Advisor Dossier: Prof. Rex (Zhitao) Ying — Yale University

Student: Weijia Zhang | M.S. CS, Yale University (Aug 2026 – May 2028, Thesis Track, Full Scholarship) Assumed goals (inferred from CV): Industry-research 60% · Academia (PhD after M.S.) 40% Report date: 2026-06-11

1. Executive Summary

Top critical risks/unknowns:

1. Near-zero frontier lab placement record (critical): Only 1 PhD graduate from this Yale lab — Aosong Feng (2024) → AWS Applied Scientist II. Zero confirmed PhD placements at OpenAI, Anthropic, DeepMind, Meta FAIR, or MSR. The Leskovec network offers theoretical access but is entirely undemonstrated for this specific lab’s students.
2. Extremely thin alumni sample (structural): Rex joined Yale in 2022. 4 years of data, 1 PhD graduate. No distribution statistics are meaningful with n=1. The “placement track record” has essentially not formed yet.
3. Research domain mismatch for Weijia: Rex’s core identity is graph neural networks and geometric deep learning. Weijia’s strengths are VLM/GUI agents, SFT pipelines, and RL for general agents. The overlap requires a deliberate bridging project.
4. NSF/NIH grant details unconfirmed: An NSF grant is mentioned on the lab site but amount, duration, and renewal status are unknown. No NSF CAREER award evidence found.

One-line verdict: Rex Ying’s lab produces strong papers and has an exceptional theoretical network, but its frontier-lab placement pipeline is unproven — which is a meaningful risk for an industry-research goal within a 2-year M.S. timeline.

Strongest pros:

• Students publishing at NeurIPS, ICML, ICLR at high rates for a young lab (Weikang Qiu: ICLR 2026 oral; Ngoc Bui: ICML 2024 + NeurIPS 2025 Spotlight; Jialin Chen: 2× NeurIPS 2023)
• Jure Leskovec (PhD advisor, Stanford, Kumo.ai) is among the most connected graph ML researchers in industry; offers a powerful but indirect network
• Rex is explicitly pivoting toward “multi-modal foundation models with reasoning and post-training” — directly overlapping with Weijia’s expertise
• Early-career faculty (4 years at Yale): very high accessibility and mentorship bandwidth
• Multi-agent LLM research is emerging in the lab (ACL 2025 paper with Jiaxuan You)

Strongest cons:

• 0 confirmed frontier AI lab placements from this lab’s Yale PhD students
• 0 confirmed PhD → industry-research placement at any top-5 AI company
• 1 PhD graduate (Feng → AWS Applied Scientist): “acceptable” but not the trajectory Weijia likely needs
• Graph neural networks remain the core identity; Weijia has no graph ML background
• Weikang Qiu’s graduation placement (likely 2026) will be the first real signal — comes after Weijia arrives

Score snapshots:

• Four-Dimension Fit Score: 63/100 → Proceed with caution (base)
• AI Industry Outcome (industry-research track): 46/100 → Significant concerns (caps final verdict)

Coverage: Technically high (for available data) but structurally unreliable — only 2 alumni rows (1 PhD, 1 postdoc). Sample size renders all placement statistics meaningless for prediction.

Concrete next steps:

1. Ask Rex directly: “What is your strategy for connecting students to frontier AI lab internships at OpenAI, Anthropic, Google DeepMind?” — his answer reveals whether he has a plan.
2. Ask him about the post-training/agent research direction: is there a project combining LLM agents with graph-structured reasoning that leverages Weijia’s RL expertise?
3. Talk to Jialin Chen (has Google internship) and Weikang Qiu (graduating soon) — best proxies for what the internship/placement pipeline actually looks like.
4. Contact Jiaxuan You (Stanford, co-author on Rex’s ACL 2025 LLM multi-agent paper) — warm intro possibility.
5. Use Rex as a secondary advisor option: if Cohan provides better frontier lab network access, consider Rex as thesis committee member for research perspective.

