Daily Report — 2026-10-15
Daily Overview
- What was done: Refined PhD Statement of Purpose and application materials, completed a high-volume Nature Methods manuscript audit, and integrated lightweight OCR and automated testing for a WeChat AI assistant.
- How it was done: Utilized iterative narrative editing, large-scale paper audits, and the integration of a CPU-based RapidOCR engine with specialized batch testing scripts.
- Impact: Established a cohesive research narrative for applications, provided a structured roadmap for high-impact manuscript publication, and achieved robust, low-latency text recognition for real-world multimodal interactions.
The day focused on advancing PhD applications through SOP refinement, optimizing a scientific manuscript based on Nature Methods standards, and significantly enhancing the AI assistant’s WeChat integration through CPU-based OCR and robust automated testing.
Tasks
Architecture & Strategy
- ✅ SOP Refinement — Edited the Statement of Purpose to ensure a cohesive narrative, specifically regarding Latent Bridge and visual prompting research.
- ✅ WeChat Media OCR Integration — Implemented a CPU-based OCR worker using RapidOCR to extract text from incoming images, stickers, and long screenshots.
- ✅ Recommender Outreach Preparation — Drafted personalized final emails for Prof. Jun Yang and Prof. Brandon Fain, including deadlines and key talking points.
- ✅ WeChat Send-Suite & Testing — Developed a specialized batch testing script (TEST_SEND.bat) for sending various media types and ran regression suites to ensure system stability.
- • Manuscript Restructuring — Reviewing proposed Results section order (7 sections) and benchmark-centric title/abstract to align with Nature Methods standards.
Implementation & Fixes
- ✅ UI/UX Refinement for Media Display — Update the web interface to show extracted OCR text and AI model interpretations as separate, distinct sections.
- ✅ Application Material Organization — Generated SOP PDF and organized required attachments (CV, Transcript, SOP) into structured desktop folders.
- ✅ Automated Reply Safety Enforcement — Modified system settings to ensure automatic replies remain disabled during testing processes.
Problems & Solutions
Critical Issues
1. LLM misinterpreting text in images and losing clarity in long screenshots due to scaling.
Solution: Decoupled text extraction from semantic interpretation by using a segmented reading approach with a local CPU OCR engine.
Key Insight: For multimodal assistants, separating ‘what is written’ (OCR) from ‘what it means’ (LLM) significantly reduces hallucinations.
2. The SOP narrative for ‘Latent Bridge’ lacked a cohesive research arc and sounded like a technical mismatch.
Solution: Rewrote the section to bridge ‘what the method does’ with ’the lesson learned’ regarding system vs. policy errors, using ‘recovery thresholds’ as a unifying theme.
Key Insight: Connecting individual research projects through shared themes creates a stronger narrative arc than presenting them as isolated successes.
3. OCR misidentifying symbols (e.g., ‘?’ as ‘2’) and testing risks interfering with live services.
Solution: Implemented an uncertainty label ‘[疑似符号，需核对]’ for low-confidence characters and created a ’test-only’ mode that pauses receiving services.
Key Insight: Building a ’test-only’ mode that can safely pause the primary service is crucial to prevent race conditions.
4. Manuscript title and results overclaimed ‘agent’ impact and used unproven metrics.
Solution: Retitled to focus on benchmarking and capability; moved agent harness to Methods to align with Nature Methods’ preference for evidence-based writing.
Key Insight: Nature Methods authors prefer ‘capability’ (what it can do) over ‘automation’ (how it was done).
General Issues
5. Managing a large workload for recommenders and technical bugs like sticker domain blocking.
Solution: Added considerate language to emails to ease workload perception and updated the domain allowlist for WeChat sticker CDNs.
Human vs AI Approaches
Strategic Level
Project Consolidation & Planning
| Role | Approach |
|---|---|
| Human | Proposed a granular, multi-step plan involving a unified workspace, function inventory, and agent-led integration using specific technical constraints. |
| AI | Suggested high-level consolidation but required human intervention to define the technical workflow. |
Difference Analysis: The human provided the structural roadmap and the ‘how-to’ for agents, while the AI remained at a general conceptual level.
SOP & Research Narrative Strategy
| Role | Approach |
|---|---|
| Human | Provided the strategic ‘why’ by connecting Markovian limitations to specific tasks and conceptualizing the ‘recovery threshold’ to unify research projects. |
| AI | Initially tried to extract technical trends and ’extrapolated’ connections that were not explicitly in the notes. |
Difference Analysis: The human provided the high-level conceptualization and logical flow that the AI lacked.
Manuscript & Architecture Design
| Role | Approach |
|---|---|
| Human | Proposed treating the agent harness as part of the pipeline (Methods) and suggested a unified ‘reception/sending’ architecture. |
| AI | Treated the agent harness as a major contribution/headline and focused on individual task implementations. |
Difference Analysis: The human focused on strategic positioning and architectural simplicity, whereas the AI focused on implementation-specific tasks.
AI Limitations
Critical Limitations
- The AI attempted to ’extrapolate’ research directions not part of the user’s actual plan, leading to inaccurate SOP content.
- The current agent-based harness effect is unmeasured in the manuscript, and 0/260 NM papers use these terms.
General Limitations
- AI failed to recognize specific characters in images before OCR implementation and could not execute batch tests on the VM autonomously.
- AI’s writing occasionally became ‘preachy’ by over-explaining ’lessons’ rather than facts, and it struggled to distinguish ‘parsing success’ from ‘content accuracy’.
Learnings
Key Learnings
- Effective research narratives are built by linking technical limitations (e.g., Markovian properties) to specific problem domains and proposed solutions.
- For multimodal assistants, separating ‘what is written’ (OCR) from ‘what it means’ (LLM) significantly reduces hallucinations and improves reliability.
- When developing automated interactors, building a ’test-only’ mode that can safely pause the primary service is crucial to prevent race conditions.
- Nature Methods authors prefer ‘capability’ (what it can do) over ‘automation’ (how it was done).
Conversation Summaries
PhD Application
✅ SOP Refinement and Recommender Strategy Iteratively edited the Statement of Purpose to integrate research on FLARE, RecoverBench, Latent Bridge, and Visual Prompting through the concept of policy recovery thresholds. Prepared and organized all materials for professor outreach.
MIHD Writing Audit
✅ Audit Report Generation Ran 260-paper audit to optimize manuscript framing. Identified critical gaps in title, abstract, and results architecture, proposing a restructuring to align with Nature Methods norms.
MihD Research
✅ Fusion Research & Literature Review Reviewed fusion research progress; 8 sub-agents completed surveys of various fusion mechanisms (e.g., cross-attention, graph fusion) citing 184 sources.
WeAgent
✅ WeChat Media Extraction & OCR Integration Resolved media identification and domain blocking bugs. Integrated a lightweight RapidOCR worker for high-accuracy text extraction and deployed a comprehensive batch testing suite for WeChat sending functionality.
PersonalAgent
✅ Consolidation Strategy Developed a strategic plan to consolidate multiple AI projects into a single, unified ecosystem using a shared workspace and specialized agent prompts.