Portfolio Pitch Deck — 4 Research Slides

For HKUST(GZ) Red Bird Challenge Camp. Minimal text, image-driven.

Slide 1 — THE GAP

Models know, but don’t act.

• RAG: Models detect contradictions, yet keep recommending dangerously.
• Agents: They recognize attacks early, yet execute them anyway.
• Insight: Safety is a routing problem, not just recognition.

Visual Suggestion

Split-screen image:

• Left: A glowing brain / neural network (labeled “Internal State ✓”)
• Right: A disconnected output wire or broken action arrow (labeled “External Action ✗”)
• Center: A gap / crack between them

Figures to Extract

• From Detecting Is Not Resolving: the multi-turn danger-rate chart (T2 danger 0.44–1.00)
• From Knowing Is Not Acting: the early-layer vs. late-layer causality diagram

Slide 2 — THE LENS

We look inside to see why.

• DISF (ACL 2026): Dual-path internal-state forcing constructs contrastive negative signals.
• LatentAudit (CoLM 2026): Residual-stream geometry reads faithfulness in 0.77 ms.
• Attribution Blind Spot (EMNLP): Internal-trajectory divergence exposes parametric memory.
• Insight: From guessing outputs → inspecting internals.

Visual Suggestion

Three vertical panels showing an X-ray / cross-section view of a model:

• Panel 1: Two parallel paths (CTX vs. NOCTX) — DISF
• Panel 2: A magnifying glass over residual-stream activations — LatentAudit
• Panel 3: Diverging trajectories inside the model — Attribution Blind Spot

Figures to Extract

• From DISF: Figure 1 (dual-path architecture with CTX / NOCTX)
• From LatentAudit: Figure 1 (pipeline overview: Monitor → ZK Proof)
• From Attribution Blind Spot: The with-context vs. without-context trajectory diagram

Slide 3 — THE LEVER

We intervene to fix it.

• FIDES (EMNLP): Token-selective decoding unlocks suppressed capability. 92–94% fidelity.
• CORDON-MAS (EMNLP): Architectural isolation blocks poison. ASR ↓ 92.4%.
• Whose Thoughts? (NeurIPS): Projection steering recovers lost routing. +17.3 pp.
• Insight: From passive detection → active control.

Visual Suggestion

A lever / gear mechanism transforming input to safe output:

• Left: Unsafe input entering
• Center: Three intervention gears (Token Selective / Isolation / Projection)
• Right: Safe output exiting

Figures to Extract

• From FIDES: The three-layer signal fusion diagram (Output / Hidden / Trajectory)
• From CORDON-MAS: The multi-agent architecture (Extractor → Auditor → Gate → Synthesizer)
• From Whose Thoughts?: The rank-k projection steering diagram

Slide 4 — THE HORIZON

We deploy where it matters.

• RETINA-SAFE (MICCAI): 12,522-sample benchmark for diabetic retinopathy risk triage.
• ZK-FPE: Blockchain-verifiable model ownership without exposing weights.
• Future: Multimodal mechanistic interpretability for accountable AI.
• Insight: Safe AI must be inspectable and accountable.

Visual Suggestion

Three horizon panels fading from lab to world:

• Left: A medical retina scan with heatmap overlay — RETINA-SAFE
• Center: A blockchain / shield icon with zero-knowledge circuit — ZK-FPE
• Right: A multimodal AI system (vision + language) with internal pathways lit up — Future

Figures to Extract

• From RETINA-SAFE: The E-Align / E-Conflict / E-Gap taxonomy diagram
• From ZK-FPE: The fingerprint verification flow diagram
• Concept art: Multimodal model internals (can be a simple diagram)

Navigation Bar (Optional)

Place at the top or bottom of every slide for continuity:

THE GAP  →  THE LENS  →  THE LEVER  →  THE HORIZON
    ■           □            □             □      (Slide 1)
    ■           ■            □             □      (Slide 2)
    ■           ■            ■             □      (Slide 3)
    ■           ■            ■             ■      (Slide 4)

Color & Typography Notes