2. Critical Problems First

3. Strong Pros and Strong Cons

Pros

• Landmark citation profile (53K+): GraphSAGE, PinSAGE, GNNExplainer are foundational papers. This makes Rex highly recognized in industry and creates warm intros to large swaths of ML talent networks, even if the specific frontier AI lab pipeline is unproven.
• Jure Leskovec connection (highest-value academic-industry bridge in graph ML): Rex did his PhD with Leskovec, was his lab’s Founding Engineer at Kumo.ai, and continues to co-author. Leskovec’s alumni are everywhere in tech (Pinterest, LinkedIn, Meta, Google, Airbnb). For industry placements in companies that rely on graph ML, this connection is unmatched.
• Current students are publishing at exceptional venues: Weikang Qiu (ICML 2024 + ICML 2025 + ICLR 2026 oral) and Ngoc Bui (ICML 2024 + NeurIPS 2025 Spotlight) show the lab can support top-tier research output in a short time.
• Explicit post-training + multi-agent direction: Rex’s lab website explicitly states “multi-modal foundation models with reasoning and post-training” as a core focus. This is the exact language that describes Weijia’s expertise. ACL 2025 paper on multi-agent LLM systems is live.
• Amazon Research Award 2024: Ongoing Amazon connection; could facilitate AWS/Amazon Science internships.
• Very high accessibility: Rex is 4 years into a faculty position. He has strong incentives to support motivated M.S./PhD students closely.
• Neil He (BS/MS → UIUC PhD with Amazon fellowship): Shows that even undergraduate researchers get strong academic outcomes.

Cons

• Industry-research placement track record is essentially nonexistent. Aosong Feng (AWS Applied Scientist) is an acceptable outcome but is not the “Research Scientist at a frontier AI lab” trajectory that Weijia likely aims for.
• The Leskovec network is second-degree: Jure’s connections are at GraphML industry teams, not necessarily at the post-training, RL, or alignment teams where Weijia is most competitive. A warm referral to Pinterest’s graph team is not the same as a referral to OpenAI’s post-training team.
• Graph ML expertise is not Weijia’s background. There’s a real risk of spending the M.S. learning graphs (to fit the lab) rather than deepening agents/RL (to maximize career options).
• The lab’s biology turn (spatial transcriptomics, EHRs, fMRI decoding) covers a significant portion of current projects. If Weijia joins and the most active project involves biology applications, she may end up doing work that’s orthogonal to her career goals.
• Weikang Qiu’s placement (expected 2025–2026) is the critical signal — but it will only become public after Weijia has already arrived. She can’t use it as validation before joining.
• No evidence of M.S. thesis students publishing first-author at top venues yet (Neil He’s work is under Rex’s supervision but he was primarily a math/MS student, not CS thesis-track).

4. Academic Profile

Position: Assistant Professor (some pages show “Professor” — rank update unconfirmed [ambiguous]), Yale CS, joined 2022 [1]. PhD: Stanford, supervised by Jure Leskovec, 2016–2021 [2]. B.S.: Duke University, CS + Mathematics, 2016 [3].

Lab: Graph and Geometric Learning Lab (G2 Lab); GitHub: github.com/Graph-and-Geometric-Learning [4].

Awards: KDD 2022 Best Dissertation Award [5]; Amazon Research Award 2024 [6]; NeurIPS 2025 Spotlight (HypLoRA, top 3%) [7].

Citation metrics: ~53,887 Google Scholar citations (provisional figure from search snippets; exact h-index unconfirmed — visit scholar.google.com/citations?user=6fqNXooAAAAJ to verify) [8]. Semantic Scholar: ~1,616 highly influential citations across ~53 papers [9]. The bulk of citations derives from Stanford-era foundational papers.

Landmark papers: GraphSAGE (NeurIPS 2017) [10], PinSAGE (KDD 2018) [11], GNNExplainer (NeurIPS 2019) [12], DiffPool (NeurIPS 2018) [13].

Research evolution:

• Phase 1 (Stanford 2016–2021): Scalable GNNs for industry (GraphSAGE, PinSAGE), GNN explainability, hierarchical pooling, identity-aware GNNs
• Phase 2 (Yale 2022–2023): Subgraph counting, hyperbolic embeddings, contrastive learning, biology applications (MuSe-GNN NeurIPS 2023)
• Phase 3 (Yale 2024–2026): LLMs + graphs (LLM as GNN, graph foundation models), hyperbolic LLM fine-tuning (HypLoRA), multi-agent LLM systems, post-training focus

Industry connection: Founding Engineer at Kumo.ai (2021), Leskovec’s graph ML startup [14]. This creates an ongoing relationship with the Stanford-to-industry pipeline.

