PhD Projects
Working lab notebook for my research projects: status, current plan, and next steps for each, with full experiment records inside every project page.
- Always-On Personal AI SurveyP0Writing · LoopAudit P0 passed
How should an always-on assistant close the loop without turning persistent access into uncontrolled authority?
- Now
- Recode the 18-system evidence matrix under the new channel taxonomy; fix roadmap figure labels
- Next
- arXiv v1 (waiting only on author list) → venue selection after the runtime pilot
11 chapters5 auditable channels2 systems auditedL5 unoccupied
- MMSkill-RLP1SFT validated · RL next
Can a visual agent learn when to act directly, invoke a text skill, use a visual skill, or combine both?
- Now
- RL rollout stack (vLLM/verl for hybrid-attention Qwen3.5); 2K RL gate (5 arms: Router-SFT / RL-only / SFT+RL / Always-Text / prompted router)
- Next
- Full RL (41,116 × 4 → ×8) + four-arm ablations; target the next ICLR cycle
+54.2pp SFT vs base164,464 counterfactual rows14.3% skill_rescue99.6% rollout readiness
- FileGramP2ICLR 2026 · revise & resubmit
Can ordinary file-system traces become grounded evidence for a personal agent instead of another opaque user profile?
- Now
- Rebuttal for submission #7267
- Next
- Decision → code & data release
3 system components1 grounding principle24/7 behavioral substrate2026 public preprint