Final-year PhD @ HKU CS  ·  Hong Kong
I prove the architectural limits of LLM reasoning, and build the systems that route around them.

I'm a final-year PhD candidate in the Department of Computer Science at The University of Hong Kong, advised by Prof. Siu-Ming Yiu. My research sits at the intersection of three threads that keep refusing to be separate:

• What transformers can actually reason about. Tight architectural bounds, plus the tool-delegation systems those bounds force you to build.
• Trustworthy LLMs in regulated settings. Compliance-grade explainability, distribution-free coverage, atomic claim verification.
• Serving infrastructure that respects both. Workflow-atomic GPU scheduling with per-tenant fairness guarantees.

Theorems tell you what cannot be done. Systems make precise what can.

The cycle runs both ways: deployment surfaces the limits worth proving, and the proofs become the constraints that keep deployment honest.

• [05.2026] 🎉 Accepted to TMLR: Tight Bounds and Fundamental Impossibility for Knowledge Editing Side Effects in Transformers. Computable bounds on edit side effects, plus an impossibility theorem ruling out perfect locality and generalization at once.
• [05.2026] 🎉🎉🎉 Two papers accepted to ICML 2026: The Deterministic Horizon: When Extended Reasoning Fails and Tool Delegation Becomes Necessary (Main) and Current XAI Methods Cannot Satisfy Financial AI Explainability Requirements (Position Track).
• [05.2026] ✨ On the postdoc market for Fall 2026. Trustworthy / compliance-grade AI, multi-agent systems & mechanism design, LLM theory, and serving systems. Reach out at bettyguo@connect.hku.hk.
• [05.2026] 📝 Serving as a reviewer for NeurIPS, EMNLP, ACM Multimedia (Main & Dataset Tracks), and UAI.
• [04.2026] 🏆🏆🏆 Four papers accepted to ACL 2026 Industry Track: FinGround (atomic claim verification), RouteNLP (conformal LLM routing), AgentEval (DAG-structured agent evaluation), and ComplianceNLP (KG-augmented regulatory gap detection).
• [04.2026] 🚀🚀🚀 SAGA accepted to HPDC 2026. It's a workflow-atomic scheduler for AI agent inference on GPU clusters, with per-tenant fairness guarantees that hold under real multi-tenant load.
• [03.2026] 📣 Adaptive Retrieval for Large Reasoning Models accepted to SIGIR 2026. When to retrieve during reasoning, with bounds, not heuristics.
• [02.2026] 💼 Conformal-bound risk management at Brain Investing is now running against live P&L. That's our HKU FinTech spin-out, and the lab's coverage work has finally made it onto a real trading book.
• [01.2026] 🛠️ Shipped multi-tenant scheduling and conformal-coverage pipelines at Stellaris AI for native-safe foundation-model deployment in regulated industries.
• [09.2025] 🎓 Began the final year of PhD at HKU CS, advised by Prof. Siu-Ming Yiu. Thesis focuses on the theory-meets-deployment cycle: bounds on transformer reasoning, and the systems those bounds force.
• [08.2025] 🏅 Continuing Cyberport Incubation (2023–2025 intake). That keeps an unbroken 2018–2025 funding run going across TSSSU, HKSTP Incu-Tech, HKU iAXON Deep Tech, and Cyberport.

Four projects worth a second look:

Paper	Venue	Code
Tight Bounds and Fundamental Impossibility for Knowledge Editing Side Effects in Transformers		ke-bounds
The Deterministic Horizon: When Extended Reasoning Fails and Tool Delegation Becomes Necessary		deterministic-horizon
When to Retrieve During Reasoning: Adaptive Retrieval for Large Reasoning Models		realm-retrieve
Current XAI Methods Cannot Satisfy Financial AI Explainability Requirements		position paper
SAGA: Workflow-Atomic Scheduling for AI Agent Inference on GPU Clusters		SAGA
FinGround: Atomic Claim Verification for Financial LLM Outputs		FinGround
ComplianceNLP: KG-Augmented Regulatory Gap Detection		ComplianceNLP
RouteNLP: Conformal LLM Routing		RouteNLP
AgentEval: DAG-Structured Agent Evaluation		AgentEval

_{Full publication list, PDFs, and BibTeX at bettyguo.github.io.}

Three lines that keep crossing in our papers. Each thread proves a bound and ships the system that meets it.