5. Alumni Outcomes and Graduation Windows

PhD Graduates (n=1, Yale)

Notes on Aosong Feng: Co-advised by Rex Ying (primary) and Leandros Tassiulas (ECE co-advisor) [15]. Dissertation confirmed in EliScholar database [16]. Outcome is a respectable applied science role at AWS but does not represent a top-tier frontier research placement.

Postdoc Alumni (n=1)

Notes: Positive academic outcome. Menglin Yang came with a PhD from CUHK (Irwin King) [17]. The HKUST(GZ) placement is a respectable tenure-track position in Hong Kong/Greater Bay Area.

Current PhD Students (Active)

M.S. Students / Undergrad Researchers

6. Placement Distribution and Attrition Analysis

PhD cohort (n=1, Yale): Insufficient for distribution analysis.

• Only data point: Aosong Feng → AWS Applied Scientist II (acceptable, not frontier)
• No faculty placements from Yale PhDs
• No frontier AI lab placements from Yale PhDs
• Median cannot be estimated; single sample is not representative

Postdoc (n=1): Menglin Yang → HKUST(GZ) faculty. Positive academic outcome.

B.S./M.S. researchers (n=1 confirmed placement): Neil He → UIUC PhD with Amazon fellowship. Strong academic outcome.

Distribution summary:

• Upper tail: Menglin Yang (postdoc → faculty) — academic pipeline
• Only PhD data point: Aosong Feng → AWS (acceptable industry)
• Lower tail: unknown (only 1 data point)
• Frontier readiness: ZERO confirmed frontier lab placements

Attrition: No evidence of non-completions or bad quits found. However, with so few students and the lab being young, this absence is expected, not reassuring.

Near-graduation unemployment risk: No evidence of post-graduation unemployment. The one sample (Feng) placed promptly at AWS. But n=1 is too small to infer anything.

Critical upcoming signal: Weikang Qiu (ICLR 2026 oral, final year) will be the next PhD graduate. His placement outcome will be the most informative data available for any student making a decision about this lab in 2026–2027.

7. Data Coverage Dashboard

Overall coverage confidence: High (for available data) Critical caveat: Coverage is “high” only because what is knowable about the tiny alumni pool is fully known. The fundamental limitation is sample size (n=1 PhD), not data gaps. High coverage confidence does NOT mean high predictive reliability here.

No coverage gate downgrade applies (coverage is technically high for available data).

What missing data would most change the verdict:

1. Weikang Qiu’s placement (graduating ~2026) — will this be frontier AI lab, academic, or industry-engineering?
2. Jialin Chen’s Google internship return offer or full-time conversion status
3. NSF grant award number, duration, and amount
4. Whether any of the 4 final-year PhD students (Jialin, Tinglin, Weikang, Ngoc) have frontier lab full-time offers by late 2026

8. Four-Dimension Risk and Fit Assessment

Goal weights (blended: 60% industry-research + 40% academia):

• Survival: 27

  Academic: 27

  Industry: 31

  Happiness: 15

Four-Dimension Fit Score: 68 × 0.27 + 65 × 0.27 + 55 × 0.31 + 68 × 0.15 = 18.36 + 17.55 + 17.05 + 10.20 = **63.2/100**

Base verdict: Proceed with caution (50–74 range)

9. AI Industry Outcome Scorecard (Industry-Research Track)

Industry-research track total: 46/100 → < 50 → caps final verdict at Significant concerns

Verified Frontier Placement Table

Frontier gate result:

• Verified frontier full-time placements: 0
• Verified frontier internships: 1 (Jialin Chen at Google; team/date ambiguous)
• Gate: 0 full-time AND ≤1 internship → caps verdict at Proceed with caution
• Combined with industry-research track < 50: final cap = Significant concerns ✓

Frontier readiness: Limited

10. Verdict and Score Reconciliation

Final verdict: ⚠️ Significant concerns (63/100 base, capped by industry-research track)

Context: The “Significant concerns” verdict reflects the current, documented state of the placement track record, not Rex Ying’s quality as a researcher or the lab’s trajectory. The lab is actively improving. Weikang Qiu’s upcoming placement and the graduation of 4 strong PhD students over 2026–2027 will materially change this verdict — but that information doesn’t exist yet.

If Weijia’s goal is primarily academic (PhD application after M.S.): Applying academia-only weights yields a Proceed with caution verdict (65/100 under pure academia weights), and Rex would be a reasonable choice if the research topic (graph + LLM + geometry) can be positioned to support strong PhD applications.

