reproducibility score of 57 averages four reviewed papers whose readiness ratings range from low (the survey) to high (learning theory). Output, impact, and activity axes remain unscored for lack of data.

Research keywords

Highlights

• A Plausible→Controllable→Actionable ladder for embodied world models

  This survey organizes embodied world models into a Plausible→Controllable→Actionable capability ladder. Evolve's review finds it well-argued and formally grounded, though validated only qualitatively, with low reproducibility readiness.

• A sharp 2−√2 barrier for consistent submodular maximization

  The paper resolves a STOC 2025 open problem, proving that with polynomial queries and constant recourse, consistent submodular maximization is capped at 2−√2 ≈ 0.586, with rigorous matching bounds.

• TrojanWorld: backdooring world-model agents via imagination steering

  The paper claims a world-model backdoor with a physical trigger that persistently steers agent behavior. Evidence is broad across simulations, but Evolve notes a single target action, small evaluations, and thin defense testing.

• Negative answer to the COLT 2015 heavy-tailed restricted-eigenvalue problem

  Using pathwise constant-width counterexamples and a sharp VC occupancy law, this theory paper answers the COLT 2015 heavy-tailed restricted-eigenvalue open problem in the negative; Evolve rates its reproducibility readiness high.

Timeline

1. 2026Authored (middle position) a survey proposing the Plausible→Controllable→Actionable ladder for embodied world models on arXiv.
2. 2026As last author, resolved the STOC 2025 open problem on consistent submodular maximization with a 2−√2 barrier (arXiv).
3. 2026As last author, introduced TrojanWorld, backdooring world-model reinforcement learning agents via imagination steering (arXiv).
4. 2026As last author, answered the COLT 2015 heavy-tailed restricted-eigenvalue open problem in the negative (arXiv).

Papers in the Evolve catalog

• 2026-09-15

  World Models for Embodied Intelligence: From Plausible to Controllable to Actionable

• 2026-09-09

  A Sharp Barrier for Consistent Submodular Maximization: Any Improvement over $2-\sqrt{2}$ Entails Exponential Queries or Linear Recourse

• 2026-09-07

  TrojanWorld: Backdooring World-Model Agents via Imagination Steering

• 2026-09-03

  Restricted Eigenvalues Beyond Gaussian Width: Threshold Occupancy under Heavy Tails

Frequent co-authors