What softmax attention can realize at inference time, and what it provably cannot. The matching upper and lower bounds become the spec for the tool-delegation layer above them.

_{📄   The Deterministic Horizon  ·   ·  Adaptive Retrieval for Large Reasoning Models  ·   ·  code: deterministic-horizon, realm-retrieve}

Explainability and verification that survive financial-services audit, not benchmark conditions. Distribution-free coverage, atomic claim verification, knowledge-graph-augmented regulatory gap detection, and provable bounds on knowledge-editing side effects.

_{📄   Tight Bounds and Fundamental Impossibility for Knowledge Editing Side Effects in Transformers  ·   ·  Current XAI Methods Cannot Satisfy Financial AI Explainability Requirements  ·   ·  FinGround, ComplianceNLP  ·   ·  code: ke-bounds, FinGround, ComplianceNLP, TrustKGRAG}

Workflow-atomic GPU scheduling with per-tenant fairness guarantees that hold under real multi-tenant load. DAG-structured evaluation harnesses and conformal routing for agent cascades.

_{📄   SAGA  ·   ·  RouteNLP, AgentEval  ·   ·  code: SAGA, RouteNLP, AgentEval}

How we approach problems, across every thread:

1. Tight bounds with explicit constants. Upper and lower bounds in the same paper. No asymptotic hand-waving.
2. Impossibility paired with construction. When a thing can't be done, that result becomes a design constraint, not a stopping point.
3. Guarantees that survive reality. Distribution-free coverage, conformal prediction, fair scheduling. No idealized assumptions.
4. Theory and the system that meets it, shipped together. The proof tells the algorithm what to achieve; the algorithm tells the proof what's worth bounding.

"Theorems tell you what cannot be done. Systems make precise what can."

85+ original public repos. Research code behind every paper, the architecture program we are betting on next, and the developer infrastructure our team relies on every day across HKU CS, Stellaris AI, and Brain Investing.
_{Browse the full index → github.com/bettyguo?tab=repositories}

_{An exploratory program: candidate sequence architectures beyond attention.}

_{One repo per paper. Theory and the system that meets it, in the same artifact.}

Reasoning, retrieval & serving

Trustworthy & regulated AI

Learning theory & systems

LLM science: behavior, collapse & brains

_{Eight live integrations across our research workflow: code, data, papers, knowledge bases.}

Repo	What it is	Lang
mcp-gateway	Any OpenAPI 3.x spec into a Model Context Protocol server. Auth, rate-limiting, OpenTelemetry baked in.
mcp-postgres	Postgres MCP server for agents. Four-tier safety: role grants, pglast AST guard, per-tx envelope, audit log. Schema introspection, EXPLAIN analysis, pgvector. PG 13 to 17.
mcp-jupyter	MCP server for Jupyter. Live kernel state (variables, dataframes, plots, tracebacks) instead of just the .ipynb JSON.
mcp-wandb-2	Analytical MCP server for Weights & Biases: hparam importance, sweep summaries, run-delta analysis, inline charts, gated Launch actions.
paperbase-mcp	Research-grade MCP composing arXiv, Semantic Scholar, and OpenAlex. Related work, citation graphs, BibTeX in your chat.
mcp-overleaf	MCP server and Skills bundle for finishing a LaTeX paper: bib cleanup, venue rule packs, latexdiff, related-work drafting.
obsidian_mcp	MCP plus 7 Claude skills for Obsidian vaults. Read, search, write, and link notes from Claude / Cursor / ChatGPT. Filesystem-direct, local-first, round-trip safe.
semantic-grep	Local semantic code search. CLI and MCP server, all on your machine.