11. Personalized Fit (Weijia Zhang × Rex Ying)

Research Overlap

Most Promising Intersection Projects

1. RL-based post-training for graph foundation models: Combine Weijia’s RL training pipeline expertise (VERL, OpenManus-RL reward signals) with Rex’s graph foundation model direction. Produce a graph FM that learns from agent feedback signals. This is a natural translation of Weijia’s skills into Rex’s domain.

2. Multi-agent LLM systems for structured data reasoning: Rex has ACL 2025 paper on “LLM multi-agent systems as scalable GNN reasoners.” Weijia built OpenManus-RL (60K stars) and GUIAgentDebugger. The intersection: applying RL training to improve multi-agent graph reasoning. Directly pitchable to Rex.

3. Scientific agent evaluation with graph-structured knowledge: Combine Weijia’s GUIAgentDebugger error taxonomy (4 categories, 29 subtypes) with Rex’s knowledge graph + retrieval work (GRIL, HypRAG). Build evaluation benchmarks for agents operating over structured scientific knowledge graphs. Less central to Rex’s research but still relevant.

Key Friction Points

1. Graph ML learning curve: Weijia has no background in GNNs, graph theory, or geometric deep learning. There is real ramp-up cost (likely 3–6 months) before she can contribute meaningfully.
2. Biology applications dominant: A significant fraction of active projects (Tinglin Huang’s spatial transcriptomics, Weikang Qiu’s fMRI decoding, Hiren Madhu’s biology GNNs) are biology-focused. If Weijia joins and most lab conversations are about gene expression or brain imaging, there’s a happiness/motivation risk.
3. Hyperbolic geometry specialization: Rex’s most distinctive current work (HypLoRA, hyperbolic LLMs) is a specialized mathematical direction. Weijia’s SFT/RL background doesn’t naturally translate here, and learning Riemannian geometry for the 2-year M.S. may not be the best investment.
4. “Post-training” means different things: Rex’s use of “post-training” may primarily mean fine-tuning pre-trained models with graph-aware techniques, not the RL-based agent training Weijia has done. Need to verify in person.

Skill Match

• Strong fit: Python/PyTorch infrastructure, agent framework development, SFT data pipelines
• Gap: Graph theory fundamentals, hyperbolic geometry, spatial biology domain knowledge
• Weijia’s unique edge Rex’s lab lacks: RL at scale (VERL), production agent systems (MSRA, OpenManus), multi-step reward design

Network Complementarity

• Rex’s network (Leskovec/Stanford → Pinterest/Meta/LinkedIn/Airbnb; Amazon; graph ML community) is complementary to Weijia’s Asia-focused network (MSRA, Tencent, OpenManus/ByteDance adjacent)
• The graph ML industry community is large and useful for engineering roles but less relevant to frontier AI research positions specifically

One-line fit verdict

Rex is a reachable, enthusiastic advisor with the right research direction language, but the placement pipeline doesn’t yet exist — and for a 2-year M.S. targeting industry research, that’s the central risk.

12. Alumni Impact and Connection Mapping (Prioritized)

Connection strength to Weijia: Weikang Qiu, Jialin Chen, Tinglin Huang = direct (same lab, will be at Yale with Weijia). Aosong Feng, Neil He = adjacent. Jiaxuan You = adjacent.

13. Funding and Resources

Funding risk for M.S. student: Low — Weijia’s scholarship is institutional. Compute risk is moderate: Rex’s lab does mostly GNN + LLM work which requires GPU but not the massive-scale RL training Weijia is used to from MSRA. Verify compute availability before committing.

14. Research Gaps

1. H-index and per-paper citation counts: Must be read directly from Google Scholar; ~53K total citations confirmed but h-index not found in any search.
2. NSF grant details: Award number, amount, duration, renewal status — all unconfirmed.
3. Jialin Chen’s Google internship outcome: Was it converted to full-time? Return offer? Team type?
4. Rex’s actual mentorship style for M.S. students vs. PhD students: No public evidence. Must ask current students.
5. Weikang Qiu’s placement (expected ~2026): This is the single most informative data point that doesn’t exist yet.
6. Kumo.ai connection depth: Does Rex actively help students access Kumo.ai resources (data, compute, internship connections)?
7. Lab size estimate: 6 active PhD students + 2–3 M.S. students is the current estimate. If Rex is also on multiple committees and doing admin work, bandwidth may be more constrained than it appears.

15. Questions to Ask

Questions for Prof. Rex Ying

1. What major problems is the lab trying to solve with LLMs + post-training, and why is this a better approach than pure graph methods?
2. Given my background in SFT data pipelines and RL for agents, what would a first project look like that fits your lab’s direction?
3. What is your policy for M.S. thesis advising — how does your time allocation between PhD and M.S. students work in practice?
4. What compute resources does the lab have? Is there a budget for large-scale RL training experiments?
5. What is your internship policy? Do you actively help students secure internships at frontier AI labs (OpenAI, Anthropic, Google DeepMind, Meta FAIR)?
6. What does the post-training work in your lab actually look like — is it RL-based agent training, or PEFT/fine-tuning of pre-trained models?
7. What is the funding model — does my thesis project depend on your grant, or is it covered by the institutional scholarship?
8. What is the typical thesis timeline for M.S. students, and do M.S. students typically publish before defending?

Questions for Current/Former Students (Weikang Qiu, Jialin Chen, Aosong Feng)

1. How does Rex engage with your day-to-day work — weekly meetings, ad hoc, or primarily at paper deadlines?
2. Did Rex actively help you secure internships, or did you primarily self-source them?
3. Is there a biology/biology-app vs. LLM-agent split in the lab’s attention — how much time does Rex spend on each direction?
4. What is the authorship policy — do students genuinely own their projects?
5. How does Rex behave when a project stalls or gets scooped?
6. What do you wish you knew before joining?
7. Are there any students who left or had difficulties? What happened?
8. How strong is Rex’s support for industry placements vs. academic placements?

High-Uncertainty, High-Impact Verification Questions

1. Ask Jialin Chen: Did Rex directly facilitate the Google internship, and is there a return offer or full-time trajectory?
2. Ask Weikang Qiu: Are you applying to frontier AI labs, and has Rex provided referrals? (This is happening right now — most current signal available.)
3. Ask Rex directly: “Have any of your current students interned at OpenAI, Anthropic, or DeepMind?” — a direct question deserves a direct answer.
4. Check NSF Award Search (nsf.gov/awardsearch) for Rex Ying — verify grant amount and expiration.

16. 12–24 Month Career Plan (Contingency)

Path A: Work with Rex Ying (viable under conditions)

Condition: Rex can articulate a specific project at the intersection of RL-based post-training AND graph/structured reasoning, with compute available.

• Pre-arrival (June–August 2026): Email Rex with a 1-page pitch: “Using RL to train multi-agent LLM systems for structured data reasoning.” Ask for a video call.
• Month 1–3: Join lab, build graph ML fundamentals (2 months of reading), connect with Weikang Qiu and Jialin Chen. Confirm project direction.
• Month 3–6: Launch project on RL post-training for graph/relational agents. Target NeurIPS 2027 or ICML 2027.
• Month 6–12: Begin internship search. Leverage Rex’s Leskovec connection for Meta/LinkedIn/Pinterest graph research teams; self-source for OpenAI/Anthropic/DeepMind.
• Month 12–18: Complete internship (Summer 2027). Return, finalize thesis.
• Month 18–24: Defend. Decide: PhD application (strong recommendation from Rex + ICML/NeurIPS paper) or industry RS role.
• Key contingency: If Rex’s network doesn’t open frontier lab doors by Month 9, activate Cohan’s Meta FAIR and Anthropic connections as parallel track.

Path B: Rex as Committee Member, Not Primary Advisor

If the first meeting reveals the project scope doesn’t work:

• Primary advisor: Arman Cohan (better frontier lab network, closer to agent evaluation research)
• Committee member: Rex Ying (brings graph ML and LLM post-training perspective)
• Best of both networks without full dependency on Rex’s unproven placement pipeline

Path C: Wait for Weikang Qiu Placement Signal

If you can delay the advisor commitment decision by 3–6 months after arriving at Yale:

• Arrive at Yale, take courses, speak to both Cohan and Rex
• By December 2026, Weikang Qiu’s placement should be public
• If Weikang goes to a frontier AI lab → Rex’s pipeline is more proven; reconsider
• If Weikang goes to AWS/industry-engineering → Cohan becomes primary choice

Sources