_{Local-first when possible; verifiable when not.}

Repo	What it is	Lang
Vannevar	Open-source agentic harness with citation-grade memory. Every fact carries a source URI, a temporal validity window, and an append-only provenance ledger. MCP-native, multi-frontend, fully self-hostable.
agent-memory	Verifiable memory for LLM agents. Every recalled claim is HMAC-signed back to its originating trajectory span.
computer_use_agent	Open-source local-VLM browser agent. AT-tree-first routing with VLM fallback, refusals enforced in code, honest benchmarks including the failure atlas.
whisper_agent	Hands-free local voice agent: faster-whisper STT, local LLM with tool use, TTS. Runs entirely on your machine.
agent-tracer-2	OpenTelemetry-native, local-first observability for AI agents. DuckDB on disk, Next.js viewer on localhost, no SaaS. Adapters for Anthropic, OpenAI, LangGraph, AutoGen, CrewAI.
local-deep-research	Self-hosted deep-research agent: multi-step query planning, source synthesis, report generation. Ollama / llama.cpp / vLLM friendly, with SearXNG, FAISS, and BM25.

_{Make model internals visible, on a laptop.}

Repo	What it is	Lang
see-the-ai-think	Watch an LLM think. Visualizes sparse-autoencoder features firing live across every token. Runs on a laptop, no GPU required.
llm-fossils	Reproducible catalog of LLM behaviors that vanished as models scaled.

_{Plus web inspectors: prompt-x-ray, tokenviewer, policy-microscope, catch-the-ai-lying.}

_{Reproducible by default. Probe for contamination, leakage, and reward hacks before declaring a number.}

Repo	What it is	Lang
agent_eval	Open-source benchmark for Claude Code skill bundles. Pass@k plus cost plus reliability, content-addressed leaderboard across Anthropic / OpenAI / Google.
bench_audit	Probes for agent benchmarks: contamination, gold-answer leaks, harness-injection vulnerabilities, reward hacking. CIs on every result.
benchprobe	Audits AI-agent benchmarks for the eight exploit families catalogued by Berkeley / RDI.
agent-backtest-lab	Statistical-rigor audit harness for LLM trading-agent frameworks, with a leakage firewall.
ai-red-team-in-a-box	Red-team toolkit for probing LLM systems.
rag-bench	Small, reproducible benchmark for RAG pipelines.
agent-arena	Arena-style framework for head-to-head agent comparison.
paper-replay	Replay and reproduce paper experiments with locked seeds, environments, and artifacts.

_{Quality layers, lockfiles, and ergonomics for the agent stack.}

Repo	What it is	Lang
promptlock	Production prompt workflow: semantic diff, eval-on-PR, lockfile, drift detection, and rollback for plain-markdown prompts in your repo.
rigging	Typed, trust-bearing, schema-mediated coupling layer that composes heterogeneous components.
skill-forge-2	Quality layer for Claude Code Skills: lint, test, and bench before you ship.
browser-skills	15 reusable, agent-agnostic browser recipes plus an MCP server. Cookie banners, infinite scroll, calendar widgets, all solved once.
diagram-skills	Generate validated diagrams across Mermaid, PlantUML, Graphviz, D2, and Excalidraw. MCP server, CLI, and Claude Code skills.
capture-engine	Capture any web page as vector PDF, standalone HTML, or high-DPR raster. Local-only, MV3.
paper_pod	Local-first audio overviews for academic papers. Take an arXiv URL, PDF, or BibTeX in, get an 8 to 15 minute two-host podcast out.
paper2repro	Paper to reproducible experiment scaffold.
test_forge	Test-generation toolkit for Python research code.

_{What we had to learn the hard way, written down for the next person.}

📓   Atlases & annotated notebooks

🗺️   Maps, lists & roadmaps

Translational work. Coverage proofs and scheduling guarantees in production, against real workloads.

• Stellaris AI. Conformal-coverage pipelines and multi-tenant scheduling for native-safe foundation models, in regulated deployments.
• Brain Investing. HKU FinTech spin-out. Conformal-bound risk management running against live P&L. The lab's coverage work, in a real trading book.

• Peer review at NeurIPS, EMNLP, ACM Multimedia (Main & Dataset Tracks), UAI.
• Mentoring LLM-infrastructure engineers at Stellaris AI on conformal coverage, agent evaluation, and multi-tenant scheduling.
• Ten consecutive years of competitive funding (2018–2025): TSSSU, HKSTP Incu-Tech, Cyberport (×2 intakes), HKU iAXON Deep Tech.

_{Dongxin (Betty) Guo  ·  The University of Hong Kong  ·  Department of Computer Science
homepage  ·  scholar  ·  orcid  ·  openreview  ·  linkedin
Last updated May 2026}